May 26 2026
How To: Optimize Your Small Business For AI-Powered Search via @sejournal, @lorenbaker

Your next customer is searching right now. Will they find you?

Waiting to show up alongside your competitors on Google Maps?

Customers are finding local businesses through AI assistants, voice search, social platforms, and review sites. You should be in these places, too.

👆 Get the exact steps to show up where your next customer is already searching. Register above to watch the full session.

Stop Being Invisible: Get Your Small Business Found Across Every Channel Your Customers Are Using

This small business marketing webinar gives you a clear, channel-by-channel system for building the kind of online presence that earns trust before the first call.

Watch Kelli Henthorn and Kevin White, Small Business Experts at Thryv, as they share a step-by-step framework for optimizing your Google Business Profile, social presence, and review strategy so customers can find you on AI, in SERP features, no matter where they’re searching.

You’ll Learn How To:

  • Show up in AI and voice search results: What signals AI assistants rely on and the specific steps to make sure your business appears.
  • Turn Google Business Profiles (GBP), social, and reviews into discovery channels: A practical framework for making each platform actively drive calls and visits.
  • Audit your digital front door: A fast way to find where your online presence is losing customers and prioritize the fixes that matter most.

Thryv’s small business experts shared proven, actionable strategies to help you build an online presence that gets found, earns trust, and turns searches into customers.

Register above to get the practical, channel-by-channel playbook every small business owner needs to show up confidently in AI search, Google, social, and reviews.

Ecommerce MGMT 0 Comments
May 26 2026
Google CEO On AI Overviews: ‘More Opinionated Than It Should Be’ via @sejournal, @MattGSouthern

Google CEO Sundar Pichai acknowledged room for improvement in AI Overviews when shown a live product-query result during a Decoder podcast interview with Nilay Patel.

Patel showed Pichai a live search result. Pichai called it “more opinionated than it should be” for the query. The interview was recorded after Google I/O 2026.

What Pichai Said About AI Overview Quality

Patel showed Pichai a “best Chromebook” search result on his phone. The AI Overview gave a confident recommendation. Below it, a Reddit result and a New York Times result each gave different answers.

Pichai responded:

“It’s probably more opinionated than it should be for the particular query you showed me. That was my reaction as a user. That’s the scope for improvement is how I would say it, in a fast-evolving space, but I would expect that to happen in the product.”

He also suggested that the result might have been personalized to Patel’s usage patterns.

Bounce Clicks & Traffic Trends

He addressed publisher traffic concerns, saying that as Google’s technology improves, low-quality clicks are being filtered out. He described this as “a natural evolution.”

“Bounce clicks are going down,” Pichai said. “And so those are all dynamics.”

Google’s VP of Search, Liz Reid, has described AI Overviews as removing “bounce clicks” rather than useful traffic. Google hasn’t shared publisher-facing data to support the claim.

Patel also read a quote from Condé Nast CEO Roger Lynch, who told his teams to plan for zero search traffic. Pichai didn’t challenge Lynch’s planning decision. He also didn’t directly address Lynch’s claim that search traffic had fallen more than Condé Nast forecast each year. He told Patel he wasn’t “in a position to tell such an iconic publisher what they should think about their business or plan.”

He also mentioned a Search feature that treats sites a user subscribes to as preferred sources.

“If you’ve subscribed to something, we reflect that as a preferred source for you as a user,” he said, calling it “a new change which we didn’t have before.”

Why This Matters

Pichai looked at a live AI Overview and called it too opinionated for the query. The comment lands in a broader debate over AI Overviews’ effect on organic clicks. A field experiment found that AIOs reduced external clicks per affected search by about 38%, but that study measured click behavior, not whether subjective AI Overview recommendations were accurate.

The bounce clicks explanation continues the pattern across Google executives’ appearances. Pichai used similar language to that Reid used in Bloomberg’s Odd Lots earlier this year. Alphabet’s Q1 earnings showed Google Search & other revenue up 19%. The company still hasn’t released traffic data publishers would need to verify the claim.

The subscription preference signal is a concrete product change worth monitoring. Pichai called this “a new change which we didn’t have before.”

Looking Ahead

Google added more link surfaces to AI Search at I/O. Pichai described the result as showing “scope for improvement” and added that he would expect such iteration to occur in the product.

SEJ covered the I/O announcements and Pichai’s separate comments on the agentic coding gap from a Hard Fork interview earlier this week.

Ecommerce MGMT 0 Comments
May 26 2026
Machine-First Architecture: How To Build Websites Machines Can Identify, Read, Cite & Use via @sejournal, @slobodanmanic

In the late 2000s, “mobile-first” emerged as a design discipline. The argument was a single sentence: don’t design for the big screen and squeeze it down. Start with the small screen, the harder constraint, the one that forces you to figure out what actually matters. If it works on a phone, it works everywhere.

Google leaned in early. By February 2010, Eric Schmidt was telling Mobile World Congress that Google’s strategy was “Mobile First in everything.” In April 2015, the Mobilegeddon update penalized non-mobile-friendly websites at scale. In October 2016, StatCounter reported mobile traffic surpassing desktop globally for the first time. A month later, Google announced mobile-first indexing. By October 2023, that migration was complete.

The web is now standing at the same kind of inflection point. Except the harder constraint isn’t a small screen. It’s no screen at all. It’s a machine.

The approach I use, Machine-First Architecture, is a full-stack methodology covering the entire arc of how machines now interact with a brand. It runs from how an organization is identified and resolved across the web, to how a website’s pages expose their data, to how content is consumed and cited, to how an autonomous agent completes a transaction on the website itself. Four pillars, in a specific order: Identity, Structure, Content, Interaction. The order matters. Each pillar depends on the one before it.

This is a website architecture discipline, not a content optimization playbook. Content is just one of four pillars. Most existing AI-search guidance, including frameworks I deeply respect, sits inside that single pillar. Machine-First Architecture extends upstream to organizational identity and downstream to autonomous agent action because that is where the actual work now is.

Last month, I outlined five layers the technical SEO audit needs to add for AI search. That piece described what to check on a website that already exists. Machine-First Architecture is the build framework the audit assumes: the architectural sequence you follow before any audit, on a website you are designing or rebuilding from the ground up. The audit catches gaps. The architecture prevents them. Reading the two together is the point: the build sequence here, the audit checklist there.

The whole journey has to be covered, and that is the part that matters most. The agentic journey is end-to-end: a machine has to identify your brand, parse your website’s structure, evaluate your content, and complete an action on your website. If any one of those steps fails, the whole chain fails. Excellent content cannot save a website with broken identity, because the machine never resolves the right entity to attribute the content to. Strong identity does nothing if the website’s structure hides the data behind JavaScript a crawler will not run. And both of those are wasted if an agent arrives ready to transact and finds a checkout flow it cannot navigate without a human.

It is important to note that machine-first does not mean human-last. Designing for the most constrained consumer (a machine that cannot interpret visual layouts, guess at meaning, or recover from ambiguity) creates a foundation that serves all visitors more effectively. Mobile-first didn’t make desktop worse. It made desktop better by prioritizing what really matters. Machine-first does the same thing for human consumers.

This is the reference version of the framework. What each pillar covers, what to build, what fails when it is missing, and what real protocol infrastructure now backs each one.

Pillar 1: Identity. Can Machines Unambiguously Identify Who You Are?

Identity must come first because AI systems cannot evaluate, recommend, or transact with a brand they cannot confidently resolve.

Google’s Knowledge Graph holds tens of billions of entities and well over a trillion facts about them, with E-E-A-T credibility signals applied at the person-entity level. AI systems consolidate brand identity by reading multiple external platforms in parallel and reconciling what they find. When your website says “AI consultancy,” your LinkedIn says “digital agency,” and your Google Business Profile says “IT services,” models either average those signals into something vague or lose confidence in the entity altogether.

Canonical Definition

A canonical definition is a single, structured, machine-readable document that defines what an organization is in fields rather than paragraphs. Think of it as your brand’s API documentation. Every bio, directory listing, schema block, and social profile description should trace back to this one canonical source.

Entity Relationships

When an AI system answers “who are the leading consultants in this space,” the model traverses connections between entities: founders, clients, industry categories, technologies, publications. The machine-first approach means actively defining and publishing those relationships as structured data, rather than leaving them implicit in blog posts.

Ecosystem Mapping

Map every platform where your brand exists or should exist. Industry directories, review platforms, podcast directories, GitHub profiles, marketplace listings, data aggregators. Each platform exposes data to machines differently. Optimize each platform’s specific structured data format rather than copy-pasting the same bio across all of them.

Version Control

Treat your canonical definition as a versioned document. When identity changes, propagate that change across every platform in your ecosystem map. Machines synthesize identity continuously, and staleness in any one source can degrade the overall picture.

Research by The Digital Bloom from December 2025 found that brands mentioned on four or more platforms are 2.8 times more likely to appear in ChatGPT responses. The architectural condition that makes that compounding effect work, in my experience, is that the platforms tell the same story, which is what the Identity pillar is built to enforce.

A note on scope. This pillar is about the identity of the brand the AI system is trying to recognize. It is not about the cryptographic identity of the AI agent accessing the website. Both matter, but they are different problems.

Output of this pillar:

  • A structured identity document serving as the single source of truth.
  • A map of every platform in your digital ecosystem.
  • A process for keeping all platforms aligned over time.

Pillar 2: Structure. Can Machines Extract Your Information?

Structure inverts the traditional web design process. Define the data model first, then wrap the design around the data.

Most websites are designed to look good to humans, with critical information locked inside visual layouts, JavaScript interactions, and design patterns that machines cannot parse. When an AI agent lands on a product page, it needs to extract the price, specifications, and availability programmatically. Structure is what makes that extraction work.

Structure overlaps with classical technical SEO and modern front-end engineering, but it is neither. Technical SEO has historically focused on what a single rendered page exposes to one crawler. Front-end engineering has focused on how that page is delivered and made interactive for human eyes. Structure, as a pillar of Machine-First Architecture, is upstream of both. It asks what data each page type exists to expose, before either the technical SEO audit or the front-end build begins. The audit checks whether the data is reachable. The architecture decides what data is there to be reached.

Data Models Before Page Designs

Before wireframing a page, define the discrete, extractable pieces of information that page must contain. The question changes from “what should this page look like?” to “what data does this page need to expose?” The page design wraps around the data model, instead of forcing the data model to conform to the design. This is the inversion that distinguishes architecture from audit. An audit can tell you whether your product page exposes price, availability, and specifications. Only the architecture step decides those are the four facts the page exists to express in the first place.

Information Hierarchy For Machines

Machine information hierarchy is structural, not visual. Machines read heading level, schema markup, semantic HTML, and position on the page, not font size, color, or visual weight. Architecturally, this means deciding what goes in the first content block of every page type before deciding how the page looks.

Relationship Architecture

This is where Machine-First Architecture diverges most sharply from how websites are traditionally built. The conventional process designs and ships pages one at a time, with the relationships between them inferred later from navigation menus and internal links. That is backward. Machines need to understand how pages relate to each other before they understand any single page: product taxonomies, service hierarchies, content-to-offering mappings, parent-child structures. Declare those connections explicitly through internal linking patterns, breadcrumb structures, and schema that names the hierarchical relationships directly. The test: Could a machine, starting from your homepage, construct a complete and accurate map of everything you offer by following structured, declared relationships? Not by guessing from menu labels. By traversing connections you have explicitly published.

One more decision belongs in this pillar: rendering. Critical data has to be present in the initial HTML response, before any client-side JavaScript runs. Build a JavaScript-heavy website where prices, specifications, and availability load after the page renders, and that data is locked away from every crawler that doesn’t execute JavaScript. Retrofitting a client-rendered SPA into something that serves data in static HTML is a very expensive failure mode. I broke down which AI crawlers render JavaScript and which ones don’t in “The Technical SEO Audit Needs A New Layer” if you want the specifics.

Output of this pillar:

  • A data model for every key page type, defining exactly what machine-readable information each page contains.
  • A relationship architecture connecting all pages.
  • A rendering strategy ensuring critical data is accessible regardless of how the page is processed.

Do not start designing pages until this work is done. The rendered page is one possible output of the data model. AI search results, voice answers, agent tool calls, and chat citations are other outputs the same data model has to serve. If the design comes first, the data model is whatever the design happened to support, which is rarely what every machine consumer needs.

Pillar 3: Content. Will Machines Rely On What You Are Saying?

Content is the pillar most existing AI-search research already targets. Kevin Indig‘s Growth Memo, Duane Forrester‘s Substack, Ramon Eijkemans’ utility-writing framework, and the ongoing work coming out of SEO Week and the BrightonSEO research community have produced rigorous data on how AI systems evaluate content. I lean on their work in this pillar more than I do in the others, and so should you.

The discipline of writing for AI extraction (answer-first writing, content extractability, citable specificity, content position) is something I get into in detail in “The Technical SEO Audit Needs A New Layer,” and the practitioners I named go deeper still. What Machine-First Architecture adds to that discipline is three architectural decisions that determine whether any of the writing-side work can succeed at all. They are: how authorship is structurally established, how time is signaled, and how the page is composed as modular knowledge units rather than a monolithic narrative.

Authorship And Attribution

AI systems evaluate authorship against the broader knowledge graph when deciding whether to cite a source. Machine-first content makes authorship explicit and structured: who wrote this, what their credentials are, where else they have published. Connected to the knowledge graph through schema markup, with sameAs links to verified profiles, with the author entity itself defined in the canonical identity document established by the Identity Pillar. This is where Identity and Content compose: the author entity referenced here is the same entity defined upstream. Authorship buried in a footer bio is invisible to that compounding effect.

Temporal Signaling

AI systems weigh recency heavily. A 2024 guide loses ground to a 2026 article on the same topic, regardless of objective quality. The distinction runs deeper than ranking. As Duane Forrester wrote, pre-cutoff and post-cutoff content occupy different systems inside the same model. Pre-cutoff content is presented confidently and without attribution. Post-cutoff content arrives with hedging language and citations. The architectural move is this: declare when specific claims were true, what data they are based on, and what has changed since original publication, at a granularity finer than the page’s publication date. AI systems can then evaluate the freshness of individual claims rather than treating the whole page as one timestamp.

Knowledge Modularity

Retrieval systems extract specific claims, answers, and data points. They do not consume content as continuous narrative. Long documents have a well-documented middle-section problem: Language models attend most strongly to the beginning and end of a document and lose fidelity in the middle. Self-contained sections are how content survives that effect. The architectural move is to design content as collections of modular knowledge units rather than monolithic articles. Each section has its own clear scope, its own question, its own supporting evidence. The page tells a complete story where each component functions independently when extracted. This is a composition decision made at the architecture level, not a writing decision made at the draft step.

Output of this pillar: a content framework where:

  • Authorship is structurally connected to your identity layer.
  • Time is declared at claim granularity.
  • The page is composed as modular knowledge units that function independently when retrieved.

Pillar 4: Interaction. Can Machines Act On Your Website Autonomously?

Interaction is the pillar where most existing AI-search frameworks stop. Visibility and citation work covers the first half of the journey: The machine finds and reads you. Accessibility work covers a different problem entirely: a human user with assistive technology making decisions in real time. The pillar that nobody else is finishing is the part where an autonomous agent has to do something on the website on behalf of a real person, with real money, with no human in the loop at the moment of action.

Leaving this last step unfinished is the costliest gap in the journey. An agent that can find your website, parse it, and decide it is the right answer will still abandon if it cannot complete the action it came to perform. That failure will be silent. You never see it in your analytics or your error log, the customer never tells you their agent gave up, and the next agent visit goes to a competitor whose interaction layer works. The full agentic journey is identification through completion, and the framework only delivers compounding value if every pillar holds.

The distinction from accessibility is important. Accessibility assumes a human is still in control: A screen reader translates the page for a person who makes decisions, interprets ambiguity, and recovers from errors. Machine interaction has no human in the loop at the point of action. The agent decides, acts, and verifies on its own.

Most of the eye-catching numbers in trade press right now (393% year-over-year jumps in AI-referred traffic, conversion lifts of 42%, peaks above 1,000% in the December holiday window) measure human traffic that came from AI-powered browsers and AI search results, not autonomous agent activity on the website. A person used ChatGPT or Atlas or Comet to find your website, then clicked through and shopped themselves. That is a real and growing share of website traffic, but it is the visibility-and-citation half of the journey, not the interaction half.

However, the logical next step for that same traffic is the machine also doing the action. The user who today asks ChatGPT to recommend a product and then clicks through to buy it will, increasingly, ask ChatGPT to buy it. The user who today asks Comet to compare hotels and then completes the booking themselves will, increasingly, hand the booking off to the agent. Each step delegates more of the journey to the machine. The Interaction pillar is the layer that has to be ready before that delegation becomes the default. That layer is currently developing, but moving very fast.

Every major AI vendor running the citation layer is also building the agent layer at the same pace, often faster. The companies that decide whether to cite your website are the same companies that decide where their agents try to act.

  • OpenAI runs ChatGPT alongside the Atlas browser, with built-in agent mode (formerly the standalone Operator product, integrated into ChatGPT in mid-2025).
  • Google folded Project Mariner into Gemini Agent and Chrome’s auto-browse capability in May 2026, and operates the Google-Agent fetcher for AI systems acting on user queries.
  • Anthropic pairs Claude with computer-use capability and the Claude-User crawler.
  • Perplexity has both its answer engine and the Comet browser.
  • Microsoft built Copilot Mode and Agent Mode into Edge for multi-step automation.

Treating AI as a pure distribution channel (optimizing for citation, stopping at “be visible in the answer”) is the most dangerous position in this discipline. It assumes the journey ends at the citation, which the vendors building the system have already publicly committed it does not. The citation and agent layers are rolling out on overlapping timelines from the same companies. The website architecture has to be ready for both.

The protocol stack supporting agent-side interaction has crystallized over the last twelve months.

  • Model Context Protocol (MCP): agent-to-tool communication. An inaugural project of the Agentic AI Foundation under the Linux Foundation.
  • A2A: agent-to-agent coordination. A separate Linux Foundation project.
  • WebMCP: agent-to-website interaction. A W3C Community Group draft.
  • Agentic Commerce Protocol (ACP): agent-initiated commerce. Co-developed by OpenAI and Stripe and launched inside ChatGPT in 2025. OpenAI scaled native in-ChatGPT checkout back in early 2026 after low adoption, and ACP now powers purchases through merchant apps integrated into ChatGPT rather than native checkout. The protocol continues, the deployment model is still being figured out.
  • Universal Commerce Protocol (UCP): agent-to-merchant commerce. Developed by Google with Shopify, Etsy, Wayfair, Target, and Walmart, and endorsed by 20+ partners across retail, payments, and processors (Stripe, Visa, Mastercard, American Express, Best Buy, Macy’s, The Home Depot, Zalando, and more). Announced at NRF in January 2026. Shopify’s implementation includes UCP-compliant MCP servers covering storefront browsing, customer account access, and developer tooling so agents can browse, compare, and place orders without screen-scraping.
  • Visa’s Trusted Agent Protocol: cryptographic identity for agent-initiated transactions. In production.

Autonomous agent transactions are not the dominant share of website traffic today, but the infrastructure is in place, the first flows are live, and the websites that wait until traffic forces the issue will be the ones rebuilding under pressure rather than designing into it. Interaction is the build-now-for-the-near-future pillar.

Discoverability Of Actions

A human can tell that a button is clickable through visual design. An AI agent has no such intuition. It needs a programmatic action manifest: Structured declarations of what actions are available on each page, what inputs those actions require, and what outcomes they produce. Schema.org actions provide one path; WebMCP provides another. Every page must answer “what can a machine do here?” as clearly as it answers “what can a human see here?”

Predictable Outcomes

Every action must return a machine-readable response confirming what happened, what changed, and what the next available actions are. An agent adding an item to a cart needs structured state confirmation: The item was added, the cart now contains three items, the total is this amount, the next available action is checkout or continued browsing. Design the state communication layer before the visual feedback layer.

Workflow Continuity

A human navigating a multi-step checkout maintains context mentally. An agent needs that context exposed as structured data: current step, prior decisions, remaining steps, required inputs, and the ability to revise without losing progress.

Error Recovery

Treat errors as structured branching points, not dead ends. When an agent encounters an out-of-stock item, “sorry, something went wrong” is useless. The error response must include structured data: The item is unavailable in size M, available sizes are S, L, and XL, a similar product is available in size M. Every error needs to be a decision point the agent can navigate without human intervention.

Trust And Verification

Humans rely on visual trust signals: padlock icons, brand recognition, professional design. Agents acting on behalf of humans with real money need machine-verifiable trust data: structured, verifiable transaction terms covering pricing, return policies, merchant verification, and guarantees that can be evaluated programmatically before committing. Visa’s Trusted Agent Protocol adds cryptographic proof-of-identity to agent-initiated transactions. The Agentic Commerce Protocol provides the merchant-side payment specification that agent checkouts run on.

Agent Policies And Permissions

When agents visit your website, you need a way to communicate what they are allowed to do. Browse only, or transact? Compare prices? Identify themselves? Rate limits? Standards work here is moving fast and not yet settled. New drafts are published every few weeks across IETF, W3C, and vendor working groups. The architectural need stays the same regardless of which draft wins: a programmatic way to declare what agents can do on your website, before they try to do it.

Output of this pillar: a functional map of every key action on the website, designed as:

  • Machine-navigable pathways with predictable outcomes.
  • Structured error recovery.
  • Verifiable trust signals.
  • Explicit agent policies.

The human visual experience is an enhancement layer on top of this.

The Four Pillars Are Sequential, Not Parallel

Build order matters. Identity first, Structure second, Content third, Interaction last.

You cannot have machine-readable Content without resolved Identity. The authorship principle (who wrote this, what their credentials are, what entities they connect to) depends on the canonical definition that Identity establishes.

You cannot expose Interaction without underlying Structure. An agent cannot complete a checkout flow on a page where the data model was never defined. The action manifest the agent reads is built on the same structural foundation that exposes price, specifications, and availability.

You cannot fix Interaction by patching it on at the end. Websites that try this end up with disconnected JavaScript widgets that simulate machine-readability without actually delivering it. Agents detect the gap, abandon the task, and leave no trace in your analytics.

Build Identity first. Layer Structure on top of it. Build Content into the Structure. Add Interaction as the operational layer once the first three are in place. Each pillar makes the next one possible.

Where To Start: One Action Per Pillar

A practical architecture move per pillar. None of these are audit checks. They are decisions you make before any audit becomes useful.

Identity. Write your canonical definition as fields, not paragraphs. What you do, who you do it for, where you operate, what makes you credible, who the key people are, what entities you connect to. Make this the source of truth that every bio, schema block, and platform listing derives from. Then Google your business name and compare what comes back against that definition. Every platform that tells a different story is a leak in your identity that the canonical document needs to resolve.

Structure. Pick your three most important page types: homepage, primary product or service, primary content. For each, list the discrete facts the page exists to expose, in priority order, before any consideration of layout or design. If you cannot list those facts, the page is being designed before the data model exists, which is the inversion you should aim to prevent.

Content. Pick the three pages most likely to be cited by AI systems. For each, establish two architectural connections: the author entity, schema-linked to the canonical identity document established by the Identity Pillar, and granular temporal signaling on specific claims, declaring when each was true and what data underlies it. The audit will catch whether the content reads well. The architecture decides whether the content is structurally connected to your identity and dated at the claim level.

Interaction. Try to complete a core action on your website (buying something, booking something, submitting a form) using only a screen reader. If you cannot get through the flow, neither can an agent. And agents do not have the patience to figure it out. They move on to a competitor.

Where Machine-First Architecture Fits Among SEO, GEO, And Accessibility

Machine-First Architecture is deliberately broader in scope than the existing AI-search guidance most practitioners are working with. Most frameworks in this space focus on a single slice of the journey: visibility, citation, content optimization, retrieval mechanics. Those are real disciplines, and they are necessary work. Machine-First Architecture is built one altitude above them: the architectural methodology that determines whether any of those tactics can land at all, plus the autonomous-interaction layer the others do not address.

Look at the scope mapping. SEO has historically covered Structure, plus parts of Identity through schema. Generative Engine Optimization covers Content, plus parts of Structure for retrieval. Accessibility covers parts of Structure and parts of Interaction, but only for human-assisted access. Both organizational Identity and autonomous-agent Interaction sit outside the primary scope of every existing discipline. Machine-First Architecture is what sits at the union.

The framework’s scope is bounded by what AI vendors and standards bodies are actively building toward consuming, not by speculation about what future AI might want. Identity protocols are landing, with Knowledge Graph consolidation already in production and verifiable-identity standards moving through W3C. Structural data extraction is mature, with all major AI crawlers parsing JSON-LD and semantic HTML. Content evaluation has documented retrieval mechanisms across position-based citation, authorship cross-referencing, and recency weighting. Interaction protocols are crystallizing as I write this. The four pillars don’t describe what to build for an imagined future. They describe what to build for the demand surface that already exists, plus a near-future surface that is already being shipped.

Duane Forrester’s The Machine Layer is the canonical guide for the visibility-and-trust side of the journey. Read it. Machine-First Architecture is what you build under that, wrapping the same content discipline inside the full architectural span, with Identity at one end and Interaction at the other.

The piece on the technical SEO audit I linked in the opening is the audit you run once the architecture is in place. The accessibility tree work I covered earlier is the rendering surface where most agentic browsers actually read your website, which is where the Structure Pillar’s information hierarchy ultimately gets evaluated.

Mobile-first took years to fully play out, but the actual transition (the point where websites that ignored it started losing) happened in months. Once Google began penalizing non-mobile-friendly websites in 2015, the window for ignoring it closed.

Machine-first is following the same curve, compressed.

More Resources:

Featured Image: Olga S L/Shutterstock

Ecommerce MGMT 0 Comments
May 26 2026
Modern Local SEO & AI Visibility: How To Get Clients Into AI Results via @sejournal, @hethr_campbell

Keyword research has a new purpose, and it’s getting local businesses into AI results.

Why are some local businesses surfacing in AI recommendations while better-ranked competitors aren’t?

Why isn’t my local client showing up in AI recommendations?

How do I get keyword research to work for AI search results?

AI recommendations for local businesses run on trust signal activity, such as keyword-rich, consistent engagement that on-page SEO alone doesn’t generate.

How To Turn Keyword Research Into AI Visibility & Recommendations

In this on-demand session, Jeff Schwerdt, CEO of Reviewly.ai, shared a practical approach to deploying keyword research into local AI trust signals. He covered where AI is pulling keyword-rich signals from, how to build and place them correctly by signal type, and how to keep that activity running consistently across every local business account.

Register above to watch the full session.

You’ll Learn:

  • How to identify what sources AI is pulling keyword-rich signals from: Reviews, responses, and GBP activity, and how keyword placement inside each one influences local AI recommendations.
  • How to build keyword-driven trust signals for a local client from scratch: Keyword selection, placement by signal type, and the response cadence that tells AI a business is active and relevant.
  • How to automate keyword trust signal activity across your full client roster: How to set up review response automation, keyword refresh intervals, and GBP activity scheduling so every client account runs on a consistent weekly cadence.

Register above to watch the recording and give your existing local SEO process a direct line into AI search results.

Ecommerce MGMT 0 Comments
May 25 2026
Google Says AI Mode Can Now Scale Faster Across Languages via @sejournal, @MattGSouthern
  • Reid said AI Mode’s multilingual model architecture has made it easier to expand across countries and languages.
  • She said Google uses existing Search ranking work to help ground AI Mode responses based on location.
  • The interview restated the I/O keynote announcements without rollout timelines.

In a post-keynote interview, Google’s Liz Reid told NDTV that AI Mode’s multilingual models have made it easier to expand across countries and languages.

Ecommerce MGMT 0 Comments
May 24 2026
All You Need To Know About Cloudflare’s Agent Readiness Score via @sejournal, @slobodanmanic

Agent-readiness crossed from concept to measurable infrastructure this week. On April 17, as Cloudflare Agents Week extended into its sixth day, the company shipped isitagentready.com, a public scanner that scores any website on how prepared it is for AI agents. Paste a URL, get a score, see which checks passed and which failed, read AI-generated guidance on how to improve. For the first time, the agent-legibility conversation moved from “is my website ready for agents” as a gut feeling to “my website scored X out of 100 in these five categories, here are the failing signals.”

The Agent Readiness Score is a real shift. It is also a structurally misleading tool if you stop reading after the composite number.

I ran the scan on this website (nohacks.co) and scored 33 out of 100, Level 2 “Bot-Aware.” The robots.txt passed. The sitemap passed. The AI bot rules in robots.txt passed. Content Signals passed. Then the score collapsed across categories where a content-only blog genuinely doesn’t need what the scanner checks for. More on that in a minute.

First, the context. Cloudflare has been shipping agent-facing infrastructure all week. The Agent Readiness Score arrived alongside Agent Memory, Shared Dictionaries, Redirects for AI Training, an LLM compression technique called Unweight, and a feature-flag tool called Flagship built for AI-generated code. Four days earlier, they shipped Project Think (a new Agents SDK), and OpenAI matched it within hours with their own Agents SDK. I wrote about that in The Agent Runtime Wars Started This Week. The readiness scanner is the logical next piece: If runtimes are the new browser layer, website owners need a way to test whether their website is legible to that layer. Cloudflare shipped the tester.

The question this article answers is narrower: What does the scanner actually check, what should you do with your score, and where is the scoring structurally misleading enough that the number by itself leads you astray?

What Cloudflare Shipped: Scanner, API, And An MCP Endpoint Agents Can Call On You

The scanner is at isitagentready.com. Paste any URL, pick a website type (All Checks, Content Site, or API/Application) to scope which signals get scanned, hit Scan. The scanner fetches the homepage and a handful of well-known paths, runs a set of checks against each, and returns a scored report with pass/fail markers, status codes, response bodies, and AI-generated guidance on what to fix.

The scanner is also available in three other ways:

  • Integrated into Cloudflare Radar, so the same checks run alongside Radar’s existing URL analysis.
  • Exposed programmatically via the Cloudflare URL Scanner API for automation.
  • Available as a stateless MCP server at /.well-known/mcp.json, so any MCP-compatible agent can call the scan as a tool and reason over the result

That last one is worth sitting with for a moment. Cloudflare shipped an agent-readiness scanner that agents themselves can call to audit websites before deciding how to interact with them. The scanner checks whether your website is ready for agents, and any agent can invoke it to decide how to interact with you before arriving. The measurement and the measured are starting to share the same surface.

Back to the practical question. What exactly does it check?

16 Checks, 5 Categories: What The Scanner Actually Tests

The scanner groups its checks into five categories. Here is what each one looks for, grouped by what the check actually means in practice.

Discoverability (3 Checks)

Whether the website publishes the basic metadata an agent needs to find what is where.

  • robots.txt exists. The classic crawl-policy file. An agent that follows robots.txt needs it to exist and parse.
  • sitemap.xml exists. Either declared via a Sitemap directive in robots.txt or available at the standard path. An agent that wants to enumerate pages uses the sitemap.
  • Link headers (RFC 8288). HTTP Link headers pointing to canonical, alternate, or related resources. Useful for agents that parse responses rather than HTML.

Content (1 Check)

  • Markdown for Agents. Content negotiation. The scanner sends Accept: text/markdown and checks whether the website returns Markdown instead of HTML. This is Cloudflare’s own proposal rather than an IETF spec, though the mechanism (HTTP content negotiation via the Accept header) is standard. Real agent runtimes prefer Markdown because it is cheaper to tokenize and easier to parse than HTML. Some early movers (Cloudflare itself, a handful of docs websites) support Markdown content negotiation; most websites do not.

Bot Access Control (3 Checks)

  • AI bot rules in robots.txt (RFC 9309). Whether robots.txt contains directives for AI-specific user agents (GPTBot, ClaudeBot, PerplexityBot, etc.).
  • Content Signals in robots.txt. An emerging spec for expressing per-URL access rules inside robots.txt. Parsed as User-agent: * followed by Content-signal: directives. Adoption is minimal right now.
  • Web Bot Auth request signing. HTTP message signatures at /.well-known/http-message-signatures-directory that let agents prove their identity cryptographically. This is the Agent Name Service side of things, Cloudflare shipped with GoDaddy earlier in Agents Week. Adoption is almost zero outside Cloudflare’s own properties.

API, Auth, MCP & Skill Discovery (6 Checks)

  • API Catalog (RFC 9727). A machine-readable index of a website’s API endpoints at /.well-known/api-catalog.
  • OAuth / OIDC discovery (RFC 8414). Standard OAuth 2.0 authorization server metadata at /.well-known/oauth-authorization-server and /.well-known/openid-configuration.
  • OAuth Protected Resource (RFC 9728). A website declaring which endpoints are OAuth-protected and how to authenticate.
  • MCP Server Card (SEP-1649). A Model Context Protocol server advertising its capabilities at /.well-known/mcp/server-card.json. SEP-1649 is a draft proposal inside the MCP spec process.
  • Agent Skills index. A list of agent-callable skills at /.well-known/agent-skills/index.json. Also emerging.
  • WebMCP (Experimental). An in-page JavaScript API registering agent-callable tools via navigator.modelContext. The scanner uses headless browser rendering to detect whether the website registers any WebMCP tools on page load.

Commerce (3 Optional Checks, Not Scored On Non-Commerce Websites)

  • x402 payment protocol. HTTP 402 Payment Required infrastructure for agent-native payments.
  • UCP profile (Universal Commerce Protocol). Google’s merchant-metadata standard at /.well-known/ucp.
  • ACP discovery document (Agentic Commerce Protocol). At /.well-known/acp.json.

The Commerce category is flagged “optional” on non-commerce websites. The scanner detects whether any ecommerce signals are present and, if not, displays the commerce checks for informational purposes without counting them in the score.

That last design detail matters. It is evidence Cloudflare anticipated exactly the problem the rest of this article is about.

Nohacks.co Scored 33/100, Level 2 Bot-Aware

I ran the scan on nohacks.co. The result was 33 out of 100, Level 2 “Bot-Aware.”

The Agent Readiness Score report for nohacks.co, scanned on 2026-04-18. Composite: 33/Level 2 “Bot-Aware.” Category breakdown: Discoverability 67 (2/3), Content 0 (0/1), Bot Access Control 100 (2/2), API, Auth, MCP & Skill Discovery 0 (0/6). Commerce checks not scored (no ecommerce signals detected). Image Credit: Slobodan Manic

A note on that number: After the first scan, I added Content Signals directives to robots.txt, which moved Bot Access Control from 50 to 100 and pulled the composite up eight points from an initial 25. Every other category below is unchanged from the first scan. I’ll come back to the Content Signals fix and why I made it at the end of this section.

Here is what drove each category score:

  • Discoverability: 67. robots.txt and sitemap.xml passed. Link headers failed because this website does not emit Link: headers in its responses.
  • Content: 0. Markdown content negotiation is not configured. The website returns HTML regardless of the Accept header.
  • Bot Access Control: 100. Both scored checks passed. AI bot rules in robots.txt (I have explicit rules for AI user agents) and Content Signals in robots.txt (I added these after the first scan). Web Bot Auth request signing is listed in this category as an informational check, but not counted toward the 2/2.
  • API, Auth, MCP & Skill Discovery: 0. All six checks failed. No API Catalog. No OAuth discovery. No OAuth Protected Resource metadata. No MCP Server Card. No Agent Skills index. No WebMCP tools on the page.
  • Commerce: not scored. nohacks.co has no e-commerce. The Commerce checks all failed, but the category is correctly excluded from the composite score.

That is a 33 on a scanner built by the company I most trust to understand where the agent-ready web is going. I consider this website reasonably well-designed for agents. The robots.txt is clean and explicit. The content is server-rendered, machine-readable HTML with clean semantic structure. The sitemap is current. The URLs are stable. If you asked me a week ago whether this website was agent-ready, my answer would be somewhere between “mostly yes” and “for what it needs to do, yes.”

And yet: 33, Level 2.

The scanner is measuring what it says it is measuring. The composite score, by itself, is still the wrong number to optimize for.

One note on the Content Signals fix, because it’s relevant to the Goodhart argument later in this article. Content Signals is a Cloudflare proposal with almost no deployment beyond Cloudflare-aligned crawlers. I debated adding it for exactly the score-chasing reason this article warns about. I decided it was defensible for two reasons. First, the fix is declarative, not decorative. The directives state real policy about what should happen with my content, and the statement has meaning even if the spec fails. That is different from adding an empty MCP Server Card to satisfy a scorer. Second, for a website that writes about agent-readiness specifically, publicly declaring content policy is editorial practice regardless of which crawler respects it. The fix was one commit to public/robots.txt and the directives are readable by any human curious enough to check.

Same Website Scores 33 Or 67 Depending On The Preset You Select

On the All Checks preset, nohacks.co scores 33 out of 100, Level 2 “Bot-Aware.” On the Content Site preset, same website, same day, different scan configuration, it scores 67, still Level 2 “Bot-Aware.” Nearly double the composite number. The 34-point gap is the difference between two scan configurations of the same scanner, not a difference between two websites.

Here is what the Content Site preset changes in the scan configuration:

The Content Site preset unchecks every item in the API/Auth/MCP/Skill Discovery category, every item in the Commerce category, and Web Bot Auth in Bot Access Control. Six scored checks remain: three Discoverability (robots.txt, Sitemap, Link headers), one Content Accessibility (Markdown negotiation), two Bot Access Control (AI bot rules, Content Signals). Image Credit: Slobodan Manic

Running that preset on nohacks.co produced this result:

Nohacks.co under the Content Site preset: 67 / Level 2 “Bot-Aware.” Four of six scored checks pass. The two failing checks are Link headers (a fix I have not deployed yet) and Markdown content negotiation (not configured). Both are real shipping signals that agent runtimes benefit from today. Image Credit: Slobodan Manic

Four of six scored checks pass. The two failures are unambiguous remediation targets: Link headers via HTTP response configuration, Markdown content negotiation via origin or CDN response logic. Both ship against real agent-runtime behavior today. Neither is a proposal-stage format that will only maybe become a standard. This is the honest reading of nohacks.co’s agent-readiness state: two specific, actionable gaps.

The Correct Toggle Is Hidden, And The Default Score Is Wrong

The scanner is doing its job. It knows a blog does not need an MCP Server Card. It knows a podcast archive does not publish an API catalog. The Content Site preset is not cosmetic. It removes irrelevant checks and gives a content website an accurate reading against standards that actually apply.

The problem is that the preset is hidden. When a user lands on isitagentready.com and pastes a URL, the default scan is All Checks. The Site Type toggle that would switch to Content Site or API/Application lives inside a Customize dropdown that most users will never open. The user clicks Scan, reads the composite number, takes a screenshot, shares it. The shareable number, the one that travels on social media, the one competitors compare across, is the All Checks composite.

For a content website that runs the default scan without reading individual checks, the composite is structurally too low. The 33 on nohacks.co is wrong for the kind of website nohacks.co is. The 67 from the Content Site preset is the accurate reading. Two numbers from the same scanner on the same website. The accurate number is behind a dropdown. The wrong number is on the front page.

Any web professional who runs the scanner and plans to share the score anywhere public needs to open Customize, select the preset that matches their website type, and re-run before sharing. Without that step, the public score will understate the website’s actual agent-readiness, and the gap between the shared number and the accurate number will be larger for content websites than for API websites (which are closer to the All Checks baseline). Read the individual checks. Do not share a composite until you know which preset produced it.

For the record: the 67 is bothering me. I am going to go get the 100. I know exactly what the Goodhart section below is about to warn against, and I am going to do it anyway. Two fixes stand between me and the 100. Both are five-minute jobs. Both map to real agent-runtime behavior (Link headers for discovery, Markdown content negotiation for efficient agent parsing), so at least the motivation is legitimate and not pure score-chasing. That caveat is also exactly what score-chasers say. Public scores are a gravitational field. Even the person writing a long article about their unreliability ends up orbiting.

Agent Readiness Measures Delivery, Not Message

Every category the Agent Readiness scanner tests is about delivery: discoverability, content negotiation, bot access, API discovery, commerce protocols. None tests the quality of the message itself.

The scanner never asks whether your headlines are clear, whether your product descriptions persuade, whether your content answers the query well, whether your writing is any good. Those are SEO and CRO questions. They occupy the discipline of making the message better. The Agent Readiness Score occupies a different discipline entirely. It asks whether an agent can fetch your content, parse the format it arrives in, authenticate against your endpoints, call your functions, pay for your outputs.

That is the distinction that matters. Classical web optimization (SEO, CRO) is about what you say and how persuasively you say it. Agent-readiness is about how you deliver what you say to a non-human reader. Two websites can publish word-for-word identical content. One serves it as server-rendered HTML with semantic markup, responds to Accept: text/markdown, exposes structured data, returns predictable response codes. The other serves it as a JavaScript-rendered single-page application with no content negotiation and an inconsistent error surface. The message is identical. The delivery is different. The agent-readiness score will be different. And it will be right to be different, because the delivery is what the agent interacts with.

This is also why agent-readiness fixes tend to be orthogonal to SEO and CRO work. You can improve an agent-readiness score without rewriting a single word of your content. You can also have world-class SEO content that scores a 10 on the agent-readiness scanner because none of your delivery pipeline was designed for machine consumers. SEO and CRO work on the content layer. Agent-readiness works on the transport and protocol layer. They are adjacent but not the same craft, and treating them as the same is the mistake that turns an agent-readiness project into a content-rewrite project and misses the actual fix.

The people who will do well over the next several years are the ones who stop arguing about which discipline matters more and start recognizing they occupy different layers of the stack.

3 Goodhart Risks Built Into The Agent Readiness Score

Goodhart’s law says that when a measure becomes a target, it stops being a good measure. The Agent Readiness Score is well-designed, but it is also now a public, shareable, compared number, which produces three predictable behavioral failures in the wild.

The first risk is that website owners will optimize for the number rather than for real agent behavior. Add an MCP Server Card that points nowhere because the scanner wants one. Publish an Agent Skills index with no actual skills. Ship a WebMCP tool that does nothing just to pass the detection check. The score goes up, and nothing changes for real agent runtimes visiting the website.

The second risk is that consultancies will start selling “Agent Readiness Score optimization” as a service, selling the score rather than the underlying architecture. The history of SEO gives us a century of data on how this plays out. PageRank became a target, and a decade of link-spam economy grew up around it. Core Web Vitals became a target, and a generation of performance-theater optimizations followed. The Agent Readiness Score is a better-designed metric than either of those were at launch, but the same gravity applies.

The third risk is that the scanner’s inclusion of emerging standards as scored signals will accelerate the adoption of those standards past the point where they are ready to carry real traffic. The scanner checks for llms.txt, a proposed format for exposing website content to language models. Llms.txt is not a ratified standard, has no governing body, and has competing proposals for how it should be structured. Including it as a scored signal gives it weight it has not earned in the ecosystem. A website owner looking to fix a failing check is the marginal adopter who tips a proposal into a de facto standard before the spec work is done.

None of these failure modes are hypothetical. They are how every public measurement score in the history of the web has played out. The Agent Readiness Score is better than most because Cloudflare is honest about what it is, because the per-check detail is available right alongside the composite number, and because the Commerce category correctly excludes itself on non-commerce websites. That honesty is a feature worth protecting. Website owners and the consultancy industry will be tempted to treat the composite number as the target anyway.

Do not do this.

6 Weekend Fixes That Map To Real Agent Runtimes

Six actions for a web professional running the scanner the weekend of its launch, ordered from highest-leverage to lowest:

  1. Run the scan on your website. It takes about 30 seconds. Note the score and open the detailed report. The detail is where the signal is.
  2. Fix the failing checks that ship against real agent runtimes today. These are the ones whose absence measurably hurts your website for agents visiting it right now:
    • robots.txt. If missing, add one. If present, make sure it contains specific rules for AI user agents (GPTBot, ClaudeBot, PerplexityBot, Google-Extended, etc.).
    • sitemap.xml. If missing, generate one and link it from robots.txt. Keep it current.
    • Markdown content negotiation. Configure your origin or CDN to return text/markdown when the Accept header requests it. Cloudflare’s own AI Crawl Control has first-class support for this. Other providers require custom server logic.
    • Structured data. Ship schema.org JSON-LD for the content types your website publishes (Article, Product, Organization, BreadcrumbList). This is not a scored check, but it is the highest-leverage fix for citation behavior across every agent runtime currently deployed.
  3. Treat the proposal-stage formats as a watch list, not a checklist. llms.txt, Content Signals in robots.txt, Web Bot Auth, API Catalog, MCP Server Card, Agent Skills, WebMCP, ACP, UCP are all real working standards in some sense. They are not shipping against real agent-runtime behavior at scale yet. Watch them. Implement them when your stack has a reason to, not because the scanner flags them.
  4. Ignore the composite number in your own tracking. Track individual check outcomes over time. A website that goes from 3 of 5 real-runtime checks passing to 5 of 5 has measurably improved, even if the composite score barely moved because the 10 proposal-stage checks still fail.
  5. Re-scan after changes. The scanner is fast, free, and available via the URL Scanner API if you want to script regression checks into your deployment pipeline.
  6. Skip the consultancies selling Agent Readiness Score optimization. The work is straightforward enough that a half-day audit and a focused remediation sprint will beat any packaged service.

The scanner is the tool. The work is still the work.

Vendor-Specific Scanners Are Coming: Track What Every Scanner Tests

The Agent Readiness scanner is standards-list-shaped: a set of checks against a fixed list of protocols and formats, some ratified (RFC 8288 Link headers, RFC 9309 robots.txt rules, RFC 8414 OAuth discovery, RFC 9727 API Catalog, RFC 9728 OAuth Protected Resource), some emerging proposals (MCP SEP-1649, WebMCP, Content Signals, Web Bot Auth, x402, UCP, ACP, llms.txt). The next thing that happens in the ecosystem is predictable: Other vendors will ship their own scanners against their own preferred lists. The overlap will be significant because most of the ratified standards are uncontroversial. The divergence will be in which proposals each vendor scores for.

That divergence is where the agent-readiness measurement story gets interesting. A Cloudflare scanner that checks for Web Bot Auth and UCP is making a bet. A Google scanner, if it ships, would check for some of the same things and some different ones (Google has UCP, does not have Web Bot Auth). A Perplexity scanner would check for yet another set. Website owners would see different scores from different scanners on the same website. The composite number, already not trustworthy, becomes vendor-specific.

The signal worth tracking is which checks show up in every scanner that ships. Those are the de facto standards. The checks that only show up in Cloudflare’s scanner are Cloudflare’s bets. Some will win. Most will not.

This is the pattern that made me comfortable publishing an article about a Cloudflare tool on the day it shipped. The Agent Readiness Score is real. The thesis behind it (agent-readiness is a measurable property) is the right thesis. The specific scorecard is version one of something that is going to have dozens of versions, each reflecting its vendor’s bets. Web professionals should engage with the version-one scorecard, fix what it correctly flags as real, watch what it flags as emerging, and keep their own running list of which checks survive across every scanner that ships in the next six months.

That running list is the real agent-readiness standard. The composite score is the marketing layer.

Run the scan. Read the report. Fix what matters. Watch what might.

More Resources:

This post was originally published on No Hacks.

Featured Image: RobinRmD/Shutterstock

Ecommerce MGMT 0 Comments
May 22 2026
3 Unrelated Stories About AI & Writing Tell The Same Story via @sejournal, @gregjarboe

I stumbled upon three separate articles about writing and AI in the same week, each from a completely different angle, and all describing the same thing.

A novelist turned MIT writing lecturer confronting students who outsourced their essays to AI. A new Graphite study showing AI-generated articles now make up roughly half of all new content on the web and have plateaued there. And fresh data from The Accountancy Partnership showing that half of freelance creatives say rising stress is affecting their work, as client budgets for human creative services shrink.

One data point is a fact. Two is a coincidence. Three is a trend.

When read together, these articles formed an argument that every SEO professional, content marketer, and creative freelancer should take seriously, acknowledging the content divide that is happening and asking, “Which side are you on?”

The First Story: What Happens When Students Outsource The Struggle

On May 10, Micah Nathan, a novelist and MIT lecturer in fiction and non-fiction writing, published a piece in The Guardian about confronting his creative writing students over their AI use. The confession session that followed, he wrote, became one of the most productive teaching moments of his eight years at MIT.

His key insight wasn’t about academic honesty. It was about what writing actually does. “Writing isn’t just the production of sentences,” he told his students. “It’s the training of endurance by way of sustained attention. It’s a way of learning what one thinks by attempting to say it. An LLM can reproduce the appearance of that activity, but it can’t replace it, because the value lies not only in the object produced but in the transformation that occurs during its making.”

He described AI prose as “faultily faultless, icily regular, splendidly null,” borrowing Tennyson’s description of a beautiful but empty face, producing what he called “simulacra of thought, generated via pattern recognition learned from millions of human-penned words, rooted in no particular experience by no particular person.”

Insightful readers, he argued, feel that emptiness even if they can’t articulate it.

For SEO professionals, this is not an abstract literary concern. It is a precise description of the content quality problem that Google’s helpful content systems have been trying to solve since 2022. The signal Google is hunting for is exactly what Nathan identifies as the thing AI cannot produce – evidence of a mind actively grappling with a specific problem from a specific experience. Pattern recognition learns from what humans wrote. It cannot replicate why they wrote it. 

→ Read More: Why Great Content Is No Longer Enough & What Beats It In AI Search

The Second Story: The Feared Takeover Hasn’t Happened – Yet

On May 15, Megan Morrone reported for Axios on new data from digital marketing agency Graphite, which analyzed 55,400 online articles and listicles published between January 2020 and March 2026, running each through three AI-detection tools. The finding was more nuanced than most AI content coverage has been about the share of primarily AI-generated content, which has held near 50% for more than a year and appears to have plateaued.

The feared takeover hasn’t materialized. AI content briefly surpassed human-authored content in late 2024, but the two have stayed roughly equal since.

The important caveat Morrone included is that many articles are no longer written purely by humans or AI. A human may use AI for outlining, drafting, rewriting, or editing, making the line genuinely blurry. Dan Klein, a UC Berkeley professor and AI model CTO, flagged the feedback loop risk. Once models train heavily on AI-generated content, the internet could become a machine that produces low-quality content that trains models that produce more low-quality content.

For SEO professionals, the plateau is reassuring and cautionary in equal measures. The volume panic was overstated. But the quality dilution problem is real and growing, and it creates the same opportunity Nathan identified from the other direction. In a web that is roughly half AI-generated content, content that carries genuine human experience and specific expertise becomes more differentiating, not less.

→ Read More: AI Platform Founder Explains Why We Need To Focus On Human Behavior, Not LLMs

The Third Story: The People Producing This Content Are Under Serious Stress

On May 13, Emma Hull at The Accountancy Partnership directly emailed me data from a new report on creative freelancers across PR, marketing, performing arts, graphic design, photography, and adjacent industries. Half of freelance creatives (50.7%) say rising stress levels are affecting their work. Half (50.2%) say client budget cuts are the biggest challenge they faced in 2025. Over two in five (43.3%) believe AI will negatively affect their sector. Nearly half regularly work unpaid hours each week.

Lee Murphy, Managing Director at The Accountancy Partnership, put it plainly: “Creative work is often closely linked to marketing budgets and discretionary spending. When businesses begin tightening costs, creative services can sometimes be one of the first areas to see reduced investment.”

The irony embedded in these three numbers together is worth reflecting on. Clients are cutting budgets for human creative work at the same time AI is generating roughly half the content on the web, while a professor at MIT is documenting the specific cognitive cost that outsourcing the writing process extracts from anyone who does it, whether a student or a professional.

The freelancers under the most pressure are the ones most tempted to use AI to produce more content faster to compensate for lower rates. The content they produce that way becomes part of the 50% that is indistinguishable from machine output. And content that is indistinguishable from machine output is exactly what the Graphite data and Google’s quality systems are training users and algorithms to discount.

→ Read More: Relying Too Much On AI Is Backfiring For Businesses

What The Pattern Actually Means

The three stories, read together, describe a market in the process of bifurcating. On one side sits high-volume, low-differentiation content produced quickly, priced cheaply, and increasingly hard to distinguish from AI output, regardless of who generated it. On the other sits content that carries specific expertise, direct experience, and the editorial judgment that Nathan’s students were trying to skip past. Content that takes longer, costs more, and is increasingly the only kind that earns meaningful search visibility and reader trust.

This is not a new argument in SEO. What is new is the empirical clarity with which three independent sources from three entirely different disciplines – literary education, web content analysis, and freelance labor economics – are all pointing at the same conclusion in the same week.

Shelley Walsh made the point in her recent Search Engine Journal piece on scaling AI content that the commodity versus non-commodity divide is where the real strategic question lives. The three stories above are evidence that the divide is already here, already measurable, and already affecting people’s livelihoods.

The writers who understand this, and produce accordingly, are the ones who will still have work worth doing when the budget cycles turn again.

More Resources:

Featured Image: SvetaZi/Shutterstock

Ecommerce MGMT 0 Comments
May 21 2026
LLM Guidance Doesn’t Transfer The Way SEO Guidance Did via @sejournal, @DuaneForrester

For roughly two decades, the SEO discipline operated on a quiet assumption that turned out to be one of its most valuable features. Guidance from one search engine traveled. If Google said sitemaps mattered, Bing said sitemaps mattered. If Bing said structured data deserved real effort, Google said the same. Practitioners optimized for Google with reasonable confidence that the work would carry across the other engines, and most of the time it did. That portability was not luck. It was the product of a structurally large overlap layer that the major search engines had jointly built, brick by brick, over twenty years.

That world doesn’t exist in LLM-land. The major providers train on different corpora, run different crawlers under different policies, route different queries through different retrieval systems, and apply different alignment processes that shape the final response in ways the upstream signals can’t predict. Guidance from any one provider, including Google’s guidance about its own Gemini products, is one data point. Practitioners carrying the SEO habit forward, the habit of treating one engine’s guidance as roughly the whole map, will optimize confidently for one platform and miss the others.

Sidebar: As I was finalizing this piece, Google published fresh guidance on optimizing for their generative AI features. Their framing is explicit: from Google Search’s perspective, optimizing for AI search is still SEO. That framing is accurate for Google Search. It does not extend to ChatGPT, Claude, Perplexity, or any other LLM, and that is precisely the trap this article is about.

The Shared Standards That Made SEO Guidance Portable

The era of portable guidance was built on actual collaboration, not coincidence. The Sitemaps protocol became the joint property of Google, Yahoo, and Microsoft in November 2006, when the three engines formally agreed to support a common protocol at version 0.90, building on Google’s earlier Sitemaps 0.84 from June 2005. Five years later, on June 2, 2011, the same three engines launched Schema.org, with Yandex joining shortly after, to create a common vocabulary for structured data markup. That was the announcement that got made on stage at SMX Advanced. I was on the Bing team at the time, and what struck me then is what still matters now. The engines were competitors, but they had decided that a shared vocabulary served them all. Webmasters got one set of rules. The web got cleaner data. The engines got better signals. Everybody won.

The pattern repeated with robots.txt, the 1994 convention that became RFC 9309 at the IETF in 2022, formalizing what every serious crawler already honored. And it repeated again, more recently, with IndexNow, the protocol Microsoft Bing and Yandex launched in October 2021. IndexNow is now supported by Bing, Yandex, Naver, Seznam, and Yep. Google has tested the protocol since 2021, but has not adopted it.

That overlap layer is exactly why Google’s guidance felt safe to follow, even if you cared about Bing traffic. The signals the engines used were not identical, but the inputs they accepted, the protocols they honored, and the standards they advertised were. Optimization had a shared substrate.

Where The LLM Stacks Actually Diverge

The LLM environment doesn’t have a shared substrate of comparable size. The differences are not cosmetic, and they are not temporary. They are baked into how the systems are built.

Start with training data. OpenAI has signed disclosed licensing deals with News Corp worth up to $250 million over five years, Axel Springer at roughly $13 million per year, Reddit at an estimated $70 million per year, plus the Financial Times, Condé Nast, Hearst, Vox Media, The Atlantic, the Associated Press, Le Monde, and others. Google has its own Reddit deal, estimated at $60 million per year, granting real-time data API access. Anthropic has not publicly disclosed equivalent publisher licensing deals, and that undisclosed status is itself the practitioner-facing point. The corpora that fed these models, and that continue to refresh them, are not the same documents. Practitioners cannot know what any given provider has paid for and what it hasn’t.

The crawler infrastructure diverges next. OpenAI runs three separate bots: GPTBot for training, OAI-SearchBot for search indexing, and ChatGPT-User for user-initiated retrieval. Anthropic runs three of its own: ClaudeBot for training, Claude-SearchBot for search, and Claude-User for user-initiated retrieval. Perplexity runs PerplexityBot and Perplexity-User. Google introduced Google-Extended in September 2023 as the user-agent that controls whether Google can use a site’s content to train Gemini, separate entirely from the Googlebot that handles traditional search indexing. There is no single AI user-agent. Every provider requires a separate rule, and the rules don’t translate cleanly across providers because the bots don’t do equivalent jobs in equivalent ways.

The retrieval architectures diverge structurally. ChatGPT has historically used Bing’s index as its primary web search source, and that connection appears to still be primary, though OpenAI continues to build out additional infrastructure alongside it. Perplexity built its retrieval system on a Vespa-based pipeline that treats documents and sub-document chunks as first-class retrievable units. Google’s Gemini uses Google’s own index plus Knowledge Graph grounding. Claude uses Brave Search as a retrieval partner. Same query, four different retrieval systems, four different views of which sources exist and which sources are worth surfacing.

Then comes the alignment layer, which is where SEO had no equivalent at all. After a model is trained on its corpus, providers run post-training to shape how the model actually behaves: tone, refusal patterns, format, safety posture, what counts as a good answer. OpenAI’s primary approach has been RLHF, or Reinforcement Learning from Human Feedback, where human raters score model outputs and the model learns to produce highly rated responses. Anthropic developed Constitutional AI, which trains models to critique and revise their own outputs against a written set of principles. These methodologies produce demonstrably different behavior in the final products. The same retrieved content, fed into two models aligned by two methodologies, can yield two materially different responses about the same brand.

When One Provider’s Guidance Demonstrably Fails To Port

The clearest single example of guidance that doesn’t port is llms.txt. Jeremy Howard of Answer.AI proposed the file in September 2024 as a markdown manifest, placed at a site’s root, that would guide LLMs to the most important content. The proposal got picked up across the SEO community. Yoast built a generator. Agencies added llms.txt creation to their service catalogs. Conference speakers declared it essential.

As of mid-2026, no major LLM provider has confirmed they consume the file. Not OpenAI. Not Anthropic. Not Google. Server-log analyses across hundreds of thousands of domains show major AI crawlers don’t routinely request /llms.txt at all. Google’s John Mueller publicly compared it to the deprecated meta keywords tag. Gary Illyes confirmed at Search Central Live in July 2025 that Google does not support llms.txt and is not planning to.

I’ve written about this elsewhere, so I won’t repeat the technicalities here. What matters for this argument is the structural lesson. Schema.org succeeded because three engines built it together and then enforced it together. Llms.txt was proposed by one researcher, picked up by tooling vendors, and ignored by the platforms it was supposed to serve. The shared-standards model that gave SEO its portable guidance is not available to LLM practitioners at the same scale, because the platforms are not building the standards together. They are building their own pipelines.

The Gemini Inversion

The cleanest illustration of how far guidance portability has degraded sits inside one company. Google publishes its own SEO documentation at Search Central, the canonical guidance the industry has followed for two decades. Those documents emphasize traditional ranking signals, E-E-A-T, content quality, technical accessibility, and structured data. That guidance is still useful for Google Search itself.

Google also makes Gemini, the model that powers AI Overviews and Google’s separate AI Mode surface. And the citation behavior of those surfaces does not appear to track the guidance the same company publishes for its own search results.

In late 2024, roughly three-quarters of pages cited in AI Overviews also ranked in Google’s top 12 for the same query. By early 2026, after Google upgraded AI Overviews to Gemini 3 in January, Ahrefs analyzed 4 million AI Overview URLs and found that only 38% of cited pages also appeared in the top 10 for the same query. A separate BrightEdge analysis put the overlap closer to 17%. SE Ranking’s post-upgrade work found that Gemini 3 replaced approximately 42% of the domains previously cited under earlier model versions and generates 32% more sources per response.

The gap widens further when you look at Google’s AI Mode, which is a separate conversational surface that runs on the same Gemini family. Semrush data shows AI Mode and AI Overviews reach semantically similar conclusions 86% of the time, but cite the same URLs only 13.7% of the time. Only 14% of AI Mode citations rank in Google’s traditional top 10.

It appears, so far, that the canonical relationship has shifted. Google’s published SEO guidance is still the cleanest path to ranking in Google Search. But that ranking is no longer a reliable proxy for being cited by Google’s own AI surfaces. The same guidance, the same content, the same domain, can produce three meaningfully different outcomes across Google Search, AI Overviews, and AI Mode, even though all three live inside the same company. The old playbook of following the search engine’s guidance and trusting that the engine’s other surfaces would behave consistently does not appear to be delivering the same returns it used to.

What Still Ports, And Why It’s Smaller Than It Looks

A universal layer does survive. Crawler accessibility still matters across every provider. Primary-source factual content still wins more citations than aggregator restatement. Clean retrievable structure still helps every system understand what a page is about. Presence on the high-authority sources that all major LLMs disproportionately cite, Wikipedia, YouTube, Reddit, major news outlets, still functions as a force multiplier across platforms. Earning visibility on those sources gives content a chance to surface in any LLM that draws on them.

But the universal layer is much smaller than it was in the SEO era. Qwairy’s analysis of 118,000 AI responses across ChatGPT, Perplexity, Google AI Mode, and Claude found that only 11% of cited domains appeared across multiple platforms. The other 89% were platform-specific. A brand that wins citations on Perplexity may be largely invisible on Claude. A brand that’s a regular reference on ChatGPT may not show up in AI Overviews at all. The same content can be the right answer for one system and the wrong answer for the system next to it.

What This Means For The Work

The practical implication is not abandoning all hope. It is that practitioners need to stop treating any single LLM provider’s guidance as the universal map and start treating it as one input among several. Read what every major provider publishes about their own systems. Test your visibility across platforms, not just on the platform you happen to use most. Treat divergence as the default and overlap as the exception, not the other way around.

This is not how SEO worked, and the difference matters. The old reflex was to optimize for Google and trust the portability. The new reality is that following one LLM’s guidance, even Google’s guidance about Gemini, will leave you optimized for a slice of the landscape and potentially blind to the rest. The discipline is being rebuilt on platform-specific work that didn’t exist in the SEO era, and the practitioners who recognize that first are going to spend the next two years setting the standards everyone else follows.

The overlap has shrunk. You now have more work than ever to accomplish.

If you have thoughts on where the divergence between providers is sharpest in your own work, reach out directly. I’d genuinely like to hear what’s showing up in the data.

More Resources:

This post was originally published on Duane Forrester Decodes.

Featured Image: Rawpixel.com/Shutterstock; Paulo Bobita/Search Engine Journal

Ecommerce MGMT 0 Comments
May 21 2026
Microsoft Clarity Now Shows Grounding Queries Behind AI Citations via @sejournal, @TaylorDanRW

When Microsoft Clarity made AI citations available to all users, it opened up a new playground for SEOs to harvest AI visibility data. Finally, we can see the exact “grounding queries” an AI engine uses to pull our content.

It raises a massive question because this is a Microsoft tool: Are the insights useless if your audience doesn’t touch the Bing ecosystem?

Microsoft Clarity Grounding Queries

When you ask Copilot a question, it translates your words into simple search terms called grounding queries to find facts on the web before it answers. You can use this data to improve your own website and content.

  • Finding gaps where your content does not match what the AI searches for.
  • Simplifying pages that the AI reads but does not link to.
  • Using these simple layouts to help your Google search results.

Copilot Vs. Gemini

Both Copilot and Gemini use retrieval-augmented approaches. Instead of generating answers using only pre-trained parameters, they dynamically query external search indexes to retrieve real-time data, which they then use as context to ground their final responses.

Feature Microsoft Copilot Google Gemini
Structure Uses a query translator, Bing index search, and OpenAI models to write the final text. Uses a query translator, Google Search, and Google’s Gemini models to write the final text.
Pulling Sources Uses the Bing index and Microsoft Graph to scan web pages, emails, and Microsoft 365 files. (With permissions enabled) Uses Google Search and Google Workspace to scan web pages, Google Drive files, and Gmail. (With permissions enabled)
Synthesising Answers Focuses on direct answers. It uses structured lists, tables, and bullet points to show facts quickly. Focuses on creative, conversational answers. It is built to handle text, images, and code at the same time.

Does Ranking In Bing Matter?

Yes (Correlation).

One of my websites was doing extremely well in Copilot, with over 36,000 citations across all queries. Now, Clarity doesn’t give you the prompts/queries themselves, but it does give you the Grounding queries (grounding queries and key phrases used to retrieve your site’s content).

Image from author, May 2026

My website has a history, running for years with a previous domain merged in 2019, and boasts over 1,000 articles. Given that Google barely sends traffic, and third-party SEO tools often label it as spam due to non-English backlinks (it covers search engines like Baidu, CocCoc, SwissCows, attracting an international audience), I never expected 36,000 citations.

So, why the Copilot love? I took the 147 grounding queries and tracked their rank in Google and Bing.

Image from author, May 2026

Of the 147 queries, Bing ranked all but 6, with the majority in traffic-driving positions (top 20). Google didn’t rank a single one.

So, If This Is Heavily Dependent On Bing Indexing, Is Clarity’s Data Useful Outside Of The Bing/Microsoft Ecosystem?

Because this is a Microsoft tool, the backend data feeding this dashboard is primarily capturing how your site is cited across Microsoft’s AI surfaces (like Copilot and Bing generative search).

It is not giving you a direct window into how OpenAI’s ChatGPT (using its own search), Google Gemini, or Perplexity are citing your links, because those platforms do not share their internal grounding logs with Microsoft.

And historically, we as an industry have been neglectful of Bing.

Even though the data collection source is skewed toward Microsoft’s AI engine, the insights themselves are highly transferable to your broader, platform-agnostic AI optimization strategies.

Can We Assume Other LLMs Retrieve Data In The Same Way?

AI engines, whether Google Gemini or Microsoft Copilot, use similar RAG frameworks to fetch data.

If the Bing ecosystem flags that a specific page on your site has a high “Share of Authority” for a complex query, it means that page is structured perfectly for AI consumption (clear tables, bullet points, direct answers). Data suggests that you can replicate that formatting across your site to appeal to Google Gemini as well.

However, this can be argued against as other research suggests that the similarity between LLMs is dependent on positional biases, and some may use the SDSR method rather than RAG.

Researchers in SEO have also found that ChatGPT has started to use Google Search as a fallback, when it was initially Bing.

In Summary

If your audience doesn’t touch the Microsoft ecosystem, this dashboard won’t give you a perfect 1-to-1 reflection of your total AI traffic, but it doesn’t make the data useless.

What grounding queries reveal is how AI systems distill user intent into retrievable search terms. That process is broadly consistent across platforms, even when the underlying indexes differ. A page earning citations in Copilot is doing something right structurally with clear answers, well-scoped topics, content aligned with how AI engines translate questions into queries. The Bing dependency tells you where the data comes from. The structural patterns tell you something more transferable.

The gap data is equally instructive. Pages your site ranks for in Bing that never appear as grounding queries signal a mismatch. Either the content isn’t structured for AI retrieval, or the topic isn’t one AI engines are actively grounding answers around.

Treat Clarity’s Citations dashboard as a useful proxy or “lab environment” and window into how LLMs read, slice/chunk, and credit your website’s content. Even if Copilot isn’t your primary AI traffic source, the patterns it surfaces are worth paying attention to.

More Resources:

Featured Image: Prostock-studio/Shutterstock

Ecommerce MGMT 0 Comments
May 20 2026
Google Shares First AI Mode Usage Data After One Year via @sejournal, @MattGSouthern

Google released a report detailing how people use AI Mode in the U.S., drawing on internal Search data and Google Trends to map search behavior one year after launch.

The report, published alongside Google I/O 2026 announcements, said that AI Mode has surpassed 1 billion monthly active users globally. Queries have more than doubled every quarter since launch.

How Query Behavior Is Changing

The report states that the average AI Mode search is three times longer than a traditional search. Both short and long queries are increasing in AI Mode, with users having conversations and asking longer questions.

Follow-up queries in AI Mode rose over 40% monthly in the U.S. More than one in six AI Mode searches are multimodal, using voice, images, or video. Image-based searches are up over 40% month-over-month since launch.

Top keywords include “information,” “identify,” “find,” “explain,” and “summarize.” Common first words are “what,” “how,” “I,” “is,” and “can,” with “I” especially notable, which may suggest people treat AI Mode more like a conversation than a traditional search.

What People Search For

Google grouped AI Mode search topics into five categories: Explore, Decide, Learn, Create, and Do. The top 10 topics include creative content, media, education, fashion, food, health, tech, travel, productivity, and development.

Brainstorming queries increased 30% faster than overall AI Mode queries since launch, with searches for “where to,” “where should I,” and “ideas for” also rising, per Google Trends.

Planning-related queries grew 80% faster over six months, with decision questions starting with “which” increasing 40%, especially “which of” and “which one.”

Shopping And Local Behavior

Shoppers start with traditional search, then move to AI Mode for deeper inquiry, especially in electronics, books, apparel, health and beauty, and automotive.

In AI Mode, store-related questions focus on “near me,” replacement parts, financing-related dealership searches, online options, and stock.

Top retail concerns include price, location, color, brand, and availability. For restaurants, users seek kid-friendly options, views, bars, vegan or vegetarian choices, and outdoor seating.

Creative And Educational Use

AI Mode’s image creation queries have more than tripled since early 2026, with users mainly requesting photos, quizzes, logos, stories, and code, as well as editing photos, documents, videos, messages, and code.

For education, top subjects include math, Spanish, history, English, and biology, while professional development searches focus on Security+, black belt, Network+, bar exam, and real estate license.

Why This Matters

The data shows AI Mode users are searching in ways that don’t map cleanly to traditional keyword patterns. Queries are longer, conversational, and increasingly multimodal. Follow-up conversations are growing, and planning and decision queries are among the strongest growth signals in the report.

If query length and follow-ups keep growing, that means thin content faces a different competition than conversational answers to multi-part questions.

Looking Ahead

Google released this report the same week it announced Gemini 3.5 Flash as the new default model in AI Mode, redesigned the Search box, and previewed search agents for this summer.

The keyword and query data covers May 2025 to April 2026 and comes from a random, unbiased sample of Google searches. The Trends data measures search interest as a share of AI Mode searches, not total query volume. AI Mode Trends data is not publicly available on trends.google.com.

Ecommerce MGMT 0 Comments