OpenAI is experiencing a widespread outage affecting two systems, APIs and ChatGPT. The outage has been ongoing for at least a half an hour as of publication date.
ChatGPT API Jobs Stuck Outage
The first issue is that batch API jobs get stuck in the finalization state. There are twelve components of APIs that are monitored for uptime and it’s the Batch part that’s experiencing “degraded” performance. The issue has been ongoing since 3:54 PM.
This post was sponsored by Weglot. The opinions expressed in this article are the sponsor’s own.
When Google’s AI Overviews launched in 2024, dozens of questions quickly surfaced among SEO professionals, one being: if AI now curates and summarizes search results, how do websites earn visibility, especially across languages?
Weglot recently conducted a data-driven study, analyzing 1.3 million citations across Google AI Overviews and ChatGPT to determine if LLMs cite content in one language, would they also cite it in others?
The result: translated websites saw up to 327% more visibility in AI Overviews than untranslated ones, a clear signal that international SEO is becoming inseparable from AI search.
What’s more, websites with another language available were also more likely to be cited in AI Overviews, regardless of the language the search was made.
The Changing Nature of Search
This shift is redefining the rules of visibility. AI Overviews and large language models (LLMs) now mediate how information is discovered. Instead of ranking pages, they “cite” sources in generated responses.
But with that shift comes a new risk: if your website isn’t available in the user’s search language, does AI simply overlook it, or worse, send users to Google Translate’s proxy page instead?
The risk with Google’s Translate proxy is that while it does the translation work for you, you have no control over the translations of your content. Worse still, you don’t get any of the traffic benefits, as users are not directed to your site.
The Study
Here’s how the research worked. To understand how translation affects AI visibility, Weglot focused the research on Spanish-language websites across two markets: Spain and Mexico.
The study was then split into two phases. Phase one focused on websites that weren’t translated, and therefore only displayed the language intended for their market, in this case, Spanish.
In that phase, Weglot looked at 153 websites without English translations: 98 from Spain and 55 from Mexico. Weglot deliberately selected high-traffic sites because they offered no English versions.
Phase two involved a comparison group of 83 Spanish and Mexican sites with versions in both Spanish and English. This allowed Weglot to directly compare the performance of translated versus untranslated content.
In total, this generated 22,854 queries in phase one and 12,138 in phase two. The methodology converted the top 50 non-branded keywords of each site into queries that users would likely search, and then these were translated between the Spanish and English versions.
In total, 1.3 million citations were analyzed.
The Key Results
Untranslated Sites Have Very Low AI Search Visibility
The findings show that untranslated websites experience a substantial drop in visibility for searches conducted in non-available languages, despite maintaining strong visibility in the current available language.
Diving deeper into this, untranslated sites essentially lose massive visibility. From the study, even when these Spanish websites performed well in Spanish searches, the sites virtually disappeared in English searches.
Looking at this data further within Google AI Overviews:
The sample size of 98 untranslated sites from Spain had 17,094 citations for Spanish queries vs 2,810 citations for the equivalent search in English, a 431% gap in visibility.
Taking a look at untranslated sites in Mexico, the study identified a similar pattern. 12,038 citations for Spanish queries vs 3,450 citations for English, showing 213% fewer citations when searching English.
Even ChatGPT, though slightly more balanced, still favored translated sites, with Spanish sites receiving 3.5% fewer citations in English and 4.9% fewer with Mexican sites.
Image created by Weglot, November 2025
Translated Sites Have 327% More AI Search Visibility
But what happens when you do translate your site?
Bringing in the comparison group of Spanish websites that also have an English version, we can see that translated sites dramatically close the visibility gap and that having a second language transformed visibility within Google AI Overviews.
Google AI Overviews:
Translated sites in Spain saw 10,046 citations vs 8,048 in English, showcasing only a 22% gap.
Translated sites in Mexico showed 5,527 citations for Spanish queries and 3,325 citations for English, and a difference of 59%.
Overall, translated sites achieved 327% more visibility than untranslated ones and earned 24% more total citations per query.
When looking at ChatGPT, the bias almost vanished. Translated sites saw near-equal citations in both languages.
Image created by Weglot, November 2025
Next Steps: Translate Your Site To Boost Global Visibility In AI SERPs
Not only does having multiple languages across your site ensure your site gets picked up for searches in multiple languages, but it also adds to the overall visibility of your site as a whole.
The study found that translated sites perform better across all metrics. The data shows that translated sites received 24% more citations per prompt than untranslated sites.
Looking at this by language, translation resulted in a 33% increase in English citations and a 16% increase in Spanish citations per query.
Weglot’s findings indicate that translation acts as a signal of authority and reliability for AIOs and ChatGPT, boosting citation performance across all languages, not only the ones content is translated.
Image created by Weglot, November 2025
AI Search Rewards Translated Content as a Visibility Signal
Traditional international SEO has long focused on hreflang tags and localized keywords. But in the age of AI search, translation itself becomes a visibility signal:
Language alignment: AI engines prioritize content matching the query’s language.
Traffic control: Proper translations prevent Google Translate proxies from intercepting clicks.
Semantic reach: Multilingual content broadens your surface area for AI training and citation.
Put simply: If your content isn’t in the language of the question, it’s unlikely it will be in the answer either.
The Business Impact
The consequences aren’t theoretical. One case in Weglot’s dataset, a major Spanish book retailer selling English-language titles worldwide without an English version of its site, shows the impact.
When English speakers searched for relevant books:
The site appeared 64% less often in Google AI Overviews and ChatGPT.
In 36% of the cases where it did appear, the link pointed to Google Translate’s proxy, not the retailer’s own domain.
Despite offering exactly what English users wanted, the business lost visibility, traffic, and ultimately, sales.
The Bigger Picture: AI Search Is Redefining SEO and Translation Is Now a Growth Strategy
The implications reach far beyond Spain or Mexico, or even the Spanish language.
As AI search evolves, the SEO playbook is expanding. Ranking isn’t just about “position one” anymore; it’s about being cited, summarized, and surfaced by machines trained on multilingual web content.
Weglot’s findings point to a future where translation is both an SEO and an AI strategy and not a localization afterthought.
With Google AIOs now live in multiple languages and ChatGPT integrating real-time web data, multilingual visibility has become an equity issue: sites optimized for one language risk being invisible in another.
Image created by Weglot, November 2025
Final Takeaway: Untranslated Sites Are Invisible in AI Search
The evidence is clear: Untranslated = unseen. Website translation is high up there for AIO visibility.
As AI continues to shape how search engines understand relevance, translation isn’t just about accessibility; it’s how your brand gets recognized by algorithms and audiences alike.
For the easiest way to translate a website, start your free trial now!
Plus, enjoy a 15% discount for 12 months on public plans by using the promo code SEARCH15 on a paid plan purchase.
This week’s news in SEO brings changes and questions about control.
Google’s shopping AI moved from showing where to buy to completing purchases directly. Google added structured data for merchant shipping policies. OpenAI released GPT-5.1 with personality controls.
And the EU opened an investigation into Google’s site reputation abuse enforcement. Raising the question, should one gatekeeper control how independent media funds online journalism?
Here’s what you need to know for this week in SEO.
Google’s Shopping AI Completes Transactions Without Your Site
Google rolled out Gemini-powered shopping features that find products, compare prices, and handle checkout across multiple retailers.
AI Mode in Search can now automate checkout on participating merchants’ sites using your saved Google Pay details, so you don’t have to manually fill payment and shipping forms.
Key Facts: AI Mode shopping is launching in Search. Agentic checkout works with select retailers. An AI calling feature can confirm stock, price, and availability with local stores. All features are U.S.-only and gradually rolling out.
Why SEOs Should Pay Attention
Google’s moving from showing where to buy things to completing transactions for you. This changes what “search” means for ecommerce sites.
When AI Mode handles checkout across retailers, your website becomes optional infrastructure. Users never see your brand presentation, never encounter your upsells, never make decisions on your pages. Google’s AI extracts the transaction with your site reduced to inventory management.
The local business calling feature shows where this goes. If Gemini calls five restaurants to check availability, users never see your website, reviews, or menu.
The impact goes beyond rankings to the transaction itself. Your SEO strategy is optimized for driving traffic where users make decisions. Google’s building an environment where AI makes those decisions using your business as a data source, not a destination.
Google Adds Structured Data For Merchant Shipping Policies
Google launched support for merchant shipping policy structured data, letting ecommerce sites describe shipping costs, delivery times, and regional availability so they can surface directly in search results.
The markup can appear alongside your products and in relevant knowledge panels for eligible merchants.
Key Facts: Shipping policy structured data appears with eligible merchant listings, with no geographic limits. It supports flat rate, free, and calculated shipping, including delivery times and regional restrictions. Best used with Product structured data for search results. Validation requires Rich Results Test or URL Inspection, as no specific Search Console report exists.
Why SEOs Should Pay Attention
Shipping costs affect purchase decisions before users reach your site. Displaying this information in search results answers a primary objection at the discovery stage.
The markup lets you differentiate on shipping when competitors don’t show it. If you offer free shipping or faster delivery, you can now surface that advantage in search results rather than hoping users click through to find out.
Implementation is straightforward if you already use Product markup. Add the shipping policy structured data to your existing schema and specify rates, zones, and delivery times. This is one of the more actionable structured data updates Google’s released this year.
OpenAI Releases GPT-5.1 With User-Controlled Personality
OpenAI shipped GPT-5.1 models with customizable personality controls and improved instruction following. Users can now adjust ChatGPT’s tone through preset styles or granular characteristic tuning instead of the previous one-size-fits-all approach.
Key Facts: GPT-5.1 is now available first to paid users (Pro, Plus, Go, Business), with free access following. Adaptive reasoning optimizes processing time based on query complexity. Legacy GPT-5 models stay available for three months.
Why SEOs Should Pay Attention
You can now customize ChatGPT’s output style to match your needs, rather than editing around the default tone.
The adaptive reasoning means faster responses on simple queries without sacrificing quality on complex requests.
The three-month legacy model availability gives you time to test whether GPT-5.1 performs better for your specific use cases before GPT-5 sunsets.
EU Challenges Google’s Parasite SEO Crackdown
The European Commission opened a Digital Markets Act investigation into whether Google’s site reputation abuse enforcement discriminates against news publishers.
Key Facts: Publisher groups in France, Germany, Italy, Poland, Spain, and EU countries report significant traffic drops. DMA violations can incur fines up to 10% of global revenue, rising to 20% for repeats.
Why SEOs Should Pay Attention
The investigation exposes the tension publishers face. Google’s definition of “spam” now includes their revenue model.
Publishers aren’t defending payday loan scams on university domains. They’re arguing that sponsored content with editorial oversight shouldn’t be treated the same as affiliate coupon pages designed purely for ranking manipulation.
Google’s position treats business arrangements as ranking signals rather than judging content quality.
If the EU forces exemptions for “legitimate” sponsored content, every spammer will dress up their tactics with editorial oversight theater. The policy only works if it draws lines. But publishers aren’t wrong to question whether one gatekeeper should control how independent media funds journalism.
This Week In SEO: The Balance Of Power Is Shifting
The news from this week tells a bigger story: Search engines are no longer just organizing the web; they’re absorbing it.
The theme of the week? Power and control.
Google’s AI is deciding what users buy and what content deserves to be seen. OpenAI is letting creators shape the AI’s voice for the first time. And regulators are finally asking who gets to define “fair” visibility online.
As AI reshapes discovery, SEOs face the challenge of staying visible in a world where the search interface itself has become the destination.
This week, the paid media world focused less on new tools and more on what’s changing beneath the surface.
Google expanded Performance Max into a new channel and offered long-awaited reporting visibility. Microsoft took a firm stance on brand safety by requiring Clarity across its publisher network. And one viral LinkedIn post questioned the effectiveness of Google’s newest “AI-powered” campaign model.
Each of these stories points to the same theme: Platforms are redefining what control and accountability mean for advertisers.
Performance Max Expands To Waze And Adds Channel Reporting
Google confirmed two changes for Performance Max campaigns.
The first notable update is that for PMax campaigns using “Store Visits” as a campaign goal, your business can now show up on Waze ads inventory. The business will show up as a “Promoted Places in Navigation” pin for users.
This update is for all advertisers in the United States, and no additional setup is required.
The second update is that Google rolled out Channel Reporting for all PMax campaigns. While this has been rolling out for a few months now, not every advertiser had this available.
Why Advertisers Should Pay Attention
Local intent now includes the navigation moment. If your brand depends on foot traffic, showing up while someone is driving near a location adds a fresh, real-world touchpoint.
The channel reporting update matters just as much. It helps shift PMax conversations from “trust the system” to “here’s where the system actually worked.”
In my opinion, this is progress on transparency and reach. It also adds variables you’ll be asked to explain.
The win isn’t “more placements.” The win is being able to connect surfaces to outcomes with fewer leaps of faith.
Microsoft Clarity Now Mandatory For Third-Party Publishers
Microsoft Ads Liaison, Navah Hopkins, shared an important announcement for all 3P publishers on Microsoft:
Screenshot taken by author, November 2025
In her post, she mentions that all Microsoft Ads clicks need to make sure those pages have Microsoft Clarity enabled.
Her post got attention from the PPC industry, where she clarified in the comments that an official announcement from Microsoft will be coming out shortly. All Microsoft Ads partners have already been notified via email.
The post also sparked some questions and potential confusion about how Microsoft Ads wouldn’t be charged if they weren’t running Clarity.
Andy Hawes asked:
Thanks for this Navah Hopkins, but when you say “Any Microsoft Advertising clicks that do not have Clarity will be filtered out and result in nonbillable impressions/clicks.” Are you suggesting that if you don’t run clarity then you’re Microsoft Ads won’t cost anything? I’m assuming that is not the case? So could you explain that part please?
Hopkins clarified during the exchange:
Screenshot taken by author, November 2025
Why Advertisers Should Pay Attention
Microsoft seems to be taking a quality stance, not just making a tracking footnote.
Based on the conversation on LinkedIn, Microsoft is tying billable media to verifiable on-site experience. In theory, that should reduce questionable placements and give brands greater confidence that their ads appear in environments that meet baseline standards.
I see this as Microsoft is trading raw reach for higher trust. Advertisers should expect fewer gray-area placements and stronger conversations with brand-safety teams.
It also nudges the market toward a new normal where “transparency” includes a window into on-site behavior, not just a placement report.
The Industry Reacts To AI Max Performance Data
AI Max was another hot topic on LinkedIn this past week.
Xavier Mantica shared four months of results comparing AI Max to traditional match types.
Screenshot taken by author, November 2025
His data showed AI Max at $100.37 per conversion versus $43.97-$61.65 for most non-AI setups (and $97.67 for phrase close variants). His view: AI Max behaves like broad match with a new label, expanding beyond intended relevance and driving up cost.
As of this writing, the post has 991 engagements with over 170 comments from the PPC industry.
How Advertisers Are Reacting
Looking at the comments, it appears that many PPC pros agree that AI Max isn’t living up to the hype that Google made it out to be when originally announced.
Collin Slatterly, Founder of Taikun, shared his skeptical optimism by not just dismissing AI Max entirely, but shared it may just not be ready for its full potential:
Give it a year, and it’ll probably be ready to deploy. Feels like PMax all over again.
One of the top comments to Xavier’s post came from Mike Ryan, who agreed after analyzing 250 campaigns of his own:
Screenshot taken by author, November 2025
There were others in the comments that had the opposite take of Xavier. Denis Capko replied in the comments, stating:
Screenshot taken by author, November 2025
Why Advertisers Should Pay Attention
This debate goes beyond one account. It reflects a wider tension between volume and control.
“AI increases conversions” is only persuasive if cost, relevance, and repeatability hold up under scrutiny.
While the comments seemed overly negative to AI Max, I see it as AI Max feels more like growing pains than failure.
Automation continues to move faster than the frameworks we use to evaluate it, and advertisers are still learning how to guide it effectively.
When data quality, conversion accuracy, and negative signals are strong, AI Max can deliver meaningful scale. But without clear visibility into how the system interprets intent, results can vary widely.
Posts like Xavier’s highlight the need for transparency as much as performance. Google also benefits from that same openness: It builds trust, helps advertisers use automation more responsibly, and ultimately makes the technology stronger for everyone.
Theme Of The Week: Accountability
The updates and discussions this past week all share one thread: accountability.
Google is expanding where automation can go, Microsoft is tightening the standards for who gets to monetize it, and advertisers are rethinking how much control they’re willing to trade for convenience.
As platforms lean further into automation, the real advantage won’t come from who adopts it first. It will come from who understands it best.
Are you confident in what your automation is doing, or just comfortable letting it run?
Search Console has some pretty severe limitations when it comes to storage, anonymized and incomplete data, and API limits.
You can bypass a lot of these limitations and make GSC work much harder for you but setting up far more properties at a subfolder level.
You can have up to 1,000 properties in your Search Console account. Don’t stop with one domain-level property.
All of this allows for far richer indexation, query, and page-level analysis. All for free. Particularly if you make use of the 2,000 per property API URL indexing cap.
Image Credit: Harry Clarkson-Bennett
Now, this is mainly applicable to enterprise sites. Sites with a deep subfolder structure and a rich history of publishing a lot of content. Technically, this isn’t publisher-specific. If you work for an ecommerce brand, this should be incredibly useful, too.
I and it love all big and clunky sites equally.
What Is A Search Console Property?
A Search Console Property is a domain, subfolder, or subdomain variation of a website you can prove that you own.
You can set up domain-level or URL-prefix-level properties (Image Credit: Harry Clarkson-Bennett)
If you just set up a domain-level property, you still get access to all the good stuff GSC offers. Click and impression data, indexation analysis, and the crawl stats report (only available in domain-level properties), to name a few. But you’re hampered by some pretty severe limitations:
1,000 rows of query and page-level data.
2,000 URL API limit for indexation level analysis each day.
Sampled keyword data (and privacy masking).
Missing data (in some cases, 70% or more).
16 months of data.
While the 16-month limit and sampled keyword data require you to export your data to BigQuery (or use one of the tools below), you can massively improve your GSC experience by making better use of properties.
There are a number of verification methods available – DNS verification, HTML tag or file upload, Google Analytics tracking code. Once you have set up and verified a domain-level property, you’re free to add any child-level property. Subdomains or subfolders alike.
The crawl stats report can be an absolute goldmine, particularly for large sites (not this one!) (Image Credit: Harry Clarkson-Bennett)
The crawl stats report can be extremely useful for debugging issues like spikes in parameter URLs or from naughty subdomains. Particularly on large sites where departments do things you and I don’t find out about until it’s too late.
But by breaking down changes at a host, file type, and response code level, you can stop things at the source. Easily identify issues affecting your crawl budget before you want to hit someone over the head with their approach to internal linking and parameter URLs.
Usually, anyway. Sometimes people just need a good clump. Metaphorically speaking, of course.
Subdomains are usually seen as separate entities with their own crawl budget. However, this isn’t always the case. According to John Mueller, it is possible that Google may group your subdomains together for crawl budget purposes.
According to Gary Illyes, crawl budget is typically set by host name. So subdomains should have their own crawl budget if the host name is separate from the main domain.
How Can I Identify The Right Properties?
As an SEO, it’s your job to know the website better than anybody else. In most cases, that isn’t too hard because you work with digital ignoramuses. Usually, you can just find this data in GSC. But larger sites need a little more love.
Crawl your site using Screaming Frog, Sitebulb, the artist formerly known as Deepcrawl, and build out a picture of your site structure if you don’t already know. Add the most valuable properties first (revenue first, traffic second) and work from there.
Some Alternatives To GSC
Before going any further, it would be remiss of me not to mention some excellent alternatives to GSC. Alternatives that completely remove these limitations for you.
SEO Stack
SEO Stack is a fantastic tool that removes all query limits, has an in-built MCP-style setup where you can really talk to your data. For example, show me content that has always performed well in September or identify pages with a health query counting profile.
Daniel has been very vocal about query counting, and it’s a fantastic way to understand the direction of travel your site or content is taking in search. Going up in the top 3 or 10 positions – good. Going down there and up further down – bad.
SEO Gets
SEO Gets is a more budget-friendly alternative to SEO Stack (which in itself isn’t that expensive). SEO Gets also removes the standard row limitations associated with Search Console and makes content analysis much more efficient.
Growing and decaying pages and queries in SEO Gets are super useful (Image Credit: Harry Clarkson-Bennett)
Create keyword and page groups for query counting and click and impression analysis at a content cluster level. SEO Gets has arguably the best free version of any tool on the market.
Indexing Insight
Indexing Insight – Adam Gent’s ultra-detailed indexation analysis tool – is a lifesaver for large, sprawling websites. 2,000 URLs per day just doesn’t cut the mustard for enterprise sites. But by cleverly taking the multi-property approach, you can leverage 2,000 URLs per property.
Remove the indexation limits of 2,000 URLs per day with the API and the 1,000 row URL limit (Image Credit: Harry Clarkson-Bennett)
All of these tools instantly improve your Search Console experience.
Benefits Of A Multi-Property Approach
Arguably, the most effective way of getting around some of the aforementioned issues is to scale the number of properties you own. For two main reasons – it’s free and it gets around core API limitations.
Everyone likes free stuff. I once walked past a newsagent doing an opening day promotion where they were giving away tins of chopped tomatoes. Which was bizarre. What was more bizarre was that there was a queue. A queue I ended up joining.
Pages that sit in the Crawled – Currently Not Indexed pipeline may not require any immediate action. The page has been crawled, but hasn’t been deemed fit for Google’s index. This could signify page quality issues, so worth ensuring your content is adding value and your internal linking prioritizes important pages.
Discovered – Currently Not Indexed is slightly different. It means that Google has found the URL, but hasn’t yet crawled it. It could be that your content output isn’t quite on par with Google’s perceived value of your site. Or that your internal linking structure needs to better prioritize important content. Or some kind of server of technical issue.
All of this requires at least a rudimentary understanding of how Google’s indexation pipeline works. It is not a binary approach. Gary Illyes said Google has a tiered system of indexation. Content that needs to be served more frequently is stored in a better-quality, more expensive system. Less valuable content is stored in a less expensive system.
How Google crawling and rendering system works (Image Credit: Harry Clarkson-Bennett)
Less monkey see, monkey do; more monkey see, monkey make decision based on the site’s value, crawl budget, efficiency, server load, and use of JavaScript.
The tiered approach to indexation prioritizes the perceived value and raw HTML of a page. JavaScript is queued because it is so much more resource-intensive. Hence why SEOs bang on about having your content rendered on the server side.
Worth noting the page indexation tool isn’t completely up to date. I believe it’s updated a couple of times a week. But I can’t remember where I got that information, so don’t hold me to that…
If you’re a big news publisher you’ll see lots of your newsier content in the Crawled – Currently Not Indexed category. But when you inspect the URL (which you absolutely should do) it might be indexed. There is a delay.
Indexing API Scalability
When you start working on larger websites – and I am talking about websites where subfolders have well over 500,000 pages – the API’s 2,000 URL limitation becomes apparent. You just cannot effectively identify pages that drop in and out of the “Why Pages Aren’t Indexed?” section.
Not great, have seen worse (Image Credit: Harry Clarkson-Bennett)
But when you set up multiple properties, you can scale effectively.
The 2,000 limit only applies at a property level. So if you set up a domain-level property alongside 20 other properties (at the subfolder level), you can leverage up to 42,000 URLs per day. The more you do, the better.
And the API does have some other benefits:
But it doesn’t guarantee indexing. It is a request, not a command.
To set it up, you need to enable the API in Google Cloud Console. You can follow this semi-helpful quickstart guide. It is not fun. It is a pain in the arse. But it is worth it. Then you’ll need a Python script to send API requests and to monitor API quotas and responses (2xx, 3xx, 4xx, etc.).
If you want to get fancy, you can combine it with your publishing data to figure out exactly how long pages in specific sections take to get indexed. And you should always want to get fancy.
This is a really good signal as to what your most important subfolders are in Google’s eyes, too. Performant vs. under-performing categories.
Granular Click And Impression Data
An essential for large sites. Not only does the default Search Console only store 1,000 rows or query and URL data, but it only stores it for 16 months. While that sounds like a long time, fast forward a year or two, and you will wish you had started storing the data in BigQuery.
Particularly when it comes to looking at YoY click behavior and event planning. The teeth grinding alone will pay for your dentist’s annual trip to Aruba.
But by far and away the easiest way to see search data at a more granular level is to create more GSC properties. While you still have the same query and URL limits, because you have multiple properties instead of one, the data limits become far less limiting.
What About Sitemaps?
Not directly related to GSC indexation, but a point of note. Sitemaps are not a particularly strong tool in your arsenal when it comes to encouraging indexing of content. The indexation of content is driven by how “helpful” it is to users.
Now, it would be remiss of me not to highlight that news sitemaps are slightly different. When speed to publish and indexation are so important, you want to highlight your freshest articles in a ratified place.
Ultimately, it comes down to Navboost. Good vs. bad clicks and the last longest click. Or in more of a news sense, Glue – a huge table of user interactions, designed to rank fresh content in real-time and keep the index dynamic. Indexation is driven by your content being valuable enough to users for Google to continue to store in its index.
Glue emphasizes immediate interaction signals like hovers and swipes for more instant feedback (Image Credit: Harry Clarkson-Bennett)
Thanks to decades of experience (and confirmation via the DoJ trial and the Google Leak), we know that your site’s authority (Q*), impact over time, and internal linking structure all play a key role. But once it’s indexed, it’s all about user engagement. Sitemap or no sitemap, you can’t force people to love your beige, miserable content.
And Sitemap Indexes?
Most larger sites use sitemap indexes. Essentially, a sitemap of sitemaps to manage larger websites that exceed the 50,000 row limit. When you upload the sitemap index to Search Console, don’t stop there. Upload every individual sitemap in your sitemap index.
This gives you access to indexation at a sitemap level in the page indexing or sitemaps report. Something that is much harder to manage when you have millions of pages in a sitemap index.
Seeing data at a sitemap level gives more granular indexation data in GSC (Image Credit: Harry Clarkson-Bennett)
Take the same approach with sitemaps as we have discussed with properties. More is generally better.
Worth knowing that each document is also given a DocID. The DocID stores signals to score the page’s popularity: user clicks, its quality and authoritativeness, crawl data, and a spam score among others.
Anything classified as crucial to ranking a page is stored and used for indexation and ranking purposes.
What Should I Do Next?
Assess your current GSC setup – is it working hard enough for you?
Do you have access to a domain-level property and a crawl stats report?
Have you already broken your site down into “properties” in GSC?
If not, crawl your site and establish the subfolders you want to add.
Review your sitemap setup. Do you just have a sitemap index? Have you added the individual sitemaps to GSC, too?
Consider connecting your data to BigQuery and storing more than 16 months of it.
Consider connecting to the API via Google Cloud Console.
Review the above tools and see if they’d add value.
Ultimately, Search Console is very useful. But it has significant limitations, and to be fair, it is free. Other tools have surpassed it in many ways. But if nothing else, you should make it work as hard as possible.r
Let’s explore some potential possibilities, but before we do, let’s distinguish the differences between AI Mode and AI Overviews, as there is a clear distinction between the two.
AI Overviews Vs. AI Mode
AI Overviews are short, AI-generated summaries that appear above traditional search results for some queries that help users find information quickly. AIOs provide quick, concise answers and save users time by reducing the need to click on links, reducing clicks and traffic to brands.
AI Mode is a more advanced, interactive search experience that might replace the standard search results page in the future, as it helps with complex, multi-step, or open-ended questions by providing a more comprehensive and conversational AI-powered response if you want to follow up and learn more. Google added “AI Mode” on its search page earlier this year, looking to retain its millions of users from going away to other AI models.
What Could Potentially Happen If AI Mode Becomes Default?
If Google decides to switch to AI Mode by default, brands will definitely see a decrease in organic traffic, since users will get direct answers to their queries and won’t need to click through to websites, because they will find what they need right in the AI Mode. With AI Overviews, this is a trend that we are already seeing happening, but if AI Mode becomes the default, this will further reduce clicks.
Brands May Rely More On Paid For Visibility
Currently, the way AI Mode is designed, there are no ads and no way for Google to monetize the interface, but that is all about to change, and change extremely fast. Google’s head of Search, Liz Reid, shared a look into how the company is navigating its transition into the AI era – and how it’s thinking about keeping its multibillion-dollar ad business alive. In 2024, Google made 264.59 billion in ads, according to Statista, and it’s been growing year over year.
Screenshot from Statista, November 2025
Google is beginning to roll out ads in AI Mode, but it’s in its infancy. Google is looking into showing ads when they’re high-quality and relevant. Since queries are 2x to 3x longer than they are on main search, which means they can do better targeted, higher quality ads, according to Liz Read. Brands that can afford to be visible in AI Mode paid results will benefit from being visible, but brands that only focus on traditional SEO tactics and strategies could be left behind.
Google has also added advertisements to AI Overviews, increasing Search ad sales, so we can expect the same from AI Mode.
A Potential Shift In Visibility And Discovery
AI search is causing us to move away from traditional SEO metrics, i.e., keyword rankings and click-through rates, to brand visibility and relevance. Your brand should be cited as the authoritative source for AI answers, and if your brand is not visible as the answer, then you will lose more clicks.
Measurement
Tracking the customer journey may become harder because users interact within the AI interface rather than on your brand’s website. Traditional analytics will provide fewer insights and will cause brands to develop new metrics focused on AI citations, brand mentions, and local visibility. We are seeing this already with the emergence of AI tools, from traditional players like Semrush, Ahrefs, and new AI players like PeecAI and Profound, to name a few.
Loss Of Control Over Brand Narrative
Since AI Overviews are taking information from various online sources to build a brand’s presence, if your brand does not have a good brand strategy and has inconsistent, outdated, or poorly managed information across the web, i.e., reviews, social signals, and local listings, etc., then AI may inaccurately represent your brand across the web.
What Could Potentially Happen To Google Chrome?
If Google does go to full AI Mode by default, Chrome could potentially undergo a major transformation with deep integration of Gemini and other AI capabilities, which would change the web browsing experience from a passive tool to a proactive, intelligent assistant. According to eMarketer, Gemini is growing its user base faster than ChatGPT.
Screenshot from eMarketer, November 2025
ChatGPT has already opened its AI browser ChatGPT Atlas, which is currently only available on macOS and is challenging Google Chrome.
Screenshot from ChatGPT Atlas, November 2025
If Google Does Make AI Mode Default, What Can We Do?
Experiment with AI paid ads when they become available and put aside some budget and test the impact and return on investment (ROI) of ads in AI Mode.
Focus on making sure conversion funnels and processes are easy and provide a good user experience.
Be present and have great content everywhere your audience is. Your brand must have a strong brand presence across Reddit, Quora, YouTube, OpenAI, Perplexity, etc., and other places where end users are looking for information about your brand. For example, Apple is looking at search options on Safari, which could end its partnership with Google, but at the end of the day, we will see if Google will maintain the relationship or Apple will go somewhere else, like OpenAI, which could boost traffic and get more users using OpenAI or another large language model (LLM).
Continue to optimize for AIO by creating high-quality, authoritative content that directly answers user questions, is well-structured, and easy for AI to understand. This involves creating new content and refreshing your old content with up-to-date research, original information, and different perspectives.
Wrapping Up
The shift toward AI-powered search isn’t hypothetical anymore; it’s actually here and moving fast. With AI Overviews and AI Mode gaining traction among more than 100 million monthly users, Google is positioning itself for a future where conversational, answer-focused experiences may replace traditional search results.
If AI Mode becomes the default search for Google, it won’t just change how users search; it will fundamentally reshape how brands earn visibility, traffic, and trust online.
For brands, publishers, and SEOs, this transition presents both risks and opportunities. Organic traffic will almost certainly decline as more answers stay within Google’s ecosystem. Paid visibility in AI results will grow rapidly, favoring brands with budgets and adaptable strategies. And success will depend less on ranking for keywords and more on becoming a trusted source that AI cites, references, or recommends across platforms.
This era will demand a new kind of optimization centered on brand authority, AI citations, structured data, user trust signals, and multi-platform presence.
No one has a crystal ball and knows what a full AI Mode future looks like, but brands that adapt early will be the market share leaders, and those that wait will lose visibility, traffic, and relevance.
Last week, we hosted EmTech MIT, MIT Technology Review’s annual flagship conference in Cambridge, Massachusetts. Over the course of three days of main-stage sessions, I learned about innovations in AI, biotech, and robotics.
But as you might imagine, some of this climate reporter’s favorite moments came in the climate sessions. I was listening especially closely to my colleague James Temple’s discussion with Lucia Tian, head of advanced energy technologies at Google.
They spoke about the tech giant’s growing energy demand and what sort of technologies the company is looking to to help meet it. In case you weren’t able to join us, let’s dig into that session and consider how the company is thinking about energy in the face of AI’s rapid rise.
I’ve been closely following Google’s work in energy this year. Like the rest of the tech industry, the company is seeing ballooning electricity demand in its data centers. That could get in the way of a major goal that Google has been talking about for years.
See, back in 2020, the company announced an ambitious target: by 2030, it aimed to run on carbon-free energy 24-7. Basically, that means Google would purchase enough renewable energy on the grids where it operates to meet its entire electricity demand, and the purchases would match up so the electricity would have to be generated when the company was actually using energy. (For more on the nuances of Big Tech’s renewable-energy pledges, check out James’s piece from last year.)
Google’s is an ambitious goal, and on stage, Tian said that the company is still aiming for it but acknowledged that it’s looking tough with the rise of AI.
“It was always a moonshot,” she said. “It’s something very, very hard to achieve, and it’s only harder in the face of this growth. But our perspective is, if we don’t move in that direction, we’ll never get there.”
Google’s total electricity demand more than doubled from 2020 to 2024, according to its latest Environmental Report. As for that goal of 24-7 carbon-free energy? The company is basically treading water. While it was at 67% for its data centers in 2020, last year it came in at 66%.
Not going backwards is something of an accomplishment, given the rapid growth in electricity demand. But it still leaves the company some distance away from its finish line.
To close the gap, Google has been signing what feels like constant deals in the energy space. Two recent announcements that Tian talked about on stage were a project involving carbon capture and storage at a natural-gas plant in Illinois and plans to reopen a shuttered nuclear power plant in Iowa.
Let’s start with carbon capture. Google signed an agreement to purchase most of the electricity from a new natural-gas plant, which will capture and store about 90% of its carbon dioxide emissions.
That announcement was controversial, with critics arguing that carbon capture keeps fossil-fuel infrastructure online longer and still releases greenhouse gases and other pollutants into the atmosphere.
One question that James raised on stage: Why build a new natural-gas plant rather than add equipment to an already existing facility? Tacking on equipment to an operational plant would mean cutting emissions from the status quo, rather than adding entirely new fossil-fuel infrastructure.
The company did consider many existing plants, Tian said. But, as she put it, “Retrofits aren’t going to make sense everywhere.” Space can be limited at existing plants, for example, and many may not have the right geology to store carbon dioxide underground.
“We wanted to lead with a project that could prove this technology at scale,” Tian said. This site has an operational Class VI well, the type used for permanent sequestration, she added, and it also doesn’t require a big pipeline buildout.
Tian also touched on the company’s recent announcement that it’s collaborating with NextEra Energy to reopen Duane Arnold Energy Center, a nuclear power plant in Iowa. The company will purchase electricity from that plant, which is scheduled to reopen in 2029.
As I covered in a story earlier this year, Duane Arnold was basically the final option in the US for companies looking to reopen shuttered nuclear power plants. “Just a few years back, we were still closing down nuclear plants in this country,” Tian said on stage.
While each reopening will look a little different, Tian highlighted the groups working to restart the Palisades plant in Michigan, which was the first reopening to be announced, last spring. “They’re the real heroes of the story,” she said.
I’m always interested to get a peek behind the curtain at how Big Tech is thinking about energy. I’m skeptical but certainly interested to see how Google’s, and the rest of the industry’s, goals shape up over the next few years.
This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here.
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
AI is changing how we quantify pain
Researchers around the world are racing to turn pain—medicine’s most subjective vital sign—into something a camera or sensor can score as reliably as blood pressure.
The push has already produced PainChek—a smartphone app that scans people’s faces for tiny muscle movements and uses artificial intelligence to output a pain score—which has been cleared by regulators on three continents and has logged more than 10 million pain assessments. Other startups are beginning to make similar inroads.
The way we assess pain may finally be shifting, but when algorithms measure our suffering, does that change the way we treat it? Read the full story.
—Deena Mousa
This story is from the latest print issue of MIT Technology Review magazine, which is full of fascinating stories about our bodies. If you haven’t already, subscribe now to receive future issues once they land.
How to help friends and family dig out of a conspiracy theory black hole
—Niall Firth
Someone I know became a conspiracy theorist seemingly overnight.
It was during the pandemic. They suddenly started posting daily on Facebook about the dangers of covid vaccines and masks, warning of an attempt to control us.
As a science and technology journalist, I felt that my duty was to respond. I tried, but all I got was derision. Even now I still wonder: Are there things I could have done differently to talk them back down and help them see sense?
I gave Sander van der Linden, professor of social psychology in society at the University of Cambridge, a call to ask: What would he advise if family members or friends show signs of having fallen down the rabbit hole? Read the full story.
This story is part of MIT Technology Review’s series “The New Conspiracy Age,” on how the present boom in conspiracy theories is reshaping science and technology. Check out the rest of the series here. It’s also part of our How To series, giving you practical advice to help you get things done.
If you’re interested in hearing more about how to survive in the age of conspiracies, join our features editor Amanda Silverman and executive editor Niall Firth for a subscriber-exclusive Roundtable conversation with conspiracy expert Mike Rothschild. It’s at 1pm ET on Thursday November 20—register now to join us!
Google is still aiming for its “moonshot” 2030 energy goals
—Casey Crownhart
Last week, we hosted EmTech MIT, MIT Technology Review’s annual flagship conference in Cambridge, Massachusetts. As you might imagine, some of this climate reporter’s favorite moments came in the climate sessions. I was listening especially closely to my colleague James Temple’s discussion with Lucia Tian, head of advanced energy technologies at Google.
They spoke about the tech giant’s growing energy demand and what sort of technologies the company is looking to to help meet it. In case you weren’t able to join us, let’s dig into that session and consider how the company is thinking about energy in the face of AI’s rapid rise. Read the full story.
This article is from The Spark, MIT Technology Review’s weekly climate newsletter. To receive it in your inbox every Wednesday, sign up here.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 ChatGPT is now “warmer and more conversational” But it’s also slightly more willing to discuss sexual and violent content. (The Register) + ChatGPT has a very specific writing style. (WP $) + The looming crackdown on AI companionship. (MIT Technology Review)
2 The US could deny visas to visitors with obesity, cancer or diabetes As part of its ongoing efforts to stem the flow of people trying to enter the country. (WP $)
3 Microsoft is planning to create its own AI chip And it’s going to use OpenAI’s internal chip-building plans to do it. (Bloomberg $) + The company is working on a colossal data center in Atlanta. (WSJ $)
4 Early AI agent adopters are convinced they’ll see a return on their investment soon Mind you, they would say that. (WSJ $) + An AI adoption riddle. (MIT Technology Review)
5 Waymo’s robotaxis are hitting American highways Until now, they’ve typically gone out of their way to avoid them. (The Verge) + Its vehicles will now reach speeds of up to 65 miles per hour. (FT $) + Waymo is proving long-time detractor Elon Musk wrong. (Insider $)
6 A new Russian unit is hunting down Ukraine’s drone operators It’s tasked with killing the pilots behind Ukraine’s successful attacks. (FT $) + US startup Anduril wants to build drones in the UAE. (Bloomberg $) + Meet the radio-obsessed civilian shaping Ukraine’s drone defense. (MIT Technology Review)
7 Anthropic’s Claude successfully controlled a robot dog It’s important to know what AI models may do when given access to physical systems. (Wired $)
8 Grok briefly claimed Donald Trump won the 2020 US election As reliable as ever, I see. (The Guardian)
9 The Northern Lights are playing havoc with satellites Solar storms may look spectacular, but they make it harder to keep tabs on space. (NYT $) + Seriously though, they look amazing. (The Atlantic $) + NASA’s new AI model can predict when a solar storm may strike. (MIT Technology Review)
10 Apple users can now use digital versions of their passports But it’s strictly for internal flights within the US only. (TechCrunch)
Quote of the day
“I hope this mistake will turn into an experience.”
—Vladimir Vitukhin, chief executive of the company behind Russia’s first anthropomorphic robot AIDOL, offers a philosophical response to the machine falling flat on its face during a reveal event, the New York Times reports.
One more thing
Welcome to the oldest part of the metaverse
Headlines treat the metaverse as a hazy dream yet to be built. But if it’s defined as a network of virtual worlds we can inhabit, its oldest corner has been already running for 25 years.
It’s a medieval fantasy kingdom created for the online role-playing game Ultima Online. It was the first to simulate an entire world: a vast, dynamic realm where players could interact with almost anything, from fruit on trees to books on shelves.
Ultima Online has already endured a quarter-century of market competition, economic turmoil, and political strife. So what can this game and its players tell us about creating the virtual worlds of the future? Read the full story.
—John-Clark Levin
We can still have nice things
A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line or skeet ’em at me.)
+ Unlikely duo Sting and Shaggy are starring together in a New York musical. + Barry Manilow was almost in Airplane!? That would be an entirely different kind of flying, altogether + What makes someone sexy? Well, that depends. + Keep an eye on those pink dolphins, they’re notorious thieves.
Google DeepMind has built a new video-game-playing agent called SIMA 2 that can navigate and solve problems in a wide range of 3D virtual worlds. The company claims it’s a big step toward more general-purpose agents and better real-world robots.
Google DeepMind first demoed SIMA (which stands for “scalable instructable multiworld agent”) last year. But SIMA 2 has been built on top of Gemini, the firm’s flagship large language model, which gives the agent a huge boost in capability.
The researchers claim that SIMA 2 can carry out a range of more complex tasks inside virtual worlds, figure out how to solve certain challenges by itself, and chat with its users. It can also improve itself by tackling harder tasks multiple times and learning through trial and error.
“Games have been a driving force behind agent research for quite a while,” Joe Marino, a research scientist at Google DeepMind, said in a press conference this week. He noted that even a simple action in a game, such as lighting a lantern, can involve multiple steps: “It’s a really complex set of tasks you need to solve to progress.”
The ultimate aim is to develop next-generation agents that are able to follow instructions and carry out open-ended tasks inside more complex environments than a web browser. In the long run, Google DeepMind wants to use such agents to drive real-world robots. Marino claimed that the skills SIMA 2 has learned, such as navigating an environment, using tools, and collaborating with humans to solve problems, are essential building blocks for future robot companions.
Unlike previous work on game-playing agents such as AlphaZero, which beat a Go grandmaster in 2016, or AlphaStar, which beat 99.8% of ranked human competition players at the video game StarCraft 2 in 2019, the idea behind SIMA is to train an agent to play an open-ended game without preset goals. Instead, the agent learns to carry out instructions given to it by people.
Humans control SIMA 2 via text chat, by talking to it out loud, or by drawing on the game’s screen. The agent takes in a video game’s pixels frame by frame and figures out what actions it needs to take to carry out its tasks.
Like its predecessor, SIMA 2 was trained on footage of humans playing eight commercial video games, including No Man’s Sky and Goat Simulator 3, as well as three virtual worlds created by the company. The agent learned to match keyboard and mouse inputs to actions.
Hooked up to Gemini, the researchers claim, SIMA 2 is far better at following instructions (asking questions and providing updates as it goes) and figuring out for itself how to perform certain more complex tasks.
Google DeepMind tested the agent inside environments it had never seen before. In one set of experiments, researchers asked Genie 3, the latest version of the firm’s world model, to produce environments from scratch and dropped SIMA 2 into them. They found that the agent was able to navigate and carry out instructions there.
The researchers also used Gemini to generate new tasks for SIMA 2. If the agent failed, at first Gemini generated tips that SIMA 2 took on board when it tried again. Repeating a task multiple times in this way often allowed SIMA 2 to improve by trial and error until it succeeded, Marino said.
Git gud
SIMA 2 is still an experiment. The agent struggles with complex tasks that require multiple steps and more time to complete. It also remembers only its most recent interactions (to make SIMA 2 more responsive, the team cut its long-term memory). It’s also still nowhere near as good as people at using a mouse and keyboard to interact with a virtual world.
Julian Togelius, an AI researcher at New York University who works on creativity and video games, thinks it’s an interesting result. Previous attempts at training a single system to play multiple games haven’t gone too well, he says. That’s because training models to control multiple games just by watching the screen isn’t easy: “Playing in real time from visual input only is ‘hard mode,’” he says.
In particular, Togelius calls out GATO, a previous system from Google DeepMind, which—despite being hyped at the time—could not transfer skills across a significant number of virtual environments.
Still, he is open-minded about whether or not SIMA 2 could lead to better robots. “The real world is both harder and easier than video games,” he says. It’s harder because you can’t just press A to open a door. At the same time, a robot in the real world will know exactly what its body can and can’t do at any time. That’s not the case in video games, where the rules inside each virtual world can differ.
Others are more skeptical. Matthew Guzdial, an AI researcher at the University of Alberta, isn’t too surprised that SIMA 2 can play many different video games. He notes that most games have very similar keyboard and mouse controls: Learn one and you learn them all. “If you put a game with weird input in front of it, I don’t think it’d be able to perform well,” he says.
Guzdial also questions how much of what SIMA 2 has learned would really carry over to robots. “It’s much harder to understand visuals from cameras in the real world compared to games, which are designed with easily parsable visuals for human players,” he says.
Still, Marino and his colleagues hope to continue their work with Genie 3 to allow the agent to improve inside a kind of endless virtual training dojo, where Genie generates worlds for SIMA to learn in via trial and error guided by Gemini’s feedback. “We’ve kind of just scratched the surface of what’s possible,” he said at the press conference.
That’s a big deal, because today’s LLMs are black boxes: Nobody fully understands how they do what they do. Building a model that is more transparent sheds light on how LLMs work in general, helping researchers figure out why models hallucinate, why they go off the rails, and just how far we should trust them with critical tasks.
“As these AI systems get more powerful, they’re going to get integrated more and more into very important domains,” Leo Gao, a research scientist at OpenAI, told MIT Technology Review in an exclusive preview of the new work. “It’s very important to make sure they’re safe.”
This is still early research. The new model, called a weight-sparse transformer, is far smaller and far less capable than top-tier mass-market models like the firm’s GPT-5, Anthropic’s Claude, and Google DeepMind’s Gemini. At most it’s as capable as GPT-1, a model that OpenAI developed back in 2018, says Gao (though he and his colleagues haven’t done a direct comparison).
But the aim isn’t to compete with the best in class (at least, not yet). Instead, by looking at how this experimental model works, OpenAI hopes to learn about the hidden mechanisms inside those bigger and better versions of the technology.
It’s interesting research, says Elisenda Grigsby, a mathematician at Boston College who studies how LLMs work and who was not involved in the project: “I’m sure the methods it introduces will have a significant impact.”
Lee Sharkey, a research scientist at AI startup Goodfire, agrees. “This work aims at the right target and seems well executed,” he says.
Why models are so hard to understand
OpenAI’s work is part of a hot new field of research known as mechanistic interpretability, which is trying to map the internal mechanisms that models use when they carry out different tasks.
That’s harder than it sounds. LLMs are built from neural networks, which consist of nodes, called neurons, arranged in layers. In most networks, each neuron is connected to every other neuron in its adjacent layers. Such a network is known as a dense network.
Dense networks are relatively efficient to train and run, but they spread what they learn across a vast knot of connections. The result is that simple concepts or functions can be split up between neurons in different parts of a model. At the same time, specific neurons can also end up representing multiple different features, a phenomenon known as superposition (a term borrowed from quantum physics). The upshot is that you can’t relate specific parts of a model to specific concepts.
“Neural networks are big and complicated and tangled up and very difficult to understand,” says Dan Mossing, who leads the mechanistic interpretability team at OpenAI. “We’ve sort of said: ‘Okay, what if we tried to make that not the case?’”
Instead of building a model using a dense network, OpenAI started with a type of neural network known as a weight-sparse transformer, in which each neuron is connected to only a few other neurons. This forced the model to represent features in localized clusters rather than spread them out.
Their model is far slower than any LLM on the market. But it is easier to relate its neurons or groups of neurons to specific concepts and functions. “There’s a really drastic difference in how interpretable the model is,” says Gao.
Gao and his colleagues have tested the new model with very simple tasks. For example, they asked it to complete a block of text that opens with quotation marks by adding matching marks at the end.
It’s a trivial request for an LLM. The point is that figuring out how a model does even a straightforward task like that involves unpicking a complicated tangle of neurons and connections, says Gao. But with the new model, they were able to follow the exact steps the model took.
“We actually found a circuit that’s exactly the algorithm you would think to implement by hand, but it’s fully learned by the model,” he says. “I think this is really cool and exciting.”
Where will the research go next? Grigsby is not convinced the technique would scale up to larger models that have to handle a variety of more difficult tasks.
Gao and Mossing acknowledge that this is a big limitation of the model they have built so far and agree that the approach will never lead to models that match the performance of cutting-edge products like GPT-5. And yet OpenAI thinks it might be able to improve the technique enough to build a transparent model on a par with GPT-3, the firm’s breakthrough 2021 LLM.
“Maybe within a few years, we could have a fully interpretable GPT-3, so that you could go inside every single part of it and you could understand how it does every single thing,” says Gao. “If we had such a system, we would learn so much.”
