Labrador, ChatGPT's index: what it reads of your site

🇫🇷 Lire en français : Labrador, l'index de ChatGPT : ce qu'il lit de votre site

Barely using Google any more? Neither is ChatGPT. In the Think mode of its free tier, the one the general public uses, ChatGPT draws 75 % of its sources from Labrador, OpenAI’s in-house search index, and only 3 % from Google, according to the study published by the agency Resoneo in August 2026. What Labrador keeps of your page amounts to a title and roughly 200 characters. Ranking first on Google no longer guarantees being cited. I applied these findings to kimoun.com: here is what I found, and what it changes for a Caribbean business in 2026 and 2027.

Key points

• Labrador is OpenAI’s proprietary search index. It supplies most of the sources for the free version of ChatGPT, which barely consults Google.
• ChatGPT almost never reads your page. It decides whether to cite you based on your title and about 200 characters taken right after your H1, your meta description being ignored.
• On kimoun.com, a share button was eating a quarter of that budget. I changed the site’s templates to test the study’s conclusions.
• For a local business, ChatGPT’s answer comes through map and review feeds, not through your site. The Google Business Profile remains the first lever.

Ranking first on Google no longer says anything about what ChatGPT knows of you. What counts is the first two hundred characters of your page.

— Olivier Watte, known as Oliver · founder of Kimoun

What is Labrador, ChatGPT’s index?  

  Tip

Labrador is OpenAI’s proprietary search index: ChatGPT queries it without going through Google or Bing.

Between 21 May and 21 July 2026, a technical field named result_source revealed where each ChatGPT result came from. Three of its values pointed to providers that scrape Google results. The fourth, labrador, pointed to OpenAI’s own index. Resoneo had time to capture 1,249 answers before the field disappeared.

Labrador is not Bing in disguise: a quarter of its titles run past the 75 characters at which Bing always truncates its own. The Peec AI investigation describes a family of indexes: web, news, local, shopping, images. During Google’s antitrust trial, Nick Turley, who heads ChatGPT, testified that OpenAI had been building its own index since 2023, after Google turned it down.

One caveat matters: OpenAI has never officially confirmed Labrador. Everything that follows comes from reverse engineering, serious and documented, but dated.

Does ChatGPT still use Google?  

Yes, but mostly for paying subscribers. Resoneo asked the same questions across ChatGPT’s different tiers: the answering mode picks the engine.

Source: Resoneo, August 2026 update. Share of URLs retrieved per conversation.

Without the Think button, the picture is starker still. On a free account, general information questions, questions about a business and questions about a product go almost entirely to Labrador. Only questions whose answer changes within the day, a score or a ruling, open the door to Google.

So you can rank first on Google and be absent from the answer ChatGPT gives your customers. The trend points that way: in mid-August, Peec AI spotted an internal experiment replacing Google results with the in-house index on roughly 8 % of conversations.

What does ChatGPT see of your site?  

  Tip

ChatGPT almost never reads your page: it decides whether to cite you from a title and about 200 characters.

For each result, ChatGPT receives a title, a URL and a short extract. In July, it opened a page barely more than once in 80, according to Resoneo. The extract Labrador keeps follows precise rules:

• It is cut off just past 200 characters, and frozen at the moment OpenAI’s crawler visited your page.
• It is taken from the start of the page’s visible text, not from your metadata. Your meta description is ignored.
• It starts at your H1 heading: when a page has one, the extract contains it in 83.6 % of cases.

In median values, the budget breaks down like this: 7 stray characters before the H1, 51 characters of H1, then 146 characters of actual content to convince ChatGPT to cite you.

The kicker, the date, the site name and above all the alt text of the first image can eat up to 50 of them. And one page in seven examined by Resoneo has no marked-up H1 at all: the extract then starts wherever the template happens to begin.

What was Labrador reading of kimoun.com?  

I reconstructed the extract Labrador would get from kimoun.com articles. The result was not flattering.

  Note

Reconstruction on the article about llms.txt files, following the rules measured by Resoneo. Before: “Rendre son site citable par une IA : robots.txt, llms.txt et le reste Partager via Kimoun Lien copié dans le presse-papier robots.txt, .md, licences, llms.txt : ce qui rend vraiment un site lisible par” After: “Rendre son site citable par une IA : robots.txt, llms.txt et le reste robots.txt, .md, licences, llms.txt : ce qui rend vraiment un site lisible par une IA — et ce qui relève du mythe, testé depuis la G”

The share button labels took up 52 characters, a quarter of the budget. The summary sentence stopped before its promise, and the standfirst never made it into the extract.

Following this study, I changed the templates of kimoun.com to test Resoneo’s conclusions. The share button now comes after the description, which is written as a self-contained extract: the one for this article fits, title included, in 202 characters. I will publish the measured results.

I also owe a correction. In May, I wrote in my guide to making your site AI-citable that there was no second door reserved for AI. That holds for Google. For ChatGPT, that door exists: it is called Labrador. It simply opens with the same keys.

What about a local business in the Caribbean?  

For a question about a business or a place, ChatGPT answers with a map fed by data feeds: Yelp, TripAdvisor and data sourced from Google Maps. Your website plays only a secondary role. And as I explained when covering the ChatGPT and Yelp deal, Yelp is almost empty in Guadeloupe. ChatGPT therefore falls back on Google Maps data.

For a restaurant, a tradesperson or a guesthouse, the first job remains the Google Business Profile: complete, current, consistent with the site.

What to do this week  

Six actions are enough, without rebuilding your site.

1. Check that every page has one marked-up H1. It is the anchor point of the extract: without it, you control nothing.
2. Write the 150 characters after the H1 as your answer. Who, what, where, with one concrete fact.
3. Clear the runway between the title and the text: buttons, kicker, date, alt text of the first image.
4. Let OAI-SearchBot through in your robots.txt. It is the crawler that feeds the index; GPTBot only serves training.
5. Mind your internal links. In Resoneo’s tests, OpenAI’s crawler followed pages reachable by a link, not those listed only in the sitemap.
6. Test it yourself: ask free ChatGPT, with the Think button, “Who is the best [your trade] in [your town]?”.

  Warning

The noindex directive and Cache-Control: no-store are ignored by OpenAI’s cache, where copies stay for more than 90 days (Resoneo and Oncrawl). They do not remove a page from ChatGPT.

Should you optimise “for Labrador”?  

No. What we know of Labrador rests on an observation window that has already closed, and the system moves fast: between July and August 2026, the paid thinking mode halved its searches per conversation.

The foundations hold whatever the engine: a page that answers in its opening lines, sourced figures, and a reputation built somewhere other than your own site. The Princeton study on GEO measures up to 40 % more visibility for content that cites statistics and sources. And our survey of 189 overseas agency sites shows the official site is the number-one source only 55 % of the time.

Labrador does not change the nature of the work described in Google’s official guide to AI search. It hardens the rules, and it dispels one illusion: that a good Google ranking is enough. That is what I check with you in my SEO and GEO work.