Your ChatGPT Search Listing Is One Title and 200 Characters

OpenAI keeps a card catalog of the web, and your page’s card is small. The index behind ChatGPT’s search runs under the codename labrador, and for each page it stores three things: the URL, the full title, and a snippet that cuts off just past 200 characters. That’s the finding from Resoneo, a French search consultancy that captured 1,249 ChatGPT answers in July, spanning 88,000 search results and 26,900 distinct pages. The firm read what the system returned before the model wrote a word.

The consequence sits in where those 200 characters come from. The snippet is pulled from the top of your rendered page, anchored on the H1, plus whatever visible text sits right after it. Not from the field marketers have spent two decades polishing.

Olivier de Segonzac, co-founder of Resoneo, published the full breakdown in Search Engine Land on August 17. His line on that field is five words long: “The meta description is ignored entirely.”

The model never sees most of your page at the moment it decides what to use. As the study puts it: “When ChatGPT answers using the web, it does not go and read your page.” It reads the card. If your page opens with navigation links instead of a claim, the card is navigation links. De Segonzac’s team found the failure case in the wild: “We measured one case where the snippet was 100% table of contents and 0% content.”

There’s a second finding underneath the first, and it splits the audience in two. In ChatGPT’s free instant mode, labrador supplied 63.6% of search results in Resoneo’s corpus, and for stable questions about products and businesses it was close to all of them. In the paid thinking mode, the mix flips: 74.6% of results came from an engine called bright, which serves scraped Google results, with labrador falling to 23.6%. The study is blunt about what that means: “In thinking mode, ChatGPT runs three quarters of the time on bright. So on Google.”

So a free user and a paid user asking the same question are searching different webs. One runs on OpenAI’s own small-card index. The other mostly runs on Google’s rankings, bought through scraping intermediaries. If your brand shows up in one and not the other, that’s why, and no dashboard currently tells you which one your buyers use.

Two more mechanics matter for anyone doing AI visibility work. First, there’s no velvet rope. Search Engine Journal reported that hundreds of outlets with no OpenAI content deal were served by labrador exactly the way licensed partners were. Inclusion doesn’t require a contract. Second, ChatGPT keeps a shared cache of full page copies, and it refreshes them on demand rather than on a schedule. De Segonzac again: “Popular pages stay fresh. Unpopular pages age indefinitely.” Some cached copies in the study had gone more than 90 days without an update. Pages over 4MB were rejected outright.

The window for checking any of this is already closing. Around July 21, according to Resoneo, OpenAI removed the field that identified which engine served each result. The measurement that produced these numbers can’t be repeated. Whatever AI visibility vendors tell you about ChatGPT from August onward, they can no longer see the sourcing label Resoneo saw.

Why build it this way? Serving a chatbot at conversation speed means the search step has to be cheap and fast, and a title plus 200 characters is what cheap and fast looks like at web scale. Google spent 25 years and unmatched infrastructure indexing full pages. OpenAI took the shortcut every startup takes: store less, decide faster, buy the rest from scrapers.

For a marketing team, the to-do list writes itself, and it’s short. Read your key pages the way labrador does: title, H1, and the first 200 characters of visible text. If the claim a buyer needs isn’t in that window, it doesn’t exist in ChatGPT’s free tier, where citations already decay fast. Front-load the sentence that says what you do, for whom, at what proof point. And treat your Google rankings as double duty, because the paid tier is quietly running on them.

The publishing industry spent two years arguing about whether AI systems should pay to read the web. Resoneo’s data shows something plainer. For most answers, ChatGPT reads 200 characters of the web, and almost nobody has checked which 200 theirs are.