ChatGPT Does Not Open Your Page in 93% of Fast Answers

When someone asks ChatGPT about your product without turning on the reasoning mode, the model usually does not visit your website at all. According to analysis published on Habr, based on a corpus gathered by the French team RESONEO, ChatGPT opens the page live in about 7% of fast answers. In the other 93% it uses a card that was cut long before the question was asked. That changes what you should put where on a page.

What the model actually sees

The card has two parts: the full page title (the researchers recorded titles up to 289 characters) and a snippet of about 200 characters. The snippet is cut once, at crawl time, starting from the H1 and whatever sits visually next to it — breadcrumb, category, date, author, table of contents, image alt text. There is no selection by relevance, because at crawl time the question does not exist yet.

That has two consequences. First, anything below the first screen — your specification table, delivery terms, price block — does not reach the model in fast mode. Second, the same 200-character fragment is served for a question about price and a question about delivery times. Until the page is crawled again, it has no other version.

The meta description is ignored on this path. It was written for search results, where the snippet is rebuilt per query; in the OpenAI index that slot is already filled with text from the page. In the RESONEO captures the meta description was used in about one case out of three, and only for results pulled from Google search results by scraping.

The markup problem

Where the page is opened for real — the reasoning mode and those 7% of fast answers — the HTML is converted to text. Script tags, iframes and JSON-LD are stripped out entirely. Structured data you added specifically for machines does not reach the model in any field. Image alt text and text hidden by CSS, on the other hand, survive.

So a fact that exists only inside a JSON-LD block does not exist for ChatGPT in either mode. A page can pass a structured data validator, produce a tidy snippet in Google, and stay mute in an AI answer. The study also notes a hard 4 MB limit per page: above it the request returns HTTP 400 and the page is not read at all — no truncation, just nothing.

There is also a caching layer. A fetched copy is fresh for roughly 30 minutes, after which a stale version is served while an update runs in the background. The corpus contained copies more than 90 days old, and this layer ignores Cache-Control: no-store and noindex.

Why it matters for a small business

Most owners assume that a good-looking page plus correct markup is enough to be quoted. The finding is that in the majority of quick answers nobody looked at your page today; the model repeats a fragment captured on an unknown earlier date. If your price, your service area or your delivery rules changed since then, the answer is wrong, and you will not see it in analytics either. Clicks on visible links carry a utm_source of chatgpt.com, but the pages the model opened and actually read in reasoning mode carry no UTM tag at all, because no human clicked. Your "traffic from ChatGPT" row describes the tail of the process, not your presence in answers.

One more useful detail: the reasoning mode is not a paid privilege. A free account with the Think button returned 35.3 URLs per conversation against 35.1 for the paid reasoning mode. One click roughly doubled the number of sources, from 15.1 to 35.3.

What to do this month

• Read the first 200 characters after your H1 on your key pages. If they are a breadcrumb, a date and an author byline, that is how your site introduces itself in 93% of answers.
• Put price, date, author and what the product is for into ordinary sentences in the body text, close to the H1, and repeat the same fact in the alt text of the nearby image. Keep your structured data — Google and Yandex read it — but never let a fact live only there.
• Check your robots.txt per crawler. GPTBot collects training data; OAI-SearchBot decides whether you appear in citations and search snippets; ChatGPT-User comes on a user's initiative; OAI-AdsBot checks ad landing pages. A blanket User-agent: * with Disallow: /, or a copied list of AI bots, closes the search crawler along with the training one. OpenAI systems pick up robots.txt edits in about a day.
• Do not rely on your sitemap to get new pages crawled. In the 16-day observation, GPTBot downloaded the sitemap daily but did not visit URLs that existed only there — 0 of 8 — while 8 pages with a normal inbound link were crawled the same day.
• Check page weight. Anything approaching 4 MB of HTML risks being refused outright.

None of this requires a new website. It requires moving the three or four facts a customer asks about out of tables, tabs and markup, into plain sentences near the top of the page, and making sure the crawler that produces citations is allowed in. The rest is patience: the card is rebuilt only when the page is crawled again.