Does ChatGPT read your page before it cites it? In 1,249 captured answers, 1 page in 80 was opened — and an opened page is cited 74% of the time against 7% for the rest

What RESONEO measured, and how

The study is published on RESONEO's research site as "What ChatGPT pulls, what it shows, what it cites" (July 2026, updated August 2026; its co-founder Olivier de Segonzac summarized it for Search Engine Land on 2026-08-17, with the conflict disclosed). The method is network capture, not answer reading: a Chrome extension recorded the raw data stream behind 1,249 real ChatGPT answers (682 instant, 567 thinking; just under 500 on free accounts, just over 750 on paid) across a fleet of accounts and countries, yielding roughly 88,000 search results and 26,900 distinct pages from 6,400 domains. Until 2026-07-21 the stream carried an undocumented `result_source` field naming which engine returned each result (760 of the 1,249 answers carry it; the field went bare on 07-22 and the team re-identified engines from formatting signatures afterwards). RESONEO also ran canary pages on its own domain with server logs, which is how the cache-retention and crawl-discovery findings below are grounded in log lines rather than inference. It credits earlier network-traffic work by Suganthan Mohanadasan and others as the starting point.

The four engines behind a ChatGPT answer

Two findings sit behind that table. First, labrador is neither rebadged Bing nor a licensed-publisher catalogue: Bing caps displayed titles at 75 characters and a quarter of labrador titles run past that (the longest seen: 977), only one labrador URL in twenty appears on Bing's first page, and outlets with no OpenAI licensing deal receive exactly the same format, length and freshness as partners. RESONEO never observed a `bing` engine on any of its accounts and hypothesises it is exposed only to Team plans. Second, the split depends on the question, not just the tier: on stable questions labrador's free-tier share rises to 100%; on business-and-place queries it was 100% across 310 snippets, product 97.6%, news 52.6% labrador / 46.1% Google. The August update re-ran the July prompts across every configuration: the free "Think" button pulls as many URLs per conversation as paid thinking (35.3 vs 35.1) but from the opposite corpus — 74.7% labrador on free Think, 75.3% scraped Google on paid thinking. "A good Google position still helps, but it does not guarantee visibility on the free Think path."

The pyramid: pulled, promoted, cited, opened

The median answer lists about twenty pages at the bottom of the panel, promotes five to the top, cites three in its text and opens none — nine answers in ten open nothing. But opening is decisive: on the 440 opened pages exposed as result objects, 326 were cited (74%), against 4,303 of 57,853 merely-retrieved pages (7%). What gets opened is specific: regulatory and official material, primary sources, pricing pages, newsrooms — "pages deep in the tree, almost never home pages" (most-opened domains: openai.com, anthropic.com, learn.microsoft.com, digital-strategy.ec.europa.eu). The August update lifts the paid lock slightly: with the free Think button, a free account now opens 0.28 pages per conversation. Two mechanics matter for anyone reading their analytics. Instant mode cites from memory a lot: 85% of instant-mode citations point to a URL that appears nowhere in the retrieval net (31% in thinking), some from widgets, "the rest looks like brand homepages written from the model's own memory." And ChatGPT appends `utm_source=chatgpt.com` to 95% of displayed links (every one in instant mode) — but never to the pages the model opened and read, so "the citations worth the most … are precisely those your analytics tool will fail to attribute to ChatGPT."

What the model actually sees of your page: 202 characters, from the H1

This is the operational half. The labrador snippet is not query-dependent and not built at search time — seven regional variants of the same press release, queried identically, produce seven different anchors; pages seen under several queries return an identical snippet. It is frozen at crawl time, cut just past 200 characters, and taken from the start of the rendered body, not from metadata: your meta description is ignored (the Google scrape picks it up one time in three). RESONEO fetched 534 pages ChatGPT had actually cited and compared them with the stored snippet: of the 463 with an H1, 387 snippets contain it (83.6%), and in 297 of those the H1 is the very first character (77%); when there is a prefix it is 20 characters at the median. One page in seven has no H1 markup at all, at which point the anchor becomes unpredictable. Hence the study's one-line advice: "Your grounding budget in instant mode is your full title plus roughly one hundred and fifty useful characters starting at your H1. Mark up an H1, clear the runway between it and the first paragraph." Three more limits are worth writing down: pages over four megabytes are not read at all ("no partial read, no truncation"); pages reach the model converted to Markdown, with JSON-LD stripped on an on-demand read (RESONEO is careful to say this says nothing about what the index does with structured data); and there is a two-layer cache — a page is not re-fetched within 30 minutes, and cached copies were still being served 91 days after the last logged OpenAI robot visit, with a server log to back it. On discovery: GPTBot downloaded RESONEO's sitemap roughly once a day and followed none of the eight URLs listed only there, while crawling all eight pages reachable through a link the same morning; Bingbot, ClaudeBot and GoogleOther did the exact opposite. "At OpenAI, discovery runs on links."

What our own run this morning adds: the homepage share, engine by engine

RESONEO's "brand homepages written from memory" reading is a hypothesis it flags as untestable from its data. Our weekly self-audit is a different instrument on a different path — 12 small-business buyer questions × 4 engines × 3 samples through the vendors' APIs (OpenAI's web-search tool, not the ChatGPT client; every one of the 279 OpenAI citation URLs this week carries `?utm_source=openai`, not `chatgpt.com`) — but it stores every cited URL, so we cut this morning's run (2026-08-17, 13:11 UTC, run `jh72qsgx…`, 144 of 144 answers completed) by URL depth. ChatGPT cited a homepage (path `/`) in 74 of its 279 citations, 26.5%, in 20 of 36 answers, across 54 distinct domains. Perplexity: 8 of 710 (1.1%). Claude: 5 of 266 (1.9%). Gemini: 0 of 382. That is the fourth consecutive weekly run reproducing the outlier we measured on the previous three (28.4% across 07-28/08-03/08-10), and it is consistent with — not proof of — RESONEO's reading that ChatGPT's instant-style citations lean on remembered brand roots while the other engines link the page that answered. Our own four ChatGPT citations this week are the same shape: `askedabout.com/` in all three samples of "How can I check what ChatGPT says about my business?" and one sample of "Which tools show how ChatGPT, Perplexity and Gemini recommend businesses?" — the homepage, never a /news or /data page, while Perplexity's three citations of us were all a deep comparison page. Applying the study's snippet rule to the two pages of ours that were cited this week: our homepage's H1 is 39 characters and the 202-character window from it reads "What is AI saying about your business? Your customers now ask ChatGPT, Perplexity and Gemini who to buy from. A free sixty-second check shows whether AI recommends you — or quietly sends them elsewhere"; the comparison page's H1 is 69 characters and its window is the price sentence. Both pages are 64 KB and 97 KB — a long way from the 4 MB wall. One more line from our own logs, in the RESONEO spirit of saying what we can see: in the ~14 hours of today's crawler window, OAI-SearchBot fetched 51 distinct paths on askedabout.com and `/robots.txt` three times, GPTBot fetched exactly one page (the homepage), and no OpenAI agent requested `/sitemap.xml` once. Artifact: `content/news/chatgpt-homepage-share-run3-2026-08-17.json`.

Why this matters if you care about AI visibility

• Your page is a three-line card until it is opened — write the card. In the common case ChatGPT never reads past a title, a URL and ~202 characters starting at your H1. If your H1 is a brand slogan and the first paragraph is a hero line, that is your entire grounding text on OpenAI's own index — and the meta description you polished is not in it. Put the answer, and the question it answers, in the H1 and the sentence after it. This is the same mechanism our slug study and our own first citation pointed at from the other side: the pages that get cited state the question they answer.
• Which engine sees you depends on how the user asked, and the free tier now runs mostly on an index Google rank does not buy you into. Paid thinking mode is three-quarters scraped Google, so ranking still carries there. Free instant and free Think are two-thirds to three-quarters labrador, an in-house index whose admission criteria RESONEO could not crack in sixteen days of monitoring (no page entered) — but whose discovery is link-driven, not sitemap-driven. Being reachable by a link from a page OpenAI already knows is the observed path in; a sitemap entry alone was followed zero times out of eight.
• A homepage citation is the weakest kind, and ChatGPT is the engine that hands them out. Our four-run series has ChatGPT at 26–28% homepage citations against ~0–2% for the other three; RESONEO reads instant-mode citations as largely memory-and-widget links, and finds that pages the model actually opens are deep pages, almost never home pages. If the only citation you can earn is your root URL from ChatGPT, three engines have nothing to link and one is linking a page it did not read. The fix is a deep page that answers the question — then check which engines name you, and which link a page rather than a slogan.