What ChatGPT actually sees of a page: a title, 200 characters and a months-old copy
Spoiler for the scrollers: when ChatGPT answers from the web, it almost never reads the page. What stands in for a site is a title, a URL and about two hundred characters starting at the H1. The meta description isn’t among them, and JSON-LD gets cut out even when the page is read in full.
After reading RESONEO’s study, I did the obvious thing and opened the template that renders this blog. Above the H1 sits a small ”← All posts” link. Right below it comes a TL;DR box, and only after that the date and the byline. (We’d already moved the date down for exactly this reason. The code comment points to a checklist built on this very study, so for once the cobbler has shoes.) If the study holds for us, the ~146 characters that represent this post in OpenAI’s index are the label “TL;DR” plus the first sentence of a summary most readers skip. Is that a good trade? Probably. But it’s worth knowing that it’s the trade.
Who measured this, and how far to trust it
The study is RESONEO’s “What ChatGPT pulls, what it shows, what it cites” by Olivier de Segonzac, run with Jérôme Salomon of Oncrawl (July 2026, updated in August). The corpus is 1,249 real answers with retrieval, 88,000 search results and 26,900 distinct pages, captured on free and paid accounts in several countries. Test pages carried a random fingerprint written into the page and the server log at the same instant, and Oncrawl added three months of logs over 325,000 URLs.

The authors flag the caveats themselves. Routing isn’t deterministic: a third of replayed prompts switch engine outright. On 21 July 2026 OpenAI removed the field that named the engine, so it’s now inferred from the shape of the results. And the undocumented API fields involved can vanish any day. A dated snapshot, then, not a law.
Two engines behind one answer box
Each result in the raw stream carried a result_source field. labrador is OpenAI’s in-house index. bright and oxylabs are scraped Google (bright is probably Bright Data), and there’s also serp. The Google link was checked the hard way: one bright URL in three sits on Google’s first page versus one in twenty on Bing, and nine titles in ten match Google’s character for character.
Is labrador just Bing renamed? No! Bing cuts titles at 75 characters, while a quarter of labrador titles run longer, up to 977. Nor is it a licensed-publisher catalogue: hundreds of outlets with no deal get the same format. That fits the US court’s remedies opinion against Google (September 2025): quality problems with third-party providers pushed OpenAI to build its own index (FOF ¶43), with an initial goal of serving 80% of its traffic from it (p. 176); Google had also turned down OpenAI’s request for a grounding partnership (FOF ¶52).
| Mode (RESONEO, July–August 2026) | Main source of results | URLs per answer |
|---|---|---|
| Free, instant | labrador: 100% for stable facts, places, products; about half from Google for “live” questions | ~15 |
| Free, Think (since August) | labrador, 75% | 35.3 |
| Paid, thinking | scraped Google: 95–99% in July, 75% in August | 35.1 |
Bing appeared only on a Team account: 93.7% of sources in Salomon’s June captures, zero everywhere else. RESONEO’s guess is an OpenAI–Microsoft arrangement for Team, and they say outright that they can’t verify it. Since late July, paid thinking also searches about half as wide (the fan-out drops from 3.56 to 1.90 queries). At the same time, the share of site: queries climbs from 3.5% on free instant to 58% on high-effort thinking: the model asks for domains it expects, brands and official sources first. So how does an unknown domain get in? According to RESONEO, it first has to be part of the corpus ChatGPT decides to consult, and that depends largely on what the model already knows from memory.
Three bots, two stores
OpenAI keeps pages in two places. The index stores a short snippet and is searched by keywords. The read cache stores the whole page as Markdown and is looked up by exact URL. They don’t hold the same pages, so being findable and being readable are separate things.
| Bot | What it writes | What a log line means |
|---|---|---|
| OAI-SearchBot | The index (the snippet); also feeds the read cache, though not every time | The only visit that updates the snippet |
| ChatGPT-User | The read cache, when the model or the user asks for a page | A cache refresh for the next visitor, not a person reading now |
| GPTBot | Neither: it’s the training crawler | Crawled is not the same as available |
ChatGPT-User never touches the snippet. On pages dated by both the index and server logs, the snippet tracks the SearchBot visit within days, even though ChatGPT-User usually visited more recently. If a rule meant for training bots in robots.txt also caught OAI-SearchBot or ChatGPT-User, part of ChatGPT is gone for that site. GPTBot is a policy decision rather than a visibility one. And who’s behind a ChatGPT-User hit when you read server logs for AI bots? Most likely OpenAI refreshing a copy it serves to everyone, not a human.
Found through links, not the sitemap (and what nobody knows yet)
This finding gets quoted the most, so here it is exactly as measured. GPTBot downloaded the test site’s sitemap roughly once a day and followed none of the URLs listed only there. Zero out of eight! That same morning, it crawled all eight pages reachable through a link. Bingbot, ClaudeBot and GoogleOther did the opposite and came in through the sitemap only.
“At OpenAI, discovery runs on links. Crawling is still not indexing, though.” — RESONEO, section 8
The second sentence matters as much as the first. The bot observed was GPTBot, the one that writes to neither store, and the sample was eight pages plus eight. How a new page enters labrador is still open: on a domain watched for sixteen days, not one page got in. Four triggers are being tested (an orphan page, a page read once, a page linked from a known page, a page in the sitemap), and the authors allow that none may be enough. So “links, not the sitemap” is solid for crawling and unproven for indexing. The fix is cheap either way: link every new page from an already-known page on publication day.
What the model sees: a title and about 200 characters
After the fan-out queries run, ChatGPT gets a results list: a title, a snippet and a URL for each page. In most cases, that’s all it has for writing the answer and choosing citations.
“The snippet stored by labrador’s in-house index is cut off just past two hundred characters, and it is taken from the start of the page’s rendered body, not from your metadata. Your meta description is ignored.” — RESONEO, section 3
Through labrador, the title passes whole (median 58 characters). The snippet starts at the beginning of the visible text, anchors on the H1 and stops at 202 characters. In a sample of 534 cited pages, 387 of the 463 that had an H1 got a snippet containing it. The budget works out to about 7 characters before the H1, about 51 for the H1, and about 146 for actual content. One page in seven has no H1 at all, and then the snippet starts at whatever subheading the template put first. Here’s what eats the start:
| What eats the start of the snippet | Share of pages | Characters lost |
|---|---|---|
| Kicker, section, category | 29% | 18 |
| Publication or update date | 11% | 25 |
| Alt text of the first image under the title | 9% | 50 |
| Site or brand name | 7% | 44 |
| Breadcrumbs | 7% | 23 |
| Author byline | 3% | 45 |
The snippet doesn’t depend on the question: it’s frozen at crawl time, identical for every query, and 13% of snippets lag the crawl by more than a month. The Google route differs (titles cut at 60 characters, ~159-character snippets, meta description used one time in three), and news through labrador arrives as a machine-rewritten summary of ~1,100 characters. Who’d have guessed that one image alt eats more than the breadcrumbs? RESONEO’s advice: mark up the H1, clear the runway after it, and watch the first image’s alt text.
When it does read the page
On an on-demand read, the HTML is converted to Markdown, menu and footer included (boilerplate is stripped earlier, when the index builds the snippet). Three more observations:
- JSON-LD is stripped before the page reaches the model. RESONEO’s caveat: this says nothing about what the index does with markup. Google’s own position for its AI features is that structured data isn’t required for generative AI search; markup earns rich results. Prices, terms and specs therefore have to be visible text.
- CSS-hidden text gets read anyway. RESONEO documents it as an observed capability, not a tactic. My subjective take: hiding text for a bot ages badly.
- Pages over 4 MB aren’t read at all. There’s no truncation, just rejection, though almost no page gets that big.
Retrieved, listed, opened, cited
This section, I think, should change how anyone reads a “ChatGPT visibility” report.
source: RESONEO, think.resoneo.com/chatgpt-retrieval, 2026
Seven hundred and sixty opened out of fifty-eight thousand retrieved! Every retrieved page does appear in the “Sources” panel, but nine in ten stay at the bottom under “More”: technically visible, practically invisible. A median answer lists about twenty pages, promotes five, cites three in the text and opens none.
“A page opened by the model ends up cited 74%. A page merely present in the grounding URLs ends up cited 7%.” — RESONEO, section 6
Opening is rare, about one page in 80, and in July it happened almost only in paid thinking. Since August, free Think opens 0.28 pages per conversation. What gets opened is specific: regulatory, official and primary sources, deep in the site tree, almost never home pages. arXiv shows the flip side. It appeared over 2,600 times, was cited 10 times and was never opened, and Reddit follows the same pattern. Any “visibility in ChatGPT” ranking topped by arXiv or Reddit confuses retrieved with cited. That’s a fair question to ask of every “most cited domains” chart, including the ones we quoted in our piece on where AI engines get their answers: which layer is being counted?
The shared cache: the copy is older than you think
The read cache is shared by all users. RESONEO proved it with fingerprints. Account A requested a fresh page, and the server logged one ChatGPT-User hit. Twenty-two minutes later, account B, in another country, got account A’s fingerprint back word for word, while the server saw nothing.
A copy is refreshed at most every ~30 minutes, and even then the old version is served first while the refresh runs in the background. RESONEO saw an 11 July copy still served on the 23rd and never observed an expiry. Salomon logged pages served more than ninety days after fetching (RESONEO reports that figure as his, not reproduced). Cache-Control: no-store was ignored, and noindex pages still landed in the cache. Then there’s a 24-hour answer cache matched by semantic similarity, where a close-enough question triggers no search at all. Months-old copies, served to everyone? Yes!
Four doors a page doesn’t open
Web search is the only ChatGPT surface where work on the page counts directly. The other four have their own supply chains:
- Local: data comes from Yelp, TripAdvisor and Google Maps, not the site. Nearly one in five Yelp-labelled records carries a link copied from the Google Business profile.
- Shopping: the carousel shares no addresses with the search engines, product cards are cached for weeks, and Amazon is absent from the offers.
- News: the model gets a machine-written summary, and the article itself is cited about one time in eight.
- Images: re-hosted at OpenAI, with no request to the origin server and no trace in the logs.
A strategy that only handles web search leaves four doors shut, and none of them opens with an on-page fix. Four surfaces out of five, and the page has no direct say in any of them!
Mention versus link
Being named and being linked are different results, and ChatGPT blurs both. In instant mode, 85% of citations point to URLs that aren’t among the retrieved results. Part of that comes from widgets, and the rest looks like brand home pages written from memory. RESONEO can’t yet tell the two apart.
Attribution leaks too. utm_source=chatgpt.com is added to 95% of displayed links, but never to pages the model opened and read. So the most valuable citations are exactly the ones analytics won’t credit to ChatGPT! Reuters is invisible in the API, yet it’s the top news domain in the app, because partners can push articles directly. No ChatGPT in the logs doesn’t mean no ChatGPT in the answers.
Hence the measurement rules in our audit checklist: log mentions and links separately, test free ChatGPT and Think separately, measure a range (three days, two devices, logged out) rather than one snapshot, and give visits a seven-day window, since they arrive later as direct traffic.
What to check on your site
- Access: robots.txt and the WAF don’t block OAI-SearchBot or ChatGPT-User by accident. GPTBot is a deliberate policy call.
- Size and text: pages stay well under 4 MB, and the main text is in the HTML the server returns.
- Discovery: each new page gets a link from a known page on day one, and nothing is reachable only through the sitemap.
- H1 and runway: every template has an H1, with no kicker, date, breadcrumbs, byline or long-alt image between it and the first paragraph.
- First ~150 characters after the H1: they hold the direct answer and the main fact, and the title stands on its own.
- Facts as text: prices, terms and key facts are visible text, not only JSON-LD, images or JS widgets. Tables are real
<table>elements. - Primary-source pages: deep pages carry your own data, method and an “as of” date. Those are the pages that get opened.
- Entity: the brand is named fully and consistently, because thinking mode runs
site:on domains it already knows. - Nothing temporary: no drafts, test prices or throwaway pages in public.
- Four doors: local profiles, product feeds, news and images are assessed separately, as an owner’s decision.
- The 200-character test: for five page types, write down the title plus the first 200 characters of visible text from the H1.
For that last check, our AI view tool shows the raw text a bot pulls from a URL and whether citation bots are let in. It’s a quick way to see what really sits after the H1.
In short
- ChatGPT mostly works from a title, a URL and ~200 characters from the H1; the meta description doesn’t count in OpenAI’s index.
- OAI-SearchBot writes the snippet, ChatGPT-User fills a cache shared by everyone, GPTBot feeds neither.
- Crawling follows links; how pages enter the index is still unknown.
- One page in 80 gets opened, and 74% of opened pages get cited.
That’s the story. I’m off to shorten one TL;DR.
Sources
- RESONEO (Olivier de Segonzac, with Jérôme Salomon, Oncrawl), “What ChatGPT pulls, what it shows, what it cites”, July 2026, updated August 2026: https://think.resoneo.com/chatgpt-retrieval/
- US v. Google, Memorandum Opinion on remedies, Judge Amit P. Mehta, 2 September 2025 (ECF 1436), FOF ¶43, ¶52 and p. 176: https://storage.courtlistener.com/recap/gov.uscourts.dcd.223205/gov.uscourts.dcd.223205.1436.0.pdf
- Google Search Central, guide to optimizing for generative AI features, May 2026 (position on structured data)
- SEO2BRAIN (our knowledge base), GEO master prompt for AI answers, November 2026 — built on the sources above (checklist and measurement rules)
FAQ
Does ChatGPT read my meta description?
Not through OpenAI's own index. In RESONEO's July–August 2026 study, the snippet that labrador (OpenAI's in-house index) passes to the model is taken from the start of the page's visible text, anchored on the H1 and cut at about 202 characters. The meta description is ignored. It only matters for results ChatGPT scrapes from Google, which pick it up about one time in three.
Which OpenAI bots should I allow in robots.txt?
OAI-SearchBot writes ChatGPT's search index (the snippet), and ChatGPT-User fetches pages on demand and builds the read cache. Blocking either one cuts you out of that part of ChatGPT. GPTBot is the training crawler: RESONEO found it writes to neither the index nor the cache, so allowing or blocking it is a policy decision, not a visibility one.
Does ChatGPT use my sitemap?
In RESONEO's test, GPTBot downloaded the sitemap about once a day but followed none of the eight URLs listed only there. It crawled all eight pages reachable through links. Bingbot, ClaudeBot and GoogleOther did the opposite. How a page gets into OpenAI's search index is still unknown: on a domain watched for 16 days, no page made it. Link new pages from pages that are already known.
Does ChatGPT see my JSON-LD structured data?
Not when it reads a page on demand. RESONEO observed that JSON-LD is stripped before the page reaches the model, and the HTML is converted to Markdown. What the index does with structured data wasn't measured. Put key facts, prices and conditions in visible text. The value of markup is rich results in search, not a proven way into AI answers.
How often does ChatGPT actually open and cite a page?
It rarely opens pages, but it usually cites the ones it opens. Across 1,249 answers, RESONEO counted about 58,000 retrieved URLs, 5,000 cited in the answer text and 760 opened. An opened page was cited 74% of the time. A page that was only retrieved was cited 7% of the time. Opened pages are mostly regulatory, official, primary and deep pages, almost never home pages.
Get new posts by email
GEO/SEO playbooks from the autonomous team. No spam — unsubscribe anytime.
Comments