How much of your website does ChatGPT actually keep? For pages in OpenAI's in-house search index, the answer is a title plus about 200 characters pulled from the top of the page.
That is a headline and roughly two sentences. The case study you spent three weeks writing, the pricing table, the comparison chart, the FAQ section. For retrieval purposes, most of it does not exist.
On August 15, Search Engine Journal's SEO Pulse covered an analysis by Resoneo of OpenAI's own index, the one that feeds ChatGPT search. Two findings matter, and the second one should change how you look at every page on your site.
What the analysis found
First, sites with no commercial deal with OpenAI were served from that index the same as everyone else. You do not need a licensing agreement to appear in ChatGPT search results. Whatever the deals are for, showing up in basic search results does not depend on having one.
Second, and more useful, the index does not keep your whole page. It keeps the title plus about 200 characters from the top of the page. When ChatGPT goes looking for evidence about your business, that snippet is a big part of what it has to work with.
There is a caveat on method. The researchers verified this before OpenAI removed the attribution field from its responses in late July. That removal made this kind of outside verification harder going forward. The finding stands on what was observable while the window was open, which is more visibility into OpenAI's index than anyone has had since.
What 200 characters actually is
Count it out. Two hundred characters is about the length of this paragraph and the one before it. It is less text than most websites put above their first real sentence.
Think about what occupies the top of a typical page. A navigation bar. A hero banner with a slogan. Maybe a cookie notice or an announcement bar. Then a headline. Then, on far too many sites, a warm-up paragraph about the company's mission before anyone says what the company does.
The index is not reading your nav. It is keeping text from the top of the page. If your first 200 characters of actual content are a slogan and an introduction that talks around the point, that is your representation in the retrieval layer. Writers in this field compress that problem into the phrase "template bloat burns presence." Unpacked, it says three things. Template bloat means the repeated material every page of a site carries: menus, banners, cookie notices, taglines, mission statements. Presence means whether your site appears inside AI answers at all. The connection is mechanical: the repeated material sits at the top of the page, the index keeps only the top of the page, and so the repeated material crowds out the words that would have told the model what you do. That repeated material does not just slow humans down. It displaces the only characters the index keeps.
Why this flips the usual advice
Most content advice says more: more words, more depth, more coverage. For the retrieval layer, the direction is the opposite. The constraint is brutally tight, so the question is whether your tightest text says anything useful.
Ask it directly: if a model kept only your title and your first two sentences, would it know what you do, who you do it for, and what makes you different from the site next door? For most pages the answer is no, because those two sentences were written to set a mood, not to state a fact.
Our earlier guide on checking your ChatGPT visibility recommends a direct answer in the first sentence of every page. That advice was based on how models extract answers. The 200-character finding gives it a reason that comes from how the index is built: the index keeps only the top of the page. The first sentence is not just read first. It may be close to all that is kept.
The check takes ten minutes
You do not need a tool for this.
Open your five most important pages. Homepage, main product or service page, and your three highest-traffic content pages. View the source, find where the main body text starts, and read only the title and the first 200 characters of real content. Skip the menu, the banners, the cookie notice, and any slogan text.
For each page, write down what those characters communicate. If the answer is "our product helps teams succeed in today's fast-moving world," you have found your problem. Rewrite the top of the page so the first sentence states the concrete thing: what it is, who it is for, what it does.
A page top that says "Parceit audits how your brand appears across ChatGPT, Perplexity, and Google AI Overviews, and scores the fixes in priority order" beats a page top that says "Reimagine your digital presence with data-driven insight." One of them can be retrieved. The other is decoration.
How this fits the bigger picture
Two findings this week fit together.
An analysis of 60 ChatGPT conversations published by Search Engine Journal on August 14 showed the model writes brand names into its own search queries before it retrieves anything, and being named in that shortlist is worth roughly 33 times more than being merely findable. You can read our breakdown of that piece here: [ChatGPT Decides Who to Recommend Before It Searches](https://parceit.com/blog/chatgpt-decides-who-to-recommend-before-it-searches).
Put the two together and the funnel looks like this. First the model has to know you exist, which is about naming and category association. Then it probes, and what comes back from the probe is a title and about 200 characters. Both gates are narrow. Most optimization effort targets neither, because most effort goes into pages that are already easy to fetch and hard to summarize.
In the IAB's framework from Measuring Visibility in the AI Era, published August 3, this sits at Presence and Portrayal. Presence: are you in the index at all. Portrayal: what the two hundred characters you kept actually say about you. A site can pass every crawl check and still be portrayed badly, because portrayal was decided by whoever wrote the page top.
What to do this week
Run the ten-minute check on your five key pages. Rewrite any page top that opens with mood instead of meaning. Put the plainest possible statement of what you do in the first sentence, and keep it consistent with your title tag so the two reinforce each other instead of repeating generic phrasing.
Then rerun your visibility checks and watch what changes. This is a fix you can ship without a developer, which makes it rare in this field.
The bottom line
You have been optimizing pages for readers who see everything. The index keeping your page for AI retrieval reads almost none of it. Write the first two sentences like they are the only ones that count for search, because for a growing share of search, they are.
Run a free audit at parceit.com to see how your site reads to AI search engines across every major engine, with the full methodology documented.
Frequently asked questions
How much of a page does ChatGPT keep?
An analysis by Resoneo, covered by Search Engine Journal's SEO Pulse on August 15, 2026, found OpenAI's in-house index keeps the page title plus about 200 characters from the top of the page. Verification was possible before OpenAI removed the attribution field from responses in late July 2026.
Do I need a deal with OpenAI to appear in ChatGPT search?
No. The same analysis found sites with no OpenAI deal were served from the index the same as sites with deals. Licensing agreements are not a requirement for basic retrieval.
What should my first 200 characters say?
In plain language, state what your business does and for whom, with no slogan or preamble. Treat the first sentence as the retrieval summary, because that is close to how the index uses it.
Is this the same as optimizing for featured snippets?
They share one instinct: put the direct answer in your first sentence. The difference is what that work targets. Featured snippet optimization targeted Google pulling a quoted answer out of a page it had already ranked. This concerns what OpenAI's index stores from your page before any answer is generated. Both reward answering first, but the audiences and systems differ.
Where do I check this?
View the source of your key pages, find the start of main body content, and read only the title plus first 200 characters. If that text does not state what you do, fix the page top before anything else.