AI Preview

What Does ChatGPT Actually Read From a Web Page?

ChatGPT does not always read a whole web page, by our own measure across 455 tests on a real AI engine. For a quick question it can judge a page by a short preview of the headline and opening lines, and when it searches first it can judge a page by one short passage.

That preview can decide whether the page gets opened at all. OpenAI's pages cited here do not say how much of a page it reads, so treat this as a measured prediction and not as a statement of how OpenAI works.

How We Measured This

We recorded what a real AI engine searches for and opens, ran 455 tests, and compared seven ways of building page previews. The one we chose agreed with the real engine on which pages get cited 96% of the time. This is our own measurement, not a customer result, and we label it as a prediction in the product.

The practical result is simple: the headline, the opening lines, and one passage that stands on its own carry more weight than anything later on the page.

First, ChatGPT Has to Be Able to Reach the Page

According to OpenAI's crawler documentation, ChatGPT uses separate bots for separate jobs:

  • OAI-SearchBot surfaces sites in ChatGPT's search features. Sites that block it will not appear in search answers, although they can still appear as navigational links.
  • GPTBot is used for training generative models and can be disallowed independently of search.
  • ChatGPT-User fetches a page when a user or Custom GPT action asks for it, and is not used to decide search eligibility.

OpenAI's help center says a site must allow OAI-SearchBot and allow traffic from OpenAI's published IP ranges at the host or CDN. A robots.txt change takes about 24 hours to take effect. Ranking depends on multiple factors, and placement is not guaranteed.

Grimpe's Access tests each AI bot with real requests and names the cause of any block. See Which AI Bots Should You Allow in Robots.txt?.

Second, the Answer Has to Sit at the Top

Because an engine can judge a page by a short preview, a company history or a slow introduction uses the space the answer needs. If the preview shows no answer, the engine may never open the page.

  1. Put the direct answer in the headline and the first lines after it.
  2. Cut long introductions.
  3. Write at least one passage that makes sense if someone reads only those two or three sentences.

Grimpe's AI Preview shows the headline and opening lines ChatGPT reads first, marks where reading stops, and shows the passage it would quote for a target prompt. Every sentence comes from your page.

Third, the Page Has to Fully Answer the Question

A page can be well written and still lose if it never answers the question. In a Grimpe test on one page, the version holding the real answer was cited in 3 of 3 runs, and a same-length version with unrelated facts in 0 of 3. A complete answer is necessary, but it does not guarantee a citation. See How to Rewrite a Page for AI Citations.

Grimpe's Prompt-Specific Optimization aims a page at one question, and Rewrite rebuilds it while keeping your meaning, terminology, and data. If the page does not exist, Create writes it from your material.

Fourth, Write to the Page's Type

What counts as an answer depends on the page:

  1. A sales page states the offer, the audience, and the terms.
  2. A comparison applies the same criteria to every option.
  3. An informational page gives measured results and the method behind them.
  4. A reference page gives numbered steps.

The AI Readiness Score rates each tracked page from 0 to 100 and lists the weakest first.

Fifth, Test Before and After Publishing

A page that looks complete to an editor can still lose inside a real AI engine. Test it before it goes live, then track real answers against pages you did not touch.

Grimpe's Simulate tests a live page, a rewrite, or a draft in a real AI engine and shows which pages it found, read, and cited. Citations links each result to the real answer. See How to Test if AI Will Cite Your Page.

How to Do This in Grimpe

  1. Run Access to confirm OAI-SearchBot can reach the site.
  2. Find the weakest pages with the AI Readiness Score.
  3. Open AI Preview to see what ChatGPT reads first and where it stops.
  4. Use Rewrite or Create to move the answer up and fill in what is missing.
  5. Test the result with Simulate before publishing.
  6. Track results in the Visibility Dashboard and Citations.

Run a free scan or book a demo at grimpe.ai to see what AI reads on your pages.

Related reading
How to Rewrite a Page for AI Citations How Do I Get My Website Cited by ChatGPT? How to Test if AI Will Cite Your Page
See where you actually stand.
Run a free scan to see reachability, AI Readiness Score, and citations for your domain.
Run a free scan