Three paths, not one
Search Engine Land's August 2026 account of ChatGPT's retrieval stack describes three ways a page reaches the model: OpenAI's own index, where a short snippet is stored at index time; a cache of pages the model previously opened in full; or a live fetch during the answer by the ChatGPT-User agent. The path decides how much of the page the model sees.
| Path | What the model sees | Consequence for your page |
|---|---|---|
| Index snippet (the fast, free Instant path) | Title plus roughly 200 characters of visible body text after the H1, captured when the page was indexed | If the answer is not in that span, the fast path does not have it |
| Opened and cached | The full page, as previously fetched | Pages the model has opened are cited about 74% of the time in the analysis; retrieved-only pages about 7% |
| Live fetch by ChatGPT-User | The raw HTML at answer time, without executing JavaScript | Client-rendered content is invisible; pages over 4 MB are rejected with HTTP 400 |
What "the first 200 characters" means in practice
Count from the end of the H1. Byline, date, breadcrumbs and share buttons are dropped or waste the budget; the extraction is body text. The first sentence states the mechanism. The second states what to do. If the page opens with "In today's fast-changing landscape", the fast path has stopped reading.
The meta description does not rescue it. On the index path the analysis found it ignored entirely; the snippet is cut from visible text, not the head. Write the meta description for humans on Google, and the first paragraph for the model.
The checks, in order
- Fetch the page without JavaScript. Use curl or view source. Find the first sentence of your answer in the raw HTML. If it is not there, the page is client-rendered where it matters, and copy editing will not help until that changes.
- Copy the text after the H1 and count 200 characters. Does that span say what the page is for and how the thing works? If not, rewrite the first paragraph.
- Check the response size. Anything approaching 4 MB, including inlined scripts and base64 images, risks a 400 on the live-fetch path.
- Check the status and canonical. A 200 with a self-referencing canonical. Redirect chains and soft 404s put the page on no path.
- Give the model a reason to open the page. The jump from 7% to 74% happens when a page is opened rather than retrieved. A snippet that promises a concrete answer (a number, a table, a step list) earns the open.
What this means for how you write
- The H1 is the job. The first paragraph is the answer. Everything else is support.
- One first-party number in the first paragraph, with its date. Numbers are what a snippet can carry and a model can quote.
- No hero image between the H1 and the first paragraph. Visual order does not matter; source order does.
- No accordion or tab hiding the answer. Content in the HTML can be styled; content loaded on click does not exist.
- A table or numbered list within the first screen. It is the block most likely to be lifted when the page is opened.
How to know it worked
Re-index happens on OpenAI's schedule, so the fix is not visible the next day. Run the prompt weekly and watch the cited URL column; the change shows as your URL appearing on repeated runs. Record the date. If three weekly runs pass with no change, re-check the raw HTML before rewriting again; the most common failure is that the edit never reached the served page.