What ChatGPT actually reads before it cites your page

ChatGPT does not read the web in one clean step. A recent reverse-engineering analysis from RESONEO suggests it works through three layers: a discovery index, a shared reading cache, and a much smaller set of pages it actually opens live. That distinction matters because citation odds appear to change sharply by layer. In the July dataset, which covered 1,200 ChatGPT answers, 88,000 search results, and 26,900 distinct pages, a page that was actually opened in thinking mode ended up cited 74% of the time. A page that was merely retrieved but never opened was cited just 7% of the time. For SEO and GEO teams, the message is simple: being present is not the same as being read, and being read is not the same as being cited.

• ChatGPT appears to ground answers through an index, a cache, and occasional live page opens.
• The top of the page matters more than many teams think, especially in faster retrieval modes.
• Live page opens are rare, but when they happen, citation odds jump dramatically.
• Clicks from ChatGPT do not tell the full story because some high-value reads leave no obvious UTM trace.
• GEO now depends on discoverability, extractability, and source selection, not just rankings.

What are the three layers behind a ChatGPT citation?

The research points to a three-layer retrieval stack. That is the most useful takeaway because it explains why brands can seem visible in one test, invisible in another, and half-correct in a third. ChatGPT is not choosing from one stable pool of web pages. It appears to choose from different pools depending on mode, cost, and whether it decides a page is worth opening.

Layer one is discovery. The researchers describe an in-house retrieval hub, labeled labrador, that can return a URL, the full page title, and a short snippet of roughly 200 characters. In fast modes, that small package may be most of what reaches the model from a web page.

Layer two is reading. Once a page has been fetched, ChatGPT appears to keep a full Markdown copy in a shared cache. That means the system may be able to reuse a stored version of your page later, even if no fresh request hits your server at that moment.

Layer three is live opening. This is the expensive step. It seems to happen mainly in thinking mode, where the system can spend more time and money to fetch a page directly through the ChatGPT-User agent.

A concrete example makes this easier to picture. Imagine you publish a comparison page for “best CRM for small B2B teams.” In one ChatGPT session, only your title and a thin snippet may be used. In another, a cached Markdown copy may inform the answer. In a deeper reasoning flow, the page may be opened directly, read more fully, and then cited. Same URL, different retrieval depth, different outcome.

Why does instant mode behave differently from thinking mode?

Because the retrieval path appears to be economic before it is editorial. The research argues that faster modes rely more heavily on what OpenAI already has on hand, while slower reasoning modes invoke more expensive tools such as scraped Google results and direct page fetches. In other words, the same prompt can reach a different web depending on whether the user asks for speed or depth.

In the follow-up replay across free and paid configurations, free Think reportedly pulled 74.7% of its results from the in-house labrador layer, while paid Thinking was almost the reverse, with 75.3% coming from scraped Google results. That is not a minor implementation detail. It means two users can ask near-identical questions and receive answers built from different retrieval ecosystems.

This also helps explain why ranking well in Google is still important but no longer sufficient. If the answer path leans on Google-fed retrieval, classic SEO strength remains a major advantage. If the path leans on the in-house index, your title, snippet shape, and eligibility inside that separate discovery layer matter more. We made a related point in our analysis of ChatGPT citing fewer domains: the citation surface may look rich, while the underlying opportunity set gets tighter.

There is a second nuance here. The research found that paid Thinking became narrower between late July and late July 2026, with fewer fan-out queries, fewer URLs, and fewer distinct domains on comparable prompts. That suggests teams should stop treating “ChatGPT visibility” as one stable KPI. Retrieval conditions can change within days, not quarters.

For brands, this creates a practical split. Query classes that trigger deeper reasoning may reward classic authority and strong page detail. Faster mass-market usage may reward pages that communicate the right idea almost immediately. Both matter, but they do not reward the same page elements in the same way.

What does ChatGPT seem to keep from your page?

In the discovery layer, not much. According to the research, the in-house index stores the full page title and a query-independent snippet of about 200 characters. The snippet is built around the H1 and whatever visible text sits nearby. That can include useful content, but it can also include a category label, a byline, a date, image alt text, or even a table of contents. In one measured case, the snippet was entirely table of contents and contained no substantive body content at all.

That should change how teams think about the top of the page. If the first useful sentence appears too late, the discovery layer may have little idea what the page actually says. If the content above the H1 is noisy, the wrong fragments may become the machine-readable summary.

The cached reading layer is richer, but it has limits. The Markdown conversion described in the analysis strips scripts, iframes, and JSON-LD. Image alt text survives. Text hidden by CSS may still be extracted. The fetcher reportedly does not execute JavaScript, and pages above 4 MB are rejected outright rather than partially read.

This is where a lot of AI visibility work turns technical. A beautifully designed page can still be a weak AI asset if the key content depends on client-side rendering, if the opening section is cluttered, or if the page is so heavy that the agent fails before reading anything useful. That is exactly why technical SEO audits now need an AI-readiness layer, not just the old crawlability checklist.

A simple example: if your pricing explainer loads the main copy through JavaScript after page render, a human user may see a complete page while the fetcher sees a shell. If your FAQ lives in structured data only, the cache path described in the research may never carry that schema into the model at all. If your best proof point starts 600 words down the page, instant retrieval may never surface it.

BotRank's Take

The most important lesson here is not that ChatGPT has a quirky pipeline. It is that AI visibility has become a layered measurement problem. Brands often ask, “Are we indexed by ChatGPT?” That is now too crude a question. A page can be discoverable, never opened, and still lose. It can be fetched once, cached for weeks, and describe your brand using outdated language. It can even be cited less often than a third-party review page that frames you better than your own site does.

This is where BotRank's Source Analysis feature becomes genuinely useful. It helps teams inspect the sources and pages sitting behind AI answers, see whether those pages actually mention the brand clearly, and review how useful the cited material really is. Combined with AI Visibility tracking across multiple LLMs, that gives a more honest picture of whether your brand is merely eligible, actively reused, or consistently chosen. That is a better operating model than guessing from one screenshot or one lucky citation.

Why does this change the SEO brief for AI search?

Because GEO is not just ranking work with a new label. Generative engine optimization, or GEO, is the practice of improving how often a brand is discovered, understood, cited, and recommended inside AI-generated answers. In classic SEO, the page could afford to rely on the click. In AI search, the page often needs to carry its value before the click happens, and sometimes without any click at all.

The RESONEO findings sharpen that shift. If instant retrieval is grounded by titles and short frozen snippets, then the first job is not “rank for the keyword.” The first job is “state the answer cleanly where the retrieval system can actually extract it.” That is why we keep repeating a core GEO principle in AI visibility starts before the prompt and ends with citations. Your page is competing not only for traffic, but also for selection and reuse.

There is a technical layer as well. The research suggests that noindex does not stop cached copies from being reused once fetched. It also suggests that JavaScript-heavy pages can be invisible to direct fetches, and that oversized pages can fail completely. This does not mean every team should panic and rebuild its site. It means the old assumption, “if Google can handle it, AI systems will too,” is no longer safe.

It also changes the content brief. The full title appears to reach the model untruncated in the in-house index, while the meta description may be ignored there but still matter in Google-fed pipelines. So the right move is not to abandon traditional metadata. It is to write titles, headings, and top-of-page copy that work across more than one retrieval path.

A strong page for this environment usually has four qualities:

• A title that makes sense on its own
• An H1 and opening paragraph that state the core answer quickly
• Important facts visible in static HTML, not buried behind scripts
• Proof points and entities that are easy for a model to quote accurately

This approach works especially well for product comparisons, explainers, category pages, and research summaries. It is less forgiving for bloated pages that make users scroll through branding furniture before the real answer appears.

How should brands respond right now?

Start with a focused cleanup, not a grand rewrite. The research itself makes an important point: the plumbing is changing fast. Specific internal labels may disappear. Retrieval mixes may shift. But the structural lesson is stable. AI systems still need pages they can discover, parse, and trust quickly.

Here is a practical response plan for marketing and SEO teams:

• Rewrite titles as standalone meaning units. If the title is all the system gets at first, it must carry the subject clearly.
• Audit the first 200 characters after the H1. Remove clutter and make the opening line say something quotable.
• Reduce front-loaded noise. Dates, category labels, and oversized table-of-contents blocks can crowd out the actual message.
• Check static render quality. If the important content is not visible without JavaScript, treat that as a visibility risk.
• Watch page weight. If a page approaches the 4 MB ceiling described in the research, it may be unreadable to the fetcher.
• Inspect off-site framing. AI systems may retrieve your page, a review page, a forum thread, and a press mention together. The best-framed source can win the citation.

If you want to operationalize that work, BotRank's GEO Page Analysis can track the pages you care about, score their AI-readiness over time, and surface the technical gaps that block discovery or reuse. And if your team needs a prioritized backlog instead of another dashboard, the Recommendations workflow is built to turn those findings into concrete next steps.

There is also a strategic reason to look beyond your own pages. The research notes that being retrieved and being cited are not the same market. Some sources may help the model think without ever earning the visible citation. Others may frame the answer even when your own page is technically available. That is why small-site inclusion alone is not enough, as we explored in our look at ChatGPT's index and smaller websites.

What should you measure if clicks tell only part of the story?

Do not stop at referral traffic. One of the most useful findings in the research is the analytics blind spot: pages opened directly by ChatGPT in thinking mode may not show the familiar utm_source=chatgpt.com parameter in user-visible citations. In plain English, some of the pages most likely to influence an answer may leave weaker evidence in normal analytics reports than the pages that only got clicked.

That means your measurement stack should include at least five layers:

• Prompt-level visibility: when your brand appears, for which questions, and in which models
• Citation rate: how often your domain or brand is actually cited
• Source mix: whether AI systems prefer your site, reviews, press, forums, or aggregators
• Page readiness: whether the important pages are technically easy to fetch and interpret
• Log evidence: whether agents such as ChatGPT-User are visiting the pages you expect

This is where a recurring audit beats one-off testing. Models change. Search fan-outs change. Index coverage changes. If you only check when a stakeholder asks why a competitor was recommended, you are already late. A better rhythm is ongoing monitoring across prompts, models, and cited sources, which is the operating logic behind BotRank's platform.

The deeper point is strategic. AI search performance now depends on what happens before the click, during retrieval, and inside the answer. If you measure only traffic, you will miss the stage where brand preference is increasingly being formed.

FAQ

Does this mean Google rankings no longer matter for ChatGPT?

No. The research suggests Google-fed retrieval is still important in thinking modes. The change is that Google rankings are no longer the only gateway, and sometimes not the main one.

Should I rewrite every page around a 200-character snippet?

No. That would be too narrow and too reactive. But you should make sure the page title, H1, and opening sentence communicate the core point quickly and cleanly.

Is structured data useless for AI visibility?

No. The research suggests JSON-LD may be stripped in one cache path, not that structured data is useless everywhere. It can still help other search and retrieval systems, so the smarter move is to keep it while also making the visible HTML strong on its own.

What is the most practical next step for a brand team?

Pick 10 high-value pages, audit how clearly they state their core answer at the top, check whether they render meaningful HTML without JavaScript, and monitor how often those pages or adjacent sources appear in AI answers. If you want that process to become systematic instead of manual, BotRank is a sensible place to start.

Here is the real takeaway: ChatGPT visibility is no longer a simple indexing question. It is a layered retrieval question. The brands that win will be the ones whose pages are easy to discover, easy to read, and easy to cite when the model finally decides to open them.