# Retrieved isn't opened: the AI visibility gap most advice ignores > Pages ChatGPT opened were cited 74% of the time. Pages it only retrieved, 7%. What that gap means, what it doesn't, and the two changes we made. By Jhonty Barreto · 2026-10-27 · Take · AI assistants Source: https://jhonty.md/retrieved-is-not-opened/ ## Key takeaways - In RESONEO's July 2026 capture of ChatGPT answers, pages ChatGPT opened were cited 74% of the time. Pages it only retrieved were cited 7% of the time. - The two groups aren't like for like, so read it as what ChatGPT favours, not proof that opening causes a citation. - GPTBot and ClaudeBot don't run JavaScript. If your pricing, services or proof only appear after hydration, they never see them. - llms.txt and JSON-LD do no harm, but neither showed a measurable lift in ChatGPT citations in the studies we trust. Being retrieved by ChatGPT buys you about a 7% shot at a citation. Being opened buys about 74%. Almost every AI visibility pitch I've seen this year is aimed at the first number. The second one is where the citations are. ## Two different things happen when ChatGPT answers When an assistant answers a question about your market, two separate things happen, and most people talk about them as if they were one. First there's the set of pages it pulls in to ground the answer. For most of those, ChatGPT only ever sees what the search system hands back: a title, a URL and a short snippet. Then there's a much smaller set of pages it actually fetches and reads in full. Most of the industry is selling its way into the first set. "Get mentioned in AI answers", "show up in AI search". That's the retrieval set. It's the easy part, and it's not where most citations come from. ## What the numbers say RESONEO, a French SEO consultancy, captured ChatGPT sessions through July 2026 and tagged what happened to every page. Search Engine Land covered the results. Pages that sat in the grounding set without being opened were cited about 7% of the time. Pages ChatGPT opened and read were cited about 74% of the time. Only a small share of retrieved pages got opened at all. That's a tenfold gap. Getting retrieved buys you roughly a one in fourteen chance. Getting opened, and then being worth reading once you're open, is the jump. ## What the numbers don't say This is the bit most people sharing the study leave out, and it matters. The two groups aren't like for like. The opened pages came from ChatGPT's slower thinking mode, which in July only paying users had. The unopened group also includes pages from quick answers. So the study shows which pages ChatGPT favoured. It doesn't prove that getting opened causes the citation. I still think the distinction holds up, because it matches what we see in server logs. On the two small business sites we checked, ChatGPT went to the homepage almost every time it opened anything ([I wrote that up here](/what-chatgpt-opens-on-your-site/)). But the exact percentages come from one corpus and one month, so treat them as theirs. ## Change one: we look at every site with JavaScript off If the job is "be worth reading once opened", the first question is what the crawler actually gets when it reads. Vercel and MERJ looked at a month of AI crawler traffic across Vercel's network: about 569 million GPTBot requests and 370 million ClaudeBot requests. Neither crawler executed any JavaScript. ClaudeBot downloaded JavaScript files about a quarter of the time and still never ran them. So if your pricing, your service list or your reviews only appear after the page hydrates, they aren't there for either of them. The flat version of this claim ("AI crawlers can't see JavaScript") is wrong, though. In the same study AppleBot rendered JavaScript, and Google's crawling for Gemini goes through Googlebot, which renders. It's GPTBot and ClaudeBot that don't. What we do now, before anything else on a new site, is load the key pages with JavaScript disabled and read what's left. That's roughly the page ChatGPT reads. It's also why every page on this blog is plain HTML that reads the same with JavaScript switched off. ## Change two: we stopped selling the easy fixes The second change was taking things out. We stopped recommending llms.txt to clients who want to get cited. Ahrefs looked at server logs across 137,210 domains for May 2026. About 97% of the llms.txt files on those sites got no requests at all that month. Schema is the other thing people sell for getting cited. Ahrefs ran a matched test on 1,885 pages that added JSON-LD, against 4,000 similar pages that didn't. The change in ChatGPT citations was about 2.2%, which they treat as noise. One caveat worth knowing: every page in that test already had a lot of AI citations before the schema went on, so it says nothing about brand new pages. Neither is harmful. Schema is still worth adding for Google's rich results. It just isn't the thing that gets you cited. Next to a homepage that only makes sense with JavaScript running, it's tidying the front garden while the front door is painted shut. ## Try it on your most commercial page The ten minute version: open your most commercial page with JavaScript switched off. In Chrome that's DevTools, then the command menu, then "Disable JavaScript". Read what survives. Is the offer there? The price, or how pricing works? The proof? If the answer is a logo, a loading spinner and a cookie banner, that's your pitch to the machine, and no amount of llms.txt fixes it. What's sitting on your page after hydration that a crawler never sees? ## Sources - [Search Engine Land: RESONEO's analysis of ChatGPT's retrieval stack](https://searchengineland.com/chatgpt-retrieval-stack-index-cache-pages-485036) - [Vercel and MERJ: The rise of the AI crawler (December 2024)](https://vercel.com/blog/the-rise-of-the-ai-crawler) - [Ahrefs: We analyzed 137K sites, 97% of llms.txt files never get read](https://ahrefs.com/blog/llmstxt-study/) - [Ahrefs: does schema markup increase AI citations?](https://ahrefs.com/blog/schema-ai-citations/)