ChatGPT visibility: what actually works, measured

The subject is young, and the volume of assertion about it far exceeds the volume of measurement. This piece states only what can be verified, and flags explicitly what cannot — including where that contradicts the pages currently ranking best on the question.

How does ChatGPT pick its sources?

Two mechanisms have to be separated first, because they are constantly conflated.

Training memory. The model read the web at a given moment. It can mention your company without ever fetching your site. That is the most fragile position there is: it does not update, it does not depend on you, and it disappears with the next model.

Retrieval. When the question calls for it, ChatGPT fetches pages and answers from them. That is the only half where your site does any work, and the only half you have a grip on.

The practical consequence is counter-intuitive: a brand can be cited without its domain ever being retrieved. A serious audit separates the two; most commercial pitches merge them.

Do you need to be in Bing?

This is the question nearly every page settles too quickly, in one direction or the other.

ChatGPT search originally ran on Bing's index. Since then OpenAI has operated its own crawler and its own index, and publisher agreements have been added on top. The defensible reading today is a blended stack: the dependency on Bing has loosened rather than disappeared.

What that means concretely: registering your site with Bing Webmaster Tools is still worth the half hour, but it is no longer the single lever it was described as in 2024. A page selling you "Bing optimisation" as the key to ChatGPT visibility is selling an architecture that has changed.

Why can a site rank on Google and be invisible to ChatGPT?

This is the most important technical point here, and the only genuinely binary one.

AI crawlers do not execute JavaScript. The joint Vercel and MERJ analysis, covering more than 500 million GPTBot fetches, found no evidence of execution at all. The crawler does fetch JavaScript files — around 11.5 % of requests — but never runs them. The finding extends to ClaudeBot and PerplexityBot.

The exception is Gemini, which inherits Googlebot's rendering infrastructure.

The consequence is blunt: if your page content only appears once JavaScript has run — a catalogue loaded by API call, prices rendered afterwards, an accordion whose contents mount on open — your site can rank normally on Google, whose crawler does render, and be a blank page to ChatGPT, Claude and Perplexity.

It is checkable in one command, with no paid tool: request the page while presenting as GPTBot, strip scripts and tags, count what is left. A number close to zero means invisible.

Should you create an llms.txt file?

No. And this is where this piece contradicts head-on several of the pages ranking best on the query, which still recommend it.

The available measurements converge and are severe. Across more than 500 million AI bot visits observed over 90 days, 408 targeted an llms.txt. Ahrefs, over 137,000 sites, finds that 97 % of those files received no traffic. SE Ranking measures roughly 10 % adoption across 300,000 domains: one site in ten has posted one, and nobody reads it. No major vendor — OpenAI, Google, Anthropic, Meta, Mistral — has committed to consuming it, and John Mueller publicly compared it to the meta keywords tag.

What works instead is more mundane, which is precisely why it sells badly: an ordinary page, served as HTML readable without JavaScript, declared in the sitemap, and linked from the rest of the site. The file nobody reads becomes a URL every crawler reads.

What are the percentages you read everywhere worth?

Not much, and it helps to know what produces them.

The field is saturated with precise-looking figures — this factor supposedly accounts for 55 % of citation probability, that one for 14 %. They almost always come from tool vendors, who have a direct commercial interest in demonstrating that everything has changed and a subscription is required. Some are credited to no source at all.

The most telling detail is that one of the best-ranking pages on this query itself writes that such percentages "rest on no publicly verifiable study" — in an article that quotes one. That is not dishonesty; it is the real difficulty of the subject. Something has to be written, and independent measurement does not exist yet.

The rule I apply, and that you can apply without being technical: before accepting a figure on this subject, ask who produced it and what they sell. The robust facts in this piece — no JavaScript execution, the inertia of llms.txt, the independence of the agents — come from server log analysis and vendor documentation, not from marketing studies.

What can actually be measured, and what cannot?

Two different questions, and only one is within reach without a subscription.

"Am I cited?" — querying the models dozens of times, engine by engine, is expensive and needs a dedicated tool. And there is no single ranking: the same site can be highly visible on one engine and absent from another. Any metric aggregated over "the AIs" is misleading.

"Can I be?" — this is verifiable for free, and it is the half that yields actions rather than a score: text actually served to crawlers, agents named in robots.txt, coherent structured data, a canonical reference page, an up-to-date sitemap.

One point almost nobody makes: two named agents do not obey robots.txt. OpenAI writes that its rules "may not apply" to ChatGPT-User, and Perplexity that Perplexity-User generally ignores the file — in both cases because a person triggered the fetch. Declaring them is still worth doing, but it is an intention, not a control.

And an expensive trap: many sites blocked training crawlers in 2023 and 2024 to protect their content, and excluded themselves from citation along the way. These are not the same agents, and OpenAI states that each setting is independent of the others. A Disallow placed under User-agent: * blocks all of them.

Where should you start?

With measuring the text served without JavaScript. It is free, it takes a minute, and it is binary: as long as it fails, everything else is decoration.

Then with robots.txt, naming the search agents explicitly so the intent survives the next refactor.

That is what my diagnosis covers — and it tells you whether you can be cited, never whether you are. Anyone promising you the second measurement for free has not made it.