All writing

Why ChatGPT does not mention your business

Assistants name a handful of businesses per answer and yours is probably not one of them. The measurement that explains why, the build reasons behind it, and what to fix in the order that pays.

Clients ask me this in almost the same words every time. They type their own category into ChatGPT, read four names that are not theirs, and want to know what those four did.

The short answer is that those four were read. Yours was not.

The number that decides it

An assistant does not answer from everything it has ever seen. For a commercial question it goes and fetches a handful of pages, then writes an answer out of those. Whether your site is in that handful is the whole game, and the size of the effect is now measured.

The best-structured study I have found was published in September 2026: roughly 35,000 prompt and engine observations across 2,854 prompts and 75 projects, two major engines, four months.

  • When neither the brand nor its domain appeared in the retrieval path, the brand was mentioned 2.8% to 3.8% of the time.
  • When the brand’s own domain was cited, mention rose to 49.0% and 58.4% depending on engine.
  • With an own-domain citation plus a branded search, 91.4% and 100%.

The author disclosed a financial interest, being the chief executive of a measurement company, and labelled the model predictive rather than causal. Both are worth carrying. The direction is consistent with everything else published, Google’s own account of how grounding works included.

Three per cent against forty-nine is not a tuning problem. It is the difference between existing and not, and it turns on whether a machine could read your page in the second it was willing to spend.

What keeps a site out

This is the part nobody sells you, because it is unglamorous and it is mostly repair.

It does not render without JavaScript. A great many sites, including some expensive ones, send an empty document and draw the page in the browser afterwards. Some crawlers execute that. Several do not, and the ones that do give it a budget. When I rebuilt this site I took out jQuery, GSAP and the page runtime, twenty kilobytes of it, and the text now exists in the HTML before anything runs.

It is slow at the first byte. The eleven production sites I measured paint in a median of 0.6 seconds. A fetch that takes four is a fetch that gets abandoned, and an abandoned fetch is a source that was never there.

It never answers the question in plain words. Assistants quote sentences. A page of adjectives about craft and passion contains nothing quotable, so even a page that was read contributes nothing. The pages that get cited say what a thing costs, how long it takes, what it does not do.

The identity is ambiguous. Three spellings of the business name, a different description on every profile, no category stated anywhere. Mentions that should accumulate into one entity scatter into several half-people instead, and none of them is solid enough to be named.

None of these is exotic. All four are build decisions, and all four are checkable in an afternoon.

The half that is not your site

I will not pretend the site is the whole of it, because it is not and you will find that out eventually.

Assistants corroborate. They favour what several sources agree on, so being named by other people raises how often you enter the conversation at all. That means being in the comparison and roundup articles your buyers trigger, holding consistent profiles where your industry looks, publishing something original enough that other people cite it, being interviewed and answering questions in public.

That work is slow and it is where a real reputation lives. It is also not billable in an afternoon, which is exactly why so much AI visibility advice stops at the domain boundary and charges for the easy half.

What makes the slow half compound, though, is the thing on your own site: one name, one description, one category, profiles pointing back. Without it, six months of mentions resolve to nobody. That is brand work wearing a technical hat, and it is the piece most AI visibility offers leave out entirely. Across four roundups of agencies selling this, branding was mentioned for two out of roughly twenty-five.

What to check this week

In this order, because the order is the whole of the advice.

  1. Load your homepage with JavaScript disabled. If the page is blank, nothing below matters yet.
  2. Time the first paint. Anything over two seconds is costing you fetches.
  3. Read your own pages for one quotable sentence each. A price, a duration, a number, a limit. If there is none, write one.
  4. Make the name and the description identical everywhere. Site, LinkedIn, directories, the lot.
  5. Ask the assistants your buyers’ actual questions, once a month, and write down what comes back. Six months of that is the only way you will know whether any of this worked.

Four of those five are free and none of them needs me.

When it is worth paying somebody

When the first two keep failing and the site is the reason, which is most of the time. That is a build problem, and it is what I do: the brand so the mentions resolve to one business, and the site so it is fetchable, fast and quotable when the assistant finally comes looking.

I publish what that costs rather than quoting on discovery. If you want it priced against your own site, tell me about the project and you will have an answer in two working days.

Questions

Why does ChatGPT not mention my business?
Because your site is not in the retrieval path for the question. Assistants answer from a small set of pages they fetched for that prompt, and a site that is slow, renders only with JavaScript, or never answers the question in plain words does not make that set. Measured across roughly 35,000 prompt observations, a brand absent from the retrieval path was mentioned about 3% of the time.
Does my website actually matter for AI visibility?
Enormously, once it is retrieved. In the same study, brands whose own domain was cited were mentioned 49% to 58% of the time depending on the engine, and 91% to 100% when an own-domain citation was joined by a branded search. The gap between 3% and 49% is the website's half of the job.
Will schema markup or an llms.txt file fix it?
No, not on their own. Google has said in its own documentation that it does not use llms.txt. Structured data helps a machine agree with what your page already says, so it is worth having and it is not the lever. Being fetchable, fast and unambiguous is.
How long does this take to change?
The build half takes effect as soon as the pages are recrawled, which is days to weeks. The reputation half, being named by other people, runs on the timescale of actual relationships. Anybody promising a fast result on the second half is quietly selling you the first.

Read next

All writing, or see how I build websites.