Why your business is not showing up in ChatGPT

Someone asks an assistant for a business like yours. It names three companies and none of them is you. The useful answer is not a list of seven tricks. It is to split the question in two, because one half is measurable this afternoon and the other half nobody on earth can measure, including us.

---

Part one. The half you can establish and fix

Four things can go wrong before any model has an opinion about you at all. Every one of them is observable from outside your website with ordinary HTTP requests, none of them needs a theory about what an assistant is thinking, and between them they account for most of the real failures we see. The sequence an assistant works through when it arrives is set out in how AI reads websites.

1. Your door was closed before anyone knocked

A retrieval crawler makes an ordinary HTTP request and identifies itself with a published user-agent string such as OAI-SearchBot, PerplexityBot or Claude-SearchBot. Your CDN, your firewall and any bot-protection layer in front of your server all get to answer that request before your website is involved. A managed bot rule can return 403 or a challenge page to a named AI crawler while serving your homepage perfectly to every human visitor, and nothing in your analytics will mention it.

This is the failure people never check, because they check the wrong file. robots.txt is a policy statement you publish. Bot protection is a separate system acting above it. The two can disagree completely, and when they do, the disagreement is the finding.

You cannot be cited from a page that was never retrieved. Rule this out first, and for a surprising number of sites it turns out to be the whole answer. The Cloudflare version of this failure, and how to clear it, is covered in Cloudflare and AI crawlers.

One honest note about how this gets tested. A probe carries the crawler's published user agent, which makes the result evidence about a probe rather than a promise about the genuine crawler's full production behaviour. So a clean result is reported as "no access barrier detected", and treatment of the real crawler can still depend on IP and network-level controls outside the test.

2. The page arrived, and it was empty

The second failure is quieter. The request succeeds, and what comes back is a shell. If your site renders client-side, the HTML your server returns can be little more than <div id="root"></div> and a script tag. The page that looks finished in your browser is a blank sheet to anything reading only what the server sent.

Rendering behaviour differs from crawler to crawler and changes over time, so the durable question is not whose JavaScript engine runs today. It is whether your important content exists in the HTML returned directly by the server, without browser execution. Many automated crawlers rely heavily on that server-delivered HTML, which is what makes it the baseline worth holding.

3. A machine cannot establish what your business actually is

Now the harder one. Suppose the request succeeded and the HTML is full of your words. An assistant answering "best supplier of X for Y in Z" still has to establish who you are, what you offer, who you serve and where you operate. If it cannot establish those things from your page, it cannot repeat them. A homepage can read beautifully to a person, win awards, and state almost none of that as a fact a machine can lift.

The two that go missing most often are the two that decide those questions. Almost nobody states who their business is for, and almost nobody states what it costs, which is precisely what a person wants when they ask an assistant to choose for them.

AI Site View extracts eleven business facts deterministically, with the evidence for each one and the source it was read from, and labels anything it cannot establish Unknown rather than guessing:

• Identity. Business name, business type, description.
• Offering. Products and services, named as entries rather than implied by a headline.
• Audience. Who it is for, stated explicitly.
• Location. Where you operate or who you serve.
• Pricing. What it costs, or the honest absence of it.
• People, trust, contact and policies. Named humans, verifiable proof, real contact details, published terms.

There is no language model in that extractor, on purpose. A model always produces a confident sentence, including for an empty page. Everything is assembled from strings literally present in your HTML, which is why an UNKNOWN here means the page genuinely does not say it. That is not an accusation. It is the shortest to-do list you are going to get.

4. You are telling machines to stay out, or you go by two names

Two smaller faults, both common enough to be worth thirty seconds of your time. The first is a noindex left on the homepage, usually a survivor from a staging environment that went live with the site. It is a clear instruction to stay out, and it caps an AI Site View result at 30 for exactly that reason: while it stands, very little underneath it matters.

The second is quieter. Your structured data declares one business name and your og:site_name declares another. A machine reading both has no basis for choosing between them, and the safe behaviour is to trust neither. Those two are the only sources we allow to conflict, because a page title is a headline rather than a statement about your name, and flagging a tagline-led title as a contradiction would be scoring our own taste.

Part two. The half nobody can tell you, including us

Everything above is observable. What follows is not, and it deserves saying in the same confident voice rather than being buried in a footnote at the bottom of the page.

Nobody outside those labs can see what a model concluded about you

Whether a given assistant mentions or recommends your business is UNKNOWN, and it is unknowable from outside. What happens at answer time sits inside systems nobody outside those companies can observe, and the output moves with the phrasing of the question, the session, the region, the model version and whatever retrieval ran in that second. Typing your category into a chat window a few times is an anecdote, not a measurement. Doing it a thousand times gives you a more expensive anecdote.

We hold the same line on ourselves. AI Site View reports a deterministic technical index built from observable access and machine-readability checks, and says out loud that it is not a percentage likelihood of appearing in an AI answer. The full list of what we refuse to pretend to know is published on the methodology page.

External authority is real, and it is not in this scan

These systems do not form a view of you from your own website alone. Being written about, listed, reviewed and cited by independent sources genuinely matters, and pretending otherwise would be the more comfortable lie. It is also not something a scan of your own pages can measure. External authority measurement is not included in this scan. We would rather state that plainly than estimate it and dress the estimate up as a score.

So treat it as separate work, done off your own site. The directories and listings your buyers genuinely use. Trade and local press. Comparison and roundup pages in your category. Professional bodies and registers you actually qualify for. The places your customers already discuss the problem you solve. None of that is a growth hack and none of it happens in an afternoon, which is exactly why the tools that promise a shortcut sell you a number instead.

One practical part of it does sit inside your control: keep the facts consistent wherever you appear. The same legal name, the same address, the same one-line description. A machine reading five sources about a company that describes itself five different ways has five weak signals rather than one strong one.

Being allowed in buys nothing on its own

Allowing a crawler guarantees nothing. It removes a barrier, and that is the whole of what it does. Plenty of perfectly retrievable, perfectly machine-readable websites are never mentioned in an AI answer, because there were three better-established options for that question and the assistant only named three.

Worth separating two things people routinely merge here. Model-training crawlers such as GPTBot and ClaudeBot are a policy preference: allowing or blocking them is a decision you are entitled to make, not a defect, and it never moves an AI Site View Score. Retrieval crawlers such as OAI-SearchBot and ChatGPT-User are the ones fetching a page in order to answer a question now. A publisher who blocks training and welcomes retrieval has made a coherent decision, and a tool that marks them down for it is scoring its own opinion.

The question worth asking instead

You cannot control whether ChatGPT recommends your business. You can control whether it is even possible for it to. One of those is unknowable and the other is measurable today, and almost every business that is invisible is failing the measurable one long before the unknowable one ever gets a say.

Work the first four until they are clean, then put your effort into the fifth, and treat anyone selling you certainty about the sixth with the suspicion they have earned. If all four technical answers are healthy and you are still not being mentioned, you have at least eliminated everything you own, and you know the remaining work is off your website rather than on it. That is a real position to be in. It is not where most sites start.

Questions people actually ask

Why is my business not showing up in ChatGPT?

The question splits in two. What can be established: whether retrieval crawlers were refused at the door by a CDN, firewall or bot-protection rule, whether your important content exists in the HTML your server returns without browser execution, whether a machine can establish the basic facts of your business from that HTML, and whether a homepage noindex or two conflicting declared names are working against you. Those four are measurable from outside your site today and account for most real failures. The rest of it happens inside systems nobody outside those companies can observe.

Can any tool tell me whether ChatGPT recommends my business?

No. Nobody outside those labs can measure what a model privately concluded about a company, and answers move with phrasing, session, region, model version and whatever retrieval ran at that moment. A tool selling a ChatGPT visibility percentage or an AI authority score derived from a scan of your own page is selling a number it cannot support, because your HTML does not contain that information. We label recommendation behaviour UNKNOWN and report what we actually observed instead.

My robots.txt allows AI crawlers, so why am I still invisible?

Because robots.txt is a policy statement you publish, and bot protection is a separate system that answers the request before your website is involved. It can return 403 or a challenge to a named AI crawler while robots.txt welcomes it and your homepage loads perfectly for people. Our own two sites did that across 22,143 pages. The only way to know is to send a request with each crawler's published user agent and compare the result with what robots.txt states.

Does allowing GPTBot get my business into ChatGPT answers?

No. Allowing a crawler removes a barrier, it does not buy a mention. GPTBot is a model-training crawler, and allowing or blocking it is a policy choice rather than a defect, which is why it never moves an AI Site View Score. The crawlers that matter for answering a question now are retrieval agents such as OAI-SearchBot and ChatGPT-User. Removing every barrier makes inclusion possible. It does not make it happen.

Do third-party mentions matter, or is my own website enough?

They matter, and external authority measurement is not included in this scan. We say that rather than estimate it. Your own website is where you make yourself retrievable and machine-readable, which is necessary and not sufficient. Being written about, listed, reviewed and cited by independent sources outside your domain is separate work, done off your own site, and no page-level score can stand in for it.

Every result AI Site View reports is a point-in-time observation, measured with published-user-agent probes, stamped with the date and the methodology version that produced it. Where the evidence stops, the report says UNKNOWN.