Why ChatGPT does not know your company, and what fixes each cause

How does an assistant know anything about a business?

Two ways, and they fail differently. The first is training data: the text a model learned from before its cutoff. A business that was written about often, on pages that were public before that date, may be in there in some form. A business that was not, or that changed its address since, is either absent or out of date, and the model cannot tell which.

The second is retrieval. When a question calls for current facts, the assistant searches the web and reads pages before it answers, the pattern the glossary describes under retrieval-augmented generation. OpenAI's crawler documentation names the two tokens that do this work for ChatGPT: OAI-SearchBot builds the index that ChatGPT search draws on, and ChatGPT-User fetches a page when a user's request calls for it. Perplexity, Google and Anthropic each publish a pair of their own.

So "ChatGPT does not know my company" is one of two sentences. Either the model never learned you, which you cannot change quickly, or the retrieval step found nothing usable, which you can. Everything below is about the second sentence, because that is where the levers are, and because a fix there also becomes the text the next model is trained on.

The six causes, side by side

In our audit of 9,483 small business websites (40 trades, 90 metros, September 2026) we fetched the homepage, the contact page and the about page of each site the way an assistant's crawler does: no JavaScript, one identifying user agent, robots.txt obeyed. The table is what we found behind each cause. The sections that follow take them in the order you should check them.

Cause one: the crawler never got in

Start here because it makes every other fix invisible. A page that returns a "verify you are human" challenge to an AI crawler is a page the assistant never read, however good the words on it are. 7.2% of the sites in the audit served our crawler a challenge page instead of the site. The setting that raises a challenge lives in a CDN or a security plugin, which is not where anyone at a business looks when they wonder why an assistant ignores them.

The other three doors are smaller and easier. 3.2% disallowed AI crawlers in robots.txt on purpose; whether that is what they meant is the subject of our robots.txt article. 3.0% rendered their content only in JavaScript, so a fetcher that runs no scripts received an empty shell. 2.1% kept the part that answers questions, the price list or the service menu, inside a PDF. The crawler access check reads your robots.txt for all nineteen tokens in a few seconds; the bot wall you have to test by fetching a page as a crawler would and reading the status code.

Cause two: the site says nothing a question can land on

An assistant lifts answers from pages that already contain answers. A homepage that says "quality service since 1998" contains none. 61.5% of the audited sites had no question-and-answer content of any kind, and the pages were not short: the median homepage ran to 880 words. The words were there; they were not in the shape of a reply.

Half the sites (50.7%) came close and missed. They had at least one heading phrased as a question, "Why choose us?", and then three paragraphs of atmosphere under it that no machine could isolate as the answer. The shape that works is plain: the heading is the question a buyer types, the next forty to eighty words answer it completely, and the detail follows for whoever wants it. What that looks like for a service page is in writing to be cited.

Cause three: it cannot tell who and where you are

Before an assistant names "Smith Plumbing" it wants to be sure there is one Smith Plumbing, at one address, with one number. 43.1% of the audited sites were inconsistent about exactly that across their own three main pages. 36.9% published more than one phone number. 34.1% had no postal address anywhere on the homepage, the contact page or the about page that a machine could parse. An entity that contradicts itself is one an assistant can decline to name without being wrong.

The fix costs an afternoon. Choose the one form of the name, the address, the phone and the hours; write them identically on every page, in the footer and in the contact block; then say the same thing in markup. schema.org's LocalBusiness, or the specific subtype for your trade, with address, telephone, openingHoursSpecification and areaServed filled in, is the machine-readable version of the sentence "this is who we are and where". Google's local business documentation lists the properties it reads. Only 17.8% of sites had no markup at all; the trouble was that most of the rest described the website, not the business.

Cause four: nothing outside your site backs you up

An assistant deciding between three businesses it can read will lean on what others say about them. Most small businesses have that evidence and hide it: 64.7% of the audited sites showed no review signal a crawler could read, no rating markup and no review quoted in the page text, while their Google profiles held dozens of reviews the site never mentioned. Only 8.9% carried rating schema.

Put the reviews where a machine looks. Quote real ones on the page as text and name the platform they came from. Keep that platform's stars out of your markup: Google's review snippet rules say not to aggregate ratings from other websites, and a business showing reviews of itself on its own site gets no review stars either way. Beyond reviews, a brand mention on a page that is not yours, a directory entry that agrees with your site, a thread where your trade is discussed, is the corroboration you cannot write yourself and should stop leaving to chance.

Cause five: what it knows about you is out of date

Sometimes the assistant does know you, from training, and what it knows is the address you left two years ago. A model answering from memory has no way to notice that the world moved; that is what a knowledge cutoff means, and a confident wrong answer of this kind is one form of hallucination. The audit did not measure staleness, because a crawl of your site cannot see what a model remembers.

The fix is the same as for every other wrong fact: change the source the assistant will read next. With retrieval on, the current page wins over memory for the facts on it. Without retrieval, the current page becomes the text the next training run learns from. Either way the work is on your own pages first, and what to do when the wrong fact lives somewhere you cannot edit is in when AI says false things about your company.

Cause six: the question you tested was not one that produces names

Assistants hedge. Asked "who is the best dentist in town", an assistant will often decline to rank and offer criteria instead, and no business is named. In our probe we put one real buyer question to Claude, with web retrieval on, for 9,240 businesses. 5,928 of those answers gave a clearly differentiated list of names; the rest hedged. Across all 9,240, 14.9% of the businesses were named; among the differentiated answers, 17.7%. That is one assistant, one question each, one day, so treat it as direction and not as a rate.

The practical point: if you typed your company name and the assistant shrugged, you learned little. Ask the questions your buyers ask, the problem question, the comparison, the one that names a service and a place, and ask them more than once, because the answers vary between runs. Our own measurement handles that variation with a fixed panel of twelve buyer questions, each asked once per reading of one assistant, Claude, with live web search on. Each reading is dated, and the weight sits on the pattern across the twelve questions, and with monitoring across months, not on one answer; how the score is built explains the rest.

How to find your cause in an hour

1. Ask ChatGPT, Claude, Gemini and Perplexity three of your buyers' real questions, twice each, and write down who is named. If you appear nowhere, keep going; if you appear with wrong facts, cause five applies as well.
2. Run the crawler access check on your domain, then fetch your homepage with a crawler's user agent and look at the status code. A challenge page or a 403 is cause one, and the fix is in your CDN's settings.
3. View the source of your busiest service page with scripts disabled. If the words you sell with are not there, cause one again; if they are there but no heading is a question with an answer under it, cause two.
4. Open your homepage, contact page and about page side by side and compare the name, address, phone and hours. Any difference is cause three.
5. Search your own pages for the word "review". If the reviews live only on Google, cause four.
6. Whatever you fix, measure again with the same questions next month. One run is a coin flip; the trend is the finding.

None of these steps is technical enough to need an agency, and none of causes two through four needs more than a few hours from whoever maintains the site. What an agency adds is the measurement: a panel that does not change, run the same way every month, so you can tell the difference between an assistant that changed its mind and one that never read your page.