Ask an engine what your company does and you get one of three answers: nothing, something right, or something confidently wrong. The third is the expensive one, because nobody knows to check it. We ran the check on ourselves twice this month and published both results, including the one that made us look bad.
Zero and one are not neighbours
There are two bad scores on an AI-visibility check and they are different problems.
Zero means the engines have never heard of you. That is a starting position, and publishing fixes it.
One means they answer confidently and wrong: your old price, your retired product, your dead positioning. That is not a starting position, it is a liability that compounds, because a stale answer gets copied into the next answer and the copy is what your buyer reads. Nobody complains to you about it. They simply do not arrive.
The method, and what it cannot reach
Thirty buyer questions, asked the way a person actually types them rather than as keywords: what is it, what does it cost, is it any good, is it safe to pay, best tools for the job it does, how it compares to the names the buyer already knows.
Each answer is scored on one axis. Not flattering, not long. Accurate.
The caveat belongs before the findings, not after. Consumer chat assistants require a login and we do not automate logged-in sessions, so ChatGPT, Gemini, Claude and Copilot were not queried; one engine returned a 403 to our request. Everything below came from the web index and from live pages. An index summariser is an honest proxy for the class of answer that retrieval-based assistants produce, but it is not a quotation from a named commercial engine, and we do not attribute a sentence to an engine we never asked.
If you run this yourself, write your caveat before your findings. A number without its method is decoration.
Finding one: your price is quoted from a page you no longer control
Ask an engine what a product costs and the page it cites is often not the product's own site. It is a directory entry, a comparison post, a review from last year.
In our case it was worse, because the page it cited was ours. On 3 August the index answered that we cost $199 a month – a price from a page we deleted in July. Our real entry price is $25 a month on the annual plan. Ten days later, after two published articles and a sitemap ping, we asked again. The answer had not moved: still $199, still "AI employees", still "managed through Telegram, WhatsApp or Slack" – a product we retired in May.
The live pages were correct the whole time. One cached page was not, and one cached page was all the index had.
Finding two: "is this company legit" belongs to review sites
Ask whether a company can be trusted and the answer arrives from Trustpilot or ScamAdviser, not from the company. That is reasonable – a company vouching for itself is not evidence – but it means an empty profile on those sites is not neutral. It is a slot, and slots get filled by whoever turns up first.
This is the cheapest finding on the list to act on, and the one most companies never check.
Finding three: category questions are won by publishing, not by existing
Asked to name tools for the job we do, we appeared zero times out of eighteen.
That was the correct result. At the time of the measurement we had published nothing that would earn a place in that set. Category presence is not a setting you switch on; it is the accumulated weight of pages worth citing. Which is exactly why the baseline is worth publishing before the improvement rather than after.
Finding four: one page can speak for your whole company
The sharpest thing the second measurement showed was not a wrong number. It was that one page of ours was in the index at all – our FAQ, cached under a title we replaced in July. Not the homepage. Not the pricing page. Not the comparison page we wrote specifically to answer the question a buyer asks about us.
So the entire company was being described by a support document from a product that no longer exists, and every correct signal we had published was unreachable through search.
Two other things fell out of the same run. Our founder's name existed on the site only inside structured data, so every engine answered "who founded ewpire" from LinkedIn rather than from us – summarisers read the page, not only the schema. And our llms.txt, which was correct, had nothing linking to it: a file no link points at is a diary, not a signal.
What we changed the same day
Because this is our own product, the repair is part of the report:
- The founder is now named in visible prose on the about page, not only in the markup.
llms.txtand its long form are announced in the head of every page, so a crawler has a path to them without a visible link cluttering the footer.- The whole sitemap went to IndexNow, which reaches Bing, Yandex, Seznam and Naver in one call – and Bing matters more than its search share suggests, because assistants read it.
- Google does not accept IndexNow, so the six pages that carry the commercial facts went through URL Inspection by hand.
What could not be fixed by editing anything: the cache itself. A correct page with a wrong cached copy is repaired by a recrawl, not by a rewrite. That is a queue you join, not a button you press.
Three checks to run on your own company this week
- Ask an engine what you cost. Note the number, then note which page it cited. If the citation is not yours, that is your first job.
- Ask whether your company is legitimate. See who answers on your behalf.
- Ask for the best tools in your category. Count how many times you appear. Zero is information, not failure.
If an answer is stale, open the live page before you edit anything. Half the time the page is already right and the index is holding an old copy – and editing a correct page does nothing at all.
The metric worth tracking
Counting how often an engine names you is the vanity version of this. It rewards being loud.
Correctness-rate – the share of those mentions that describe you accurately – is the one that matters. Being mentioned is easy. Being mentioned correctly is the whole job.
We re-run the same questions on ourselves and publish the delta whichever way it moves.
Where ewpire fits
Traffic runs this check for you: several frontier models are asked the buyer questions about your brand independently, and you get back what each said, where they agree, and where they contradict each other – a dissent map for how you are described, plus the pages responsible. It then generates the AI-readable presence that gives an engine something current to cite. It increases the probability that assistants describe you accurately; it never guarantees a citation, a ranking, or visibility.
Ideation and Validation are the two steps before it, on the same account and the same credit balance.
What it costs
Signing up gives you 20 free credits, no card, which stay on your balance for 30 days – enough to run a brand check and read the result. Paid plans start at $25 a month on the annual plan. Full breakdown on pricing.
Run a brand check – 20 credits free, no card
This is structured deliberation, not advice, and no outcome is guaranteed. You decide with your eyes open.
FAQ
How do I check what AI says about my company without any tools? Open a search engine and ask it four questions in plain language: what your company does, what it costs, whether it is legitimate, and what the best tools in your category are. Read which pages the answers cite. That takes ten minutes and tells you most of what a paid audit would.
Why does the engine show my old price when my site is correct? Because the answer comes from a cached copy, not from your live page. Search engines re-crawl on their own schedule, and a page that has not changed for a long time can sit unvisited for weeks. Editing the live page again will not help; requesting a recrawl might.
Is being absent from AI answers worse than being described wrongly? No. Absence is a starting position that publishing fixes. A confident wrong answer travels: it gets copied into the next answer and read by buyers who never contact you to check. Fix the wrong answer first.
How often should I re-run the check? Monthly is enough for most companies, and immediately after anything that changes your public facts – a price change, a repositioning, a product you retire. Keep the same questions each time, or the delta means nothing.
Measured 3 and 13 August 2026, on ourselves. Method and caveats as stated above.