Your buyer asked AI for a recommendation.It named three companies. Not yours.
A bounded piece of work that answers one question with evidence: when your buyer asks AI for a recommendation, who gets named, and why is it not you?
You will never see this happen on your own. No ranking drops, no traffic alert, no report. Nobody tells you about a shortlist you were not on.
What you get
A two-page document, written for the person who has to act on it, not a dashboard to log into.
The scoreboard
Every brand the AI named, how often, on which platform, and which one each platform recommended first.
The sources
The specific sites the answers were built from. This is usually the finding that changes the conversation.
Why you are missing
Checked site by site, not assumed. We name any source we could not check rather than implying we checked everything.
What could change it
Every item verified against your own pages before it is written, so nothing on the list is something you have already done.
How we tested
The full method, so you can repeat it yourself or challenge it.
No login, no dashboard to learn, no data feed to maintain. One audit covers one buyer question, which is deliberate and the opposite of how the tools work.
The audit is the commodity. The assessment is the product.
An audit gathers the state of play. It takes a picture: what is there, what is missing, what the numbers say. That part is largely mechanical, and increasingly anyone can do it.
An assessment interprets the picture. It says what the result means. Recommendations follow the assessment, never the audit.
Why the order matters
A wrong assessment does not stay contained. A misread cause becomes a confident plan, the plan becomes months of work, and the work moves nothing.
The audit itself can be flawless and the engagement still fails.
Ordering the tests is not the skill. Reading them is. And reading them well is a function of how many cases you have seen, which is not something a tool accumulates and not something a new entrant can shortcut.
Two ways to be invisible, and they need opposite fixes
AI visibility has two halves. One is recall, where an assistant names a company from what it already holds, without searching. The other is retrieval, where it searches, opens pages and cites them. Commercial tracking tools measure retrieval only. This audit measures both.
What a tracking dashboard can and cannot see
The two halves of AI visibility. Recall produces no search event, so nothing is logged and no tool can observe it. On some assistants recall is the larger half, which is how a company can look healthy in a dashboard and be absent from most answers its buyers see.
The two failures need opposite work. A recall gap is decided off your website, by what other people have published about you. A retrieval gap is on it: the page does not answer the question, the facts contradict each other across your own properties, or the information sits in a format an assistant cannot lift.
Most companies assume the second and spend a year on the first half of the problem, rewriting pages for an engine that was never going to consider them. The audit separates the two before anyone spends anything, which is the most expensive mistake it prevents.
Some questions your own website cannot win
Before acting on any gap, there is a prior question that decides whether the work is even possible: is the buyer choosing a supplier, or trying to understand something?
Provider-choice questions
The buyer is picking a named supplier. Who makes the best one, who should we buy from, which brand is most reliable.
Your own domain can realistically win these citations. On-site work is viable.
Advice-seeking questions
The buyer wants to understand something before choosing anyone at all.
These are largely closed to brand domains, however well the pages are built. The only route is off-site: becoming the source that the cited sources rely on.
Share of AI citations going to brand-owned domains
Share of AI citations going to brand-owned domains. All seven are high-stakes categories, so the difference is not stakes, it is intent. Source: Semrush AI Visibility Index, “Your Money or Your Life” mini study, September 2026, 22 consumer topics, US data with a UK cross-check, across ChatGPT, Google AI Overviews, Gemini and AI Mode. Brand-owned counts a company’s own site selling its own product or service; multi-brand retailers and marketplaces count as independent. Those topics are consumer, not B2B, so the figures are directional for our clients rather than applicable. The mechanism is what we adopt, not the numbers.
Why this overrides everything else
A company named from memory but not cited on an advice-seeking question should not be sent to do structural work on its own website.
The work will be done correctly and the citation will still not arrive, because that is not where answers to those questions are built. This is the single most expensive misdiagnosis in the category, and it is the one the audit is designed to catch first.
Not every gap costs the same to fill
An absence is not a single thing. Where your competitors sit in the answer decides whether filling the gap is a short project or a long campaign, and a report that does not separate them will point your budget at the hardest questions first.
Highest value: nobody is named
The assistant has a detailed opinion about what to look for and none about who. An empty position. The work is to become the first retrievable name.
High: only software is named
The category has been ceded to tools. A service answer is differentiated by default, and it defines the positioning for you.
High: different names every run
Unstable, which means contested with no champion established yet. Winnable while it stays that way.
Lowest: the same two or three names every run
Entrenched incumbents. This is displacement rather than entry, and it is expensive and slow. Worth knowing before you commit, not after.
Filling an empty position and displacing an incumbent are different projects. The audit tells you which one you are looking at.
Why testing your own brand name misleads you
The first thing almost everyone does is type their own company name into ChatGPT and read what comes back. It is a reasonable check and it measures the wrong moment.
1. The problem question
No category yet, no brands. The buyer is learning, and has not heard of anyone.
2. The shortlist question
What are the best options for this job. This is the moment that decides who gets considered.
3. The brand question
Is this company any good. It only exists because the shortlist already put your name in front of them.
Being described accurately when someone asks about you by name is worth knowing, and it is the easiest question in the set to pass. It is not where new business comes from. We run the questions with your name deliberately left out, because that is where you are discovered, and where your competitors are being recommended in your place.
How we run it
Your buyer’s words, with your brand name left out. A buyer who does not know you yet will not type it.
ChatGPT, Claude, Perplexity and Grok. Identical wording every time, private sessions, no history, web search on.
If the answers are built from five review sites, we search each one for your product and report what we found.
We tally only what appears in the answer itself. Suggested follow-up questions do not count as a mention, and we say so on the page rather than quietly including them.
Before concluding anything we name at least two explanations for what we are seeing and go and get the field that tells them apart. We never write that no coverage exists when what we mean is that we did not find it.
Why twelve runs and not one. AI answers are non-deterministic, so a single query is a sample of one rather than an observed fact. An absence that leads the document has to hold across all twelve, or it does not lead.
What a result looks like
From a recent audit, with the client and the category removed.
Answers naming each brand, out of twelve
All four platforms built their answers from the same handful of independent review sites. The client was absent from every one of them. Their product sat squarely in the price band the AI was recommending, and two of those sites already covered their other product lines.
So the finding was not that the product was wrong, or that the website was weak. The website was fine. The shortlist was being decided somewhere the client had never been looked at, and no amount of work on their own pages would have moved it.
What the audit does not claim. It measures presence in answers, not revenue. It is a sample, not a census, and different wording of the same question surfaces different competitors. And an empty position is not automatically worth filling: a question can go unanswered because nobody is asking it with money behind it. The audit surfaces the opportunity. It does not decide that the opportunity is worth having.
Why we repeat it every thirty days
One run is a snapshot. The same question, the same four platforms, the same conditions, every thirty days, turns it into a trend that shows four things a one-off cannot:
- how your competitors move, answer by answer
- gains and losses, for every brand including yours
- new entrants arriving in the answers
- whether your own work is actually reaching the answers
Most companies discover their position once, act on it, and never check whether the action worked. The repeat is where the audit stops being interesting and starts being useful.
What this is not
Not a tool subscription
You are not buying a dashboard, a seat or a data feed.
Not a visibility score
A number between 0 and 100 is not a finding.
Not a website audit
We may look at your pages to verify a recommendation, but your website is checked after the shortlist, not before it.
Not a retainer
A bounded piece of work with a defined output, useful on its own whether or not anything follows it.
Who this is for, and who it is not
It fits a company selling something a buyer researches before choosing: manufacturers, distributors, considered consumer products, B2B services. It is most useful when you already do well in search and suspect, correctly, that something else is now happening earlier in the process.
It is a poor fit if nobody asks AI for a recommendation in your category yet, or if you want a number for a board deck rather than a finding to act on.
What usually happens next
Most audits point in one of two directions, and they are different jobs.
You are not on the list at all
The shortlist is being decided by sources you are absent from. That is Generative Engine Optimization, and it is off-site work.
You are on the list but passed over
The engines can see you and cannot use you properly. That is Answer Engine Optimization, and it is on-site work.
Often it points at neither, and the honest answer is that your search foundations need to come first. We will say so. There is no obligation to continue, and the document is yours either way.
The argument behind this service is set out in From Searching to Asking.
Questions we get asked
Q. How long does an AI Visibility Audit take?
Most audits are delivered within three business days of agreeing the question. The twelve runs are done in one sitting, and the time goes into checking the cited sources one by one.
Q. Do you need access to my website, analytics or search console?
No. The audit runs entirely on public information, the same information the AI engines use. We may ask for a product page URL so we can verify a recommendation before writing it.
Q. How many questions does one audit cover?
One. Covering several questions means running several audits, each with its own twelve runs and its own document, because a finding averaged across questions stops being actionable.
Q. Can you audit a competitor instead of us?
Yes, and the method is identical, because the audit never requires the subject to cooperate. Competitor audits are most useful when you already know you are absent and want to understand what the named brand has that you do not.
Q. What if the answer changes between your run and mine?
It probably will, and that is exactly why we run twelve. AI answers are non-deterministic, so a single query is a sample of one, and any absence we report has to hold across all the runs before it reaches the document.
Q. If you find negative coverage about us, does it go in the document?
We tell you, and we keep it out of the document. The audit is a working file you may circulate internally, and third-party criticism of your product belongs in a conversation where it can be discussed rather than in a page that can be forwarded without context.
Q. Does my brand need to appear in AI answers already?
No, and an audit returning zero mentions is usually the more valuable one. It establishes the starting point and identifies exactly which sources would have to change for that to move.
Q. Can we choose which platforms you test?
Yes, though we would keep all four. Dropping one makes the audit slightly cheaper to produce and considerably weaker to act on, because a brand missing from four assistants is a finding while a brand missing from one is a quirk.
Q. What happens if the audit says we do not have a problem?
We write that, and the engagement ends there unless you want the repeat. An audit that manufactures a problem in order to sell the next piece of work is worth less than nothing, because you would act on it.
Q. Can this run alongside our existing agency or in-house team?
Yes. The output is a document, not a dependency, and it is deliberately written so whoever does the work can act on it, whether that is us, your agency or your own team.
Start with a conversation
Thirty minutes on the question worth asking, and what the answer is likely to show. If the honest answer is that you need search foundations first, we will say that rather than sell you the more fashionable service.