The surface nobody publishes to
Ask an AI what a company does and the answer is assembled from whatever the world has written about it, not from anything the company published about itself. For a growing number of buyers that answer is the first contact they have with a business, since it shapes a shortlist before anybody visits a website. A mature specialist market already works this layer, with dedicated tooling and longitudinal data.
The standard, done properly
A controlled matrix, not a conversation
A fixed set of prompts, written once, run identically every time. Answers move with phrasing, personalization, location and timing, because ad-hoc questioning produces anecdotes rather than evidence. Four kinds of prompt at minimum: what the business does, category prompts naming no brand at all, comparison prompts naming your competitors, and prompts carrying genuine buying intent.
Competitors named before querying
Which keeps the exercise focused and makes share of voice measurable rather than impressionistic.
Several systems, with disagreement treated as data
They draw on different material, so agreement means something consistent is reaching all of them, while divergence means no coherent account of the business exists anywhere.
Three numbers recorded
How often the brand appears at all, its share of mentions against the named competitors, and how many distinct category descriptions came back. That last is a census in the same sense as a typeface count, and just as hard to argue with.
Sources traced
Where an account came from is more actionable than the account. Frequently a directory listing, an old press release, a review aggregator, or a competitor’s comparison page.
Continuous rather than annual
Published analysis of large citation samples finds most cited sources change within a fortnight. A single reading here has a short life, so this layer needs a shorter re-examination interval than the rest of the audit.
Where most of the field starts from the wrong premise
The prevailing framing is inaccuracy detection and correction. The vocabulary is hallucination, error, fix. All of it starts from the premise that the client’s self-description is the truth, while any divergence is a mistake to be repaired. We do not start there. A machine’s account is built from what the world has actually written, which makes it evidence and occasionally better evidence than a company’s own copy. So where the systems disagree with your stated position, both readings get argued: the strongest case that you are right and the account is stale, and the strongest case that the account is right and your self-description has drifted from what you now actually are.
"If several independent systems describe a business as the budget option while the business claims premium, one of those is wrong, and it is not automatically the machines."
The three-way comparison
Three accounts placed beside one another: what the machines say, what the brand publishes, and what its own people say when asked separately. Where all three diverge, the problem is not visibility, because nobody inside or outside the business shares an account of it, and no amount of optimization fixes that. We have not found anybody else running this comparison, and it is frequently the finding a client remembers longest.
Claims the world makes that you do not
The machine layer is also searched for things said about the business that are true, favorable, and absent from everything it publishes. Those are claims the world already makes on your behalf though you have never made them yourself, which makes them the cheapest positioning available.
Where the value sits
Continuous monitoring across many platforms is a specialist product and a useful one, though every such tool returns the same thing: a number. Reading the number is separate work. Appearing in three of eight prompts means nothing until somebody establishes whether the machines have you right, whether your position has drifted, and which of the two should move. No tool performs that judgment. This pass exists to make it. Where a platform produced evidence in your report it is credited by name, and any published study supplying a benchmark is attributed with its date and sample.
Where perception work connects
Whether the underlying claims are true is covered in claim verification. What your own people say is covered in internal alignment. Where the machines are drawing their material from is covered in digital footprint. And making a page legible to these systems in the first place is covered under organic search marketing.
How the engagement runs
Screen shared sessions, with prompts, systems and dates recorded verbatim so the run is repeatable. If you have never asked an AI what your company does, that is a ten-minute exercise and it is frequently uncomfortable.