What an AI entity audit actually contains

If "AI search visibility" sounds like a category that doesn't quite have a method yet, it's because it doesn't. Everyone's still working it out.

But the diagnostic underneath any sensible programme is the same one journalists do when they're checking who knows what: ask the source, write down what they say, compare it against what's actually true. The AI engines are just sources we're asking on behalf of every prospective client who's started asking the same questions.

Here is the shape of an entity audit, and why each piece matters.

What an entity audit is

It's a structured probe of what the major AI engines currently know - and currently believe - about your agency, your founder, and your category. Run as a snapshot, with the prompts logged, the answers logged, and the diffs across engines recorded.

It serves three purposes:

1. Baseline. Without one, you can't see whether the programme is working. The engines refresh on their own schedules; without a fixed prompt set, you're guessing.
2. Diagnosis. Every wrong answer points at a fixable thing on your site, schema or earned-media footprint.
3. Prioritisation. Most agencies have three or four fixes that would move the needle. The audit tells you which ones.

A good audit is boring to look at. It's a spreadsheet. Each row is a prompt, each column is an engine, each cell is a verbatim answer plus a note on what's accurate, what's wrong, and what's missing.

The shape of the prompt set

There is no fixed canon, and the exact prompts matter less than the categories they cover. The set has to probe three things.

Entity awareness. What does the engine know about your agency, your founder, and the specific services you sell - and how confidently does it say it?

Discovery behaviour. When a prospective buyer types the kind of query they actually would type, does your agency appear, and in which company?

Authority and category placement. Does the engine place you in the right category at all, and against the right peers?

Twelve to fifteen well-chosen prompts across those three areas is enough for a first pass. The choice of prompts matters more than most agencies expect, because a prompt set that misses the buyer's actual language will produce a clean-looking audit that misdiagnoses the problem. Which is where most self-run audits quietly go wrong.

Run them against ChatGPT, Claude, Perplexity and Gemini. Save the date, the model version, and the verbatim answer.

What the diffs tell you

The pattern of mistakes points at the fix.

Entity prompts wrong or thin → on-site canonical descriptions are inconsistent. The engines synthesise from what's published. If your homepage says one thing, your About page says another, and your LinkedIn says a third, the engines pick the strongest signal - which is usually the one with the most third-party reinforcement, and which may not be the one you want them to pick.

Discovery prompts ignore you → your specialism isn't documented enough. The engines can only recommend you for what they have evidence of. If you say you specialise in something on the homepage but nowhere else, the engines won't trust the claim.

Authority prompts misplace you → the citations are pointing somewhere else. The engines weight third-party mentions heavily. If most of your earned media is in a different category than where you want to be, the engines will follow the citations, not the homepage.

Which fix to run first, and in what order, is where the diagnostic starts to earn its keep. Fixing the wrong layer first can burn a quarter.

What schema is doing under the hood

The hygiene fixes - the cheap ones - are usually on-page. Structured data (Organisation, Person, and their sameAs signals to the outside databases) is what removes ambiguity from the engines' synthesis step. It is not magic; it is just a structured copy of facts you already publish, in the shape the models are looking for.

The reason schema helps more than it looks like it should is that it disambiguates when the free text is contradictory - and most agency sites are contradictory somewhere. Getting the schema right without first reconciling the underlying facts across the site produces a strong signal for the wrong description of your agency, which is worse than no schema at all.

What the audit isn't

It isn't an SEO audit. SEO optimises for being one of ten blue links; AI visibility optimises for being the synthesised answer. The mechanics overlap but the goalposts are different.

It isn't a one-time exercise either. The engines change. Your category changes. Your earned-media footprint changes. An audit that's older than six months is mostly storytelling.

What to do with the result

The output of an audit should be a short prioritisation memo, not a deck. A small number of on-site fixes that move ChatGPT and Claude quickly, a small number of off-site pushes that move Perplexity, and a re-run cadence that lets you see whether the answers are actually changing.

Which fixes belong in which bucket, and how aggressive to be, depends on the state of the current audit and the competitive citation space. Most agencies do an audit once, get scared, and stop. The ones that move are the ones that keep running it until the answers change.

That's how this category gets won. Slowly, on purpose, in public. If it would be useful to have the audit run for your agency, or a working session on the prioritisation memo, get in touch.