What a Real AI Visibility Audit Should Include

Almost everyone selling an AI visibility audit right now is selling the same shape of thing: a score, a list of competitors, and some advice. The scores look alike. The advice looks alike. The prices do not.

What separates them is not the cover page. It is whether the thing underneath can be checked by the person who bought it.

This is the list we would hand a buyer who was about to pay someone else. It is written so you can hold any vendor to it, including us.

What should a real AI visibility audit include?

Eight parts: a frozen question set, more than one assistant, more than one pass, the sentence behind every named rival, a named competitor set, one specified page to build, a stated method, and a re-runnable baseline.

Every one of those is checkable by a non-technical buyer in about ten minutes. That is deliberate. An audit you cannot check is a report you have to take on trust, and the whole reason a business orders one is that it is tired of taking things on trust.

It starts with a frozen question set

The questions are the instrument, so they have to be fixed before the first run and shown to you. Everything measured afterward is measured against them.

A good set is written the way buyers speak, not the way the industry writes. "Who does small group strength coaching in Denver" is a question. "Fitness coach visibility ranking factors" is not something a human being types. Ask to read the 25 questions before anybody runs anything, and if a single one of them contains your business name, ask why: questions that already name you are testing recall, not discovery.

Freezing them matters because of what comes later. A re-run at day 90 only means something if it asks the same questions. A question quietly reworded between the two runs turns an honest comparison into a decorative one.

Why one run is not a measurement

Because the same question, asked twice, can produce two different worlds. This is the finding that changed how we build audits, and it is the one we would most want a buyer to know.

SAME 25 QUESTIONS, ONE HOUR APARTSOURCE: OUR OWN TEST RUN, 2026-08-01 · CHIP COUNTS ILLUSTRATIVE

First pass, names returned

Names both passes agreed on

Nothing. The two lists had no name in common

Second pass, an hour later

The middle band is what a single run cannot give you. Names are only worth reporting where two independent passes, and two different assistants, produced them both.

AnswerHalo answerhalo.com

CLICK TO ENLARGE

On August 1st 2026 we ran one identical set of 25 questions twice, an hour apart, with nothing changed in between. The two lists of rival businesses that came back had no overlap at all. Not a small overlap. None.

Carry the limits of that with the number: it is one question set, in one niche, on one day, and it is our own test rather than a published study. It is enough to establish the shape of the problem, which is that a language model asked to name businesses will produce a fluent, confident, different answer on request. A vendor who runs your questions once and prints the result has handed you one roll of a die and called it a measurement.

The same logic applies across tools. Assistants disagree with each other as readily as they disagree with themselves, which is why a single-tool audit describes that tool rather than your business. They are not even built the same way: Perplexity's help center describes a product that searches and cites sources for every answer, while Google's guidance on its AI features describes a system that issues several related searches behind one question. Two different machines, asked the same thing, reading different pages.

Evidence, not assertion

Ask for the sentence each competitor name was judged from. A list of names is something you can only agree with. A quoted sentence is something you can check.

This is the difference between reviewing an audit and rubber-stamping one. When a report says a rival was named in your place, the useful version shows you the line where the assistant said it, so you can see whether the assistant actually recommended them or merely mentioned them in passing. Those are very different findings and they look identical in a bar chart.

There is research pointing the same direction from the content side. The Princeton-led study GEO: Generative Engine Optimization tested content changes against generative search answers and found that adding citations, quotations and statistics raised how visible a source became. What engines reward in a page is exactly what a buyer should demand in a report: attribution rather than assertion.

The eight parts, checked against any vendor

Here is the whole list with the tell for each one. Print it, and use it on whoever you are considering.

Two things belong on the other side of the ledger, because a good audit refuses to promise them. No honest vendor can guarantee a specific citation by a specific date, because they do not control the assistant. And a promise measured in pages per week is measuring their output, not your result.

Also worth saying plainly: plenty of businesses do not need to buy one of these at all. If a free readout shows one obvious gap, fix the gap. That case is set out in when a free check is enough, and the ongoing version of the question is covered in monitoring tools versus done-for-you work. Whether the paid version earns its price is argued through in is a paid AI visibility audit worth it.

If you want the free version first, that is the sensible order. The free check asks ChatGPT 25 real buyer questions from your niche at $0 and emails a plain-English readout: your score, who got named in your place, and the one fix to start with. It runs as soon as you ask for it, and the report lands within the hour between 7am and 5pm Central, or by 10am the next morning outside those hours.