Before you buy one
What an AI visibility audit should contain
This category is young and largely unpoliced. Most published numbers in it come from companies selling the service. Here is what a defensible audit records, and the questions that separate measurement from a sales document.
What has to be recorded
An answer without its conditions cannot be checked, repeated or compared. A report should state all of the following for every result it shows.
The questions, written out
Not categories or themes. The exact wording put to the engine, because phrasing changes the answer. A report that shows you a theme has hidden the instrument.
The engine and the model version
Named. Answers change between model versions, so a result without a version cannot be compared with a later one.
How many times each question was asked
A single run is one observation. Answers vary between runs, so a claim built on one run of one question is a coin toss reported as a finding.
We asked ChatGPT "my furnace is 20 years old and I need to replace it in Cincinnati. Who should I call?" five times on 24 August 2026, through its API. The five answers cited 16 different domains, and none of them appeared in all five. Of the six questions we repeated five times that day, three had no domain common to every ChatGPT answer.
The date, and the failures
Both the attempts and the successes. If 40 questions were planned and 31 returned usable answers, the report says so. Reporting only what worked inflates every rate in the document.
Whether it was an API or the consumer app
These are different products and can give different answers. A result sampled through an interface should say which one, and should not be described as what a customer sees unless that is what was measured.
What an audit cannot establish
Some claims are not available at any price, and seeing them offered is useful information about the provider.
How often real people ask a question inside an assistant is not published by anyone. There is no dataset for it. A provider quoting in-assistant search volume is quoting something that does not exist.
Cause is the other one. A change followed by a different answer is not proof the change worked, because answers move on their own. Separating the two needs a comparison group that was left alone, and most reports do not have one.
Counts of answers are not market share, not impressions, not traffic and not customers. They describe the questions that were asked.
Questions worth asking any provider
Which engines did you sample, and which did you leave out, and why. How many times was each question asked. Will you show me the raw answers rather than a summary. What is your comparison group. Which of your recommendations is supported by something you measured, and which is general practice. What would you refuse to promise.
The last one is the most revealing. A provider who will not name a limit has not found one yet.
Where the field actually stands
We reviewed the available material on this subject and found one peer-reviewed source across three full courses on it. Practitioners with well over a decade in search said independently that most of the accepted best practice in this area has never been tested. Treat confident advice here, including ours, as something that should come with its evidence attached.
Vector Index AI sells one audit built to this checklist: 25 questions we select from your business, asked four times each on ChatGPT, Gemini, Claude, Perplexity and Google AI Overviews, with three change specifications delivered in writing, for $750.
Our scope and prices · How we measure · Being recommended, not just mentioned · Answer engine optimization services
Questions we get asked about this.
What is an AI visibility audit?
A record of how named AI engines answered a set of buyer questions about a category, showing where a business appeared, where it did not, and which sources the answers drew on. A defensible one stores the raw answers and the conditions they were produced under.
What should an AI visibility audit include?
The exact questions as they were asked, the engine and model version, how many times each question was run, the date, the failed attempts as well as the successful ones, and whether the sampling used an API or a consumer app. Without those, a result cannot be checked or repeated.
Can an audit tell me how many people ask AI about my industry?
No. No public dataset of in-assistant question volume exists. Sampling measures what answers contain, not how often a question is asked, and any provider quoting in-assistant search volume is quoting a number that is not available.
Can an audit prove a change caused more citations?
Not on its own. Answers vary without anyone doing anything, so a before-and-after alone cannot separate the change from the drift. Establishing that needs a comparison group left deliberately untouched, and even then it is directional rather than proof.
Does Vector Index AI sell an AI visibility audit?
Yes, the AI Shortlist Blueprint. It is $750, prepaid, for one brand, one main site and one market: 25 questions, four repeats each, on five named engines, every answer stored, and three change specifications delivered in writing within five business days of submitting your intake form. It does not include implementation or any guarantee of citations.
How is this different from an SEO audit?
They overlap substantially in the underlying work. The difference is the observation: an answer-engine audit records what engines said and which sources they used, which is a different measurement from rankings or traffic. We explain the overlap on our AEO and SEO page.
We publish what we find.
We're measuring AI answers in Cincinnati and will publish what we find, including our own results, including when they're unflattering. No pitch attached.
One email a month at most, and you can leave with one click. We do not sell your address and we will not add you to anything else.