How to Audit What AI Says About Your Business

Most teams check their Google rankings every week and have never once asked an assistant what it tells customers about them. The audit takes an afternoon, and the citation list usually explains everything.

By Bruna Sotte de Carvalho, published August 24, 2026, 6 minute read

Most marketing teams check their Google rankings every week. Very few have ever opened ChatGPT and asked it the question their best customer would ask.

That is a strange gap, because the answer is already being given. Somewhere today, a person described their problem to an assistant and got back a short list of companies to call. Your business was on that list, or it was not, and nobody on your team saw it happen.

You cannot buy a place in that answer. You can, however, find out what it currently says, and most of the work fits in one afternoon.

This is the practical follow-on to how AI changed the way we search the internet. That piece explains why discovery moved. This one is the audit that tells you where you stand inside it.

Why This Became a Marketing Job

The instinct is to file this under technology. It is not a technology problem. What an AI system says about your company is assembled from your website copy, your reviews, your press coverage, your directory listings, and whatever third parties have written about your category. Every one of those inputs belongs to marketing.

The audit below is deliberately manual. Tooling for this is young and most of it is expensive, and doing it by hand the first time teaches you more about your own positioning than a dashboard would. Read the Answer Engine Optimization entry first if the term is new.

Step One: Write the Question List Before You Open Anything

The most common mistake is to start by typing your company name. That tells you almost nothing, because a customer who already knows your name is not the customer at risk.

Write ten to fifteen questions a real prospect would ask, in their words, before they know you exist. Cover four kinds:

Category questions. "What is the best commercial cleaning company in Orlando?"
Problem questions. The symptom, not the service. "My warehouse floor keeps failing inspection, who fixes that?"
Comparison questions. "What is the difference between hiring an agency and hiring a freelancer for SEO?"
Objection questions. The thing that stops a deal. "Is it worth paying more for a licensed installer?"

If you have customer personas or a voice of customer file, pull the phrasing straight from it. Invented questions produce invented findings.

Step Two: Ask Every Assistant, Twice

Run the same list through ChatGPT, Gemini, Claude, Perplexity, and a plain Google search so you can see the AI Overview. Use a signed-out or temporary session so your own history does not flatter the result.

Ask each question twice, in separate sessions. These systems are not deterministic, and a company that appears once out of two runs is in a genuinely different position from one that appears every time. Note both.

Step Three: Score It Simply

Resist building a scoring model. Four columns in a spreadsheet is enough:

Mentioned. Did your company appear at all?
Accurate. Is what it said about you actually true?
Recommended. Were you presented as an option, or only as a passing reference?
Sourced. Which pages did it cite, and are any of them yours?

That last column is the one people skip and the one that pays. The citations show you which pages are actually feeding the answer. Very often they are not your website. They are a directory, a review site, a local news piece, or a competitor's comparison page. That tells you where to work.

Step Four: Fix in Order of Damage

Not every finding is equally urgent. Work in this sequence:

Wrong facts first. An assistant confidently stating that you closed, moved, or do not offer a service costs you customers today. Correct the source it is reading, which is usually a stale directory listing or an outdated page on your own site.
Absence second. If you never appear for a category question, the issue is usually that no page of yours plainly answers it. Write that page, in the format covered in writing for people and machines at the same time. The SEO checklist covers the technical half.
Weak positioning third. Appearing but sounding interchangeable with four competitors is a message problem, not a search problem. That is the work in positioning strategy.

What a Real Finding Looks Like

An example from the shape these audits usually take, with the details changed.

A regional service company ran fifteen questions. Their brand name returned an accurate summary every time, which was reassuring and almost useless. The category question, the one a stranger would ask, returned five competitors across two runs and never returned them at all.

The citation column explained why. Four of the five competitors were cited from the same two regional directory listings and a "best of" roundup on a local news site. The company had a listing on one of those directories, abandoned in 2019, with an old address and two services that no longer existed.

Nothing about their website was broken. Their rankings were fine. They were absent from the answer because the sources feeding the answer had stale or missing information about them, and no page on their site plainly answered the category question in the words a customer would use.

The fix took a month and none of it was technical: correct the directory records, get added to the roundup, and publish one page that answers the category question directly. That is a typical result. The problem is rarely where teams expect it to be, which is exactly why the audit is worth running before the strategy meeting.

Step Five: Make It Boring and Repeatable

An audit you run once is a curiosity. An audit you run every quarter is an instrument.

Keep the question list fixed so the results are comparable, store each run as a dated file, and diff it against the last one. Ninety minutes a quarter is enough for most small teams. Put it in the content calendar alongside everything else you schedule, or it will not happen.

What This Audit Cannot Do

Be honest about the limits, especially when you present the findings to a leadership team.

It is a sample, not a ranking. There is no position one to hold, and no click to count, which is why the zero-click search makes traffic a poor proxy for visibility here.
Answers vary between runs, between users, and between weeks.
You cannot pay to change the result, and anyone selling guaranteed placement in AI answers is selling something else.
Improvement is slow, because the inputs are slow. Reviews, coverage, and rewritten pages take months to be reflected.

None of that makes the audit less useful. It makes it a positioning instrument rather than a performance dashboard, which is the more valuable of the two anyway.

Key Takeaways

Ask the questions your customers ask, not your own company name, or the audit will flatter you.
Run every question through several assistants, twice each, because these systems are not deterministic.
The citation list matters more than the answer, since it shows which pages are actually feeding the machine.
Fix wrong facts before absence, and absence before positioning.
Repeat quarterly with a fixed question list so runs are comparable over time.
Chapter sessions on our events calendar go deeper, and the AEO field guide is in the resource library.

Topics: ai search, aeo, brand visibility, audit, llmo, local seo