Measurement

How to Run an AI Visibility Audit, Step by Step

A practical AI visibility audit method: pick buyer questions, ask ChatGPT, Gemini and Google AI Overviews more than once, then score mentions and citations.

An AI visibility audit measures how often AI assistants name a brand when people ask the questions buyers ask, and who they name instead. You pick 20 to 30 buyer questions, ask each one on ChatGPT, Gemini and Google AI Overviews more than once, and record whether the brand is named, where in the answer, which competitors appear and which websites are cited. The result is a mention rate per engine, a share of voice against competitors and a list of the sources the assistants rely on.

This guide walks through the method so you can do it by hand with a spreadsheet. At the end we show how AI Visibility Desk runs the same steps for you. If you need the background first, start with what AI visibility is.

Why run an AI search audit at all?

Buyers increasingly read an answer instead of a list of links. When Pew Research tracked the browsing of 900 US adults in March 2025, people who saw a Google AI summary clicked a regular search result in 8% of visits, against 15% when no summary appeared. If the answer names three competitors and not your client, ranking well below it does less than it used to.

Classic SEO tools do not tell you what those answers say. An AI SEO audit fills that gap: it shows whether the brand is part of the answer, and gives you a baseline before any generative engine optimization work starts.

Step 1: Define the brand, its other names and its competitors

Write down:

  • The brand name and every other name it goes by. Product names, abbreviations, the old name after a rebrand. If an answer says “FB” or the product name instead of the company name, you still want to count it.
  • The website domain. You will need it to tell when the assistants cite the brand’s own pages.
  • Three to five competitors, with their websites. Pick the ones a buyer would actually compare against, not the biggest names in the industry. Five is enough to make share of voice meaningful without turning the spreadsheet into a wall.

Watch for names that are ordinary words. A brand called “Wave” or “Notion” will match sentences that have nothing to do with the brand. Decide now that you will read those matches rather than trust a text search.

Step 2: Choose the buyer questions

This is the step that decides whether the audit is useful. Write the questions a prospect would type before they know which brand they want. Do not include the brand’s name: a question that names the brand will almost always produce an answer that names it, which tells you nothing about visibility.

Spread the questions across the kinds of intent buyers have:

Question type Example (accounting software)
Best-of “What is the best accounting software for freelancers?”
Comparison and alternatives “What are good alternatives to QuickBooks for a one-person business?”
Price-led “What is the cheapest invoicing tool that handles VAT?”
How-to or problem-led “How do I track expenses and invoices in one place as a contractor?”
Use case “Accounting software for a design studio with three people”

Aim for 20 to 30 questions. Pull wording from sales calls, support tickets, Reddit threads and the “People also ask” box in Google. Write them the way people talk, including the longer, specific ones; Pew found AI summaries appeared on 53% of Google searches of ten words or more.

Separately, ask two or three questions that name the brand (“What do people say about Brand X?”). Keep them out of the scores. They are there to hear how the assistants describe the brand when asked directly.

Step 3: Pick the engines

Cover the assistants your client’s buyers use. For most audits that means:

  • ChatGPT, which for many buyers is the assistant they mean when they say “I asked AI”.
  • Gemini, Google’s assistant.
  • Google AI Overviews, the summary at the top of Google results. Not every search shows one, so record “no overview” separately from “overview that does not name the brand”.

Use a logged-out or private browser session and the same location for every check. Personalisation and location change answers, and you want results anyone could reproduce.

Step 4: Ask each question more than once

AI answers are not stable. In a 2025 study run by SparkToro and Gumshoe, volunteers ran 12 recommendation prompts 2,961 times, and the researchers found less than a 1 in 100 chance that ChatGPT or Google’s AI would return the same list of brands twice. The order changed even more often. Their conclusion was that how often a brand appears across many runs carries real signal, while its position in any single answer does not.

So ask each question at least twice on the chat assistants. If the two runs disagree (named once, not named once), that disagreement is information: the brand is on the edge of the answer for that question. Note how many question-and-engine pairs disagreed; it tells the client how much of the score is settled and how much could swing.

Step 5: Record each answer

For every answer, log:

  1. Whether the brand is named (yes, no, or “no answer”).
  2. Which competitors are named.
  3. The order: is the brand the first of your listed brands to be named, second, third?
  4. The sources cited, by domain and, where shown, by page.
  5. A short passage around the mention, so someone else can check the match.
  6. Any brands named that are not on your competitor list.

A spreadsheet with one row per question, engine and run works. Expect it to be long: 20 questions asked twice on ChatGPT, twice on Gemini and once on Google is 100 rows.

Step 6: Score the results

Mention rate per engine

For each engine: answers that name the brand divided by answers returned. Leave out checks that produced no answer (a Google search with no AI Overview, or a request that failed) rather than counting them as misses. An overall score can be the average of the three engine rates, so one engine with more checks does not dominate.

Position in the answer

For answers that name the brand, count how many of your listed competitors were named before it. If none were, it ranks first. Report the average rank and the share of answers where the brand came first. Treat it as supporting detail, given how much ordering varies.

Share of voice

Count the answers that name each brand, add them up across the client and its competitors, and divide each brand’s count by the total. We cover the formula, a worked example and the traps in our guide to AI share of voice.

Questions where the brand is missing everywhere

List the questions where the brand was not named in any run on any engine. These are usually the most useful lines in the report, because each one points at a piece of content or coverage that does not exist yet.

Head-to-head and question types

For each competitor, count the questions where only they were named, only the client, both, or neither. Then group questions by type (best-of, comparison, price, how-to, use case) and compare mention rates. A brand that wins comparisons but disappears on price questions has a specific, fixable gap.

Step 7: Analyse the citations

The sources the assistants cite show where their information comes from. For each cited domain, note how often it appears and what kind of site it is: the client’s own site, a competitor’s site, a review site such as G2 or Capterra, a forum such as Reddit, video, a reference site such as Wikipedia, or editorial and comparison content.

Two lists matter most:

  • How often the client’s own site is cited when the client is named. If assistants name the brand but cite other people’s pages, the brand’s own content is not doing the work.
  • Third-party sources cited in answers that name a competitor but not the client. These are the pages to target for coverage, reviews or corrections. Our article on getting cited in AI Overviews goes into what tends to earn a citation.

Also check that the client’s site lets the AI search crawlers in. OpenAI states that sites that block OAI-SearchBot will not be shown in ChatGPT search answers, and Google says a page must be indexed and eligible for a snippet to appear as a supporting link in AI Overviews.

Step 8: Read what is said about the brand

A mention is not always good news. Read the passages around each mention and mark them positive, neutral or negative, with what the brand is praised or criticised for. Group similar points (“easy to use”, “expensive for teams”) so the client sees themes rather than a pile of quotes. Be clear in the report that this is a reading of the text, not a measurement.

Step 9: Write it up

A good audit report has:

  • The headline mention rate and the rate per engine.
  • Share of voice against the listed competitors.
  • The questions where the brand is missing, with who was named instead.
  • The top cited sources and the gaps among them.
  • What the assistants say about the brand.
  • Three to five concrete recommendations tied to those findings.
  • The method: number of questions, runs per engine, date, and how many checks returned nothing.

See what goes into an AI visibility report for a fuller outline.

How long does a manual audit take?

Asking 20 questions five times each, reading every answer and logging brands and sources is easily most of a working day, before any analysis. It is a good way to learn how the assistants behave. It is a slow way to produce audits for several clients a month.

How AI Visibility Desk automates the audit

AI Visibility Desk runs the method above for agencies and publishes the result under the agency’s own brand.

  • Setup. Enter the client’s website and the tool reads the site, searches the web and fills in the brand, category, competitors (up to five) and buyer questions for you to review and edit. Plans allow up to 10, 20 or 30 questions per audit.
  • Runs. Each question is asked twice on ChatGPT, twice on Gemini and once on Google. Two extra questions name the brand to hear what is said about it; they are reported separately and never scored.
  • Turnaround. In our measurements ChatGPT answers arrive in about 20 seconds, Google in about 80 seconds and Gemini in 8 to 12 minutes. Gemini is the least reliable: some checks never come back. Those are left out of the scores and the report states how many.
  • Scores. Mention rate per engine and overall, average position and how often the client is named first, share of voice, head-to-head against each competitor, mention rate by question type, and questions where the client is missing everywhere.
  • Citations. Top cited domains by type, cited pages, how often the client’s own site is cited, and sources that cite competitors but not the client. A check of the client’s robots.txt for AI crawlers and of the homepage.
  • Review. Every answer is shown with a short passage, so a wrong match (the “Wave” problem) can be corrected by hand. Brands the assistants named that you did not list are shown so you can add or hide them. A model can read the mentions for sentiment and group them into themes.
  • Report. Preview, then publish to a private link carrying the agency’s name, logo and colours. Print to PDF from the browser, or download every answer as a CSV.

It covers ChatGPT, Gemini, Google AI Overviews, Google AI Mode, Perplexity and Claude, and on paid plans an audit can re-run weekly or monthly to show the trend. It does not cover Copilot, Grok or Meta AI.

Try it on a client

If you want to see the method run on a real brand, the free plan includes one audit with up to 10 questions. See the plans or read how the AI visibility checker works.

Frequently asked questions

What is an AI visibility audit?

An AI visibility audit measures how often AI assistants such as ChatGPT, Gemini and Google AI Overviews name a brand when buyers ask category questions. It records where the brand appears in each answer, which competitors are named instead, and which websites the assistants cite as sources.

How many prompts do you need for an AI visibility audit?

Twenty to thirty buyer questions is a workable size for one brand. Fewer than ten makes each answer swing the score too much. Ask every question more than once on engines that vary, because the same question can return a different list of brands from one run to the next.

Can I do an AI search audit by hand?

Yes. Open each assistant in a logged-out or private session, ask your list of questions, and record in a spreadsheet whether the brand and each competitor is named and which sources are cited. It works, but 20 questions asked five times each is 100 answers to read and log.

How often should an AI visibility audit be repeated?

Answers change as the models and the web change, so treat one audit as a snapshot. A sensible rhythm is once a quarter, or after a significant content or PR push, using the same questions and competitors so the two results can be compared fairly.

Written by the AI Visibility Desk team

We build AI visibility audits for agencies and run them across ChatGPT, Gemini, Google AI Overviews, AI Mode, Perplexity and Claude every day. Guides are checked against what the audits show and updated when the engines change. About us.

Run your first audit free

Enter a client’s website, let the AI fill in competitors and buyer questions, and get a branded report across six AI engines. No card needed.

Start free