Before spending anything on AI search visibility, find out what the assistants already say about you. It takes about fifteen minutes, needs no paid tools, and produces the baseline that every later change is measured against. The method is simple: write down the questions your buyers actually ask, put them to ChatGPT, Gemini, Perplexity and Google, score what comes back, then check what the rest of the web says about your company. Most teams discover one of three things, and each points at different work.
Why the baseline has to come first
Without a recorded starting point, you cannot tell whether later work did anything. AI answers also vary between runs, between models and between signed-in users, so a single screenshot taken after a change proves nothing. The only way to see movement is to ask the same questions the same way, write down the answers, and repeat it on a schedule.
There is a second reason. The audit tells you which problem you have. A company that is never mentioned has an access or identity problem. A company that is mentioned but described wrongly has a specific outdated page somewhere. A company that is cited for facts but never recommended has a corroboration problem. These need different budgets, and guessing between them is expensive.
What you need before you start
- Accounts on ChatGPT, Gemini and Perplexity. Free tiers are enough for this
- A browser in a logged-out or private window, so personalisation and chat history do not shape the answers
- A spreadsheet with one row per question
- Fifteen minutes, or thirty if you want to run each question more than once
Step 1: build the question bank
Write 15 to 20 questions in the words a buyer would use, not the words your marketing uses. The test of a good question is that someone who has never heard of your company could ask it. Cover four kinds:
| Type | Shape of the question | What it reveals |
|---|---|---|
| Category | “What is the best way to handle X?” | Whether you appear at all in your own subject |
| Vendor selection | “Who are the leading providers of X in India?” | Whether you are recommended, and who is named instead |
| Comparison | “X or Y, which is better for a mid-sized company?” | Which competitive set the assistant puts you in |
| Brand | “What does [your company] do?” | Whether the description of you is accurate and current |
Keep your company name out of the first three types. A question that names you will usually surface you, which tells you nothing about discovery. If you need help generating the sub-questions buyers ask around a topic, query fan-out keyword research covers the method.
Step 2: run the questions
Paste each question into each assistant and record the result. Do not refine the prompt, do not add context, and do not correct the assistant. You are measuring what a buyer would see on their first attempt.
For each answer, capture three things: whether your company is named, which URLs are cited, and which competitors appear. Copy the answer text into the spreadsheet rather than summarising it, because the wording matters when you look for the source of a wrong description later.
Run the brand questions twice if you have time. Answers that change between runs on the same question are telling you the assistant has weak or conflicting information about you, which is itself a finding.
Step 3: score each answer
A four-point scale is enough and keeps the scoring fast and repeatable.
| Score | What it means |
|---|---|
| 0 | Not mentioned anywhere in the answer |
| 1 | Mentioned in passing, or cited as a source without being named in the text |
| 2 | Named as one option among several |
| 3 | Named as a leading or recommended option, or quoted directly |
Add a separate accuracy flag for the brand questions: correct, outdated or wrong. Keep these apart from the score. A company can be recommended enthusiastically and described incorrectly at the same time, and those are two different jobs.
Your headline number is the share of question and platform combinations scoring 1 or higher. With 20 questions across four platforms that is 80 cells, and the percentage that are not zero is your mention rate. Record it with the date.
Step 4: check your footprint off your own site
Assistants build their picture of a company from the whole web, not only from your website. Two searches show you what they have to work with. In Google, search your company name with your own domain excluded, using the minus operator in front of a site search. That returns the directories, articles, profiles and forum posts that mention you. Then search your company name alongside your category to see whether you appear in any list of providers.
| What you find | What it means | First move |
|---|---|---|
| Almost nothing off your own domain | No corroboration, which is why citations do not become recommendations | Directory listings, then trade media coverage |
| Profiles exist but the details differ | The assistant cannot be sure the sources describe one company | Make name, location, founding year and description identical everywhere |
| Mentions exist but never link to you | Unlinked mentions still help identity, but you are not capturing the reference | Ask for a link where the relationship allows it, and keep the mention accurate |
| Outdated information ranks well | This is usually the source of a wrong AI description | Update or remove the page, then re-test the brand questions |
For listings specifically, the review sites assistants cite covers which ones are worth the effort, and brand SERP management covers what shows for your company name.
Step 5: read the results
Three patterns cover most outcomes, and each has a different first action.
| Pattern | What it points at | Where to start |
|---|---|---|
| Mention rate near zero across all platforms | Access or discovery. Often the assistant cannot read the site, or no page targets the question | Test AI crawler access, then check whether any page answers the question at all |
| Cited for facts but never recommended | Identity and corroboration. You are useful but not yet a known entity | Consistent company facts, author pages, directory and trade coverage |
| Mentioned but described wrongly | One or more outdated pages, on your site or elsewhere | Trace the wording back to its source page and fix that page first |
A fourth outcome is worth naming because it is easy to misread. If one platform mentions you and the others do not, the cause is usually crawler access specific to that company rather than content, since the content is the same for all of them. The seven-point diagnostic for a missing mention walks through the causes in order.
What this audit does not tell you
It measures what assistants say, not why your site is or is not retrievable. If the audit points at access or extraction problems, the next step is a technical pass over the site itself: rendering, indexing, schema and passage structure. Our GEO Readiness Score covers that side as a 100-point check, and the 2026 SEO audit covers the wider technical ground.
It is also a snapshot, not a trend. One run tells you where you stand today. The value comes from the second and third runs.
How often to repeat it
Monthly is enough for most companies, on the same question set, logged out, with the results appended to the same sheet rather than overwritten. Re-run it sooner after any significant change: a site migration, a rebrand, a new service line or a fix to a wrong description. For automating the repeat runs and the reporting, see tracking brand mentions for free and building an AI visibility dashboard.
The thirty days after the audit
- Fix anything the audit showed as factually wrong about your company, starting with the page the wrong wording came from
- Clear any crawler access problem, since nothing else works until pages can be read
- Make your company facts identical across your site and every profile you control
- Rewrite the three pages that map to your highest-scoring buyer questions so the answer comes first
- Claim or correct the two or three directory listings that appear most often in your category
- Put the next audit date in the calendar before you close the spreadsheet
Resist the urge to do everything. The audit exists to tell you which of these matters most for your situation, and doing the first two badly makes the rest pointless.
Frequently asked questions
Do I need paid tools for this?
No. The free tiers of the assistants and a spreadsheet are enough for a baseline. Paid monitoring tools save time once you are running the test monthly across many questions, but they are not needed to find out where you stand.
Why run the test logged out?
Because a signed-in session carries your history and personalisation, which can surface your own company for reasons that have nothing to do with how a stranger would see it. Logged out is closer to a buyer’s first encounter.
Will the answers be the same every time?
No, and that is expected. These systems are not deterministic, and some of them search the web live. Variation between runs on the same question is a signal in itself: stable answers mean the assistant is confident, and unstable ones usually mean thin or conflicting information about you.
How many questions should the bank contain?
Between 15 and 20 for a first audit. Fewer than that and one odd answer distorts the result. Many more than that and the test stops being repeatable monthly, which matters more than breadth.
Should I include competitor names in the questions?
In the comparison questions, yes, because that is how buyers ask them. Keep your own name out of the discovery questions, since naming yourself guarantees a mention and hides the thing you are trying to measure.
What if an assistant is unavailable when I run the audit?
Note it in the sheet and run that platform later, rather than substituting a different model or a different question. The comparison across months only holds if the question set and the platforms stay the same.
Who should run this?
One named person, every month. The method is simple enough for anyone in a marketing team, and the value comes from consistency rather than skill, so handing it around between people tends to break the comparison.
Sources
- Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, 22 July 2025
- Google Search Central documentation on AI features in Search
- OpenAI, Anthropic and Perplexity published crawler documentation