Ai Search

How to Audit Your AI Visibility Yourself

You can check your own AI visibility in an afternoon with a browser, a private window, and a spreadsheet. The method is simple: ask the AI engines the questions your buyers actually ask, then record whether each one names you, links to you, or skips you for a competitor.

This guide is the exact do-it-yourself audit. You will build a prompt set, run it cleanly across ChatGPT, Perplexity, Gemini, and Google AI Overviews, and score what comes back. No paid tool is required to start. What a manual check cannot tell you, and where a repeatable process beats a one-time look, comes at the end.

Citation, mention, or absent: what you are scoring

Get the vocabulary straight before you run anything, because the whole audit hinges on it. Three outcomes matter, and they are not the same thing.

Citation: the engine names your business and links to a page on your own site as a source. This is the strongest signal. It means the model retrieved your page and leaned on it to build the answer.

Mention: the engine names you in the text but links nowhere, or links to a third party instead. You exist in what the model knows, but your site is not the source feeding the answer.

Absent: a question that should surface you returns competitors instead. This is the outcome that quietly costs you work, and it is the one most owners never check.

Here is the mechanism that makes the distinction real. Most AI answers are not recited from memory. When you ask a current, local, or specific question, the engine runs one or more live searches, pulls back a handful of pages, and grounds its answer in them. The pages it actually uses become the citations shown to the buyer. Your audit exists to find which questions trigger that retrieval and whether your pages make it into the set.

Build a prompt set your buyers would actually type

A useful audit tests real demand, not vanity queries. Searching your own brand name proves almost nothing, because the engine will nearly always find you. Test the questions a stranger with a wallet would ask before they know your name.

The four prompt types to cover

Category plus place: best [service] in [your city], or top [service] companies near me.

Problem-first: the symptom a buyer describes before they know what to buy, such as my [problem], who do I call.

Comparison: [competitor] alternatives, or [service A] versus [service B].

Qualifier: the words buyers actually weigh, like affordable, licensed, or same day.

Write ten to twenty of these and then freeze the list. The fixed prompt set is the real asset here. Reusing the identical list next quarter is the only honest way to tell whether anything moved.

Run the prompts across the engines, cleanly

Each engine retrieves differently, so test them one at a time and control for contamination that would hand you a flattering, fake result.

Keep the session clean

Use a logged-out or incognito window. A logged-in account carries memory and history that personalize the answer, so you end up seeing your result instead of a cold buyer's.

If you must stay signed in, switch off chat memory and any custom instructions first.

Run each prompt once and take the first answer. Do not refine it or argue with the model. You want the cold response a real buyer gets, not one you coaxed.

The engines worth covering

ChatGPT with search enabled, which cites linked sources inside or beneath the answer.

Perplexity, which is citation-first and numbers its sources, making it the fastest place to see who gets credited.

Google AI Overviews, the generated box above the blue links, which pulls heavily from the pages already ranking.

Gemini, which leans on Google's index and the knowledge graph behind it.

Record it the same way every time

Open a spreadsheet. One row per prompt, one column per engine. In each cell, log one of four codes so the pattern is impossible to miss later.

Cited: named and linked to your site.

Mentioned: named, but with no link back to you.

Displaced: a competitor is named and you are nowhere.

Wrong: you are named, but a detail is off, like a stale service area or an old phone number.

Patterns surface fast once it is written down. If competitors own the category prompts while you appear only on brand prompts, you have a retrieval problem, not a reputation problem. If you are named with no link, the model trusts the idea of you but not your pages. If a fact is wrong, something the engines are copying is wrong, and that source can usually be found and corrected.

Diagnose the gaps you found

A blank cell is a question, not a verdict. The usual causes are concrete and checkable, and most of them are things you can inspect today.

Crawler access: read your robots.txt and look for rules blocking GPTBot, OAI-SearchBot, PerplexityBot, or Google-Extended. If an engine's crawler is disallowed, it cannot retrieve your pages to cite them, no matter how good they are.

Unstructured content: if the answer to a buyer question is buried mid-paragraph instead of stated plainly under a clear heading, the model has to work to extract it, and often will not.

Weak entity presence: engines cross-reference who you are against the knowledge graph and outside sources. Thin or inconsistent business information makes you a weaker candidate to name.

No outside corroboration: one site calling itself good is weaker than directories, reviews, and discussions that independently say the same. Engines weight agreement they did not get from you.

No machine-readable guide: an llms.txt file at your domain root lists your key pages in plain text so a model can find your best answers without crawling your whole site. Its absence will not block you, but it drops a hint you could be giving.

What a do-it-yourself check cannot tell you

A manual audit is a snapshot, and AI answers are noisy by nature. The same prompt can return different sources from one hour to the next, because retrieval is live and model versions shift underneath you. One afternoon tells you roughly where you stand, not whether you are climbing or sliding.

It also cannot fully strip out personalization, measure your share of voice over time, or separate one bad sample from a genuine decline. That gap is the whole difference between a spot check and monitoring: a fixed prompt set, run on a schedule, logged the same way every time, so a change you see is real movement and not sampling noise. If you would rather have that process run for you and the gaps it exposes fixed on the same engagement, that is exactly what a formal audit and the ongoing work are built to do.

Questions

How to check my AI visibility myself?

Ask the AI engines the questions your buyers actually type, run them one at a time in a logged-out window across ChatGPT, Perplexity, Gemini, and Google AI Overviews, and record in a spreadsheet whether each one cites you, mentions you, or returns a competitor instead. A browser, a private window, and a fixed prompt set are all you need to start, and freezing that prompt list is what lets you compare one quarter to the next instead of reading noise.

Do I need a paid tool to check my AI visibility myself?

No. A browser, a private window, and a spreadsheet are enough for a first real audit. Paid tools earn their keep later, when you want the same prompt set tracked automatically, across more engines, on a schedule. The manual method gives you the same core answer at no cost.

How many prompts should I test?

Ten to twenty is plenty for a first pass. Cover category, problem-first, comparison, and qualifier questions so you are testing real buyer demand. Keep the list fixed so you can rerun the exact same prompts later and compare honestly.

Why does ChatGPT name me one day and not the next?

Retrieval is live and the underlying index shifts, so the same prompt can surface different sources from one hour to the next. That is why a single snapshot is a rough read rather than a trend. Running a fixed prompt set on a schedule is the only way to tell real movement from noise.

ChatGPT mentions my business but gets the facts wrong. What do I do?

A wrong fact means the engine is copying bad data from a source you can usually trace, like an outdated directory listing or an old page. Correct it at that source and on your own site, and make the right version easy to extract under a clear heading. The engines tend to follow the corrected record over time.

Should I test my own brand name?

Test it once to confirm the engines know you exist, but do not count it as visibility. Buyers who already know your name are not the opportunity. The questions that matter are the category and problem prompts a stranger types when they have never heard of you.