A GEO audit is a one-time check of how AI assistants and AI-generated search answers talk about your brand right now: whether they name you, whether they cite a page of yours, and who they mention instead. You can run a rough version by hand in an afternoon by asking the assistants the questions your buyers actually ask. A paid AI visibility tool becomes worth it once you need that check repeated on a schedule, at a sample size larger than you can type by hand, across the engines you actually care about.
People search this as geo audit, ai visibility audit, ai search audit and geo score, and they are asking the same underlying question: where do I stand right now, before I spend money on a tracker or an agency? That is a different job from AI search optimization itself - an audit does not change anything, it just gives you an honest baseline so the changes you make afterward have something to be measured against.
What a GEO audit actually checks
Strip away the tooling and a GEO audit answers five plain questions. Each one maps to something you can check without a subscription, though a paid tool automates the checking and repeats it at a scale no person can match by hand.
| What you're checking | The question | Why it matters | Manual way to check it |
|---|---|---|---|
| Mention rate | Does the assistant name your brand at all, for the questions your buyers ask? | Zero mentions means you are invisible to that surface regardless of anything else | Ask ChatGPT, Perplexity and Gemini the same 10-15 buyer questions and log a yes/no per answer |
| Citation | When you are mentioned, does the assistant link or cite a page of yours? | A named-but-uncited brand gets no click and no trust signal from the citation itself | Check whether the answer includes a source link or footnote pointing at your domain |
| Share of voice | Who else gets named alongside you, and how often do they win the top spot? | Tells you who you are actually losing the answer to, not just whether you appear | Log every competitor name that shows up across the same 10-15 prompts |
| Sentiment and accuracy | Does the assistant describe you correctly, or repeat stale or wrong facts? | A wrong price or a discontinued feature quoted as current costs trust with a buyer who checks | Read each mention for factual errors against your current pricing and feature set |
| Consistency across surfaces | Do ChatGPT, Perplexity, Gemini and Google's AI Overviews agree with each other? | Disagreement usually means your public facts (pricing, category, name) are inconsistent somewhere | Repeat the same prompts across at least three assistants and compare the answers side by side |
A manual GEO audit you can run today
None of this requires a paid account. It requires patience and a spreadsheet, and it will tell you more than a single dashboard number would.
1. Write down the questions your buyers actually type, not your own brand name
Searching your own brand name is the least useful prompt you can run - it mostly tests whether the assistant has heard of you. The check that matters is whether you show up for the comparison and problem-shaped questions a buyer types before they know your name: "best tools for X", "is [competitor] worth it", "alternatives to [category leader]". Pull 10-15 of these from your own sales calls, support tickets or the queries already showing impressions in Search Console.
2. Run the same list across at least three assistants
Ask each question in ChatGPT, Perplexity and Gemini in fresh, logged-out sessions where possible, and note per question: did they name you, did they cite a page, who else got named, and top spot or footnote. Perplexity's citations are the easiest to read since almost every answer carries visible source links; ChatGPT and Gemini cite less consistently, so absence of a link there is weaker evidence than absence in Perplexity.
3. Repeat each prompt two or three times before you conclude anything
The single biggest trap in a manual audit is treating one answer as the truth. Ask the identical question again in a new session and the mention, the citation, and even the competitor list can shift - these are generated answers, not a stored ranking. A brand that is mentioned in two of three runs is materially different from one mentioned in zero, but you cannot tell the two apart from a single pass. This is also the main reason a dashboard that automates hundreds of runs finds a signal a spreadsheet of one-off checks tends to miss.
4. Score yourself against the table above, not against a single number
A raw "geo score" out of 100 is a compression of the five rows in the table above into one figure, and compressing them hides which one is actually broken. Being named often but never cited is a citation problem worth fixing on its own; being cited correctly but rarely named is a discovery problem with a different fix. Keep the breakdown even if a tool later hands you a single score.
Where a manual audit runs out of road
A spreadsheet of 15 prompts run three times across three assistants is roughly 135 data points, gathered by hand, for one point in time. That is enough to tell you whether you have a problem. It is not enough to tell you whether last month's content push moved the needle, because you would need to repeat the same 135-point exercise on a schedule and compare the two. It also will not cover AI Overviews and AI Mode at any scale, since those need to be triggered by real search queries rather than a chat prompt, and it will not separate a genuine trend from the normal answer-to-answer variance covered above. That gap - repeatable sampling, at a size a person cannot type by hand, across more engines than three - is what a paid AI visibility tool is actually for.
When it is time to move from a manual audit to a tracked one
Run the manual version first, on both yourself and your two or three closest competitors - the comparison is usually more instructive than the absolute number. Move to a paid tool once you need to answer three questions the manual audit cannot: is this getting better or worse month over month, does a content change actually move the mention rate, and how do we compare on a sample large enough to trust. If budget is the blocker rather than need, the free and trial tiers on the market are built for exactly this - running a real tracked audit before committing to an annual contract - and our independent reviews break down what each one actually measures versus what it claims to.
One more thing worth knowing before you run any of this: "GEO" itself is not a marketing coinage - it comes from a specific 2023 research paper that tested nine ways of rewriting content and measured which ones actually changed visibility. See what generative engine optimization means for what that study found, and which tactic (keyword stuffing) it showed does not work.