You can audit what the AI engines say about your brand in about twenty minutes, for free, using nothing but the engines themselves and a spreadsheet. The method: write ten questions your buyers actually ask, open ChatGPT, Perplexity, Google AI Overviews, and Gemini in a clean session with no history, ask each question once exactly as written, and record four things every time. Were you named. Who else was named, in what order. Which sources were cited. And is anything said about you wrong. Then repeat it on the same day every month, with the same questions, in the same order. One reading tells you where you stand today. The monthly repetition is what turns it into evidence, because AI answers vary between runs and only a trend line separates real movement from noise. Everything below is the detail that makes each step worth doing.
Step one: write ten questions in buyer language
The questions decide whether the audit is worth anything, so do not reuse your keyword list. Search keywords are compressed; questions put to an assistant are long and conversational. Write ten across three shapes. Three or four category questions, asked by someone who does not yet know any brand names, such as “who does technical SEO for e-commerce brands in Bangkok” or “what should I look for in a GEO agency”. Three or four comparison questions, asked by someone narrowing a shortlist, such as asking for the alternatives to a competitor you keep losing to. Two or three direct brand questions, such as “what is Once Be Found” and “is Once Be Found any good”. The category questions tell you whether you exist. The brand questions tell you whether the engine has you right.
Step two: open a clean session on each engine
Personalisation will lie to you if you skip this step. ChatGPT remembers past conversations and will lean toward what it already knows about you, so use a temporary chat, which neither reads nor writes memory. For Perplexity and for Google AI Overviews, use a private or incognito window and stay logged out, so your account history and saved location do not shape the answer. Do the same on Gemini, and on Microsoft Copilot, Claude, or Grok if your buyers use them. Being logged out matters for a second reason beyond clean data: it puts you in roughly the position of a stranger researching your category for the first time, which is exactly the person you are trying to learn about.
Step three: ask once, and do not help it
Paste the question exactly as written, read the answer, and stop. No follow-ups, no clarifications, no “actually I meant”. The temptation to nudge the engine until it finally mentions you is strong, and the moment you nudge, the reading is worthless, because no real buyer will nudge on your behalf. One question, one answer, move on. If an answer comes back about something else entirely, record that as its own finding rather than rephrasing: an ambiguous question is a real thing your buyers run into, and that ambiguity may be the reason nobody in your category gets named on it.
Step four: record four columns, not a general impression
This is the step people skip and the one that makes the exercise repeatable. Open a spreadsheet, in Google Sheets or anything else, and give every question and engine pair its own row, so ten questions across four engines gives you forty rows. Column one: were you named, yes or no. Column two: every other brand named, written down in the order they appeared, because order carries meaning in a generated answer. Column three: the sources cited, which Perplexity and ChatGPT both show. Column four: anything factually wrong about you, such as a stale price, the wrong city on your Google Business Profile, a service you stopped offering two years ago, or a founder who has left. That fourth column is usually the one that produces immediate work.
Step five: read the citations, not just the brand names
The list of cited sources is the part most people glance past, and it is where the real work comes from. If the same handful of domains keeps appearing across your ten questions, that is your category's source set, and it is usually a mix of independent reviews, forum threads, trade publications, and a couple of competitors' own pages. Reddit, YouTube, Wikipedia, LinkedIn, and review sites such as Trustpilot turn up far more often than most brands expect. You now have a concrete target list, because to be in the answer you generally need to be described on the pages the answer is built from. That is a very different project from writing another blog post, and a considerably more useful one.
What this method cannot tell you
Be honest with yourself about what forty rows in a spreadsheet are. They are a sample of one, taken once. AI answers are generated rather than looked up, so the same question asked twice in one hour can return different brands in a different order, and a single absence is not proof of a problem. You also have no demand data behind any of it: nothing here tells you how many people ask that question, and a question you invented may be one nobody asks. And you are seeing one country, one device, one language, at one moment. If you sell in two languages, run the whole panel twice, once in each, because an engine answering in Thai leans on Thai-language sources and can name a completely different set of brands than the same question in English. Those limits are real, and they are survivable, because the answer to them is repetition rather than software. A full GEO audit adds the things a panel cannot see from the outside, such as whether the crawlers can reach you at all, but the panel is where everyone should start.
Make it a habit, not a project
Run it on the same day each month and resist changing the questions, because every change resets your baseline and you lose the comparison that made the exercise worth doing. Hold the set steady for at least two quarters, and add a question only when the business genuinely changes, such as a new service or a new market. Keep every month's sheet rather than overwriting it. If the panel eventually stops fitting, because you are tracking four markets or fifty questions or three brands, that is the point where paid monitoring starts to earn its cost, and honestly not much before it.
The short version
Ten buyer questions, four engines, a clean session, one ask each, four columns, repeated monthly. That is the entire method and it costs nothing but the twenty minutes. Run it once and you will know whether you exist in AI answers. Run it three months running and you will know whether anything you are doing is working. If you would rather have the first reading done for you, our free AI visibility report is the same panel run across the engines and written up within a day, with the source set and the highest-impact fixes named.