ChatGPT, Perplexity, and Google AI Overviews each pull from different sources, and a business can be visible in one while invisible in another. That single fact turns “are we showing up in ChatGPT?” into the new “are we on page one?”, and answering it takes an actual measurement process instead of a one-off vibe check. Here is the monthly method we run for clients, step by step, so you can run it yourself if you want to.
Why does AI-search visibility need its own tracking?
AI assistants answer buying questions with a short list of names, and if yours is missing there is no page two to fall back on. Classic rank trackers cannot see inside those answers, so you need a separate, repeatable check. Without one, you only find out you are invisible when a prospect tells you an assistant sent them somewhere else.
Traditional SEO tools measure positions on a results page. An assistant’s answer has no positions, just a recommendation, a few alternates, and whatever facts the model believes about each business. Those facts can be wrong, stale, or borrowed from a competitor, and nothing in your analytics will flag it. The only way to know what the machines say about you is to ask them the way a buyer would, on a schedule, and write down what comes back. We covered how the engines assemble those answers in how AI assistants pick their recommendations; this post is about measuring the output.
What is the fixed-prompt-set method?
We write 15-30 questions a real buyer would ask an AI assistant, then run the identical set every month across ChatGPT, Perplexity, and Google AI Overviews. The set stays fixed so month-over-month movement means something. Change the questions and you have a new experiment, not a trend line.
Building the set is the part worth slowing down on. We pull prompts from three places: what clients say prospects actually asked before calling, the queries already driving conversions in search data, and the comparison questions buyers ask late in a decision. A useful set for a Tampa service business mixes several shapes:
- Category prompts: “best [service] in Tampa,” “top-rated [service] near me”
- Problem prompts: “who should I hire for [specific problem]”
- Comparison prompts that name real competitors
- Trust prompts: “is [your business] legit,” “reviews of [your business]”
That last group matters more than most owners expect, because assistants answer those questions whether or not you would like them to. We wrote the prompts for a national audience the same way; the local modifier is the only thing that changes.
One practical note on running the prompts: use fresh sessions with no history, because a logged-in account with months of context gets personalized answers that do not represent what a stranger sees. Same wording, clean session, every engine, every month.
How do you score the answers?
Each response gets scored on three questions: were you mentioned, were you recommended, and was the information accurate? Mentioned means your name appears anywhere in the answer. Recommended means the assistant put you forward as a pick rather than listing you among options.
The gap between mentioned and recommended is where most of the strategy lives. Plenty of businesses appear in answers as one name in a list of five, which is progress but rarely wins the customer. A recommendation, “for that, I’d look at X,” is the outcome that moves revenue.
Accuracy is the score that surprises people. Wrong hours, dead addresses, and a competitor’s phone number all show up more than you would hope. An assistant confidently telling prospects you close at 5pm when you are open until 9 costs you customers you never knew existed. We log every factual claim the engines make and flag anything wrong, because each error points at a stale source somewhere on the web.
We keep the scoring in a plain spreadsheet: one row per prompt per engine, columns for mentioned, recommended, accurate, and cited sources. No special tooling required. The discipline is the tooling.
How do you trace where the answers come from?
When an engine cites its sources, those citations are your to-do list. Perplexity and Google AI Overviews link sources directly; ChatGPT shows them when it browses. The directories, review sites, and articles that keep appearing are the surfaces shaping your machine reputation, and fixing them is the actual work.
Sources cluster into a few types, and each gets a different fix. Directory listings with wrong data get corrected at the source. Review platforms that dominate citations deserve an active review strategy. Third-party articles and roundups that engines lean on are outreach targets. And gaps, questions where engines cite nothing about you at all, are content you have not written yet. If the tracing turns up a thin or inconsistent footprint, the fixes look a lot like the groundwork in our checklist for showing up in ChatGPT: consistent facts, real answers on your own site, and a presence on the surfaces the engines trust.
How do you report it alongside SEO?
We put AI-search results on the same page as classic SEO metrics: rankings and clicks next to AI mentions, recommendations, and accuracy. One view, both worlds. Movement in one usually precedes the other, and seeing them together is how you catch it early.
That pairing also keeps the work honest. A month where organic clicks climb but AI mentions stay flat tells you the engines have not caught up to whatever earned the clicks. The reverse, mentions rising before traffic does, is the early signal that the footprint fixes are landing. Either way, the monthly cadence matters because engines update, sources shift, and answers move. That volatility makes a lot of owners want to skip measuring altogether. We read it the opposite way: when the answers move this much, measuring on a schedule is the job.
What have we learned running this?
The businesses that win AI answers are boringly consistent everywhere: same name, same facts, same story on every surface of the web. The ones that lose are often better businesses with messier footprints. The machines cannot recommend what they cannot verify.
The second lesson is that accuracy problems outnumber visibility problems early on. Most first-run reports for established businesses find at least one wrong fact circulating in the engines, and fixing those errors is usually faster and cheaper than earning new mentions. Start there.
This tracking is part of our SEO and AI search service. If you want to know what the machines currently say about your business, that is a fifteen-minute check with an occasionally uncomfortable answer. Ask us.
Common questions about AI tracking
How often should we run the prompt set? Monthly. Engines change fast enough that quarterly checks miss the story, and weekly runs add noise without adding signal for most businesses.
Can we track this with rank-tracking software instead? Some tools now sample AI answers, and they can help at scale. The fixed-prompt method still matters because your buyers’ actual questions, run consistently, beat a generic keyword list run by a tool.
What if we never show up in any answer? That is a starting point, not a verdict. Trace what the engines cite for your category, fix your footprint on those surfaces, and re-run the set. Movement typically shows up within a few monthly cycles, though the pace varies by market.
Does this replace SEO tracking? No. It sits next to it. The same clean site structure and consistent facts that earn classic rankings feed the AI engines, so the two reports tend to move together.