How to measure your dealership’s AI visibility: questions, repeat runs and margins

How do you measure a dealership’s AI visibility?

Fix a set of local buyer questions, ask each one several times of each assistant, and record who gets named, what gets cited and what the answer gets wrong about your store. Keep the raw answers, repeat the same questions monthly, and read the results next to Search Console, Bing Webmaster Tools and GA4.

Think of it as two kinds of evidence. The first is a sample of what assistants say when a buyer in your town asks a question, which only you can collect by asking. The second is what Google and Microsoft report about their AI features sending impressions, citations and visits to your site. Neither one is complete on its own. If the vocabulary is new, start with AEO vs SEO for car dealers, which shows where AI answers sit next to regular search.

The same discipline sits behind AEO and GEO for car dealers: measure first, change things, then measure again the same way so the before and after compare cleanly.

Why does one test mislead?

Because AI answers vary. The same question can name different stores from one run to the next, from one person to the next and from one town to the next, and OpenAI tells users that search results and citations can be wrong or out of date. One screenshot is a single draw from a system that keeps moving.

Location is one reason. OpenAI says ChatGPT may estimate a user’s general location from their IP address and share it with its search providers to localize results; its example turns “good restaurants near me” into “top restaurants San Francisco” (OpenAI Help Center). Your sales manager at the store and your GM at home in the next county can get different answers to the same words.

Accuracy is another. OpenAI’s own help page warns: “Search results and citations can be incomplete, outdated, or incorrect.” An answer that names you with last year’s hours, or leaves you out because a directory lists you as closed, is still the answer a buyer sees. For how one engine picks stores, read how ChatGPT recommends dealerships.

A hypothetical makes the point. A used-car store in a metro area asks ChatGPT once for the best place to buy a used SUV nearby, sees its own name and calls the job done. Asked twice more that afternoon, the same question could name two competitors instead. Repetition is the only defense against a lucky or unlucky single answer.

How many times should each question be asked?

At least three times per assistant, on the same day, with web search on. Three runs will not remove the variation, but they show whether a result holds: named in three of three runs is a different finding from named in one of three. More runs narrow the range further and take more time.

AutoLander’s free scan uses this method. It asks ChatGPT and Claude, each with web search on, up to 20 local buyer questions, 3 times each, which comes to as many as 120 answers. A person on our team checks every match before the report goes out, and the score out of 100 carries a margin because answers change from run to run.

Two habits make repeat runs useful. Start a fresh chat each time, so an earlier conversation does not steer the answer. And name the town in the question the way a buyer would, which gives a fairer read than relying on wherever your office computer happens to be.

Can any tool see inside Google’s AI?

No. Google states that “No third-party tool has access to our internal ranking or AI systems.” Any AI rank or visibility number you see from a vendor, ours included, comes from the outside: someone asked questions and recorded the answers. That can still be useful, as long as the method is shown to you.

The warning comes from Google’s guide to its generative AI features, which tells site owners to be wary of third-party tools that promise ranking success or claim to use internal Google metrics, because an outside tool cannot see those systems.

So ask any vendor three things: which assistants were asked, how many times each question ran, and whether you can read the raw answers. A number without those three answers tells you very little.

Know the limits of your own sample too. AutoLander’s free scan measures ChatGPT and Claude only. It does not measure Google AI Overviews, AI Mode, Gemini, Perplexity or Microsoft Copilot, which is why the Search Console and Bing reports belong in the same monthly review.

How often should you re-measure?

Monthly, on the same question set, with the same assistants and the same number of runs, so each month compares cleanly with the last. Weekly checks on a handful of high-value questions can catch sudden changes, but the monthly run is the one to trend. Change the questions only when your market or inventory mix changes.

Expect clicks and visibility to move separately. Bing notes that content can show up as impressions, citations or grounding references in Copilot without a click, and says “A decline in clicks does not always indicate a loss of visibility” (Bing Webmaster Guidelines). Jordi Ribas, Microsoft’s corporate vice president of search and AI, wrote in February 2026 that visibility is no longer defined only by rankings or clicks (Microsoft Bing blog).

A simple monthly sheet does the job: one row per question, one column per assistant and run, a cell for who was named and a cell for what was cited. Add the Search Console, Bing and GA4 numbers at the bottom. When something changes, the raw answers explain why.