
AI visibility has quickly become one of the most talked-about topics in automotive marketing. It has also become one of the easiest places to create false precision.
Dealers are being shown mention rates, citation scores, share-of-voice dashboards and sentiment reports that appear to describe exactly how they are performing inside ChatGPT, Google AI Overviews, Gemini and other AI experiences.
Some of those metrics are useful today. Some are useful only when treated as sampled estimates. Others are still too unstable at the individual-dealer level to support confident conclusions.
To separate what can actually be measured from what merely looks measurable, we asked David Gruhin to walk through the current state of AI-search measurement and the implications for dealerships evaluating vendors, dashboards and 2027 visibility strategies.
Want to execute better AI visibility tracking in your store?
Download the Hrizn Labs white-paper
David Gruhin: I would separate the metrics into three groups.
The first is AI referral traffic: visits reaching the dealer’s site from an AI assistant and what those visitors do once they arrive. Total visits, unique visits, form submissions, clicks to call, total actions and action rate can all be reported with reasonable reliability today.
The second is AI presence: mention rate, citation rate and share of voice. Those can be useful, but they are estimates. They become much more defensible when the location is fixed, the same prompts are run repeatedly and the result is evaluated over a meaningful window rather than from one run.
The third is sentiment. At an individual-dealer level, I do not think the sample is dependable enough yet in most cases. There simply are not enough observations behind the number.
Gruhin: Because an AI answer is generated fresh every time.
The assistant may rewrite the original question into one or more searches of its own. It retrieves pages and local business data. It then generates an answer from what it retrieved and attaches sources. Variation can enter at every stage.
That makes the measurement fundamentally different from checking a conventional search result.
There is also an important distinction between a test environment and an actual shopper. Real consumer answers can depend on location, account history and the conversation that came before the question. A scheduled test does not have the same context. A provider’s developer interface may also be a different product from the consumer application itself.
Gruhin: No. It can be very useful if you are honest about what it represents.
A fixed prompt list is a probe. It is not a sample of actual consumer demand because the AI platforms do not publish the full set of prompts consumers type or the volume behind them.
That means a presence test can help answer questions such as: Do we consistently appear around this topic? Which competitors are being retrieved? Which pages are getting cited? Where do we seem to have a content or authority gap?
Those are useful diagnostic questions.
The problem comes when a sampled test is presented as though it were a precise measure of how often every customer in the market sees the dealer.
Gruhin: More than most daily dashboards are using.
Take a simple example. If you run ten prompts once and the dealer appears in four answers, the dashboard may report a 40% mention rate.
But with only ten observations, the approximate 95% margin of error around a rate near 50% is about plus or minus 31 percentage points. One different answer moves the reported rate ten points.
So a dealer moving from 40% this week to 50% next week may not have improved at all. The change may simply be run-to-run variation.
For a more defensible program, I would fix the location, repeat each prompt multiple times and report the observation count beside the rate. In the memo, we used roughly 300 observations per dealer in a reporting period as an example of a much more stable design. That gets the margin closer to approximately six points rather than thirty.
Gruhin: I would not manage the dealership against daily movement in sampled presence metrics.
If the underlying measurement is probabilistic, generating it every day does not make it more precise. It can simply make the noise more visible.
A weekly or monthly window is much more useful. Keep the methodology consistent, run enough observations and look for durable patterns rather than reacting to every change in a generated answer.
Gruhin: First-party evidence deserves a lot of weight.
One important development in 2026 is that Google and Microsoft began providing site owners with actual exposure data from their AI experiences.
Google Search Console now reports generative-AI exposure tied to pages, countries, devices and dates. Bing Webmaster Tools provides AI-performance information including citations and cited pages.
That is different from running a synthetic prompt and asking whether the dealer appeared. It is evidence from the platform that the dealer’s content was actually surfaced.
I would combine that first-party exposure data with AI referral traffic, then use sampled prompt testing to understand where opportunities or gaps may exist.
Gruhin: It is one of the strongest metrics available because it represents actual customer activity.
GA4 now has an AI Assistant channel that can identify traffic arriving from supported AI assistants. That allows a dealer to measure the visits and then look at what those visitors do: forms, calls, other actions and conversion behavior.
There are still limits. Visits without a referrer may be classified as Direct, and traffic from Google AI Overviews and AI Mode is not included in that AI Assistant channel. Those visits remain within Google organic traffic.
But we are still dealing with observed customer behavior rather than a simulated answer. That makes referral activity particularly valuable.
Gruhin: This is one of the most important measurement problems in the transition.
When Google answers more of the question directly on the search-results page, fewer people need to click through to the underlying website.
That means organic clicks and CTR can fall even while a dealer’s content is gaining more exposure inside AI answers.
The memo looked at several outside data points. Pew found substantially lower click-through behavior when an AI summary appeared. Ahrefs has also reported large CTR declines where AI Overviews are present.
The exact magnitude varies by study and query type, but the direction matters: informational searches are increasingly vulnerable to zero-click behavior.
For dealers, that means a strategy can potentially improve AI visibility while the old organic dashboard appears flat or worse. If leadership looks only at clicks, success can start to look like failure.
Gruhin: No. Traffic still matters because it represents an opportunity for the dealership to create a customer action.
What changes is the assumption that more visibility must always produce more clicks.
I would report AI exposure alongside organic traffic rather than treating either one as the entire picture. Then I would look downstream at form submissions, calls and other meaningful actions.
The dealership needs to understand both influence and response.
They are becoming less synonymous.
Gruhin: Sentiment is probably the metric I would treat most cautiously at the individual-dealer level.
You can only classify sentiment when the dealership is actually mentioned. If a dealer appears in four of ten responses, you may be trying to build a sentiment score from four generated answers.
That is far too little information to support the precision the dashboard may imply.
There is also a relationship with existing reputation signals. The wording of the answer may be heavily influenced by the reviews and local data the assistant retrieved.
At launch, I would be much more comfortable evaluating sentiment at a larger program or regional level, or leaving the dealer-level field blank until enough observations exist.
Gruhin: Ask about the methodology before looking at the score.
Which platforms are included? Which prompts? How were those prompts selected? What location was used? How many times was each prompt run? What counts as a mention? What counts as a citation? Is the consumer application being observed, or is the test running through a developer interface?
Then ask how many observations sit underneath the rate.
Those questions matter because two vendors can both report something called “AI share of voice” while using different prompts, different locations, different run counts and different counting rules. The numbers may look comparable even though the measurement systems are not.
Standardizing the method is essential before those scores are used to compare dealers or vendors.
Gruhin: I would build it in layers.
Start with AI referral traffic and the actions those visitors produce.
Add first-party AI exposure data from Google and Microsoft wherever it is available.
Use presence metrics as sampled estimates, with enough repeated observations to make the trend meaningful.
Keep measuring the underlying organic, local and reputation signals because AI retrieval still depends heavily on search indexes, local listings and the information already available about the dealership.
And I would be careful about turning sentiment into an individual-dealer KPI before the sample supports it.
The point is not to create one magical AI score.
It is to build enough evidence from different sources to understand whether the dealership is becoming easier for these systems to retrieve, cite and ultimately connect with customers.
Gruhin: Do not confuse precision of presentation with precision of measurement.
A dashboard can display a number to one decimal place even when the underlying sample is far too small to support that confidence.
AI search absolutely deserves measurement. Customers are using these systems, dealers want to appear in the answers, and the platforms themselves are beginning to expose more useful data.
But we should be disciplined about what each number can actually tell us.
Some metrics are ready to manage.
Some are ready to diagnose.
Some are not ready yet.
Knowing the difference is part of being ready for what comes next.
AI visibility is becoming measurable, but not every metric carries the same level of confidence. Start with the evidence available today, establish a defensible baseline and build the search, local and content authority that makes the dealership easier for customers and AI systems to understand.
Start With Your AI SEO + GEO Visibility Audit →
Build a Defensible AI Visibility Measurement Framework →
Understand the Reality of AI Visibility →
Use IdeaCloud to Build Around Observable Customer Intent →
Read: Dealership Visibility Is No Longer a Marketing Channel →
Explore the 2027 Customer Experience Reset — Playbooks, Field Guides & Thought Leadership →
AI measurement will get better. Build the methodology that is honest about what it knows today.
We Rise Together.