Google AI Overviews answer the question above the results. A reader can finish their search without scrolling to a single link, which means your position in the blue links is now a separate question from whether you were part of the answer at all.
That makes AI Overviews worth tracking. It also makes them awkward to track, because almost every instinct carried over from rank tracking is wrong here.
Why an AI Overview is not a ranking
A traditional search result is retrieved from an index. Ask the same question twice, ten minutes apart, and you get essentially the same ten links in essentially the same order. That stability is what makes a rank tracker meaningful: position four means something, and position four tomorrow means the same thing.
An AI Overview is generated. Google assembles it fresh, from sources it selects at that moment, and the result varies. Two checks minutes apart can name different brands, cite different pages, or produce no Overview at all.
This has an uncomfortable consequence. A single check of an AI Overview is close to meaningless. If a tool checks once and shows you a green tick, it is reporting the outcome of a coin flip as though it were a fact. The only honest way to describe something that varies is a rate across repeated observations, with the number of observations attached.
We set out the general version of this rule in our measurement methodology: counts always carry their denominators, and below five observations we show a fraction rather than a percentage, because one run is a draw and not a rate.
The four states worth measuring
Most tools collapse AI Overview tracking into one number: are you in it or not. That throws away the distinction that tells you what to do next. There are four outcomes, and they need different work.
No Overview appeared
Google answered with ordinary results and no Overview at all. This is common, and it varies by query, by location and over time.
This is not a query you are losing. There is nothing to win on it yet. If you count it as an absence you will depress your own numbers with queries that were never available, and you will chase a problem that does not exist. We record the absence separately and exclude it from the denominator.
An Overview appeared and you were absent
An Overview ran, named other brands, and did not mention you. This is the real gap, and it is the number most worth watching.
You were cited but not named
Your page is in the source panel. A competitor is named in the sentence a reader actually reads.
This one surprises people, and it is the most commonly misdiagnosed state in the whole category. The instinct is to write more content. That is usually the wrong response: Google already found your page, already judged it relevant enough to retrieve, and already linked it. What it did not do is repeat your brand as the answer. That is a framing problem rather than a coverage problem, and another three articles on the same topic will not fix it.
You were named
Your brand appears in the answer text. Worth recording where in the answer, because a first mention and a mention in the final clause are not equivalent.
What to actually measure
Four things, per query, per day.
Whether an Overview appeared at all. The base rate. If Overviews appear on three of your twenty tracked queries, your addressable surface is three queries, not twenty, and every percentage should be against that denominator.
Whether your brand was named. The headline. As a share of the checks where an Overview appeared, never as a share of all checks.
Whether your domain was cited. Tracked separately from naming, because the gap between the two is the single most actionable signal available.
Which competitors were named. A share of voice with nobody to share it with is just a number. Being named in two Overviews out of twenty means one thing if nobody else is named either, and something entirely different if one competitor is named in eighteen.
Location changes the answer
AI Overviews vary by where the search comes from. An Overview in Sydney can name different brands to the same query in New York, and the citation set can differ even when the named brands do not.
So location is part of the measurement, not a setting you configure once and forget. A tool that does not tell you where a check was run is not telling you what the number means. We store results per location rather than merging them, which sometimes means more rows and less tidy averages. That is the correct trade.
AI Overviews and AI Mode are different surfaces
Both are Google. They are not the same thing and should never be merged.
AI Overviews sit above ordinary results on a normal search. AI Mode is a separate conversational surface with its own retrieval behaviour. We have measured the two naming different brands for the same question on the same day.
Averaging them into a single Google score would smooth away exactly the disagreement worth acting on. If AI Mode names you and the Overview does not, that gap is a specific, investigable thing. A blended number makes it vanish. The same argument applies across every engine, which is why we report each platform separately rather than producing one score.
How often to check
Daily is the right default for a set you care about.
Weekly is too slow to separate a real movement from ordinary variance, because the variance between individual runs is large enough that two weekly checks can differ substantially with nothing having changed. Hourly is waste: you pay for volume that tells you nothing a daily series does not.
Give it about two weeks before you read a direction into anything. Below that you are looking at noise with a trend line drawn through it.
Common mistakes
Checking once and believing it. The single most common error, and the one most tools encourage.
Counting no-Overview queries as losses. Depresses your numbers with queries that were never winnable and sends you chasing a phantom.
Merging citations and mentions into one score. They have different causes and different fixes. Collapsing them destroys the information.
Ignoring competitors. Your own number in isolation cannot tell you whether two out of twenty is bad.
Rewriting the query. Wording changes the answer. Change a tracked question and you have started a new series, not continued the old one. If you must change it, treat it as new.
Where to start
Pick ten to thirty questions your buyers actually ask, phrased the way a person would type them. Fix the wording. Pick your locations. Run daily, record all four states, and track your competitors alongside yourself.
Then wait two weeks before drawing conclusions. The first day tells you where you stand. The first fortnight tells you which way it is going, and only the second of those is worth acting on.
If you want this run for you, that is what our AI Overview tracker does, and how it works covers the mechanics end to end.