AI Overviews answer questions, not keywords

An AI Overview trigger rate is the share of searches for which Google decides to show a generated summary above the results, and the single strongest predictor of that decision in recent measurement is whether the query is phrased as a question.

That matters for anyone estimating how exposed their pages are to AI Overviews. Most tracking starts from a keyword list exported from an SEO tool, and keyword lists are mostly short noun phrases. If the trigger depends on phrasing, a keyword list measures the part of search where AI Overviews rarely appear and reports a number that looks reassuringly small.

#What the measurement found

Haofei Xu, Umar Iqbal and Jacob M. Montgomery at Washington University in St. Louis ran 55,393 trending queries through Google over a 40-day window from March 13 to April 21, 2026. The queries came from Google's publicly available trending searches, so they reflect what people were actually typing rather than a list assembled for the study.

Across the whole set, 13.7% of queries returned an AI Overview. That average hides the split that matters. The authors classified a query as question-form if its first word was one of 15 interrogatives (who, what, where, when, why, how, which, is, are, was, can, do, does, did, has). Question-form queries triggered AI Overviews 64.7% of the time, against 9.5% for everything else, a gap of 6.8 times.

The type of question moved the rate further. Queries starting with "how" triggered an AI Overview 84.3% of the time and "why" 73.4%, while "who" came in at 47.9% and "did" at 39.8%. Open questions that ask for an explanation get a summary most reliably. Closed lookups, where a single fact answers the query, get one less often.

Length mattered even without a question word. Among non-question queries, the rate climbed from 9.9% for single-word queries to 38.7% for queries of six words or more. A long, specific query behaves more like a question than a short one does, even when it is not phrased as one.

Topic mattered too, in the other direction. Politics (7.5%) and Law and Government (9.6%) fell below the average, which fits Google's own description of the feature as selective.

#Google says the same thing, less precisely

Google does not publish trigger rates, but its guidance points the same way. The AI features documentation says AI Overviews appear only when Google's systems decide they are additive to classic results, and that as a result they often do not trigger. The same page describes AI Mode as built for queries that need further exploration, reasoning or complex comparisons.

A two-word product term rarely needs anything added to ten blue links. A question about how something works usually does. The study puts numbers on that intuition.

#The trigger rate is also moving

Any single snapshot has a short shelf life. Semrush tracked over 10 million keywords through 2025 and found the share triggering AI Overviews went from 6.49% in January to a peak of 24.61% in July, then back to 15.69% in November. Over the same year, informational queries fell from 91.3% of AI Overview triggers in January to 57.1% in October, while navigational queries rose from 0.84% to 10.33%.

These two datasets should not be compared directly. One is a keyword database weighted toward terms SEO tools track, the other is a stream of trending searches. What they share is the lesson that the rate depends heavily on which queries you sample and when. A number without its query set attached is not an exposure estimate.

#Why a keyword list undercounts exposure

Keyword research tools group searches into short head terms because that is how search volume has long been reported. When those lists are reused to measure AI Overview exposure, three distortions follow.

The phrasing is wrong. Your buyers ask "how do I migrate from X to Y" and "why does my build fail on Z". Your tracking list contains "X migration" and "Z build error". The study's numbers suggest the first pair triggers AI Overviews far more often than the second.

The length is wrong. Tracked terms are short because short terms carry the volume. The long, specific queries that trigger summaries more often are spread across many low-volume variants that tools either drop or bundle.

The mix is wrong. Keyword lists overweight commercial and navigational terms because those are the ones teams bid on and rank-track. AI Overviews are expanding into those intents, as the Semrush data shows, but explanatory questions remain where they fire most reliably in the Washington University sample.

The result is a low exposure number built from the queries least likely to show a summary. That number then gets used to argue that AI Overviews are a minor factor for the site.

#What to change in how you track

Build a second query set alongside the keyword list, written in the form people use when they ask. Start with the questions your support team, sales calls and site search already capture. Keep the head terms for classic rank tracking. Use the question set for AI Overview and AI Mode tracking.

Tag every tracked query by form (question or not), leading word, length and topic. Report trigger rates per tag, not as one blended number. A drop in the blended rate can simply mean the mix shifted toward short terms.

Check which of your pages already answer the "how" and "why" questions in your category. Those are the queries where a summary is most likely to sit above your result, so they are where being cited matters most. Google states there are no special requirements to appear in AI Overviews beyond being indexed and eligible for a snippet, so the work is content coverage, not markup. Our notes on long-tail question coverage and query fan-out cover how to plan that coverage.

Record the date with every measurement. The 2025 swing in Semrush's data from under 7% to nearly 25% and back to about 16% happened within one year, so a quarterly comparison without dates and query sets attached can show a trend that is really a change in Google's rollout.

#The practical takeaway

An AI Overview exposure figure is a property of the query set as much as of the site. Measure with the questions your audience actually asks, report rates by query form, and treat any single blended trigger rate as a sampling decision rather than a finding.