What triggers one
Google generates an AI Overview when its systems judge that a synthesised answer helps more than a list of links. In practice that clusters around a recognisable set of queries.
It suppresses them, or shows them far less, where the stakes or the intent make a summary a bad idea — clear transactional intent, navigational queries, some sensitive health, finance and news topics, and anything too fresh or too ambiguous to summarise safely.
- Definitional and explanatory — what is X, how does Y work. The hardest hit category by a distance.
- Comparative — X versus Y, best type of Z, where an overview can assemble a table.
- Multi-step and procedural — how to do something with several stages.
- Long-tail conversational — the long, specific phrasing people learned from talking to assistants.
- Rarely — brand names, product purchases, addresses, phone numbers, or anything where a click is obviously the point.
The click damage, using numbers that hold up
The strongest public evidence comes from Pew Research Center's 2025 analysis of real Google users' browsing behaviour. On visits to a results page containing an AI summary, users clicked a traditional result on 8% of visits. Without a summary, 15%. That's close to half the clicks gone, from a methodology that watched actual behaviour rather than modelling it.
Coverage is where the published numbers stop agreeing. Different trackers report AI Overviews on anywhere from roughly a fifth to nearly half of queries, and both can be honest — the answer depends entirely on the keyword set sampled. A tracker weighted toward informational long-tail queries will report high coverage; one weighted toward commercial and brand terms will report low. Country, device and vertical move it again. Treat any single confident percentage as a statement about that sample rather than about the web.
What this means for your own site is measurable, and it doesn't require a study. Take your Search Console Performance data, split queries into informational and commercial buckets, and compare click-through rate at stable positions across the last 18 months. Your own click-through rate decay is the only figure worth planning against.
Which query types are already gone, and which are safe
Averages hide the shape of this. The loss is not spread evenly across your site — it's concentrated in one kind of page, and if that's the kind of page your content strategy is built on, the average understates your problem badly.
| Query type | Exposure | What to do |
|---|---|---|
| What is / definition | Severe — the summary is the whole answer | Keep for authority and citation, stop forecasting traffic from it |
| How to, simple procedures | Severe | Add depth a summary can't carry: video, templates, real examples |
| Comparison and best-of | High | Publish original testing and data an overview must attribute |
| Pricing and cost | Moderate | Real numbers, current, specific to your market |
| Local and near-me | Low | Google Business Profile work still pays |
| Brand and navigational | Low | Largely untouched |
| Transactional and buy-intent | Low | Where the remaining clicks concentrate |
What actually gets cited
The links inside an AI Overview overwhelmingly come from pages already visible for that query — page-one organic results, plus the occasional deeper page with a passage that fits the sub-question the summary is answering.
This is the most reassuring finding available, because it means the work hasn't changed as much as the vocabulary suggests. Ranking well remains the entry requirement. What's added on top is making individual passages liftable — a clean definition under a question-shaped heading, a comparison in a table, a number attributed to a named source. That craft is covered in answer engine optimization.
One structural annoyance: Google folds AI Overview clicks and impressions into your normal Search Console totals with no separate breakout. You can't isolate them, so you infer from click-through rate changes at stable positions. It's inference, and anyone selling you a precise AI Overview report is selling you a model of a number nobody publishes.
Where to move the effort
The correct response isn't to stop publishing. It's to stop publishing the pages whose traffic has been absorbed and to redirect that budget toward pages that still need a click to complete.
In practice, on the retainers we run, that means fewer definitional articles chasing volume and more pricing pages, comparison pages with original data, location pages, and content that carries something a paragraph can't — a calculator, a template, a walkthrough, an opinion worth arguing with.
It also strengthens the case for measuring outcomes rather than sessions. A 30% fall in informational traffic with flat enquiries is not a bad quarter; it's the same business with a smaller audience of people who were never going to buy. That's exactly why our guarantee is written against qualified organic leads frozen on day one rather than against traffic — the traffic metric was already fragile before AI Overviews, and zero-click search has been eroding it for a decade.