What sets one off
Google generates an AI Overview when its systems judge a synthesised paragraph more useful than ten links. That judgement is fairly predictable once you've watched a few hundred SERPs.
It also withholds them. Clear buying intent, navigational queries, some health, finance and legal topics, breaking news, and anything too ambiguous to summarise safely tend to get the classic results page.
- Definitional queries — what is X, what does Y mean. Worst hit by a distance, because the summary genuinely is the whole answer.
- Procedural queries — how to do something in a few steps, where a numbered list finishes the job.
- Comparative queries — X versus Y, types of Z, where an Overview can assemble a table from three sources.
- Long conversational phrasing — the sentence-shaped queries people picked up from talking to assistants.
- Rarely — brand names, product purchases, phone numbers, addresses, or anything where clicking through is obviously the point.
The click damage, using the one number that holds up
The strongest public evidence is 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%. Roughly half the clicks, measured by watching what people did rather than modelling what they might do.
Coverage — how often an Overview appears at all — is where published numbers stop agreeing, and the disagreement is honest rather than sloppy. Trackers report anywhere from a fifth of queries to nearly half, and both can be true, because it depends entirely on which keyword set was sampled, in which country, on which device. Treat any confident percentage as a fact about somebody's keyword list, not about the web. The number that matters is yours, and you already have the data to calculate it.
The 2026 stabilisation, and why it isn't a recovery
Through 2024 and into 2025, publishers watched informational traffic fall off a cliff and assumed it would keep falling. It largely didn't. The common shape in the Search Console accounts we look at is a sharp drop, then a flat line, and sometimes a partial climb back — the thing people now call a rebound.
Three ordinary mechanics explain most of it, and none of them is Google reversing course. The drop was a one-time reset, not a slope: once the queries an Overview can fully answer were absorbed, there was nothing left to absorb, and what remains are queries a paragraph can't finish. The sites that visibly recovered changed their mix, swapping definitional articles for pricing, comparison and product pages — a different set of pages doing the work, which looks identical on a traffic chart. And impressions kept growing while clicks fell, so some of the apparent 2026 improvement is just the comparison period settling.
Nobody has published a methodologically sound figure for the size of the rebound. If an agency quotes you one to a decimal place, ask where it came from.
Work out your own exposure before you panic
Averages hide the shape of this. The loss isn't spread across your site — it's concentrated in one kind of page. If that's what your content plan is built on, the industry average badly understates your problem. If it isn't, you may have lost almost nothing.
Open Search Console, set the range to the last 16 months, export the Queries report, and sort into these buckets. The middle column is the filter that finds each one.
- Add up the clicks in the top three buckets as a share of your total. That percentage is your real exposure.
- Under 20%, this is a rounding error and you should stop reading about it. Over 60%, your content strategy needs a rewrite, not a tweak.
| Bucket | How to find it in Search Console | Exposure |
|---|---|---|
| Definitional | Query contains: what is, meaning, definition, explained | Severe — assume the click is gone |
| Procedural | Query contains: how to, steps, guide, tutorial | Severe on simple tasks, moderate on complex ones |
| Comparative | Query contains: vs, versus, difference, best, types of | High, unless you publish original testing |
| Commercial / pricing | Query contains: cost, price, pricing, charges, quote | Moderate — Google is cautious with money queries |
| Local | Query contains: near me, or a city name | Low — the map pack still owns this |
| Brand and navigational | Query contains your brand name | Low — largely untouched |
| Transactional | Query contains: buy, order, book, demo, download | Low — this is where the remaining clicks pooled |
How Google picks the sources it cites
This is the reassuring part, and it gets buried under the panic. The links inside an AI Overview come overwhelmingly from pages already visible for that query — page-one organic results, plus the occasional deeper page carrying a passage that answers a sub-question the summary needed.
So the work hasn't changed as much as the vocabulary has. Ranking well is still the entry requirement. What's added is making passages liftable: a clean 40-to-60-word answer under a question-shaped heading, comparisons in a real table, numbers attributed to a named source rather than to "studies". That craft picked up the name answer engine optimization, but it's mostly good technical writing with a citation habit.
There's no mechanism to bid, pay or apply for citation. Guaranteed AI Overview placement is the guaranteed-rankings trick in newer packaging.
What to change on informational pages
The wrong response is to stop publishing. The right one is to stop publishing pages whose only job was collecting a click that no longer arrives, and move that budget to pages that still need a click to complete. On the retainers we run, that has meant five changes:
- Stop forecasting traffic from definitional pages. Keep the good ones — they earn citations and build topical authority — but budget them as brand assets, not traffic.
- Add the thing a paragraph can't carry. A calculator, a template, a real price, a walkthrough with screenshots. Summaries restate facts; they can't hand over a spreadsheet.
- Front-load a citable answer. One tight paragraph under the heading, written to be lifted. If you're going to be quoted, be quoted accurately and by name.
- Rebalance the calendar toward buying intent. Pricing, comparison with original data, location and product pages. That's where the clicks pooled.
- Report leads, not sessions. A 30% fall in informational traffic with flat enquiries isn't a bad quarter — it's the same business minus an audience that was never going to buy.
How to track AI Overview presence for your own queries
Google folds Overview clicks and impressions into the standard Search Console Performance report with no breakout. You can't isolate them. Anyone showing you a precise AI Overview traffic report is showing you a model of a number Google doesn't publish — and honest reporting says so out loud.
You can triangulate, which is enough to make decisions:
- Track click-through rate at stable positions. If a query sat at position 3 all year and its CTR halved, something above it is eating the click.
- Use a rank tracker that flags SERP features. Most major trackers now record whether an Overview appeared for a keyword on a given day, and whether you were cited. That flag is the closest thing to direct measurement available.
- Spot-check manually, logged out. Presence varies by country, device and account. Your own logged-in results don't represent your market.
- Watch impressions and clicks separately, by bucket. Rising impressions with falling clicks is the Overview signature. Falling impressions is a ranking problem and a completely different fix.