AI Overviews and what happened to your click-through rate

That last sentence is the whole method, and it costs nothing.

I know the number you want. A percentage you can put in a board deck. I do not have one that is honestly sourced and I will not invent one, so what follows is a way to find your own. That is better anyway, because your query mix is not anyone's average.

What is an AI Overview?

It is the generated summary that appears above the ordinary results for some queries, assembled from pages Google's systems retrieved and citing several of them.

Two facts about it matter operationally. First, it is built on core Search ranking. Google states its generative features retrieve from the Search index and use query fan-out, so the pages eligible to appear are the pages already competing in Search. Second, a page needs only to be indexed and eligible for a snippet to be technically eligible for inclusion. There are no additional technical requirements.

So an AI Overview is not a separate product you optimize separately. It is a new presentation layer over the ranking you already have.

The change is not to how you earn visibility. It is to what visibility is worth. A result that used to be a click may now be a sentence in a summary with your name on it.

How often do they even appear?

Less often than the discourse implies, and the variation by query type is enormous.

Ahrefs analysed 146,122,391 desktop SERPs sampled in September 2025 and found AI Overviews on 20.5% of them, just under 30 million results. The composition of that 20.5% is the useful part.

• 99.9% of AIO-triggering keywords were informational. Commercial keywords 4.3%. Transactional 2.1%. Navigational 0.9%.
• Non-branded queries triggered them 1.9 times as often as branded ones, 24.9% against 13.1%.
• Query length ran almost linearly. One-word queries 9.5%. Seven words or more 46.4%. Question-phrased queries 57.9%.
• By subject, Science 43.6%, Health 43.0% and medical topics 44.1% at the top. Shopping 3.2% and Real Estate 5.8% at the bottom.
• Local searches: 7.9%.

Ahrefs sells SEO software, so this is vendor research, though the method and sample are disclosed and the population is large.

Read that table against your own query list before you decide how alarmed to be. Short transactional phrases carry structurally small exposure. Long question-shaped informational phrases carry structurally large exposure. The averages between those two cases describe nobody.

Why can't anyone tell you the real impact?

Because the measurement problem is genuinely hard, and the published numbers do not agree.

Ahrefs compared 150,000 keywords showing AI Overviews against 150,000 matched non-AIO informational keywords, using aggregated Search Console click-through data from March 2024 against March 2025, and reported a 34.5% decrease in clicks. A later cut of the same work revised that to 58%. SparkToro, working with a Similarweb panel, put the drop at roughly 60% when an Overview is present. Those are three figures from two vendors and they span a factor of nearly two.

Publisher-level traffic tells a similar story with a similar spread. Comparing US organic Google traffic in June 2025 against June 2026: USA Today roughly 50% down, CNN 25% down, Politico 23% down, Business Insider more than 85% down. Same period, same surface, wildly different outcomes, because those four sites do not publish the same kind of page.

The wider context is a real trend measured badly. SparkToro's panel put 68.01% of US Google searches ending without a click in 2026, against 60.45% in 2024, and reported that per 1,000 searches only 276 clicks now reach the open web, down from 374. That is a 26% reduction in two years. Google has publicly disputed SparkToro's methodology, and a separate 2026 measurement using a different definition of a click put zero-click at 27.6% instead. Most of that gap is definitional, specifically whether a click to another Google property counts as a click. Use the direction. Do not quote the decimal.

Overview presence is also a moving target. The same search can produce one for one person and not for another, so any study of this is sampling a population that changes while it is being counted.

The honest position: click-through rates for informational queries have clearly come under pressure, the size of the effect varies enormously by query type and publisher, and any single headline number should be treated as an advertisement.

How do you measure it on your own site?

What does Google itself report, and what does it withhold?

Since June 2026 Google publishes Search Generative AI performance reports in Search Console, giving what it calls dedicated views of impressions within generative AI features such as AI Overviews and AI Mode.

Read the dimension list closely, because the omissions are the story. Available: impressions, pages, countries, devices and dates. Not available: clicks, click-through rate, position, and the user's prompt. AI Mode and AI Overviews are not separated from each other. And the data is a view onto the existing performance report rather than a separate dataset, so it is already counted in your overall numbers and must not be added to them.

Two practical cautions. Secondary coverage puts the data start at 18 May 2026 with no historical backfill, so a year-over-year comparison inside that report is not available yet. And Search Engine Land has already reported a data bug in it, so sanity-check anything surprising before it reaches a client.

The larger gap is not in the report at all. Clicks from AI Overviews and AI Mode arrive in your analytics as ordinary organic traffic. Google confirms they are included in overall search traffic within the Web search type. No referral parameter, no separate channel, no way to isolate them. The biggest AI surface in local search is fundamentally unattributable, and no tool has solved that because there is nothing in the data to solve it with.

Can you just opt out of AI Overviews?

Not without paying for it in ordinary Search, and the tactic most often recommended does not do what people think.

The claim in circulation is that adding `Google-Extended` to robots.txt removes you from AI Overviews. It does not. Google describes `Google-Extended` as a robots.txt token, not a crawler with its own user agent, and its scope is whether already-indexed pages may be used for training and grounding in Gemini. There is no Gemini crawler to block. Google grounds those features on content fetched by ordinary Googlebot.

What Google says does limit AI Overviews and AI Mode is `nosnippet`, `data-nosnippet`, `max-snippet` or `noindex`. Those are the same blunt snippet controls that govern classic Search. Use them and you lose your ordinary result snippet at the same time, which for most local businesses is a worse trade than the one they were trying to avoid.

There is no selective opt-out. There is only a snippet, present or absent, on both surfaces at once.

Which queries lose the most, and which hold?

Informational queries lose. The Ahrefs trigger data makes that structural rather than anecdotal: 99.9% of AIO-triggering keywords were informational. Anything answerable in three sentences was always a fragile click, and now it is often not a click at all.

Queries with intent attached hold up better, and the trigger data agrees again. Commercial keywords triggered an Overview 4.3% of the time and transactional keywords 2.1%. Someone comparing two providers, checking whether you are open, looking for a price, or ready to book still needs to arrive somewhere. A summary cannot take a booking.

Local queries behave differently again, and mostly better, sitting at 7.9%. Local intent works differently in AI search for that reason, and it is why panic imported from publisher discourse tends to be miscalibrated for a local operator.

Service area businesses should be especially careful reading national commentary, since their baseline visibility already works differently from a storefront's, and ranking without a storefront has its own rules.

What to do this week

1. Pull twelve months of Search Console query data and find every query holding position while losing clicks. That list is your actual exposure. 2. Filter it to informational, question-shaped, non-branded queries, which is the validated proxy for AIO exposure. 3. Run the five-step arithmetic above, including step five. Count how many of those clicks were ever going to produce revenue. 4. Take your three biggest informational losers and rewrite the pages to answer in the first three sentences, then continue into something a summary cannot replace. A price, a process, a comparison, a booking. 5. Check whether your query mix skews short and transactional or long and question-shaped. That single split predicts your exposure better than any published average. 6. Since fewer visitors means each one matters more, tighten what happens after the click. Asking for a review at the right point in the customer journey is part of that, because reviews feed the same systems taking the clicks.

You will also see a second effect while you are in there. A single question now triggers a spread of related searches behind the scenes, and query fan-out changes which of your pages get pulled in. Some pages start receiving impressions for phrasings you never targeted.

Want to know which of your queries are actually bleeding?

Export your Search Console queries for the last twelve months and send the file to eric@seod.com. I will find the queries holding position while losing clicks, rank them by lost clicks, and tell you which ones were ever going to make you money. No charge, no pitch, and you get the analysis in plain language.

I offer this because the export takes you two minutes and the analysis takes me fifteen, and almost nobody does it before deciding how worried to be.

Otherwise, keep reading the AI search library.