AI Overviews and What They Do to Your Traffic

You have probably been sent a headline claiming that AI summaries have destroyed organic traffic, and another claiming the effect is overstated. Both were written with conviction. The honest position, if you look at what has actually been published, is that the direction is fairly clear and the size is not.

That is an unsatisfying answer. It is also the correct one, and knowing where the uncertainty sits is what lets you make a sensible decision rather than a panicked one.

What the independent research found

The most frequently cited independent study comes from the Pew Research Center, published in July 2025. Pew tracked the browsing of a panel of around 900 US adults across roughly 69,000 Google searches carried out in March 2025.

Two findings from it are worth holding onto. When an AI summary appeared, users clicked a traditional search result on 8% of visits, against 15% of visits where no summary appeared. And only 1% of visits involving a summary produced a click on one of the sources cited inside it.

Note what that study is and is not. It is a real behavioural sample rather than a modelled estimate. It is also US-only, drawn from a single month over a year ago, and it measures a moment in a product that has changed repeatedly since. Various SEO platforms and agencies have published their own analyses reporting steeper falls, but those use different samples, different definitions of a click-through and, in several cases, come from companies selling the remedy. Treat direction as better established than magnitude.

What Google says in response

Google’s head of search, Liz Reid, has publicly pushed back, stating that total organic click volume from Google Search to websites has been relatively stable year on year, and arguing that the clicks that do happen are higher quality — people arriving after an AI answer have already filtered themselves and are less likely to bounce straight back.

The company did not publish figures alongside that claim. So you have one side with a named methodology and a limited sample, and the other side with the full dataset and no disclosure. You are entitled to weigh that as you see fit.

Both claims can also be simultaneously true. Aggregate click volume across the whole web can hold steady while the distribution shifts underneath it — towards large platforms, forums, video and Google’s own properties, and away from the middle of the market. “Stable in total” is not a promise about your site.

Where the effect actually concentrates

The pattern that appears consistently across the analyses, whatever their disagreements about size, is that it is uneven.

Queries phrased as questions with a short factual answer are hit hardest. Definitions, conversions, opening hours, “what is”, “how many”. If your traffic was built on that kind of page, a summary answers the reader before they reach you.

Queries where the searcher needs to compare, book, buy, see prices, check availability or trust a specific supplier hold up considerably better. An AI answer can describe how to choose a roofer. It cannot send someone a quote.

That distinction should be doing more work in your content planning than any headline percentage.

What your own data can and cannot tell you

In June 2026 Google added generative AI performance reports to Search Console. This was welcome and it is also limited in a way worth understanding before you build a decision on it.

The report shows impressions from AI Overviews and AI Mode, broken down by page, country, device and date. It does not currently show clicks, click-through rate, average position or queries. AI Overviews and AI Mode are blended into a single surface, so you cannot separate them. Google has said more metrics will follow without committing to a date.

In practice this means you can see whether your pages are being surfaced inside AI answers. You cannot yet see what that surfacing costs or earns you. Anyone presenting you with a precise figure for your AI Overview traffic loss is estimating, and should say so.

Reading the signal you do have

Four things in your own account are more informative than any industry study:

1. Impressions against clicks, by page type. Impressions holding steady while clicks fall is the fingerprint of answers being satisfied above your link. Both falling together is a different problem — usually ranking, not AI.
2. Split your pages into informational and commercial. Run the comparison separately for each. If the fall is concentrated in your explainers and your service pages are level, the story is clear and the fix is a shift of emphasis, not a rescue mission.
3. Enquiries per session. If sessions fell and enquiries did not, you lost the readers who were never going to contact you. That is a survivable trade.
4. Your baseline. Compare year on year rather than month on month, because seasonality will otherwise tell you a story that is not there.

Getting these splits set up properly is ordinary analytics work, and it is worth doing before you change anything.

What is worth doing about it

Publish things a summary cannot replace. First-hand assessments, pricing and process transparency, worked examples, comparisons that require judgement, anything anchored to a place or a specific situation. This is the same advice that applied before AI Overviews existed; the change is that the penalty for generic content is now immediate rather than gradual.

Reduce your dependence on any single channel. If a majority of your enquiries arrive through organic search, that was already a concentration risk. Email lists, repeat custom, referral relationships and paid search all reduce it.

Fix the conversion side. If fewer people arrive, the value of each arrival goes up. A site that converts poorly was always leaking; it now leaks from a smaller pool.

What is not worth doing

Do not delete pages because they lost impressions. Do not rewrite your site to sound more like an encyclopedia entry in the hope of being quoted. And be sceptical of anyone offering a fixed percentage of AI-driven loss recovered, or a guaranteed place in AI answers. The reporting to prove either claim does not currently exist, which makes the promise unverifiable by construction.

If you want a straight read on how your own numbers have moved and which parts of your traffic were genuinely exposed, that is a sensible first conversation to have. Start with a look at your site rather than with a strategy for a problem you may not have.