Pew Research Center watched 900 American adults run 68,879 Google searches and recorded what they did next. On the searches where an AI summary appeared, people clicked a regular result on 8 percent of visits. Where no summary appeared, they clicked on 15 percent. That is the cleanest measurement anyone has published of what the answer layer does to a click, and it is almost never the number you see quoted.
What you see quoted instead is usually one of two other figures, and they do not mean the same thing. Everyone who sells something online has now been told that AI Overviews traffic loss explains their falling clicks. The claim is broadly true. The arithmetic behind it is a mess, because three separate studies measuring three different things have been stirred into one percentage.
- The three figures most often quoted measure a person's behaviour on a visit, a ranking position's click rate, and a whole sector's referral volume. None of them is interchangeable with another.
- Pew found people clicked a link inside the AI summary on 1 percent of visits, which means being cited in an answer is not a traffic channel and should never be sold as one.
- Ahrefs measured a 34.5 percent fall in position one click through rate against a forecast, not against the raw before and after, and that distinction is where most misquotes begin.
- The publisher figure of 33 percent covers news sites losing traffic for several reasons at once, so applying it to a shop selling products is a category error.
- Google still sends roughly 500 times the referrals ChatGPT does from search alone, so the answer layer on the search engine is the story, not the chatbots.
- The queries most exposed are the ones whose answer was always a single fact, and that traffic was never loyal. What survives is anything a reader cannot finish without doing something.
This piece sets the three studies beside each other with their methods attached. Every figure carries its sample and its date, because all of them will move and the framework around them will not. If you want the mechanics of finding these numbers in your own account, our guide to reading AI search impressions in Search Console covers the measurement. This one covers why the measurement looks the way it does.
The three numbers, and what each one counts
Put the studies in a row and the disagreement stops being mysterious. They disagree because they asked different questions of different populations over different periods.
| Study | Sample and period | Headline figure | What it actually measures |
|---|---|---|---|
| Pew Research Center | 900 US adults, 68,879 searches, March 2025 | 8 percent click rate with a summary against 15 percent without | Human behaviour on a single visit, across all query types |
| Ahrefs | 300,000 keywords, March 2024 against March 2025 | 34.5 percent lower click through rate at position one | The click rate of one ranking position, against a forecast |
| Reuters Institute with Chartbeat | 2,576 news sites, November 2024 to November 2025 | 33 percent fall in organic search referrals globally | Total referral volume to one sector, from all causes |
| Google Search Central | No sample published | Clicks from pages with AI Overviews are higher quality | A qualitative claim about visit depth, not about volume |
Read the right hand column twice. A behavioural rate, a positional rate, a sector total and a quality claim are four different objects. Stacking them into one list of alarming percentages is exactly what most of the statistics roundups on this topic do, and it is why a reader comes away believing traffic fell by 34 percent and 47 percent and 33 percent simultaneously.
Why do the three numbers disagree?
Because two of them are rates and one is a volume, and because only one of them controls for what would have happened anyway. Those two differences account for nearly all of the apparent contradiction.
Start with the Pew Research Center study of browsing behaviour around AI summaries. Its subjects were people, not keywords. It observed 68,879 searches in March 2025, found summaries on 12,593 of them, and compared what the same population did in both situations. The 8 against 15 gap is a per visit probability. It tells you nothing about which keywords, which industries or which ranking positions, and it does not try to.
The Ahrefs analysis of 300,000 keywords and position one click through rate asked a narrower question. It took 150,000 keywords showing AI Overviews and 150,000 informational keywords without them, then compared aggregated Search Console desktop click rates for March 2024 against March 2025. The crucial step is the one most summaries drop. Informational keywords without AI Overviews also fell over that year, from 0.056 to 0.031. So the researchers used that general decline to forecast where the AI Overview keywords should have landed, which was 0.040, and compared the actual 0.026 against the forecast. The 34.5 percent is a gap against a counterfactual, not a raw year on year drop. The raw drop, from 0.073 to 0.026, is far larger, and quoting that instead would blame AI Overviews for a decline that was happening across the board.
The third figure comes from a different universe entirely. The Reuters Institute survey of 280 media leaders across 51 countries, run between 18 November and 20 December 2025, carries Chartbeat data showing organic search referrals to 2,576 news sites down 33 percent globally and 38 percent in the United States between November 2024 and November 2025. That is a volume, for one sector, from every cause at once. News sites are also losing traffic to changed Discover behaviour, to social platforms throttling links, and to a long decline in news avoidance that predates any answer layer. The report itself says the impact is uneven, with lifestyle and utility publishers worst hit.
That breakdown is the reason a single percentage can never answer the question for your site. When your clicks fall, at least four things are moving at once, and only one of them is the answer layer. Your ranking may have slipped. Your query mix may have drifted toward questions that summaries cover heavily. Your whole category may be shrinking. Attributing the entire fall to AI Overviews is as wrong as attributing none of it.
How much of the fall is actually the answer layer?
For an individual site, the honest answer is that you cannot know precisely, but you can bound it. The Ahrefs method is the one to copy, because it is the only one of the three that separates the effect from the background.
The procedure is simple enough to run on your own data. Split your queries into those that trigger an answer and those that do not, then compare how each group moved over the same period. If both fell by similar amounts, your problem is not the answer layer. If the answer group fell substantially further, the gap is your estimate. It will not be exact, and it does not need to be. It needs to be good enough to stop you rebuilding a content strategy around a cause that was not the cause.
One caution on impressions. A common pattern is impressions holding steady while clicks fall, and it is often read as proof that the answer layer is eating the click. It is consistent with that, but it is also consistent with your pages ranking in positions that never earned clicks anyway. Impressions are a weak instrument here, which is part of why the exposure of your query types matters more than any single headline number.
Is being cited inside an AI answer worth anything?
Not in clicks. Pew measured link clicks inside the summary itself at 1 percent of visits, and that number deserves far more attention than it gets.
An entire consulting category has grown up around becoming the cited source in AI answers. The pitch is reasonable on its face. If the answer is going to be given without a click, better to be the source it is given from. What the 1 percent figure says is that the citation is not a traffic channel, and anyone selling it as one is selling something the data does not support.
That does not make citation worthless. It makes it a brand placement, and it should be justified the way brand placements are justified, on presence and influence rather than on sessions. If your business depends on being the name a buyer already trusts by the time they reach a comparison, appearing in the answer is worth something real. If your model needs the visit to make money, it is not a substitute for the visit. Keeping those two cases apart is the single most expensive distinction in this field right now, and the same logic applies to the broader shift in how customers now find a business.
Pew also found that 26 percent of visits with a summary ended the browsing session entirely, against 16 percent without. The summary does not only redirect the click. On a meaningful share of visits it ends the journey, which is a harder thing to win back than a click that went to a competitor.
What Google says, and where it is not wrong
Google's own documentation makes a claim worth taking seriously rather than dismissing. Its guidance on AI features and your website states that when people click from result pages carrying AI Overviews, those clicks are higher quality, meaning the visitor is more likely to spend longer on the site. It also says people are reaching a greater diversity of websites, and that no special optimisation is needed to appear.
No sample is published for the quality claim, so it cannot be checked. But it is plausible in a way that matters for anyone selling something. If a summary answers the idle question, the person who clicks through anyway has a reason to. Fewer visits with a higher intent is not the same disaster as fewer visits at the same intent, and a shop measuring revenue rather than sessions may find the damage smaller than the traffic chart suggests. Check your conversion rate over the same period before you panic about the click count.
The documentation is also the place to learn what controls you actually have, which is fewer than people assume. The preview controls are the old ones: nosnippet, data nosnippet and max snippet, plus noindex if you want out altogether. There is no dedicated switch for the answer layer. Google describes a query fan out technique behind its responses, which is why the links shown often do not match the classic ranking for the typed query.
Which of your queries are most exposed?
The ones whose answer is a fact. Exposure tracks almost perfectly with how completely a paragraph can satisfy the person who typed the query.
Pew's data supports this directly. Summaries appeared on 60 percent of searches phrased as questions and on 53 percent of searches of ten words or more, against 8 percent of one and two word searches. Long, natural language questions are where the answer layer lives. Short navigational and brand terms are largely untouched.
| Query type | Answer layer coverage | Click realistically retrievable | What your page must offer instead |
|---|---|---|---|
| Definitional and what is | Very heavy | Little | Original measurement, or a view the summary cannot assemble |
| How to and procedural | Heavy | Some | A tool, a template or a calculator the reader operates |
| Comparison and buying | Moderate | Most | Primary testing, real prices, stock the reader can buy now |
| Troubleshooting a specific error | Moderate and rising | Some | Depth beyond the common case, and a community that answers back |
| Local and transactional | Light | Nearly all | Accurate hours, inventory and the ability to complete the purchase |
| Queries needing fresh data | Light but growing | Most | Recency the model cannot have, updated on a visible schedule |
Run your own pages against that table before you change anything. Most shops discover that the pages losing clicks are their thinnest explanatory posts, and that product, category and local pages are close to untouched. That is a far less frightening diagnosis than a 33 percent headline, and it points at different work. If a large share of your traffic is local, the more useful read is our piece on how a shop stays visible in local AI search.
What an answer cannot finish for the reader
Every durable recommendation in this area reduces to one property. The page has to require something from the reader that a paragraph cannot supply on their behalf.
Original measurement means numbers you produced. A model can summarise a study it has read; it cannot summarise the one you ran last month on your own order data. A calculator means the reader supplies their own situation and gets an answer specific to it, which is structurally impossible to pre-answer. Primary reporting means you went and found out. A community means the value is other people, refreshed. And anything requiring the reader's own inputs is safe for the same reason a calculator is.
Notice what is absent from that list. There is no formatting trick, no schema type and no word count. The advice circulating about restructuring pages into question and answer blocks to be more extractable is not wrong exactly, but it optimises for being quoted, and being quoted is the 1 percent channel. Writing something that cannot be quoted in full is the stronger position. This is the same conclusion our analysis of what actually ranks when the content is AI generated reached from the other direction.
What to do, in what order
First, measure your own gap rather than importing anyone else's. Split queries by whether they trigger an answer, compare the movement, and get a number that belongs to your site.
Second, check revenue alongside sessions. If conversion rate rose while sessions fell, some of what you lost was traffic that was never going to buy, and the correct response is smaller than the traffic chart implies.
Third, sort your pages by the exposure table above and stop reinvesting in the top rows. A definitional post that used to earn clicks is not coming back through better writing. Move that effort down the table.
Fourth, build at least one thing on your site that a reader has to operate. For a shop this is usually simpler than it sounds: a size or fit finder, a delivery cost estimator, a stock checker, a comparison that uses the visitor's own constraints. Owning the code and the data for that kind of tool is much easier when you are not renting your storefront, which is part of the argument for running on an ecommerce site whose pages and data you control.
Fifth, keep a channel that does not route through a search engine. Email remains the obvious one. The publishers in the Reuters Institute survey expect search traffic to fall by more than 40 percent over three years, with some expecting worse than 75 percent. Those are forecasts from people with an interest in the answer, so treat them as a direction rather than a measurement. The direction has been consistent for two years.
Can you opt out of the answer layer without losing the search traffic?
Not cleanly, and that is the trap. The controls Google documents operate on snippets in general, not on AI features specifically, so turning them on withdraws you from the ordinary result at the same time.
This is the part of the guidance that surprises people. There is no setting that says show my page in the classic result but keep it out of the summary. The available instruments are nosnippet, which suppresses the preview text; max snippet, which caps its length; data nosnippet, which marks specific passages; and noindex, which removes the page from search altogether. All of them apply through the standard robots directives for Googlebot.
Setting nosnippet on a page that still ranks means the ordinary listing loses its description too, which in practice costs more clicks than the summary was taking. A max snippet limit set aggressively has the same shape of problem in a milder form. The honest summary is that withdrawal is available and rarely worth it, which is a different conclusion from the one reached about training crawlers, where the decision genuinely can go either way depending on the business. We worked through that separate decision in our piece on whether a shop should block or allow AI crawlers, and the two questions should not be answered together.
There is one case where a partial control earns its keep. If a single page carries the answer to a question that drives real commercial traffic, and that answer sits in one paragraph, marking just that paragraph with data nosnippet keeps the rest of the page previewable while making the key passage harder to lift. It is narrow, fiddly and worth testing on one page before anywhere else. Measure it against clicks rather than against whether the summary still appears, because the summary will usually still appear, assembled from somewhere else.
The counterweight worth remembering
Amid all of this, one number from the Reuters Institute report keeps the panic proportionate. Google still delivers around 500 times as many referrals as ChatGPT from search alone, and around 1,300 times as many once Discover is included.
The traffic is not moving to chatbots. It is being absorbed by the answer layer sitting on top of the search engine that already had it. That matters because the two stories imply opposite responses. If readers were migrating to chat assistants, the job would be to get visible inside those assistants. Since they are mostly still on Google and simply clicking less, the job is to be the result worth clicking on a page that is trying to make clicking unnecessary.
It also means the measurement you already have is still the right measurement. Search Console continues to describe most of your discovery. It is not yet time to rebuild your analytics around a channel that delivers a fraction of a percent of what the old one does.
The framing that will outlast every figure here
All of the numbers in this article will be stale within a year. The structure underneath them will not, and it comes down to a single observation: the queries the answer layer takes are the queries whose answer was always a fact that anyone could state.
That traffic was never loyal. A person who typed a question and needed one sentence was never going to become a customer at any meaningful rate, and the click they used to give you was a toll they paid to get the sentence. The answer layer stopped collecting that toll on your behalf. Losing it feels like a loss because it was counted, not because it was worth much.
What remains is the set of questions whose answers cannot be finished without the reader doing something: entering their measurements, seeing your prices, joining the people who already bought, trusting that you tested the thing yourself. Building for those is a content strategy rather than a search tactic, which is why it survives the next ranking change and the one after that. The sites that come through this period in good shape will mostly be the ones that were never really in the business of answering questions in the first place.