Your Search Console graph shows an AI Overviews traffic drop that makes no sense: impressions flat, average position barely moved, clicks falling off a shelf. Nothing broke and nothing got penalised. Your content is still ranking exactly where it always did, and the click just stopped arriving.
I covered this briefly in my last post as one of three causes of a traffic drop. It deserves its own piece, because it’s now the most common one I see, and because the published research on it’s a mess.
Six credible sources have measured this. They report six different numbers, and two of them point in opposite directions. If you have spent an evening reading them trying to work out what happened to your site, you already know the problem.
So let’s put all six on one table, work out why they disagree, and find the one that’s actually about your site.
Six studies, six different numbers
| Source | What it found | What it actually measured |
|---|---|---|
| Pew Research Center 900 US adults, March 2025 | Clicks on a normal search result fell from 15% to 8% when an AI summary appeared. Only 1% clicked a link inside the summary itself | Real browsing behaviour across 68,879 searches |
| Agarwal & Sen 1,065 US users, Jan–Feb 2026 | AI Overviews cut outbound clicks by about 38%. Zero-click searches rose 33% | A randomised field experiment: cause, not correlation |
| AI Mode study 1,100 US users, March 2026 | Forcing every search into AI Mode cut the share reaching a website by 18.8 percentage points | A forced condition, not normal user behaviour |
| SparkToro / Similarweb US panel, early 2026 | 68.01% of Google searches end without a click, up from 60.45% in 2024 | Every Google search, AI Overview or not |
| Chartbeat Publishers, 2026 | Google Search referrals to publishers down 40% year on year | News and media publishers only |
| Semrush 200,000+ keywords | Zero-click rates fell, from 38.1% to 36.2%, after AI Overviews appeared | A keyword corpus, before and after |
Read that table as a business owner and it tells you nothing. One source says clicks halved. Another says zero-click got slightly better. A third says two thirds of all searches never send a click, which sounds catastrophic until you learn that was already true before AI Overviews existed.
Why the numbers disagree
They’re not really contradicting each other. They’re answering different questions, and the headlines flatten that away.
They measure different things. SparkToro counts every Google search, including navigational ones where nobody was ever going to click a publisher. Pew measures what a person did on a results page. Agarwal and Sen measure what changes when you take the AI Overview away. Semrush measures a keyword corpus, so its figure moves with which keywords are in the set.
The people are different too. Chartbeat’s 40% is news publishers, who are the most exposed group on the internet because their traffic is almost entirely informational. If you sell a service in Mohali, that number was never about you.
And they ran at different times. Pew’s browsing data is from March 2025. Agarwal and Sen ran in early 2026. AI Overviews expanded enormously in between. A 2025 figure and a 2026 figure aren’t comparable, and quoting them side by side is how people end up with a number that’s either far too calm or far too alarming.
The useful question isn’t “how much traffic do AI Overviews take?” It’s “how much do they take from pages like mine?” Those have very different answers, and the second one you can actually measure.
The study that answers the question properly
One of these six is built differently from the rest, and it’s the one worth reading.
Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon University ran a randomised field experiment, published as Google AI Overviews and Publisher Traffic: Evidence from a Field Experiment. Everything else on the list observes what happened. This one intervenes and measures the difference.
The design: 1,065 US participants ran their normal searches through a browser extension between 7 January and 10 February 2026. Some saw Google as usual. For others, the AI Overview was removed from the page. Same people, same queries, one variable changed.
The result: external clicks per search rose from 0.38 to 0.61 when the AI Overview was taken away. That’s roughly a 38% reduction in clicks to websites when an overview is present, and a 33% rise in searches that end without anyone visiting a site. AI Overviews appeared on about 42% of the queries observed.
This matters because it’s causal. Correlation studies can’t separate the AI Overview from everything else Google changed that year. This design can.
The honest caveats, which the authors state themselves: it’s US-only, desktop Chrome, and the paper doesn’t measure revenue impact, only clicks. A second experiment in March 2026 forced 1,100 users into AI Mode and found an 18.8 percentage point drop in searches that reached a website. But forcing people into a mode they had barely used (it was 0.6% of their searches beforehand) tells you about the interface, not about what people choose.
Take the 38% as your working number, then adjust it for your own site, which is what the rest of this post is about.
Which of your pages are actually exposed
This is where the averages stop being useful and your own site starts to matter.
AI Overviews are overwhelmingly triggered by informational searches. Semrush’s analysis of over 200,000 keywords put it at roughly 88% informational, with commercial and transactional queries making up around a tenth between them. That single fact decides your exposure.
| Page type | Exposure | What it means for you |
|---|---|---|
| “What is” and “how does” articles | Very high | An AI summary answers this completely. The click was never coming back |
| Definition and glossary pages | Very high | The overview is your page now. Stop measuring these on traffic |
| Symptom, condition and problem content | High | Heavily summarised, and in health topics Google is cautious about what it will state |
| Comparison and “best X” content | Moderate | Summarised, but buyers still click through to verify before spending |
| Product and service pages | Low | Someone ready to buy or enquire needs your site. An overview can’t complete the transaction |
| Location pages and local searches | Low | The map pack still owns this, and proximity beats summarisation |
| Branded searches | Very low | They already know who you’re. They are looking for you specifically |
Look at that table and then at your own site. If your organic traffic was mostly top-of-funnel blog content, you have taken a real hit and no amount of rewriting will undo it. If your traffic comes from service pages, product pages and local searches, your exposure is genuinely small, and a drop probably has a different cause.
This is also why the 40% publisher figure frightens people who should not be frightened. News publishers are almost entirely in the top three rows. Most businesses aren’t.
How exposed is your industry
Exposure isn’t evenly spread. Two businesses can lose the same share of impressions and have completely different problems, because what the searcher wanted next is different.
| Industry | Typical exposure | Why |
|---|---|---|
| News and media | Severe | Almost all traffic is informational. This is where the 40% figure comes from |
| Health and wellness information | High | Symptom and condition queries summarise well, though Google stays cautious on medical claims |
| Finance and insurance explainers | High | Definition-heavy content, and the money pages sit behind a long research journey |
| SaaS and B2B | Moderate | Top-of-funnel content is exposed; comparison and pricing pages still convert clicks |
| eCommerce | Low to moderate | Category and product searches still need your site. Buying guides take the hit |
| Ayurveda and supplements | Mixed | Ingredient and condition content is exposed, product and buying searches aren’t. More on ayurveda and healthcare SEO |
| Local services | Low | Proximity decides the result. The map pack isn’t being summarised away |
| Legal services | Low to moderate | “What is” legal content is exposed; “find a lawyer in my city” isn’t |
Find your row, then check it against your own Search Console rather than trusting it. An industry average is a starting hypothesis, not a finding.
Measure your own exposure in fifteen minutes
At this point you can stop reading studies about other people’s websites. Your own Search Console has a better answer.
- Open Performance → Search results and switch on all four metrics: clicks, impressions, CTR and average position. Set the date range to the last 16 months so you can compare seasons rather than weeks.
- Look at CTR, not clicks. This is the whole diagnosis. If impressions held or grew while CTR fell, you’re being summarised, not outranked. If average position also fell, you have a ranking problem and this post isn’t about you.
- Go to the Queries tab and sort by impressions. Work down the top thirty and ask of each one: could a paragraph answer this completely? Those are your exposed queries. Count them as a share of your impressions and you have your exposure number.
- Go to the Pages tab and compare this quarter against the same quarter last year. Informational pages falling while commercial pages hold is the signature. Everything falling together is something else.
Google’s own guide to debugging Search traffic drops is worth twenty minutes before you draw conclusions, and I walked through the full diagnostic process in my post on diagnosing a traffic drop.
| What you see | What it’s |
|---|---|
| Impressions flat or up · CTR down · Position flat | Being summarised. This post applies |
| Impressions down · CTR flat · Position worse | A ranking loss. Different problem, different fix |
| Informational pages down, commercial pages flat | Classic AI Overview exposure pattern |
| Everything down together, sharply, on one date | An algorithmic update. Check the dates |
Query fan-out: why one keyword is now several
There’s a mechanism underneath all of this that changes how keyword research works, and most coverage skips it.
AI Overviews and AI Mode don’t simply answer the query you typed. Google’s own documentation describes a technique it calls query fan-out: the system issues multiple related searches behind the scenes and assembles an answer from all of them. Google says this surfaces a wider and more diverse set of links than a single query would.
You can be cited for a search you never targeted. If your page answers one sub-question well, it can be pulled into an overview for a broader query you hold no ranking for at all. Depth across a subject now beats optimising one page for one phrase.
Keyword-level tracking gets less honest. Your rank tracker reports a position for the query you typed. It can’t see the fan-out behind it, so it can’t tell you why a page you rank third for stopped receiving clicks. That’s why Search Console impressions and CTR have become more useful than any rank tracker for this particular problem.
The practical response is to cover a subject properly rather than a keyword narrowly , which is what earned rankings before any of this existed anyway.
What actually works now
Three things, in the order I would do them.
1. Become the source the overview cites
Citations aren’t clicks. But a citation puts your brand name in front of someone at the exact moment they’re learning about your category, and roughly a quarter of AI Overviews now carry external links where previously almost none did.
Google is unusually direct about how this works. Its documentation on AI features and your website states there are no additional requirements and no special optimisations for appearing in AI Overviews or AI Mode. The same crawlability, structure and usefulness that earns a ranking earns a citation.
What that means in practice:
- Answer the question in the first hundred words. A model lifting a clean, self-contained passage needs one to exist
- Use headings that match how people ask, then answer directly underneath rather than building up to it
- Give it something only you have. Original data, a real number from your own work, a documented process. Summaries of summaries don’t get cited
- Make claims checkable. Cite sources, name your authors, date your content honestly
- Get the structured data right so your claims are machine-readable
2. Win the searches an AI answer can’t finish
This is the higher-value move, and most people skip it because it’s harder than rewriting a blog post.
An AI Overview can tell someone what an SEO audit involves. It can’t tell them whether their site needs one. It can explain what ayurvedic formulations do. It can’t ship the product. Any search where the answer depends on the person’s specific situation, or ends in a transaction, still sends the click.
So shift the content budget. Less “what is X”. More comparison content, more decision content, more pages built around the searches people make when they have already decided to spend money.
3. Convert the clicks that still arrive
If you get 38% fewer visitors, the arithmetic leaves you one lever: each remaining visitor has to be worth more.
That means the page they land on does its job. A clear offer, an obvious next step, a form that isn’t buried, an email capture worth the address. Most sites I audit have spent years optimising for traffic and about twenty minutes on what happens after it arrives. That ratio has to change.
What recovery looks like, and how long it takes
Traffic lost to an AI Overview doesn’t come back the way traffic lost to an algorithm update does. There’s no reassessment to wait for. The query has changed shape permanently, and the click you used to get is now being satisfied on the results page.
| What you’re doing | Realistic outcome | Timeline |
|---|---|---|
| Restructuring pages so they can be cited | Citations, brand exposure, a modest click recovery | Four to twelve weeks |
| Building comparison and decision content | New clicks on searches that still click | Three to six months |
| Improving conversion on the traffic you keep | More revenue from fewer visitors | Immediate, and the highest-leverage move |
| Waiting for informational traffic to return | Nothing. It isn’t coming back | — |
Clients push back on that last row harder than on anything else here. A definition page that lost most of its clicks to a summary hasn’t been penalised and can’t be edited back to health. Its job has changed from earning traffic to earning a citation, and it should either be measured that way or retired.
The businesses adapting well have stopped treating this as a drop to recover and started treating it as a permanent change in what organic traffic is worth, then adjusted what they spend to get it.
What doesn’t work
- Chasing the ranking harder. You already have the ranking. Position one under an AI Overview still loses the click. There’s nothing left to win there
- Rewriting content that’s still ranking. The most expensive mistake available. You spend a quarter and risk a position that was never the problem
- Blocking Google’s AI crawlers. You lose the citation and, depending on what you block, the ranking with it. Publishers who tried this mostly quietly stopped
- Publishing more of the same. If informational content is the exposed category, more informational content is more exposure, not less
- Assuming the 40% figure is about you. It’s about news publishers. Measure your own site
One more thing, happening right now
While you’re looking at your graph, note the date. Google began rolling out its fourth spam update of 2026 on 24 September, and unlike the earlier ones that finished in two or three days, this one has a rollout window of up to two weeks. It’s global and applies to all languages.
Which means any drop you’re seeing this week has two candidate explanations, not one, and they need opposite responses. Check the Google Search Status Dashboard for the rollout state, and work through the diagnostic in my post on telling those causes apart before you touch anything.
A slow rollout also means slow, partial movement. Don’t draw conclusions from a Tuesday.
Frequently asked questions
Do AI Overviews really cut traffic by 40%?
For news publishers, roughly yes. Chartbeat put publisher referrals down 40% year on year. For everyone else, no. The best causal estimate is about a 38% reduction in clicks on searches where an AI Overview appears, and overviews appeared on around 42% of queries in that study.
Your own figure depends entirely on how much of your traffic is informational. Measure it rather than adopting someone else’s number.
My rankings are the same but traffic is down. Is it AI Overviews?
Probably, if impressions held steady while CTR fell and average position didn’t move. That combination has very few other explanations. If average position also slipped, you’re looking at a ranking loss instead, which needs a completely different response.
How do I get cited in AI Overviews?
Google states there are no special optimisations. Answer the question clearly near the top of the page, structure content so a clean passage can be lifted, publish something only you have, and make claims verifiable with named authors and real sources. The work that earns rankings is the work that earns citations.
Should I stop writing blog posts?
No, but stop writing the ones a paragraph can replace. Definition posts and basic explainers have lost their commercial logic. Comparison content, decision content and anything drawing on your own data or experience still earns clicks, because a summary can’t substitute for it.
Does being cited in an AI Overview help if nobody clicks?
It helps less than a click and more than nothing. Pew found only 1% of users click a link inside an AI summary. What you get is brand exposure at the moment someone is forming a shortlist. Worth having, not worth building a strategy on.
Is zero-click search new?
No, and this is the most misread statistic in the field. Around 45% of searches ended without a click back in 2016, long before any AI summary existed. The figure reached 68% in early 2026. AI Overviews accelerated a decade-old trend rather than starting one.
What about ChatGPT and Perplexity?
Same discipline, different surface. They need crawlable, well-structured, genuinely useful pages with claims they can verify. The difference is measurement: there’s no Search Console for them, so you track brand mentions and referral traffic instead of rankings.
The short version
An AI Overviews traffic drop is real, it’s measurable, and it’s smaller than the headlines suggest unless you’re a publisher. The best causal research puts it at roughly 38% fewer clicks on the searches where an overview appears, and it’s concentrated almost entirely on informational content.
Your product pages, service pages and local searches are mostly untouched. Your “what is” articles aren’t coming back. Knowing which of those describes your traffic takes fifteen minutes in Search Console and is worth more than every industry statistic in this post.
If your rankings held and your traffic fell and you can’t tell which of these is happening, send me the site and what you found. You will get an honest read on what your actual exposure is and what I would do about it, within 24 hours and at no cost. If it turns out your traffic problem is technical rather than AI, that’s what the audit is for.
Leave a Reply