In February 2026, Ahrefs published a study that ought to have changed how every marketing budget in the world is written. Looking at top-ranking pages, it found that the presence of a Google AI Overview correlates with a 58% reduction in click-through rate. That is not a rounding error and it is not a seasonal dip. It is nearly double the 34.5% decline the same kind of analysis showed in April 2025, which means the trend is not settling — it is steepening.
Pew Research put a blunter number on it: users click a result 8% of the time when an AI Overview is present, versus 15% when it is not. Chartbeat measured a 33% decline in Google search referrals across more than 2,500 news sites during 2025. Digital Content Next members reported losses ranging from 1% to 25%, with some publishers exceeding 75%.
Why this matters for founders and operators
The instinct is to read this as a publisher problem. It is not. If you sell anything to anyone who researches before they buy, your funnel has a hole in it that did not exist two years ago, and the hole is at the very top where you cannot see it in your analytics. Your rankings can be flat, your impressions can be up, and your sessions can still be falling — because the answer got delivered without the visit.
This is precisely the shift that created the market for AI search agencies. The work is no longer only about earning a position in a list of links. It is about being the source the answer is assembled from. Those are related disciplines, but they are not the same discipline, and the tactics diverge quickly.
Google's response, and why it is not enough
On 6 May 2026, Google announced five changes aimed squarely at this criticism: a "Further Exploration" links section at the end of AI Overviews, hover context on inline links, subscription labels that highlight publications the user already pays for, forum perspectives surfacing Reddit and first-hand sources, and expanded product review cards on shopping queries.
Read carefully, every one of those is a way to put more links inside the answer rather than a way to send the user out of it. That is not cynicism — more surface area for citation genuinely does help sites that get cited. But it rewards a specific kind of page: one that is easy for a model to attribute a claim to. Generic marketing copy does not get cited. Original data, clear comparisons, first-hand accounts and structured facts do.
The antitrust pressure is real
Penske Media has filed an antitrust suit. The European Commission has opened an investigation into whether AI Overviews breach competition rules. In the United States, the Department of Justice won its search antitrust case and a federal judge barred exclusive search distribution contracts. None of these will resolve quickly, and none of them will restore the clicks in the meantime.
Planning your 2027 revenue around a legal outcome is not a strategy. Planning it around being citable is.
What to actually do about it
- Separate impressions from sessions in your reporting. If they have decoupled, you are being summarised, not ignored. That is a different problem with a different fix.
- Find the queries where you are the answer but not the destination. Search Console still shows the impression. Pair it with a prompt test in ChatGPT, Claude and Gemini to see whether you are named.
- Build the pages that get cited. Original numbers, honest comparisons, and specifics a model can attribute to you by name.
- Get into the third-party sources assistants lean on. Directories, comparison write-ups and review sites carry disproportionate weight because they read as independent.
If you want help doing this properly rather than experimenting for two quarters, you can get matched with a vetted AI agency, or browse providers who specialise in exactly this work in our AI search agency category. More background reading lives on our AI insights blog.
What the decline looks like across sources
| Source | Measured effect | Period |
|---|---|---|
| Ahrefs | 58% CTR reduction on top-ranking pages with an AI Overview present | February 2026 |
| Ahrefs (prior) | 34.5% CTR reduction | April 2025 |
| Pew Research | 8% click rate with Overview vs 15% without | 2025-2026 |
| Chartbeat | 33% decline in Google referrals across 2,500+ news sites | 2025 |
| Digital Content Next | 1-25% traffic loss typical; some members above 75% | 2025-2026 |
Two things stand out when you line them up. The first is that every independent measurement points the same direction, which is unusual in this industry and worth taking seriously. The second is that the spread is enormous — a 1% loss and a 75% loss appear in the same dataset. That spread is the actionable part. It says the outcome is not determined by the technology; it is determined by what kind of content you publish and whether an assistant has a reason to send someone to it.
Three measurement mistakes that hide the problem
Mistake one: reporting sessions without impressions. If your dashboard only shows traffic, a 20% decline looks like a ranking problem and you will spend a quarter fixing rankings that were never broken. Pull impressions alongside. Flat impressions with falling clicks is a summarisation problem, and no amount of on-page optimisation addresses it.
Mistake two: averaging across query types. AI Overviews do not appear uniformly. Informational and definitional queries attract them heavily; transactional and navigational queries much less. Averaging your whole site together produces a mild-looking number that conceals a catastrophic one on your top-of-funnel pages and a healthy one on your bottom-of-funnel pages. Segment by intent before you conclude anything.
Mistake three: treating brand and non-brand identically. Brand queries are relatively insulated, because someone searching your company name usually wants your site specifically. Non-brand discovery queries are where the damage concentrates. If brand traffic is propping up your total, you are watching your discovery channel die behind a number that looks fine.
A thirty-day diagnostic
- Week one: export twelve months of Search Console data and plot impressions against clicks for your top 100 non-brand queries. Flag every query where impressions are flat or up while clicks fell more than 20%.
- Week two: take the twenty highest-value flagged queries and run each as a natural-language prompt in ChatGPT, Claude, Gemini and Perplexity. Record whether you are named and which URLs are cited.
- Week three: build the citation map. For each cited URL that is not yours, identify whether it is a directory, a comparison article, a review site, a forum thread or a competitor's original research. That distribution tells you exactly where the influence sits.
- Week four: pick the two source types you can realistically influence within a quarter and start there. For most B2B companies that is directory and comparison placements first, original data second.
This takes one person about a day a week and produces a document nobody else in your company can produce. It is also, not coincidentally, exactly the artefact the better AI search agencies sell as their opening engagement.
Frequently asked questions about AI Overviews
Do AI Overviews appear on every query? No. Coverage skews heavily toward informational and definitional searches and thins out on transactional and navigational ones. This is why a site-wide average understates the damage to discovery content and overstates it for bottom-of-funnel pages.
Can you opt out of AI Overviews? Only by blocking Google from using your content in ways that also affect ordinary search visibility, which for most businesses trades a bad outcome for a worse one. The practical route is to make the content worth citing rather than to withhold it.
Does being cited in AI Overviews send meaningful traffic? Some, and less than a ranking position used to. But a citation also carries an endorsement effect that a blue link never did, and it compounds: pages assistants already trust get cited again. Treat citations as brand placement with a traffic tail, not as a traffic channel with a brand tail.
Is this only a Google problem? No — the same dynamic runs across every assistant, which is why measurement should never be limited to one surface.
The bottom line
A 58% CTR decline on your best-ranking pages is not a channel in decline; it is a channel that changed what it rewards. The traffic did not evaporate — the intermediary got better at answering, and the reward moved from being findable to being quotable. Companies that rebuild their content around being quotable will look, in eighteen months, like they got lucky. They will not have got lucky. They will have read the study.
Sources worth reading in full: Ahrefs for the CTR study methodology and Pew Research for the click-rate comparison.