Short answer: AI Overviews take the answer, not the impression. Publisher pages keep appearing and stop being clicked, which is why reporting that leads with impressions or average position looks stable while sessions fall. Research reported by Search Engine Land puts organic search clicks down roughly 42% since AI Overviews began expanding. The content that still earns a click is what an overview cannot reproduce: original data, first-hand reporting, tools, and opinion with a name attached.
The uncomfortable part for publishers is that this is not a ranking problem. You can hold position three and lose most of the traffic that position used to deliver, and no amount of conventional SEO work will recover it, because nothing is broken.
The measured impact
Define Media Group research, reported by Search Engine Land, found organic search clicks down approximately 42% since AI Overviews began expanding — starting at around a 16% decline immediately after launch and deepening as coverage grew. The same study put AI Overviews appearing in roughly 15% of news queries.
Two things are worth separating in those numbers. The 15% figure means most news queries still do not trigger an overview, so this is not yet an extinction event for news search. The 42% figure means that where overviews do appear, and across the broader query mix a publisher depends on, the effect on clicks is severe.
The same research found breaking-news traffic up 103% and Discover up 30% over the window, which is the other half of the picture: the traffic did not vanish so much as relocate to a surface with different rules.
Why impressions hold while sessions fall
This is the mechanic that breaks most publisher reporting.
Search Console counts an impression when your result appears. An AI Overview sits above the results and answers the question. Your page still appeared, so the impression is still counted — the reader simply had no reason to continue. Position is unchanged. Impressions are unchanged. Click-through rate collapses.
The practical consequence is that a monthly report built on impressions and average position will show a healthy, stable channel during a period when the channel stopped delivering readers. If your dashboard leads with either metric, it is currently hiding the problem you most need to see.
Lead with clicks and sessions. Then track click-through rate as a trend per page, because that is where the erosion actually shows up first.
What Google says you should do about it
This deserves a direct quote, because the market around it is noisy and much of what is being sold contradicts the primary source.
Google states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." To be eligible, "a page must be indexed and eligible to be shown in Google Search with a snippet, fulfilling the Search technical requirements." And explicitly: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." (Google Search Central.)
That is worth sitting with if you have been quoted for an AEO package built around new markup or a special file. For Google's AI surfaces, the stated requirements are the ordinary ones: crawlable, indexable, useful, with structured data that accurately matches what is visible.
Two honest qualifications. First, this is Google's account of Google's features; ChatGPT, Perplexity and Copilot select sources their own way and are less documented. Second, "no special optimisation required for eligibility" is not the same as "nothing affects selection" — the controlled research in GEO: Generative Engine Optimization (Aggarwal et al., KDD 2024) found that adding statistics, citing sources and adding quotations improved a source's visibility by roughly 30 to 40 percent, with lower-authority domains gaining most. Those are content properties, not markup tricks, and they are worth pursuing on their own merits.
What still earns the click
| Content type | Exposure to overviews | Why |
|---|---|---|
| Consensus explainer ("what is X") | Severe | Precisely what an overview is built to replace |
| Reference and definition pages | Severe | Answerable in two sentences |
| Aggregated listicles | High | Synthesisable from other sources |
| Breaking news | Low | Recency and volume; overviews lag |
| Original data and research | Low | Cannot be generated without your dataset |
| First-hand reporting and interviews | Low | The source is you |
| Tools and calculators | Minimal | Requires interaction, not an answer |
| Named opinion and analysis | Low | A summary of an argument is not the argument |
The pattern is consistent. Anything a model can assemble from what it already holds is exposed. Anything that requires access, effort, presence or a named perspective is not.
Rebuilding the evergreen mix
Evergreen content used to be the revenue floor between news spikes. That role is under pressure, but the answer is not to abandon it — it is to change what evergreen means in your newsroom.
Stop commissioning consensus explainers. If the answer is available from ten other sources, the overview will state it and nobody will click. This is the hardest change to make because explainers are cheap to produce and used to work.
Convert reference pages into something with a reason to be visited. A page that only states a fact is replaceable. The same page with your own data, a comparison the reader can act on, or a tool is not.
Invest in the smallest amount of original data you can sustainably produce. A quarterly survey of your own audience, a dataset you assemble from public records, or a tracker you maintain will out-earn a dozen explainers — and it is citable, which increasingly matters as much as clickable.
Put names on things. A named author with expertise on the beat is both an E-E-A-T signal and the reason an argument cannot be paraphrased into an overview without loss.
Three things publishers get wrong about this
Treating it as a penalty
The instinct after a traffic decline is to look for a fault and fix it. There is no fault here. The page still ranks, the site is still healthy, and nothing was demoted — the answer is simply being delivered before the click. Publishers who spend a quarter auditing technical SEO in response to overview erosion find nothing wrong, because nothing is wrong, and lose the quarter.
The diagnostic that distinguishes the two: in an overview erosion, position and impressions hold while click-through rate falls. In an actual ranking problem, position falls too. Check that before commissioning an audit.
Buying markup to fix a content problem
A large amount of what is currently sold as AI optimisation is schema work and file generation. Google's own documentation says neither is required for its AI features. If a proposal's central mechanism is markup, ask what evidence supports it beyond the vendor's own case studies — and compare that against the primary source above.
The things that genuinely appear to influence selection are content properties: real numbers, cited sources, direct quotations, and a named author. Those cost editorial effort rather than implementation budget, which is precisely why they are less frequently sold.
Optimising the wrong pages harder
When traffic falls, the reflex is to improve the pages that lost the most. But the pages that lost most are usually the consensus explainers, which are the least recoverable category on the list above. Effort spent making a definition page better is effort spent competing directly with the thing that replaced it.
The better move is to shift the mix. Accept the decline on replaceable content, stop commissioning more of it, and move that budget into the categories overviews cannot reproduce. That is a commissioning decision, not an SEO task, and it usually needs an editor rather than an agency to carry it.
Measuring the erosion
- Report clicks and sessions first. Demote impressions and average position; they will mislead you here.
- Track click-through rate per page over time. A page whose CTR halves while its position holds has been overview-eroded, and that is your clearest diagnostic.
- Segment evergreen from news. They are now two different businesses with different trends, and blending them hides both.
- Add assistant referral traffic as its own channel, understanding it is a floor rather than a measure — most citations produce a read and no click.
- Do not expect a recovery. This is a structural change in how answers are delivered, not a penalty that gets lifted once you have made amends. Plan the content mix accordingly, and set expectations with stakeholders early rather than after two quarters of missed forecasts.
For the surface where publisher traffic actually grew over the same period, see Google Discover for news publishers. For what to do when Discover itself moves against you, see diagnosing a Discover traffic drop.
We help news websites and digital publishers rebuild the content mix and the reporting around it, usually alongside search and answer-engine work and growth programs. If your impressions look fine and your revenue does not, get in touch.
