The standard account of AI and search traffic is a single downward line. Referrals fall, publishers suffer, content marketing gets harder. It is not wrong, and it is almost useless, because a single line cannot tell you which of your pages to defend and which to stop making.
The Association of Online Publishers commissioned the study that answers the more useful question. Six months of traffic across eight major UK publishing groups, 10.8 billion pageviews, around 80% of participants' total traffic once long-tail URLs were stripped out, segmented across 147 domains by both topic and content type. Conducted by DJB Strategies, with a behavioural component run by Ipsos across 1,000 UK adults.
Segment that dataset and the single line disappears. What emerges is a redistribution. Some categories lost more than half their Google referrals. Others gained. And the split between them follows a logic you can act on.
The headline number, and the projection people quote instead
Across all participants, organic Google Search referrals fell 7.1% between the first and second quarters of 2025. That is the measured finding, and it is worth sitting with the timing. The same period saw Google report 12% growth in Search revenue while expanding AI Overviews.
The number that gets quoted more often is different. If the 7.1% quarterly decline holds, the study projects that pageviews delivered by Google would halve by the third quarter of 2027, before accounting for any compounding effect from habitual AI use on return visits.
Note what that second number is. A projection produced by extending an observed trend, clearly labelled as such by its authors. It is not a measurement, and treating it as one is the first mistake most summaries of this study make. The 7.1% is evidence. The halving is a scenario.
The decline is also not evenly distributed, which is the part that makes the study worth reading rather than citing. News publishers averaged roughly 5% decline per quarter. The consumer and business-to-business publisher segment averaged around 18%. The study attributes that gap to tighter guardrails applied to news inside AI Overviews compared with the evergreen content that AI systems aggregate more freely.
By topic, the spread runs from minus 28% to plus 37%
Segmented by subject matter, the picture stops looking like decline and starts looking like reallocation.
Sports referrals rose 37%. Travel rose 22%. Health and wellness rose 7%. News itself ticked up 2%. Against that, technology fell 28%, finance and business fell 23%, and politics fell 7%.
Read those together and a pattern shows up. The categories that fell are the ones where a competent summary substitutes adequately for the source: what a technology does, how a financial product works, what a policy means. The categories that rose are the ones where the value is in the specific instance rather than the general explanation. A match result, a particular hotel, a personal account of a condition.
That is not a claim about what AI can theoretically do. It is a description of what happened to 10.8 billion pageviews in Britain over six months.
The Extremes, and What Separates Them
A 117-point spread between the worst and best category in the same dataset, over the same six months, in the same country.
Food and drink how-tos: minus 54%
The clearest loss in the study. A recipe or a technique is a bounded procedure with a correct answer, which is exactly the shape of content an AI summary reproduces without loss. Nothing about the source is necessary to the reader once the steps are on the results page.
Guides and how-tos overall: minus 20% or worse
Across all topics, not just food. This is the single largest content-type decline and the category most digital marketing programmes have historically produced most of. The whole top-of-funnel explainer model sits inside this bar.
Opinion and commentary: plus 4%
Modest, and the direction is the finding. Content carrying an explicit point of view resisted substitution best, because a position belongs to whoever holds it. An AI summary can report that a view exists. It cannot be the person who holds it.
News interviews: plus 63%
The best performing combination of topic and format in the entire dataset. An interview is primary source material that did not exist until somebody obtained it. There is nothing to summarise from elsewhere, because there is no elsewhere.
The dividing line is not quality, it is substitutability
Every losing category shares one property: a competent summary serves the reader as well as the original did. Every gaining category shares the opposite: the specific instance, the obtained material or the held position is the thing being sought, and it cannot be aggregated out of other sources. That is a testable property of a content brief, which means it can be applied before the piece is commissioned rather than discovered in the analytics afterwards.
Sources: AOP Artificial Intelligence Publisher Impact Study, conducted by DJB Strategies, reported 2026 • Six months of traffic across eight UK publishing groups, 10.8 billion pageviews, approximately 80% of participant traffic after long-tail exclusion, 147 domains • Confidence: REPORTED
Created by Arfadia • arfadia.com/blog
The full segmentation, in one place
Both cuts of the data, with the caveat that these measure publisher referrals rather than commercial site traffic, and are reported rather than independently verified.
| Segment | Change in Google referrals | Why it moved that way |
|---|---|---|
| News interviews | +63% | Primary source material. Nothing exists to summarise from elsewhere. |
| Sports | +37% | Specific, time-bound and continuously renewed. A summary goes stale within hours. |
| Travel | +22% | Decisions turn on the particular place, image and price rather than a general explanation. |
| Health and wellness | +7% | High stakes drive source verification, and users click through to check who is saying it. |
| Opinion and commentary | +4% | A held position cannot be aggregated. The author is part of the product. |
| News (as a topic) | +2% | Tighter guardrails applied to news inside AI Overviews than to evergreen content. |
| Interviews (as a format) | +1% | Broadly flat across all topics, with news interviews carrying the upside. |
| Politics | -7% | Explanatory coverage substitutes more readily than reported coverage. |
| Guides and how-tos | -20% or worse | A bounded procedure with a correct answer. The archetypal substitutable format. |
| Finance and business | -23% | Heavy in definitional and explanatory content, which summarises cleanly. |
| Technology | -28% | The worst-hit topic. Specification and how-it-works content is precisely what AI answers do well. |
| Food and drink how-tos | -54% | Topic and format compounding. The clearest single loss in the study. |
The behavioural half, which explains the traffic half
Traffic data tells you what happened. The Ipsos component of the same study, run across a nationally representative sample of 1,000 UK adults, tells you why.
Where an AI Overview was present, clicks were suppressed by 18%. That is a UK behavioural measurement with a stated sample, which makes it more useful for British planning than the larger figures circulating from other markets and other query sets.
The mechanism the study identifies is what it calls stop here behaviour. The user retrieves the information they wanted and does not interact further. Not frustration, not a failed search. A completed one, ending on the results page.
But here is the finding that cuts against the fatalistic reading. Around 34% of respondents indicated they would click at least one of the links shown in an AI Overview. Which means AI Overviews do appear to be at least partly designed to encourage referrals, and citation placement is worth competing for rather than writing off. A third of a suppressed audience is still an audience, and it is an audience arriving with the summary already read and the question already framed.
The study also surfaces a trust paradox worth naming. AI answers are only as trusted as the news brands they cite. The platforms suppressing referrals depend for their credibility on the same publishers whose traffic they are absorbing. That is an unstable arrangement, and the regulatory response covered in the generative engine optimisation work we do for the UK market exists partly because it is unstable.
Why stable rankings can still lose you traffic
There is a second mechanism that the referral numbers alone do not reveal, and it explains a specific complaint that has confused a lot of teams: rankings unchanged, traffic dropped, no obvious cause.
Authoritas found that AI Overview placements are substantially more volatile than organic ones. Over a two to three month period, roughly 70% of the pages cited in AI Overviews changed. Critically, those changes were not linked to traditional organic rankings.
Read that carefully, because it has an operational consequence. Citation and ranking are separate selection processes running on separate cadences. A page can hold position three all quarter and lose its citation in month two, and nothing in a rank tracker will show you when it happened or why. Sudden unexplained traffic drops alongside stable blue-link positions are not a measurement error. They are what this looks like.
Which is also why treating AI visibility as something to address after rankings plateau gets the sequence wrong. The two systems are not sequential, they are concurrent, and only one of them is visible in the tools most teams already own.
Four Tests That Predict Whether a Page Survives Substitution
Derived from what actually separated the gaining categories from the losing ones. A brief passing none of these is buying traffic that is being repriced downward.
Does it contain data that did not exist before
Original research, a survey you ran, first-party benchmarks, an audit of your own category. This is the content class AI summaries cannot replace, because replacing it would require producing it. It is also the class both Google and AI systems increasingly select.
Does it carry a position somebody holds
Opinion and commentary rose 4% while how-to content fell more than 20%. A summary can report that a view exists but cannot be its holder. Named authorship with real credentials is the mechanism, not a byline for form's sake.
Is the value in the instance or the explanation
Sports, travel and interviews gained. Technology and finance explainers lost. If the reader needs the particular case, the particular place or the particular person, a general answer does not serve them. If a general answer does serve them, they will take it on the results page.
Does the query still require a click to complete
Commercial and transactional intent is far less affected than informational intent, because the transaction cannot happen on the results page. Protecting high-intent pages is the least glamorous and highest-return move available.
And a warning about the fifth test people invent
Volume is not a test. The instinct when referrals fall is to publish more, and the AOP data says the categories that fell hardest are precisely the ones most easily produced at volume. Guides, explainers and how-tos are cheap to scale and are being repriced fastest, which makes scaling them the one response guaranteed to make the problem worse. Google's 2026 spam update direction points the same way.
Sources: AOP Artificial Intelligence Publisher Impact Study, content type and topic segmentation, 2026 • Ipsos behavioural component, 1,000 UK adults • Authoritas AI Overview citation volatility analysis • Confidence: REPORTED throughout
Created by Arfadia • arfadia.com/blog
The wider picture, labelled by geography
The AOP study is UK-specific, which is rare and valuable. Most of the robust research on this question is not, and mixing the two without labels is how misleading claims get built. So here is the surrounding evidence with its geography attached.
Globally, Chartbeat data reported via Axios in March 2026 found publisher referrals down roughly 33%, with the United States at 38%, small publishers down about 60% and large publishers down about 22% over two years. That is a global and US measurement, not a British one.
On click-through rates, Seer Interactive tracked organic CTR on AI Overview queries falling from 1.76% to 0.61% between June 2024 and September 2025, then recovering partially to around 2.4% by February 2026. That partial recovery is a nuance almost nobody reports, and it is US and global data rather than UK.
In the UK specifically, the Professional Publishers Association reported year-on-year CTR declines of 10 to 25% despite stable rankings, which lines up with the citation volatility mechanism described above.
And one figure that requires care. A study by the agency Tank, released in October 2025 and covering 800 companies across 16 UK sectors using Ahrefs traffic estimates, found average monthly organic traffic growth falling from 26.3% to 3.7% after AI search features rolled out, with ranking pages down 11.1% having grown 14.1% the prior year. Hospitality was worst affected at minus 6.7% growth, information technology most resilient at plus 2.1%.
That is a change in a growth rate, not a change in traffic level. It was widely headlined as UK traffic collapsing 86%, which merges a relative change in growth with the volume of traffic itself and is not what the study measured. Tank is also a commercial agency with an interest in AI search anxiety as a service driver, which does not make the finding false but does belong in the citation.
What to do differently on Monday
Three changes follow from this data, and none of them is publish more.
First, audit your content inventory against substitutability rather than against traffic. Pages ranking well in the categories that fell hardest are not safe, they are exposed. The traffic they currently deliver is the thing being repriced, and a rank tracker will report health right up until it does not.
Second, shift commissioning toward the classes that held. Original data, first-hand expertise, named authorship with real credentials, a stated point of view, and the specific case rather than the general explanation. That is a change to briefs and budgets, not to keyword lists.
Third, protect commercial-intent and branded pages deliberately, because a transaction still requires a click and branded demand is where AI visibility shows up as a measurable outcome. Report branded search volume alongside non-branded organic clicks and keep the two separated, since merging them hides exactly the signal you need.
None of that removes the need to measure citation directly, which is a different discipline running on a different cadence and is where our UK generative engine optimisation work sits. But the content decisions above are worth making regardless, because they are decisions about what to commission rather than about which platform to track. The AOP data does not say stop making content. It says the category matters more than the volume, and it is specific enough about which categories that the argument can finally be settled with evidence rather than instinct.
Frequently Asked Questions
What did the AOP Artificial Intelligence Publisher Impact Study actually measure?
Two things, using two methods. The traffic component analysed six months of Google Search referrals across eight major UK publishing groups, covering 10.8 billion pageviews and approximately 80% of participants' total traffic once long-tail URLs were excluded, segmented across 147 domains by both topic and content type. The behavioural component was a nationally representative Ipsos survey of 1,000 UK adults. The study was conducted by DJB Strategies for the Association of Online Publishers. Its headline finding is a 7.1% fall in organic Google referrals between the first and second quarters of 2025.
Will Google traffic to UK publishers really halve by 2027?
That is the study's own projection rather than a measurement, and the distinction matters. If the observed 7.1% quarterly decline were sustained, pageviews delivered by Google would fall to half their first-quarter 2025 level by the third quarter of 2027, before accounting for any compounding effect from habitual AI use on return visits. It is a scenario produced by extending a trend, clearly labelled as such by its authors, and treating it as an observed outcome is the most common error in summaries of this research.
Which content types lost the most Google referrals?
Guides and how-tos fell by more than 20% overall, making it the largest content-type decline in the study. Combining topic and format produces sharper results: food and drink how-tos fell 54%. By topic, technology fell 28%, finance and business fell 23%, and politics fell 7%. The common property across all of them is substitutability, meaning a competent summary serves the reader as well as the original did.
Which content types gained referrals?
News interviews rose 63%, the best performing combination in the dataset. By topic, sports rose 37%, travel rose 22%, health and wellness rose 7%, and news itself rose 2%. By content type, opinion and commentary rose 4% and interviews overall rose 1%. What these share is that the value sits in the specific instance, the obtained material or the held position, none of which can be aggregated out of other sources.
Why did news publishers hold up better than business-to-business publishers?
News publishers averaged roughly 5% decline per quarter against around 18% for the consumer and business-to-business segment. The study attributes the gap to tighter guardrails applied to news content inside AI Overviews compared with the evergreen content that AI systems aggregate more freely. Evergreen explainer content is both easier to summarise and less constrained by those guardrails, which is a structural disadvantage rather than a quality one.
How much do AI Overviews suppress clicks in the UK?
By 18%, according to the Ipsos behavioural component of the AOP study, measured across a nationally representative sample of 1,000 UK adults. This is the most useful figure available for British planning because it is UK-specific with a stated sample size, unlike the larger click-decline figures circulating from other markets and other query sets. The mechanism the study identifies is stop here behaviour: the user retrieves what they wanted and does not interact further.
Is it worth trying to appear in AI Overviews if they suppress clicks?
Yes, and the same study supports that. Around 34% of respondents indicated they would click at least one of the links shown in an AI Overview, which suggests the format is at least partly designed to encourage referrals. A third of a suppressed audience is still an audience, and it arrives having already read the summary and framed the question, which typically means better-qualified rather than worse.
Why does our traffic drop while our rankings stay stable?
Because citation and ranking are separate selection processes running on separate cadences. Authoritas found that AI Overview placements are substantially more volatile than organic ones, with roughly 70% of cited pages changing over a two to three month period, and those changes were not linked to traditional organic rankings. A page can hold its organic position all quarter and lose its citation partway through, and nothing in a rank tracker will show when or why. Unexplained traffic drops alongside stable blue-link positions are what this looks like.
Did UK organic traffic really collapse by 86%?
No, and that headline misreads its own source. A study by the agency Tank, released in October 2025 across 800 companies in 16 UK sectors using Ahrefs traffic estimates, found average monthly organic traffic growth falling from 26.3% to 3.7% after AI search features rolled out. That is a change in a growth rate, not a change in the level of traffic. Reporting it as an 86% collapse merges a relative change in growth with traffic volume itself. The same study found ranking pages down 11.1% after growing 14.1% the prior year, with hospitality worst affected at minus 6.7% growth and information technology most resilient at plus 2.1%.
Should we publish more content to compensate for lost referrals?
This is the response the data most directly argues against. The categories that fell hardest, guides and explainers and how-tos, are precisely the ones cheapest to produce at volume, which makes scaling them the one response guaranteed to deepen the problem. Google's 2026 spam update direction points the same way. The productive shift is toward content classes that held: original data, first-hand expertise, named authorship with real credentials, a stated point of view, and the specific case rather than the general explanation.
How should we split reporting to see this effect in our own data?
Separate branded from non-branded organic demand and never merge them, because merging hides the signal. Report impressions, clicks, click-through rate, conversions and revenue as distinct measures rather than collapsing them into traffic. Track query classes where search result features materially change click opportunity. And protect high-intent commercial and transactional pages deliberately, since those are the queries where a click remains necessary to complete the task and are therefore least exposed.
Sources & References:
- Association of Online Publishers, "Artificial Intelligence Publisher Impact Study", conducted by DJB Strategies, author David Buttle, reported via Press Gazette and InPublishing 2026. Traffic component: six months across eight major UK publishing groups, 10.8 billion pageviews, approximately 80% of participants' total traffic after long-tail URL exclusion, 147 domains segmented by topic and content type. Behavioural component: nationally representative Ipsos survey of 1,000 UK adults. Confidence: REPORTED, observational study rather than a randomised causal experiment.
- Headline finding: organic Google Search referrals fell 7.1% between Q1 and Q2 2025 across all participants, in a period coinciding with Google reporting 12% growth in Search revenue and expanding AI Overviews. Segment averages: news publishers approximately 5% quarterly decline, consumer and B2B publisher segment approximately 18%, attributed to tighter guardrails on news inside AI Overviews relative to evergreen content.
- Projection: if the observed quarterly decline is sustained, pageviews delivered by Google would fall to half their Q1 2025 level by Q3 2027, before accounting for compounding effects of habitual AI use on return traffic. This is the study's own forward-looking inference, not a measured outcome.
- Segmentation by topic: sports +37%, travel +22%, health and wellness +7%, news +2%, politics -7%, finance and business -23%, technology -28%. Segmentation by content type: opinion and commentary +4%, interviews +1%, guides and how-tos -20% or worse. Combined topic and format: news interviews +63%, food and drink how-tos -54%.
- Ipsos behavioural component: AI Overview presence suppressed clicks by 18%. Approximately 34% of respondents indicated they would click at least one link shown. The study identifies "stop here" behaviour, where a user retrieves desired information and does not interact further. The study also documents a trust paradox in which AI answers are only as trusted as the news brands they cite.
- Authoritas analysis of AI Overview citation volatility: approximately 70% of pages cited in AI Overviews changed over a two to three month period, with those changes not linked to traditional organic rankings. This explains sudden traffic drops occurring alongside stable organic positions.
- Chartbeat data via Axios, 17 March 2026: publisher referrals down approximately 33% globally and 38% in the United States, with small publishers down approximately 60% and large publishers approximately 22% over two years. Geography: global and US, not UK.
- Seer Interactive: organic click-through rate on AI Overview queries fell from 1.76% to 0.61% between June 2024 and September 2025, with partial recovery to approximately 2.4% by February 2026. Geography: US and global. Confidence: REPORTED.
- Professional Publishers Association: year-on-year click-through rate declines of 10 to 25% despite stable rankings. Geography: UK. Confidence: REPORTED.
- Tank study, released 30 October 2025, reported by PPC Land 13 November 2025: 800 companies across 16 UK sectors, using Ahrefs traffic estimation. Average monthly organic traffic growth fell from 26.3% to 3.7% after the August 2024 AI Overviews launch and July 2025 AI Mode launch. Ranking pages fell 11.1% on average after growing 14.1% the prior year. Hospitality worst affected at -6.7% growth, information technology most resilient at +2.1%. This measures a change in growth rate, not a change in traffic level; reporting it as an 86% traffic collapse conflates a relative change in growth with traffic volume. Tank is a commercial agency with a stated interest in AI search as a service driver. Confidence: REPORTED, using third-party traffic estimates rather than verified analytics, and insufficient for causal attribution.
- Google's 2026 core and spam update direction, together with independent analysis, points toward original data, first-hand experience, named authorship with credentials, answer-first structure and distinct point of view, and away from generic informational content produced at volume.
- No credible public study isolating AI Overview effects on UK retailer organic traffic using retailer analytics and a disclosed causal or matched-query design was located in this research. Publisher studies must not be relabelled as retailer evidence.