Google referrals to publishers fell about a third in the year to late 2025; chatbot traffic is nowhere near offsetting the loss yet.
Publishers are the canary in this particular mine. They depend on search referrals more directly than almost any other category, they measure them obsessively, and they feel changes first. So when a survey of nearly three hundred news executives across fifty-one countries reports Google referrals down about a third in a year, with chatbot traffic nowhere near filling the hole, it is worth reading carefully — both for what it says about publishing and for what it says about everyone else, a little later.
This is a survey of senior news industry leaders across a wide international sample, reporting what they are observing in their own referral data and how they expect the picture to develop. That makes it a different kind of evidence from the click-through studies elsewhere in this hub — those measure behaviour directly; this collects what practitioners see across many properties.
The design has real strengths. It captures conditions across many markets and business models rather than a single panel, and it reaches organisations with unusually good visibility into their own traffic. It also captures expectation and response, which behavioural data cannot. Its limitation is equally clear: survey responses are self-reported, and what people believe is happening is not always precisely what is.
The headline is a fall of roughly a third in Google referral traffic over the year to late 2025. For businesses whose economics rest on audience arriving from search, that is not a trend to monitor — it is a structural problem, and the survey captures an industry treating it as such.
Several forces contribute rather than one. AI Overviews satisfy a share of queries on the page. The results page has grown more crowded, pushing links down. And more information-seeking is starting inside assistants that may never surface a publisher link at all. The survey does not decompose the decline into these components, which is one of its genuine limits, but the aggregate direction is unambiguous.
The finding that matters most is not the decline itself but the absence of an offset. Referral traffic from chatbots and AI assistants is growing quickly in percentage terms, and it remains a small fraction of what has been lost. Fast growth from a very small base does not replace a third of a large number, and the survey makes clear that publishers are not experiencing this as a channel shift.
This is the crucial distinction. A channel shift means audience moving from one route to another, which is disruptive but survivable by following them. What the data describes is closer to disintermediation — the audience is still getting the information, and the visit is simply not happening. There is no equivalent destination to follow them to, which is why the response cannot be to optimise for a new referral source.
Chatbot referral growth of a couple of hundred per cent sounds like a channel arriving. Applied to a base that begins below one per cent of total traffic, it produces something still below one per cent. Percentage growth from a tiny base is one of the most reliably misleading statistics in circulation, and it has been used repeatedly to suggest the transition is further along than it is.
The honest framing needs both numbers together: the growth rate and the base. Chatbot referrals are growing quickly and remain immaterial to the traffic economics of most publishers. Both halves are true, and reporting either alone produces a distorted picture — optimistic if you cite only growth, dismissive if you cite only the base.
Chatbot referrals are growing fast from a base so small that they replace almost none of the lost search traffic. The audience did not move to a new channel; the trip to the site stopped being necessary.
The obvious objection is that publishers are a special case — unusually dependent on search, unusually exposed to informational queries, unusually vulnerable to having their content summarised. All true, which is exactly why they register the effect first and most severely.
That makes them an early indicator rather than an irrelevance. Any business with a meaningful informational content operation — the guides, comparisons, and explainers that bring people in before they are ready to buy — owns content with the same exposure profile. The severity will be lower and the timing later, but the mechanism is identical, and publishers are demonstrating what it looks like at full strength.
What it cannot tell you: precisely how much of the decline is attributable to AI answers versus other causes, since survey responses report an outcome rather than decompose it. Self-reported data also carries recall and salience bias — a widely-discussed decline may be reported more readily than a stable picture.
The survey also captures response, and the direction is consistent: reducing dependence on referred search traffic by building direct relationships. Newsletters, apps, membership and subscription models, events, and audio all recur — each an attempt to own the audience relationship rather than rent access to it through an intermediary.
This is a rational response to disintermediation specifically. If the problem were a channel shift, the answer would be to follow the audience; because the visit itself is disappearing, the answer is to build routes that do not depend on a visit being prompted by a third party. That logic is not confined to publishing, and it is the most transferable part of the survey.
When the visit stops happening, traffic reporting goes quiet while your brand may still be in the answer. DUNkē tracks citations across eight AI engines — per prompt, against competitors — so influence stays measurable.
A decline of this kind produces a specific reporting failure. Traffic falls, so traffic-based reporting shows deterioration — while the content may be being read more than ever, in summarised form, with the brand cited. Under conventional analytics, high influence and zero visits is indistinguishable from irrelevance.
That is not a reason to abandon measurement; it is a reason to add a dimension. Citation presence, share of voice against competitors, branded search movement, and direct traffic trends together describe influence that referral counts cannot see. Without them, the reporting will keep describing a business that appears to be failing while it may be reaching more people than before.
If purely informational content is the most exposed category, the implication is not to stop producing it — it is what makes you a source worth citing and what builds the topical authority that citation depends on. The implication is to stop expecting it to pay for itself in sessions.
The rebalance that follows is toward content the answer layer cannot displace: material requiring your data, your tools, your judgement, or your ongoing relationship. And toward direct routes — the same logic publishers are applying — so that the value of being known does not depend entirely on an intermediary sending someone over.
This is survey evidence, not measurement. It reports what a large, senior, international sample observes and believes, which is genuinely informative and is not the same as a direct behavioural study of referral logs.
It also cannot separate AI-driven decline from other causes operating in the same window — algorithm changes, platform policy shifts, and broader consumption changes all overlap. Treat the direction and severity as credible, and be cautious about attributing the whole decline to a single cause.
The most useful thing about this study is that it describes a destination rather than a forecast. For one industry, the shift from referral to answer has already largely happened, and the substitution everyone assumed would arrive has not. That is information other sectors get to have before the same thing reaches them at their own pace.
The lesson is not that search is finished — publishers still receive substantial search traffic and the decline is a proportion, not a collapse. It is that the assumption of substitution should be abandoned. Planning that assumes a new referral channel will replace the old one is planning for something that, in the one industry far enough along to know, has not occurred.
Behavioural measurement is generally preferable to self-report, so it is worth being explicit about why a survey earns its place in a research hub otherwise built on log data. The answer is coverage and context: no single organisation has visibility into referral trends across 51 countries and hundreds of publishers, and no log dataset captures what leaders intend to do in response.
Surveys of senior practitioners also aggregate a great deal of first-party measurement indirectly — each respondent is reporting from their own analytics rather than guessing. That makes this closer to a meta-observation of many measurements than to opinion polling. The limitations remain real, but the alternative is not better evidence; it is no evidence at this breadth.
The survey reports an outcome and cannot decompose it, which is its most significant analytical limitation. Several forces plausibly contributed within the same window: AI Overviews satisfying queries on the page, a more crowded results page displacing links, ordinary algorithm updates, shifting consumption toward social and video, and information-seeking moving into assistants entirely.
Attributing the whole decline to AI answers would therefore overstate what the data supports. The honest position is that AI answers are a substantial contributor within a set of overlapping pressures, and that the direction is unambiguous even if the decomposition is not. Anyone citing this study as proof of a specific AI-attributable percentage is reading more into it than it contains.
It is worth being precise about the substitution failure, because it is the most consequential finding and the most frequently softened. The comparison is not between two channels of similar scale where one is growing and one shrinking. It is between a large established source of traffic that has lost a third of its volume and a new source that remains a rounding error in most publishers’ totals.
For the arithmetic to work, chatbot referrals would need to grow by orders of magnitude rather than by percentages, and there is no structural reason to expect that — the entire design of an answer engine is to resolve the question without a visit. Referral traffic is a byproduct of that design, not its purpose, which is why treating it as an emerging channel to be optimised misreads what it is.
The transferable finding for non-publishers concerns a specific asset class: the library of informational content that most content-led businesses have accumulated. Guides, explainers, glossaries, and how-to material were built to capture demand at the top of the funnel, and they are the content most exposed to being summarised.
The response is not to stop producing it — that content is what makes you a citable source and what builds topical authority. It is to change what you expect from it. If a guide’s job was to deliver sessions that convert downstream, and it now delivers citations instead, it is still working; the measurement has to follow. Businesses that judge that content purely on session volume will conclude it has failed and cut exactly the material that earns them presence in the answer layer.
The response the survey documents — newsletters, apps, membership, events, audio — shares one property: each establishes a route to the audience that does not require an intermediary to prompt a visit. That is the correct structural answer to disintermediation, as distinct from the answer to a channel shift.
The same logic applies well beyond publishing. Any business that has grown by renting attention from a search intermediary faces a version of this exposure, and the mitigations are the same in kind: owned audience relationships, direct channels, and reasons for people to come to you rather than to be sent. This is the most durable takeaway in the survey, and it is not really about search at all.
A business in this position needs measurement that can distinguish declining relevance from declining referral, because traffic data alone cannot. That means adding citation tracking to see whether your content is being used in answers, branded search and direct traffic trends to see whether awareness is holding, and self-reported attribution to catch influence that instrumentation misses entirely.
Without those, the reporting will show a business in decline while it may be reaching a larger audience than ever in a form that produces no sessions. Publishers have arrived at this problem first and are building exactly this kind of measurement. Everyone else has the advantage of being able to build it before the decline arrives rather than during it.
Balance requires acknowledging what could change. Platforms face regulatory and commercial pressure over content usage, and licensing arrangements between AI companies and publishers have been forming. Interface changes could give sources more prominence. And as assistants take on more agentic tasks, they may drive more traffic to destinations where something needs to be done.
None of these is guaranteed, and none is visible in the current data. But planning as though the present trend continues indefinitely and unchanged is as unsound as assuming rescue. The defensible position is to build for the environment as measured while remaining alert to changes in it — which is the same posture this entire hub argues for.
The most valuable property of this study for a non-publisher is that it describes a state other sectors have not yet reached. Publishers are unusually exposed, so they encountered the full effect earlier and harder, and their measured experience is a preview rather than a curiosity.
What that preview says is: expect the decline to be real and material, expect substitution not to arrive, expect the remedy to be direct relationships and measurement that captures influence without traffic. Having that in advance is a genuine advantage — it is considerably cheaper to build direct channels and citation measurement before you need them than to construct both while explaining a traffic decline to a board.
One structural point deserves emphasis because it explains why recovery is unlikely to be symmetrical. Traffic was lost from a mature channel that had been optimised over two decades, with established practices, tooling, and expertise. Whatever replaces it, if anything does, starts without any of that accumulated infrastructure.
This means even a genuine new referral source would take years to reach comparable efficiency, quite apart from reaching comparable volume. Businesses planning on the assumption that they will simply learn to optimise the new channel are underestimating both the volume gap and the capability gap. The more realistic planning assumption is that the lost traffic is not coming back in the same form.
The survey captures responses as well as conditions, and some of the early responses across the industry were mistakes worth learning from. Chasing chatbot referral optimisation as though it were an emerging channel absorbed effort for negligible return. Blocking AI crawlers wholesale removed brands from answers without restoring the traffic. And doubling down on volume publishing to compensate for declining per-piece performance accelerated the problem.
The responses that appear more durable are the structural ones: direct audience relationships, distinctive material that cannot be summarised away, and measurement that captures influence without traffic. Non-publishers reading this study have the advantage of skipping the first set of mistakes, which is a substantial part of the value of watching an industry that got there earlier.
The question of whether to block AI crawlers arises naturally from a study like this, and the survey period covered a good deal of experimentation with it. The logic is understandable — if content is being used to produce answers that eliminate the visit, withhold the content. The practical outcome is generally that the brand disappears from those answers while the traffic does not return, because the questions get answered from other sources.
That trade only makes sense where the content itself is the product being sold and its use in answers directly cannibalises a paid relationship, which is true for some publishers and few other businesses. For most, blocking forfeits presence in the answer layer without recovering anything. The decision deserves to be made deliberately rather than as a reflex to a declining traffic chart.
If there is one transferable operational lesson in this study, it is that a traffic-only reporting stack will misdescribe a business in this environment. It cannot distinguish a brand losing relevance from a brand being read more than ever in summarised form, and it will report both as decline.
Every business with informational content is now exposed to that failure mode, publishers merely first and hardest. The fix is not complicated — add citation tracking, watch branded search and direct traffic, ask customers how they found you — but it has to be built before it is needed, because retrofitting measurement during a decline means having no baseline to compare against.
The most useful posture for a business that has not yet felt this is neither complacency nor panic, but preparation. The specific preparations are cheap and take time to mature, which is exactly why they should be started before they are urgent: build a direct audience relationship, establish citation measurement so you have a baseline, and begin the presence work that makes you a source rather than a destination.
Each of those takes quarters to become useful, and each is considerably harder to start while explaining a declining traffic chart. Publishers built these under pressure; anyone reading their experience gets to build them under normal conditions. That timing advantage is the single most valuable thing this study offers to a non-publisher, and it expires quietly as the same pressures arrive in other categories.
A survey of 280 news executives across 51 countries reports Google referrals to publishers down roughly a third over the year to late 2025, with chatbot referral traffic growing quickly but remaining a small fraction of what was lost. The important finding is the absence of substitution: this is not audience moving to a new channel but visits ceasing to be necessary, which is a different problem with a different remedy.
Publishers are the leading indicator rather than a special case — any business with a substantial informational content operation owns content with the same exposure, arriving later and less severely. Their response is the transferable part: build direct routes that do not depend on an intermediary prompting a visit, add citation and demand measurement so influence without traffic remains visible, and stop expecting informational content to justify itself in sessions alone.
“Chatbot referrals grew a couple of hundred per cent and replaced almost nothing. The audience didn’t move to a new channel — the trip to the website simply stopped being necessary.” The Age’X Research Team
Get a free GEO audit — the same analysis behind every article here.