The claim that Google AI Overviews cut publisher clicks by 40% is a useful signal and a bad universal forecast. It combines several different evidence objects: a preregistered field experiment, a large commercial search panel, a third-party synthesis, and whatever a specific publisher can observe in Search Console or GA4. Those objects answer different questions.
The operational fix is a measurement ledger. Every row should preserve the population, treatment, unit, geography, date window, source, metric, causal status, and limitation. Without that ledger, a number measured on US desktop searches becomes a global forecast, a panel's exposure rate becomes a site's traffic change, and a correlation about AI citations becomes a promise of referral growth.
What the 40% result does and does not say
The arXiv paper AI in Search Reduces Publisher Referrals Without Improving User Experience describes a preregistered field experiment with 1,100 participants. Its abstract reports that removing AI Overviews and AI Mode increased click-through to publishers, while an AI Mode-only experience reduced click-through and weakened reported user experience and trust.
That is causal evidence for the experiment's treatment, population, interface, and measured outcomes. It is not a statement that every publisher loses 40% of its traffic. The treatment is a search experience, while a publisher's traffic is also affected by rankings, brand demand, device mix, country, seasonality, content type, and changes in the result page.
RevScope's November 2026 synthesis presents the result as about 40% fewer clicks to outside websites when an Overview appears. The same article reports DataForSEO panel observations across millions of keywords, with most of the visible sample concentrated in US English search. Its headline is useful for discovery. The panel coverage and the experiment should remain separate rows in a ledger.
Four evidence layers that should not be merged
| Layer | Example | What it can support | What it cannot support alone |
|---|---|---|---|
| Randomized or preregistered experiment | The N=1,100 field experiment | A treatment effect for the tested experience and population | A forecast for every site and market |
| Search panel estimate | DataForSEO or another keyword panel | How often a feature appeared in the panel's tracked queries | Your site's impression or click change |
| Search-platform reporting | Google Search Console AI feature or web reports | The pages, dates, devices, countries, and metrics exposed by the account's reporting scope | A complete causal decomposition of why clicks moved |
| First-party analytics | GA4 organic sessions or page views | Your observed organic traffic under your tagging and reporting configuration | A clean separation of AI Overview referrals from ordinary web search |
The first two layers describe an external environment. The last two describe a property. They can be compared, but they should not be added together or treated as the same denominator.
Google's guidance narrows the optimization claim
Google's AI features and your website says that existing SEO best practices remain relevant for AI Overviews and AI Mode. It also says there are no additional requirements or special optimizations required for inclusion. That directly limits the strongest versions of GEO advice. A page can be clear, crawlable, well sourced, and useful without gaining a guaranteed AI citation or referral.
Google's Search Generative AI performance reports provide a separate measurement surface for eligible accounts. The report should be recorded with its own metric definitions and date coverage. A specialized exposure report is not the same as a click report, and a missing row is not evidence of zero exposure.
The practical implication is to measure inclusion and referral as separate outcomes:
- Was the page exposed in the relevant AI feature report, where available?
- Did the page receive web-search impressions and clicks during the same window?
- Did GA4 record organic sessions or page views, using the property's own timezone and channel definition?
- Can the change be distinguished from ranking, content, device, country, and seasonal changes?
Build a page and query ledger
A useful ledger can be implemented in a spreadsheet or JSON file. Each observation should include:
observation_id
property_and_page
query_or_query_set
market_and_language
device
feature_definition
start_date / end_date
timezone
source_system
metric_and_unit
observed_value
coverage_state
causal_status
confounders
evidence_url
The coverage_state field matters. Google Search Console can lag, privacy-limit query rows, and expose incomplete daily coverage. GA4 has its own processing and attribution definitions. A seven-day comparison with two missing GSC dates is an incomplete observation, not a confirmed decline.
The causal_status field should be equally explicit:
randomized_experiment: treatment assignment was controlled for the reported comparison.panel_observation: a tracked query or user panel was observed, with coverage limits.account_observation: the publisher's own reporting changed during a window.correlation: two measures moved together without a causal design.unknown: the source or denominator is insufficient.
The word loss should be reserved for a defined comparison. If a page's clicks fell from 12 to 8 while impressions, ranking, device mix, and query mix also changed, the ledger records a change. It does not automatically record AI-caused loss.
A worked interpretation of the 40% headline
Suppose a field experiment reports that publisher click-through is 40% lower when an AI feature is present. A careful publisher workflow would write:
- The experiment observed a negative referral effect under its treatment and population.
- RevScope's panel suggests AI Overviews are common in its tracked US English query set.
- Google says AI-feature traffic and ordinary search reporting have defined platform-level reporting rules.
- Our site has a separate time series with its own missing dates, pages, countries, devices, and query exposure.
- The site-level conclusion remains unknown until the local observation is complete and alternative explanations are checked.
This is slower than copying 40% into a strategy deck. It is also the only version that can guide a real decision. If local evidence shows lower clicks on pages exposed to an AI feature, the next action could be a content, internal-link, or audience experiment. If exposure is absent or unmeasured, rewriting for GEO is an unsupported response.
What a content team can improve without inventing a GEO hack
The evidence supports ordinary quality work: answer the target question directly, use stable entities and terminology, cite primary sources, make important definitions visible in the crawlable body, and preserve dates and scope. These practices help readers and make the page easier to evaluate. They do not guarantee inclusion or citation.
The team can also improve its own measurement contract:
- Freeze a set of pages and queries before a change.
- Record market, language, device, and platform-reporting timezone.
- Keep GSC impressions and clicks distinct from GA4 organic sessions and page views.
- Preserve missing dates as unknown.
- Treat AI citations as unknown unless they are directly sampled and logged.
- Review the same article age and seasonal window before interpreting movement.
This turns a broad concern about AI search into a sequence of testable questions. It also prevents an attractive number from becoming a false precision claim about a site that was never part of the original sample.
Limits and conclusion
The field experiment, RevScope synthesis, Google documentation, and a site's own analytics are complementary, not interchangeable. The experiment gives stronger causal evidence for a bounded treatment. The panel describes a tracked search environment. Google documentation defines reporting and SEO boundaries. First-party data show what happened in one property under one measurement setup.
The right conclusion is therefore conditional: AI search can reduce publisher referrals under some tested conditions, and the risk deserves measurement. A universal 40% forecast, a guaranteed GEO recovery tactic, and a causal claim from an incomplete GSC window all exceed the evidence.
Keep the ledger. Let each source answer only the question its design can answer.