AI Overviews Click Decline: Why Three Studies Disagree
Three 2026 figures circulate as if they measure one thing. A randomised Chrome experiment put the fall in outbound organic clicks at 39.8%. Ahrefs put the fall in position-one click-through rate at 58%. ABC News reported one Australian news site down 35%. Different populations, denominators and counterfactuals. None forecasts a B2B SaaS site.
At a glance:
- 39.8% is a causal estimate — conditional on an AI Overview appearing — from a pre-registered randomised browser experiment on over 1,000 users, early 2026, by Saharsh Agarwal (Indian School of Business) and Ananya Sen (Carnegie Mellon University).
- 58% is not a publisher traffic loss. It is Ahrefs’ modelled gap in position-one click-through rate across 300,000 keywords.
- 35% is one unnamed Australian news site, year on year, in Similarweb data obtained by the ABC. Rival panel Ipsos Iris pointed the other way.
- We could not open the SSRN listing:
ssrn.com/abstract=6513059returned HTTP 403 on 4 September 2026. The 39.8% and 34.5% below are taken from two other first-party sources — the authors’ signed article, and the abstract on Agarwal’s own research page.
The Agarwal–Sen field experiment: 39.8% fewer outbound clicks
The strongest of the three: the only one with a randomised counterfactual. Agarwal and Sen describe the design in ProMarket (Stigler Center, Chicago Booth, 17 August 2026): “Using a custom Chrome extension, we randomly assigned users to either standard Google Search … or a version in which AIOs were removed in real time.” They “recruited over 1,000 users who installed this extension in early 2026” and “observed their behavior for two weeks”.
In the authors’ words, AI Overviews “reduced users’ organic clicks to third-party sites by 39.8 percent” and “increased searches where the user clicked on no links at all by 34.5 percent”. Links inside the AIO did not make up the shortfall: “citations within AIOs generated relatively little referral traffic (about 8% of all clicks)”.
Three caveats the authors state, which the coverage drops. User experience: “We found no measurable improvement in users’ perceived search quality or ease of finding information.” Click quality: “Our data show no difference across these click quality measures between the treatment and control groups, seemingly contradicting Google’s public claims that AIOs produced higher quality traffic for websites that are clicked through on Google.” And the legal framing it gets deployed in: “Our study does not address several essential elements of recent legal cases, such as market definition and exclusionary conduct, nor does it establish an antitrust violation.”
Agarwal’s own research page lists it under Working Papers, not as a peer-reviewed publication, under the title “The Impact of Google AI Overviews on Publisher Traffic and User Experience: Evidence from a Field Experiment”. The abstract printed there states both headline figures in the authors’ own words: “Conditional on appearing, AIOs reduce outbound organic clicks by 39.8% and increase zero-click searches by 34.5%, without affecting sponsored clicks or overall search frequency.” Crossref registers the same paper as DOI 10.2139/ssrn.6513059, type posted-content, subtype preprint. It is unrefereed, and that is how it should be described.
It is, however, pre-registered, which none of the other figures on this page can claim. The trial is lodged with the AEA RCT Registry as AEARCTR-0017393, first published 9 January 2026 — before fieldwork began. The registration fixes the design in advance: three arms (standard AI summaries, summaries hidden, an AI Mode interface), a planned sample of about 1,500 individuals, a two-week tracking window, an intervention period of 7 January to 7 February 2026, and ethics approval from the Indian School of Business (SB-IRB 2025-48). Those are the registered plan rather than the realised sample — the ProMarket article reports “over 1,000 users” actually recruited.
What we could not obtain is the paper itself. The SSRN abstract page returned HTTP 403 to every route we tried on 4 September 2026, and the full-text PDF that Google Scholar indexes at tse-fr.eu returned HTTP 404. So we print 39.8% and 34.5% on the strength of two first-party statements by the authors, not on a reading of the paper: we have not seen the standard errors, the exclusions or the robustness checks. Weigh the numbers accordingly.
Ahrefs’ 58%: a position-one CTR gap, not a traffic loss
The Next Web’s 6 May 2026 article by Alina Maria Stan is where many people meet the 58%. Its headline is where the number goes wrong — “Google’s AI Overviews killed 58 per cent of publisher clicks” — but the body is precise: “An Ahrefs study published in February 2026 found that AI Overviews correlate with a 58 per cent reduction in click-through rates for top-ranking pages, nearly double the 34.5 per cent decline documented in April 2025.” The attribution is to an SEO vendor’s study, not to a court filing, and the quantity is a rate.
So it is not a study of publishers, not a measurement of traffic, and not causal. It is an SEO vendor’s correlational analysis of a rate. Ahrefs’ post “Update: AI Overviews Reduce Clicks by 58%” (4 February 2026, Ryan Law with Xibeijia Guan, checked 4 September 2026) gives the method: “We selected 300,000 keywords from Ahrefs Keywords Explorer database, consisting of 150,000 keywords with an AI Overview present and 150,000 keywords with informational intent and no AI Overview present.” The input is aggregated Google Search Console data, December 2023 against December 2025, and the metric is average position-one CTR — position two is down 50.8%.
Now the trap. Ahrefs’ earlier study, “AI Overviews Reduce Clicks by 34.5%” (17 April 2025, checked 4 September 2026), found that “the presence of AI Overviews reduces the click-through rate for position 1 by ~34.5%” — a 34.5% reduction in position-one CTR. The field experiment reported a 34.5% increase in zero-click searches. Same number, opposite sign, different quantity. We have seen both cited on one slide as corroboration of each other. They are not.
ABC News, Similarweb and Ipsos Iris: two panels, two directions
James Purtill’s ABC News report of 8 October 2025 states: “Most Australian news sites are seeing declines in search traffic, with one down by 35 per cent.” The source is named plainly: “Exclusive data obtained by the ABC from SimilarWeb, a rival service to Ipsos, shows the combined readership of the top news sites in Australia is much lower than a year ago.” A third-party panel estimate, not server logs — and the 35% site is not named.
The same article carries the finding that should stop anyone quoting the 35%: “Ipsos Iris, shows some top-10 news sites have swapped rankings, but their combined readership has actually gone up in the past year.” Two commercial panels, one market, one period, opposite signs. The ABC also reports Man of Many’s director Scott Purcell saying his site’s traffic to “informational” articles was down 10–30 per cent — a within-site split more useful than any headline. Being year on year, the comparison absorbs every other change in the window too.
Google’s position, and how to weigh an interested party
Google is a party to the dispute and should be read as one. Liz Reid’s post of 6 August 2025 says “total organic click volume from Google Search to websites has been relatively stable year-over-year” and that “average click quality has increased”, where “by quality clicks, we mean those where users don’t quickly click back”. It dismisses third-party reports of declines as “often based on flawed methodologies, isolated examples, or traffic changes that occurred prior to the roll out of AI features in Search.”
The post publishes no sample, window or dataset behind those claims; the traffic statements are directional (“relatively stable”, “slightly more”), and the only quantity in it is “billions of clicks to websites every day” (checked 4 September 2026). Asked to comment for the report above, Google directed the ABC to this same post, which the ABC noted “did not provide data”. Google’s Search Central documentation (updated 10 December 2025, checked 4 September 2026) repeats the quality claim unquantified: “We’ve seen that when people click from search results pages with AI Overviews, these clicks are higher quality (meaning, users are more likely to spend more time on the site).” The experiment tested that proposition and found no difference between groups.
The same page adds the operationally important line: AI features are “included in the overall search traffic in Search Console”, reported “within the ‘Web’ search type”. Search Central documents no way to isolate AI Overview impressions within that report (checked 4 September 2026) — one reason panels and experiments dominate this evidence base. See our Search Console generative-AI guide.
Why one effect produces three numbers
| Figure | Who measured it | Population and denominator | What it supports — and does not |
|---|---|---|---|
| −39.8% outbound organic clicks; +34.5% zero-click searches, conditional on an AIO appearing | Agarwal (ISB) and Sen (CMU), pre-registered randomised Chrome-extension experiment (unrefereed working paper; AEARCTR-0017393) | Over 1,000 recruited users, two weeks, early 2026. Denominator: that panel’s searches, treatment vs control | Supports: a causal claim that AIOs divert clicks. Does not support: a per-site forecast, a news-industry figure or an antitrust conclusion |
| −58% click-through rate (position one); −50.8% (position two) | Ahrefs (Ryan Law, Xibeijia Guan), February 2026 | 300,000 keywords, half with an AIO. Denominator: clicks ÷ impressions for the top-ranked page, Dec 2023 vs Dec 2025 | Supports: ranking first is worth fewer clicks. Does not support: “publishers lost 58% of clicks” — this is a rate, not sessions |
| “one down by 35 per cent” | ABC News analysis of Similarweb data, October 2025 | Top Australian news sites. Denominator: estimated monthly search traffic, one site, year on year | Supports: some Australian news sites lost search traffic. Does not support: attribution to AI Overviews — the window holds every other change |
| “relatively stable” click volume; “higher quality” clicks | Google (Liz Reid; Search Central docs) | Unstated — no sample, window or denominator published | Supports: Google’s own account of its aggregate data. Does not support: independent verification — no figures are published to check it against |
All four sources checked 4 September 2026. Five differences generate the spread. Population: a user panel, a keyword list and news websites. Denominator: clicks per search, clicks per impression at one rank, sessions per site. Query mix: Ahrefs’ control set is explicitly informational-intent keywords, and the experiment’s effect concentrates where the AIO sits at the top of the page. Navigational and transactional queries barely feature. Window: two weeks, two years, one year. Counterfactual: only the experiment has one.
What a B2B SaaS team should actually measure
Quoting a news-publisher click-loss figure in a B2B board deck is a category error. Your buyers run navigational queries for your brand, comparison queries with commercial intent, and a long tail of specification questions — almost none of which these samples contain. Four things are measurable on your own property.
- AI referral sessions, defined once and held still — build the channel grouping before arguing about the trend: measuring AI referral traffic in GA4.
- Branded search and direct arrivals — when the click disappears, demand resurfaces as a branded query days later: zero-click B2B pipeline.
- Conversion rate of AI-sourced sessions, not their volume. The pattern is low volume, high intent: do AI-search visitors convert better?
- Share of answers that name you, sampled repeatedly on a fixed prompt set. Before adopting anyone’s published average as a target: how to read a 2026 AEO benchmark report.
Zian AI runs autonomous sales agents across phone, SMS, email and WhatsApp, and tracks its own answer-engine visibility on a fixed prompt set for the same reason: an average built from someone else’s population tells you nothing about yours. If you would rather have agents that book meetings than borrowed statistics, Apply For Partnership.
Frequently asked questions
Which of the three figures is the most reliable?
The 39.8% from the Agarwal–Sen experiment: the only one with a randomised control group rather than a before-and-after comparison. It remains an unrefereed working paper, though a pre-registered one (AEARCTR-0017393), and it measures a recruited panel’s searches, not any website’s traffic.
Does the 58% figure mean publishers lost 58% of their clicks?
No. Ahrefs’ February 2026 update measures the average click-through rate of the position-one result across 300,000 keywords, December 2023 against December 2025. It is a rate on a keyword set, not sessions on a site, and its counterfactual is modelled, not randomised. Checked 4 September 2026.
Why do two 34.5% figures appear in this story?
Coincidence, and it causes real confusion. Ahrefs’ April 2025 study reported a 34.5% reduction in position-one CTR. The 2026 field experiment reported a 34.5% increase in zero-click searches. Different quantity, direction and method. Neither corroborates the other.
Is the Australian 35% figure about AI Overviews?
Not demonstrably. It is a year-on-year change in estimated search traffic for one unnamed news site, from Similarweb data obtained by the ABC. The same ABC News report notes Ipsos Iris data showing combined top-ten readership rising over the same period.
Can I use any of these numbers in a B2B forecast?
No. A news-publisher or informational-keyword click-loss figure has no defensible mapping to a SaaS site whose valuable queries are branded, comparative and transactional. Use the studies for the mechanism, then forecast from your own baseline.
If a number cannot survive the question “what was the denominator?”, it does not belong in the deck. When you want a measurable outbound system behind the demand you can see, Apply For Partnership.