You have read that AI destroyed 58 percent of publisher traffic. Or 70 percent. You have possibly also read that Google disputes those measurements and maintains that traffic to the open web is holding steady.
Both camps cite data. Both camps have an interest in the result. And both camps measure different things while using the same word.
The problem is not that somebody is lying. The problem is that these numbers do not share a denominator, and putting them in the same sentence produces a false conclusion.
This article takes the measurements apart one by one. It is the only article on this blog that never mentions GEO, and it is probably the most useful.
The measurements, one by one
Clicks per search, measured on a user panel. The Pew Research Center tracked roughly 900 users across nearly 70,000 searches in March 2025. Without a generated summary, a search produces an organic click about 15 percent of the time. With a summary, about 8 percent. Links inside the summary are clicked about 1 percent of the time, and the session ends outright in 26 percent of cases against 16 percent. Provenance marking: independent research institute, real panel, published methodology. Google publicly disputes this methodology, which is a fact you should attach to the number every single time you quote it.
Click-through rate in first position, measured on keywords. An analysis covering 300,000 keywords, comparing March 2024 with March 2025 through Search Console data, measures a move from 7.3 percent to 2.6 percent for the top organic result when a summary is present, a relative drop of 34.5 percent. Provenance marking: SEO tool vendor publishing data about its own market, which is a direct conflict of interest. Essential restriction: the scope is informational queries only, chosen precisely because they overlap heavily with the queries that trigger a summary.
Aggregate traffic, measured at publishers. In August 2025, a US publisher trade association surveyed 19 of its roughly forty members, including major titles, across eight weeks in May and June 2025, year over year. Median decline in Google referral traffic: 10 percent overall, 7 percent for news brands, 14 percent for the others. Most members lose between 1 percent and 25 percent. Provenance marking: self-reported by publishers, through a body that negotiates with the platforms on this exact subject.
Three serious measurements. Three results that look incompatible. They are not.
Why these numbers do not contradict each other
| Measurement | Numerator | Denominator | Scope |
|---|---|---|---|
| Pew | Searches followed by a click | Searches where a summary appears | All panel queries |
| Position 1 click-through rate | Clicks on the first result | Impressions of that result | Informational queries only |
| Publishers | Sessions received from the engine | All sessions from the previous year | Whole site, all queries |
A 34.5 percent drop in click-through rate on informational queries does not produce a 34.5 percent drop in total traffic. A site also receives traffic on brand, transactional and navigational queries, where the summary appears rarely or not at all.
Passing a number from the first row off as a number from the third row is the most common error in the entire debate. Sometimes it is accidental. Sometimes it is extremely convenient for whoever commits it.
And the most moderate figure comes from the source with the least interest in playing things down: a publisher association negotiating with the platforms on precisely this subject measures 10 percent, not 58 percent.
The numbers we could not verify
Two values circulate everywhere, and neither could be traced back to a primary source.
The claim that click-through decline reached “58 percent in December 2025” appears across dozens of secondhand write-ups with no reference to an identifiable publication. The “minus 70 percent” belongs to the same category: quoted constantly, sourced nowhere.
Those numbers may well be accurate. They will not appear on this blog until someone has opened the original publication, its methodology and its sample. In an article about the honesty of numbers, doing otherwise would be absurd.
The textbook case: one indicator, five values
The most instructive example comes from an infrastructure provider that has published, since July 2025, a ratio between the number of pages a platform’s crawlers fetch and the number of visits that platform sends back. Provenance marking: the calculation formula is published in full, the data is continuous and publicly consultable, and the provider sells no visibility services. It is the opposite of a black box.
And yet the values circulating for that same indicator range from 217 to 1 up to nearly 24,000 to 1.
None of those numbers is false. They correspond to different platforms and, above all, to different measurement windows. A transparent indicator, published continuously, produces a two-order-of-magnitude spread depending on the period somebody chooses to quote.
That is exactly where the statistical manipulation in this ecosystem lives. Not in inventing data: in selecting the window that yields the most spectacular figure, followed by a chain of republication that loses the window somewhere along the way.
When you see a ratio of this kind, the only useful question is: over what period, and for which platform.
The symmetric error: trusting your own dashboard
It would be convenient to conclude that the decline is overstated and move on. That would be falling into the opposite bias, and it would cost you more than the first one.
A June 2026 study, built on a panel that joins the real browsing behavior of consenting users to their conversations with three assistants, with event-study design and matched placebos, measures that a brand recommendation made to a user who did not previously know the brand lifts Google searches for that brand by 4.3 points, visits to its site by 2.4 points, and product pages at resellers by 1.0 point. Provenance marking: arXiv preprint, observational design, not peer reviewed.
That effect leaves no referrer trace at all. Your analytics tool will file it under brand organic or direct traffic.
So referral traffic measurements structurally understate the influence of assistants, at the same time as click-through measurements overstate the collapse of total traffic. Both errors are real, they run in opposite directions, and they do not cancel each other out cleanly.
Who benefits from what
This is the most useful reading grid available, and it is not cynical. It is hygiene.
The engine benefits from the decline looking small, because an ecosystem of angry publishers produces lawsuits and regulation. So it disputes unfavorable methodologies, which is sometimes entirely justified.
Sellers of AI visibility services benefit from the decline looking maximal, because fear funds the budget. They are the ones quoting the highest values most eagerly, usually with the scope stripped off.
Publishers benefit from the decline looking severe in their negotiations, and yet their own aggregate measurement is the most moderate of the three. That is exactly what makes it credible.
And independent institutes have no interest beyond their methodological reputation, which explains why they publish numbers that both camps find inconvenient.
What you do tomorrow morning
Stop importing sector averages into your decisions. Measure at home, with three separations that almost nobody performs.
Separate brand traffic from non-brand traffic. The first is barely exposed to generated summaries, the second is heavily exposed. An overall decline that hides this split cannot be steered, because you do not know which half moved.
Separate informational queries from transactional queries. The decline measurements cover the first kind. If your traffic is mostly transactional, the numbers being waved at you do not apply to you in the same degree, and you should say so out loud in the meeting.
Track searches for your own brand name over a rolling twelve months. This is the one place where the indirect effect of assistants becomes visible, and it costs nothing.
Then, faced with any number a vendor presents, ask three questions: which numerator, which denominator, which window. Three questions, and most of the public debate on this subject dissolves.
Traffic is falling. It is probably falling by an order of magnitude closer to ten percent than to sixty. And that is still very bad news, which does not need to be exaggerated to deserve your attention.
Sources
- Pew Research Center (2025). Google users are less likely to click on links when an AI summary appears in the results
- Ahrefs (2025). AI Overviews Reduce Clicks by 34.5%
- Digiday / Digital Content Next (2025). Google AI Overviews linked to 25% drop in publisher referral traffic
- Cloudflare (2025). The crawl before the fall of referrals
- Iannelli, M. & Ai, A. (2026). From Prompt to Purchase: How AI Brand Recommendations Move Consumers on the Open Web, arXiv:2606.10907
- Conductor (2026). AEO / GEO Benchmarks Report
<strong>LaFactory</strong> measures AI visibility with a published protocol: repeated measurements, paraphrases, control group. No guaranteed placement, ever. Contact us to scope an audit.