01 — What happened

For two years, the effect of AI search on website traffic was an argument. In the first week of October it stopped being one. On 5 October, Digital Content Next published a summary of the evidence accumulated through 2026. The studies agree — and they agree on something more specific than the headline.

The first is a pre-registered randomised field experiment by Saharsh Agarwal of the Indian School of Business and Ananya Sen of Carnegie Mellon. A Chrome extension randomly assigned 1,065 US desktop users either to normal Google Search or to a version with AI Overviews stripped out, across 68,089 searches in January and February 2026. Overviews appeared on roughly 41% of queries. Where they appeared, outbound organic clicks per search ran at 0.37 against 0.62 when they were hidden — a reduction of 39.8%. The probability of a search ending in no click at all rose from 0.54 to 0.73. Placement decided almost everything: removing the top-of-page overviews, more than 87% of the total, raised outbound clicks by 88%.

Then the finding that almost nobody quotes. Informational queries made up 71% of the sample and drove the entire result, gaining 0.26 outbound clicks per search when the overview was hidden. Navigational and transactional queries — the other 29% — produced estimates too noisy to distinguish from zero.

The second experiment, posted on 18 August by researchers at the University of Pennsylvania and Northeastern University, forced participants into Google’s AI Mode for seven days. Click-through fell by 18.8 percentage points. Satisfaction dropped 0.73 standard deviations and users’ sense of control 0.66.

Pew Research Center, observing 68,879 searches by 900 US adults, found a traditional result clicked on 8% of visits where an AI summary appeared, against 15% where none did. Links inside the summary were clicked on 1% of visits.

And the network data matches. Chartbeat’s 2026 Publisher Playbook, reported on 24 September, put Google Search referrals across its publisher network down 40.2% between July 2025 and July 2026, against a 21.9% fall the year before. Search’s share of pageviews went from about 9% in July 2024 to 5% in July 2026.

The number that gets left out of the coverage: AI referrals are 0.01% of traffic in that dataset, and they fell during 2026 across ChatGPT, Claude, Perplexity, Copilot and Gemini.

02 — Why it matters

Two things follow, and neither is “SEO is dead”.

The replacement channel does not exist yet at measurable scale. Assistant usage is climbing steeply — Sensor Tower put AI assistant traffic up 86% year on year — while generative AI referrals sit at roughly 0.7% of overall web traffic and 0.01% of Chartbeat’s publisher network, declining. An industry is selling optimisation for a channel that sends almost nothing. That may change. It has not yet, and the direction inside 2026 was down.

The loss has a shape, and the shape is commercially decisive. Someone typing a question gets their answer and leaves. Someone looking for a supplier, or looking for you by name, still clicks. In randomised data that distinction is the difference between losing 40% of your clicks and losing nothing measurable.

So the honest question is not “how much traffic have we lost?” but “what kind of traffic were we getting?” Most businesses cannot answer it. That is the real problem, and it is a measurement problem rather than a marketing one.

03 — The economics

For twenty years, search supplied cheap strangers. Chartbeat’s reader breakdown shows where the withdrawal lands: loyal readers flat, returning readers down 8%, new readers down 14%. Relationships held. Strangers went.

For most businesses the revenue effect is far smaller than the traffic effect. Take an illustrative service business: 6,000 organic sessions a month, 4,200 of them landing on guides and how-to pages, 1,800 on pages describing what the business actually sells. Assume the informational pages convert to enquiry at 0.3% and the commercial pages at 6%. That is about 13 enquiries from informational traffic and 108 from commercial. Now apply the research: the informational 4,200 falls 40%, the commercial 1,800 holds. The business loses around five enquiries a month out of 121. A 28% fall in sessions, a 4% fall in enquiries.

Those conversion rates are assumptions, not findings. The ratio is the point: two businesses can read the same 40% headline and face entirely different problems.

Because the inverse is also true. A business whose organic traffic was overwhelmingly informational — one that monetises attention, sells information, or generates leads from content volume — has not lost a slice of a channel. It has lost the channel.

What has risen for everybody is the price of a stranger. Organic strangers are scarcer, so the marginal stranger now comes from paid, and paid is getting harder to steer by hand: Google and Microsoft have been retiring keyword-level control, leaving conversion data as the main lever. Higher acquisition cost raises the lifetime value a customer must reach, which makes retention the thing the arithmetic depends on.

And none of this can be calculated without source data. If enquiry source is not recorded at the point of enquiry, the analysis above is unavailable to you. You are left comparing a traffic chart to a revenue chart and guessing.

04 — What businesses should do

1. Split your organic traffic by intent before spending anything. In your analytics, separate landing pages that answer a question from pages that describe what you sell, and compare conversion to enquiry. One afternoon’s work. If the informational pages convert under half a percent and the commercial pages convert at several percent, you have lost volume, not revenue, and should act accordingly.

2. Check branded and local search separately. Searches for your name, and “near me” searches, are navigational and transactional — the categories the trials could not distinguish from zero. If those held steady while your guide pages collapsed, do not restructure your website. Nothing is wrong with it.

3. Price any AI-visibility proposal against your own referral data. Before signing a generative-engine-optimisation retainer, pull referrals from chatgpt.com, perplexity.ai, gemini.google.com and copilot.microsoft.com for the last 90 days. If that is under 1% of sessions — and every dataset above says it will be — judge the fee against that number, not against the growth in assistant usage. The channel is worth tracking long before it is worth paying to optimise for.

4. If you take 500 enquiries a month and record the source of none of them, fix that first. It is a required field on a form and a field on a record. Until it exists, every decision here is an opinion.

5. Put the effort into the pages that still get traffic. Commercial-intent traffic survived. Lifting conversion from 6% to 7% on 1,800 sessions returns more enquiries than recovering a third of the informational traffic.

6. Stop renting the whole relationship. The traffic that held up was people who already knew the name — an argument for capturing contact details and following up systematically rather than re-acquiring strangers at a rising price.

05 — Where technology fits

For most businesses here, the answer is not AI. It is three unglamorous things, in order.

Measurement. A source field captured at the enquiry form and carried intact into wherever you keep customer records. That is an integration and a form change, not a project. It is also the only thing on this list that makes the other two decidable.

Website conversion work on the commercial pages. Not a redesign. The specific pages where buying-intent traffic lands, which the research says you still have.

A CRM and follow-up automation for the relationship you now have to own. With the caveat from our own archive: a CRM is not a growth strategy, it is where the strategy gets recorded. The work is connecting enquiry capture, source, follow-up and outcome into one chain — ordinary connected-systems work.

There is a real “nothing at all” case here, and it is large. If your enquiries arrive by referral, repeat business, tender or phone, and organic search was never a meaningful share, this development is noise. Record your enquiry sources so you can prove that, and ignore the rest.

The thing not to buy is an AI-visibility dashboard. A dashboard monitoring 0.01% of your traffic is a recurring cost attached to a rounding error.

06 — RevenueStack’s take

Our view: the important result is not the 40%. It is the null result sitting beside it.

In the Agarwal and Sen experiment, removing AI Overviews did not make users more satisfied. Satisfaction, perceived information quality and ease of finding information were all unchanged — the authors call them “precisely estimated nulls”. People are not suffering a worse search engine and waiting for their clicks back. They are getting what they wanted. The Penn and Northeastern experiment found dissatisfaction only when users were forced into AI Mode, which is a finding about compulsion, not about summaries.

So this is not a distortion to be waited out, litigated, or optimised around. It is the new baseline.

The second thing we would put money on: most businesses are about to spend real money solving a publisher’s problem. The publishers have one: their product was informational traffic, precisely the category the trials show collapsing. A business selling services or products mostly does not, and the evidence says so explicitly on the queries where people are trying to buy.

The businesses that come out of this well will not be the ones who moved fastest on AI search. They will be the ones who spent an afternoon working out whether they had a publisher’s problem before they bought the publisher’s remedy.

Sources

  1. The evidence is clear: AI search is costing publishers — Digital Content Next, 2026-10-05
  2. AI Overviews cut publisher clicks 39.8% in first randomized study — PPC Land, 2026-07-04
  3. Study Confirms Google AI Overviews Cut Organic Clicks 38% — Search Engine Journal, 2026-04-27
  4. AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence — arXiv (University of Pennsylvania and Northeastern University), 2026-08-18
  5. Google Search Decline Steepened To 40% YoY For Publishers — Search Engine Journal, 2026-09-24
  6. Google users are less likely to click on links when an AI summary appears in the results — Pew Research Center, 2025-07-22