Organic sessions fell 10.5% across 54 advertisers after AI Overviews spread, but AI referrals converted at a higher rate

Digiday reported on 25 August 2026 that a marketing services group had analysed its own Google Analytics data across 54 client accounts and found organic web traffic fell on average by more than 10% after Google's AI Overviews became common on search results pages. Across the whole client set, sessions dropped 10.5% over the 14 months from January 2025 to April 2026, from 140.1 million to 125.4 million, while sessions arriving from AI platforms rose 163%.

The second half of that finding is the part worth planning around. Over the same 14 months, key events driven by AI referrals rose 335%, and the rate at which key events happened was 1.5 times higher for visitors arriving from ChatGPT, Copilot, Gemini and Perplexity than for visitors arriving from organic search. Volume went down and the commercial density of what remained went up. Both statements are true at the same time, and only reading them together describes what actually changed.

What was measured, and over what window

The dataset is one company's own Google Analytics data for 54 clients: 24 United States firms, 24 British brands and six clients operating in other markets. 29 of the 54 were larger advertisers employing at least 1,000 staff. The roster spans 19 sectors and includes 16 retailers. The measurement window runs from January 2025 to April 2026, and 46 of the 54 clients saw a dip in organic sessions.

The study leans on two metrics. The first is sessions, which is straightforward. The second is what GA4 calls key events, meaning a visitor action the advertiser has defined as meaningful. The report gives three kinds of example: making a purchase, signing up to a newsletter, or scrolling to the bottom of a page. That spread matters when reading the 335% increase, because a key event is not automatically a sale. A brand that has configured scroll depth as a key event and a brand that has configured only completed checkouts are counting very different things under the same label, and the study does not break down which definitions sat behind its clients' numbers.

September 2025 is the hinge in the data

The study places the turn in September 2025, when AI Overviews began appearing on at least 30% of United States search results pages. The group's senior SEO manager described that threshold as an inflection point for the client set. After it was crossed, referral traffic from AI sources such as ChatGPT or Gemini rose to roughly 200,000 a month on average, while organic sessions fell from around 20 million a month to under 15 million.

Those two monthly figures need a careful read. As published they describe the client set rather than any single advertiser, so the ratio is the usable part of them: AI referrals were running at a little over one percent of the volume that organic search was still delivering each month. AI referral traffic was growing fast from a very small base, and it was not replacing lost organic sessions one for one. Anyone building a 2027 plan on the assumption that AI referrals will absorb the shortfall is reading something into this data that is not in it.

The headline numbers in one place

Every figure in the table below appears in the study as reported on 25 August 2026, and all of them cover the January 2025 to April 2026 window across the same 54 accounts.

Fewer sessions, denser sessions

The mechanism the study offers for the two-sided result sits in one quote from the group's senior SEO manager: "Overwhelmingly, [Overviews] are being triggered by educational, informative discovery keywords. It is starting to trickle down to transactional, commercial [searches]."

If AI Overviews absorb mostly informational queries, the visits they absorb are the early, exploratory ones. The report's reading is that the search traffic which has been lost likely represents casual users, while the people who still click through are further into the product discovery process and therefore more likely to buy or engage. The 1.5 times key event rate is the number that carries that argument, and it is a rate rather than a volume, which is why it can rise while sessions fall.

This next paragraph is analysis rather than a finding of the study, and it is worth stating plainly. A higher conversion rate on a smaller pool is not evidence that losing the pool is costless. Informational visits are also how a brand enters consideration in the first place, and 14 months is shorter than the buying cycle in several of the categories the study covers. The study measured sessions and key events. It did not measure revenue, brand recall, or what happens to a brand's pipeline two years after it stops meeting people at the research stage. Reading "quality went up" as "the decline does not matter" goes past what the data supports.

The categories that moved most

The dataset spans 19 sectors, and the variation inside it is wider than the 10.5% average suggests.

On retail specifically, the group's senior SEO manager said: "Retail was one of the least affected by AI overviews, which makes sense with the AI overviews not triggering as much [against] transactional searches."

Treat that split as directional. Spreading 54 accounts across 19 sectors leaves very few advertisers behind each category label, so these are patterns in a small sample rather than sector benchmarks you can hold your own numbers against.

AI referral traffic is probably being undercounted

The study comes with a measurement warning attached. Its author noted that estimates of AI referral traffic may be conservative, because some AI referral visits get recorded as direct traffic when a user opens another browser window to reach a site after a search instead of clicking straight through from inside the AI interface. She said brand teams monitoring performance need to be aware of that gap, and that brands should take conversion events into account for paid media planning rather than looking only at referrals and traffic. "We need to start taking this channel seriously," she said. "It's definitely a channel you should be including in your everyday measurement KPIs."

The same report cites Sensor Tower figures putting ChatGPT at 1 billion monthly active users in May, against 56 million for Claude, and the study's own data shows ChatGPT accounting for the largest share of AI referrals both before and after the September 2025 hinge. If you are going to build one AI referral segment before you build any others, the usage numbers point at which one.

What this study cannot tell you

The sample deserves as much attention as the percentages, because this is not an independent or randomised study of the web.

• It is one company's own client roster, measured with that company's own Google Analytics access and reported by that company. Clients of a single marketing services group are not a random sample of advertisers.
• It is heavily weighted to two markets. 24 clients are United States firms and 24 are British brands, leaving six clients across every other market in the world combined.
• It skews large. 29 of the 54 employ at least 1,000 staff, so it says more about how big advertisers fared than about small sites.
• The report as published does not describe controlling for seasonality, site migrations, budget changes, or the other Google ranking updates that ran through the same 14 months. AI Overviews spreading and organic sessions falling happened in the same period. That is a correlation the study describes, not a causal test it ran.
• The 58% figure attributed to Ahrefs, which the report cites as the traffic cut AI Overviews caused to some websites, is a separate measurement of a different sample. It is not the same finding as the 10.5% average and the two should not be averaged together or used interchangeably.
• There is no Thailand data in it. Not a small amount, none. The six clients outside the United States and Britain are not identified by market in the report.

What this means for Thai marketers

Start with what is genuinely unknown. No published measurement of AI Overview coverage on Thai-language results pages appears in this study or in the sources it cites. The 30% coverage threshold is a United States figure, from September 2025, on United States results pages. Quoting it in a Thai deck as though it describes Thai search is the single easiest way to build a plan on a number that was never measured here, and it happens constantly. If someone hands you a Thai AI Overviews percentage, ask which dataset it came from before it reaches a slide.

The practical response is to stop looking for a borrowed benchmark and to measure your own. Thai and English queries on the same site can behave differently, brand and non-brand queries certainly do, and your own property is the only dataset that describes your own market. The study's real transferable lesson is methodological: it tracked the rate of key events by traffic source rather than only session volume, and that is the comparison that made the counter-intuitive result visible at all. Any Thai advertiser can run the same comparison this week, and the answer will be about Thailand rather than about someone else's client roster.

The undercounting warning transfers as well. If AI assistant visits land in your reports as direct traffic, a Thai brand looking only at the referral line will conclude AI search is irrelevant to its business when it might simply be invisible in the report. That is a measurement result, not a market result, and the two get confused easily. Teams thinking about how their brand gets surfaced inside AI answers can start with our work on generative engine optimisation, and teams whose analytics cannot yet separate these sources cleanly may need to fix the tracking layer first through a GA4 setup review.

How to derive your own number from GA4 and Search Console

None of this requires a vendor, and it can be done with data you already own.

1. Pick a baseline that predates the change you are testing. The study used January 2025 as its start and treated September 2025 as the point when conditions shifted. Choose your own comparable window rather than assuming those dates apply to your market.
2. In GA4, split sessions by source rather than looking at a single organic total. AI assistants appear under their own referrer hostnames, so check what your property is actually recording before assuming which ones show up.
3. Compare the rate of key events in GA4 by source, not the count. The 1.5 times figure in the study is a rate comparison. A count comparison would have shown AI referrals as a rounding error and hidden the finding entirely.
4. In Search Console, split queries into informational and commercial groups and track impressions against clicks separately for each. If the pattern in this study holds for your site, the gap will open on informational queries first, with click-through rate falling while position stays flat.
5. Check your direct traffic line for the same period. A rise in direct that tracks a fall in organic is the fingerprint of the undercounting problem the study describes, and it is worth investigating before you conclude that AI referrals are negligible.
6. Repeat the measurement on a fixed schedule. One reading tells you nothing about a trend, and this is a moving target.

Brands that want the informational queries back in play through their own content, rather than conceding them, will find the work sits closer to optimising for AI assistants than to conventional rank chasing.

Frequently asked questions

Did AI Overviews actually cause the traffic drop?

The study reports a correlation rather than a proven cause: organic sessions fell 10.5% over a period in which AI Overviews spread across search results, and the report does not describe an experiment isolating Overviews from seasonality, site changes or other ranking updates in the same 14 months. The senior SEO manager quoted in it links the September 2025 coverage threshold to an inflection point in the client data, which is a reasonable reading of a timeline, not a controlled test.

Is any of this measured in Thailand?

No. The dataset covers 24 United States firms, 24 British brands and six clients in other markets, and the report does not identify those six markets or state anything about Thai-language search results. Any Thai figure you see attached to this study has been added by someone else.

Does this mean I should spend less on SEO?

The study does not say that, and its own numbers cut both ways: organic search was still delivering under 15 million sessions a month across the client set after the decline, against roughly 200,000 from AI sources, so organic remained the far larger channel even at its reduced level. What the data supports is measuring the two sources separately and judging each on key event rate as well as volume.

How much of my traffic should be coming from AI platforms?

The source gives no benchmark, and the monthly figures it publishes describe a 54-client aggregate rather than a per-brand target. The honest answer is that there is no published share to aim at, which is why measuring your own trend over time is more useful than comparing yourself to a number from someone else's roster.

Do I have to change anything in my analytics setup?

The study does not prescribe a setup, but it does flag that some AI referral visits are recorded as direct traffic, which means a default report can understate the channel. Checking how your property attributes those sessions, and whether your key events are defined consistently enough to compare across sources, is the part you can act on without waiting for more research.

If you want a read on how your own organic and AI referral traffic has moved over the last year, and what your key event rate looks like by source, that analysis is worth doing on your own data before anyone quotes an American percentage at you. Our team is happy to look at it with you.