The zero-click number everyone cites is the wrong number

For four years, the SEO industry has argued about a single figure. Somewhere between 40% and 60% of Google searches end without a click, depending on which study you cite and which year it was run. The number has been used to declare SEO dead, to declare it fine, to sell tools, to justify pivots, and to fill about half the LinkedIn posts I scrolled past this morning.

Mike King and iPullRank just published an analysis of 13.1 billion real search events from the Datos clickstream panel, and the headline finding isn't that zero-click is high. Everyone already knew that. The finding is that "the zero-click rate" isn't actually one number. It's a distribution so wide that citing the aggregate is worse than useless — it's actively misleading.

I've been saying a version of this to clients for eighteen months, usually while trying to talk them out of a panic response to a stat they read in a newsletter. The Datos data now says it more precisely than I could. The strategic mistake most businesses are making right now isn't ignoring zero-click. It's treating it as a single force acting equally on their entire site.

The distribution is the story

When you cut 13 billion search events by platform, country, age, query type, and destination, the zero-click rate doesn't wobble around a mean. It shatters. Some search cohorts sit near 20%. Others sit near 70%. There is no single "zero-click rate" that describes what's happening on Google in any useful way.

This matters because almost every strategic conversation I've had about AI Overviews and zero-click behaviour in the last year has assumed a uniform effect. *Traffic is down because of AI Overviews.* Sometimes. *We need to optimise for zero-click.* For which queries? *Search is dead.* No — search is fragmenting into cohorts that behave completely differently, and averaging across them tells you nothing about your own situation.

The iPullRank finding I keep coming back to is that some SERPs have become genuinely closed loops — the query resolves inside Google and the click never happens — while others are still reliably sending traffic. These aren't the same product. Treating them as one problem gets you a strategy that's wrong for both.

"The zero-click rate" isn't a number. It's a distribution, and the distribution is the strategy.

What the cohort split actually reveals

The Datos data shows Google absorbs roughly a third of Wikipedia's would-be traffic but barely touches Reddit's. That's not because Reddit is doing better SEO. It's because the two sites answer fundamentally different query types. Wikipedia gets factual, extractable queries — the kind Google can (and does) answer inside the panel. Reddit gets experiential, opinion-shaped, "what do people actually think about X" queries — the kind Google can't summarise without losing the thing that made the click worth making.

The lesson isn't "become Reddit." The lesson is that some content types are structurally exposed to zero-click and others aren't, and the difference is baked into what the content is for. If your site is built to answer factual queries with tidy, extractable answers — the exact shape SEO has been optimising for since 2015 — you are the raw material AI Overviews are eating. If your site is built to give the reader something they can only get from you (an opinion, a specific experience, judgement, a proprietary dataset, a working relationship), the zero-click effect looks completely different.

I've watched two of my own clients navigate this over the last year. One runs a home services business — highly transactional, factual queries, all local. Their zero-click exposure is real but bounded, because "book a plumber in Salford" still needs a click to complete. The other publishes analytical content in a professional niche. Their organic traffic is down, but the traffic they still get converts at a much higher rate than it did two years ago, because the queries that used to send them curious top-of-funnel readers now get answered by AI, and only the readers who genuinely need what they specifically offer make it through.

Both are fine. Both would look identical on a dashboard that reports "organic sessions" as a single number.

The ChatGPT ads finding nobody's talking about

Buried in the Datos analysis is a detail that reframes the whole discussion. ChatGPT now pulls more of its outbound clicks from ads than from any other top destination.

The AI product the industry has spent two years calling a threat to Google's ad business is, at the click level, functioning as an ad delivery system.

Read that again. The AI product the industry has spent two years calling a threat to Google's ad business is, at the click level, functioning as an ad delivery system. Not as much as Google, obviously. But the direction of travel is not "AI kills paid search." It's "AI becomes another paid search surface, on different economics."

This matters for how you plan the next two years. If you've been holding Google Ads spend because you assumed the AI shift would depress the entire paid discovery model, you're operating on the wrong assumption. The paid model is migrating with the users. It's arriving in AI surfaces in new formats, on new terms, and — if I had to guess — with worse targeting and worse measurement for a while. But it's arriving.

I wrote last week that Google reports revenue to the penny and reports web traffic in adjectives. The corollary is that wherever revenue can flow, ads follow. The question was never whether AI search would monetise. It was how, and how quickly. The Datos data suggests: through ads, and faster than most people think.

Why entity mapping and GEO can't fix a distribution problem

Duane Forrester wrote a piece today that I'd been waiting for someone to write. The short version: entity mapping — the tactic being sold as the GEO answer to AI search — works on Google because Google has a real knowledge graph you can influence indirectly. It doesn't work on ChatGPT the same way because there's no equivalent graph to feed. The model's "knowledge" of your brand is a diffuse statistical pattern smeared across billions of weights, learned from a corpus at scale, not from your schema.

I'd extend his point. The reason so much GEO advice sounds plausible and produces so little measurable movement is that it's applying a Google-shaped tactic to a fundamentally non-Google system. And the reason the industry keeps repeating this mistake is that the vocabulary from 2021 still works — "entities," "disambiguation," "structured data" — even though the target has silently moved.

Now stack this against the Datos finding. Zero-click behaviour is cohort-dependent. AI systems are structurally different from Google. And what most agencies are selling right now is a single GEO playbook, applied uniformly across a client's entire content estate, based on tactics developed for a different system.

That's three category errors compounding. The playbook is wrong for the system, uniform where the underlying behaviour is fragmented, and derived from a metric (aggregate zero-click) that doesn't describe reality.

Most strategy decks being shown to UK businesses right now are operating on assumptions the data has just contradicted.

What the cohort view actually implies for the work

The practical shift is not complicated, but it does require throwing out a lot of dashboard reporting.

You need to know which of your queries are in the closed-loop cohort and which aren't. The closed-loop queries — factual, extractable, easily summarised — are gone or going. Not maybe, not eventually, gone. Optimising harder for them is throwing effort at a shrinking pool. The queries that still generate clicks are the ones where the reader either can't get what they need from a summary (experiential, opinion, judgement, transactional) or can't complete the task without leaving the SERP (booking, buying, contacting).

The second implication is that content strategy needs to segment by exposure. If half your content is aimed at the closed-loop cohort and half at the click-generating cohort, they need different KPIs, different production models, and different budgets. Reporting them as one bucket hides the fact that half your investment is now going into a channel that's structurally shrinking.

The third implication, and this one I keep having to argue for with clients, is that the content that survives is disproportionately the content that couldn't be written by anyone else. Not "high quality" in the generic content marketing sense. Specifically: content with a fingerprint. First-party data. A named person's judgement. A working consultant's actual experience. The stuff AI can't extract without losing what made it worth reading.

The rest is being commoditised at speed. I've written about the 89% of AI search categories with no clear owner — the window for claiming those is genuinely open, but only for content that has the fingerprint. Undifferentiated coverage of a category is now worth roughly what it costs to produce, which is approximately nothing.

The honest limits

The Datos panel is real behaviour from 9.1 million users, which is a much bigger sample than most zero-click studies. But it's a panel — clickstream data captured from opted-in users — and panels always have selection effects. Heavy internet users, US-skewed, probably over-indexed on knowledge workers. The cohort variance is real. The absolute levels might be somewhat different in your specific market.

I'd also caveat that the "ChatGPT clicks come from ads" finding is a snapshot of a fast-moving product. ChatGPT's ad model is barely a year old. The composition next year will look different. What I'd bet on is the direction: AI surfaces monetise through ads, and the share will grow. What I wouldn't bet a strategy on is the current percentage staying stable.

And zero-click isn't zero-value. Search Console impressions that don't produce clicks still contribute to brand recall, still show up in the mental availability that matters for direct traffic, still feed into whatever AI systems are learning about who you are. Reporting only on clicks understates the effect of visibility, and I've had clients pull budget from campaigns that were doing genuine work because the click column looked flat.

Where this leaves the argument

The zero-click debate has been stuck for years because the industry insisted on a single number for what was always a distribution. iPullRank's data doesn't resolve the debate. It dissolves it. There's no useful aggregate zero-click figure, because the aggregate is the average of two or three cohorts that behave nothing like each other.

The strategic question stops being "how do we respond to zero-click search" and becomes "which of our queries live in the closed-loop cohort, and what should we do differently for the ones that don't." That's a much more useful question. It's also a much harder one, because it requires actually looking at your own data instead of citing an industry-wide stat.

Most businesses won't do that work. They'll keep chasing whichever GEO tactic gets pitched to them next, on whichever podcast their CMO listened to on Tuesday. The ones that do the segmentation work will spend the next two years quietly compounding advantage against competitors who are still optimising for a number that doesn't describe reality.

That's the loop. And we built it.