Before you put a GEO line in the Q4 budget: Google's own docs say there is nothing to optimise.

Search Central states plainly that AI Overviews need no special markup, no AI text files, no schema. The one independent study of what users actually do is 18 months old and says a citation is worth a 1% click. Everything fresher is measured by the people selling the service.

It's Q4 planning season, which means a lot of small teams are about to write GEO on a budget line. Before you do, I'd spend twenty minutes on two documents, because they disagree with the pitch in a specific way.

The first is Google's own page on AI features and your website, in Search Central. It says there are no additional requirements to appear in AI Overviews or AI Mode, and no special optimisations necessary. Further down it's blunter: you don't need to create new machine-readable files, AI text files, or markup, and there's no special schema.org structured data to add. The whole page is a restatement of ordinary SEO fundamentals - be crawlable, be indexable, be useful, have your important content in text.

That's the platform saying, in its own documentation, that the product being sold to you does not exist as a distinct technical practice. You can decide Google is being coy. I'd just note that this is the one party in the conversation with nothing to sell you, and that it's the version I'd have to defend if I were wrong.

The second document is whichever GEO study is in your inbox this month. Mine was Previsible's 2026 State of AI Discovery - 6.77 million AI-driven sessions across 166 sites, November 2024 to May 2026, and genuinely the longest-running dataset of its kind. Its findings are plausible: Google's own AI surfaces drive more AI-influenced traffic than every standalone assistant combined, and ChatGPT carries 92.4% of the trackable standalone referral traffic.

I want to be careful here, because vendor data is usually deployed as a sneer and I don't mean it that way. Previsible describes itself as an AI discovery agency for enterprise GEO visibility. The report's recommendation is to build the signals that get you cited. The report is also, as far as I can tell, honest work by people who know more about search than I do. All three of those things are true at once. The problem isn't dishonesty, it's that there is no independent audit of the number, and the 166 sites are enterprise properties in SaaS, finance, legal, health and publishing. Your six-person app is not in that sample.

The one study nobody is selling.

There is exactly one large dataset here from an organisation with no product in the category: Pew Research Center's analysis, published July 2025, of 900 US adults who agreed to have their browsing metered for the month of March 2025. It covers 68,879 Google searches, of which 12,593 produced an AI summary.

Three numbers from it get quoted. Users who saw an AI summary clicked a traditional result in 8% of visits; users who didn't, 15%. Users ended their browsing session 26% of the time with an AI summary present versus 16% without. And the one that should matter most to anyone buying GEO: users clicked a link inside the AI summary itself in 1% of visits.

One per cent. That is the click value of the thing the entire optimisation industry is organised around obtaining. It isn't zero - a citation is a brand impression, and I'll come back to that - but if you're modelling GEO spend against clicks, that's your multiplier.

The part of that study nobody quotes.

Here's what changed my read, and it's in the same report. Longer and more conversational queries trigger AI summaries far more often. Just 8% of one- or two-word searches produced one. For searches of ten words or more, 53%. For searches starting with a question word - who, what, when, why - 60%.

Sit with that next to the 8%-versus-15% click figure. Those two groups of searches are not the same searches. The AI-summary group is disproportionately long, question-shaped, exploratory queries; the no-summary group is disproportionately short and navigational. Short navigational queries have always converted to clicks at a higher rate, AI or no AI. So the gap between 8% and 15% is measuring the AI summary and the difference between two populations of query, and the study doesn't let you separate them.

I'm not saying AI Overviews don't take clicks. I think they obviously do. I'm saying the cleanest independent number available is not a clean causal estimate, the report is transparent about its method, and almost every downstream citation of it drops that. Which is a good reminder that the problem with this whole debate isn't bad actors - it's that a two-input result keeps getting summarised as one scary number.

The twenty-minute exercise.

The query-length finding is the actionable one, and I've never seen it in a GEO pitch. Your exposure to AI Overviews is mostly a function of the shape of the queries that bring you traffic, not of anything on your page. So open Search Console and look at your top queries. If most of them are your product name and brand-adjacent variants, you are largely in the 8%-trigger bucket, and GEO spend is buying you very little. If most of them are ten-word questions - how do i stop overthinking before bed - you are in the 53% bucket and deeply exposed, and Google's position is still that there's nothing special to implement.

That's the whole exercise. You don't need a tool or a retainer to run it.

You can now also measure it first-party: Google introduced Search generative AI performance reports in Search Console in June 2026. Before that, AI-feature traffic was folded into the Performance report under the Web search type with no way to isolate it - which is precisely the vacuum the third-party estimates grew into. The vacuum is smaller now. Check your own console before you buy someone's model of your console.

What I'd hold onto.

The honest state of the evidence is uncomfortable: the independent study is 18 months old, US-only, one month long, and covers AI Overviews only - not AI Mode, not ChatGPT. Everything current is vendor-measured. That's not a reason to dismiss the vendor numbers; it's a reason to treat them as well-informed hypotheses with a price tag attached, rather than as findings.

The Murror version is short and slightly against my own interest in writing this. Our traffic is overwhelmingly brand and app-store, so we're barely exposed, and the twenty minutes told me not to spend the money. But I can't measure the thing that would justify spending it - whether being cited in an answer, at a 1% click rate, still puts us in someone's head for later. Brand impressions have always been the part you can't attribute, and AI answers made a lot more of your marketing into brand impressions. Anyone who tells you they've measured that in September 2026 is selling something. Possibly something good. Just ask them who counted.