Everyone is looking for the same thing. A number. A graph. A clean line that says: here, this is where AI sent you traffic, this is what it was worth, this is your return. And the reason nobody can find it isn’t because GEO doesn’t work.
They’re looking in completely the wrong place, with completely the wrong tools, asking a question that attribution was never designed to answer.
Let me explain why.
Why the data gives you the wrong answer
Clicks as the basis for attribution exist mainly because of availability bias.
When marketing measurement evolved from exposure metrics, through brand lift, eventually to clicks, the tools shaped the thinking. Last-click attribution wasn’t designed to measure influence. It was designed to measure the final navigation event before a transaction. These are not the same thing, and they never were.
We just forgot that, because for a while, with search, they more or less coincided.
AI breaks that coincidence wide open. And now the seams show.
How attribution actually works – and what it was never built to do
Think about how you actually make a decision to buy something.
Your friend tells you about a restaurant. You remember it three weeks later, book a table through Google, the restaurant’s organic listing gets the win. Your friend gets nothing in any attribution model.
You watch a YouTube video where someone reviews a running shoe. You think about it for a few days, then search for the brand directly. Google gets the credit. YouTube gets nothing.
You log into Netflix. It recommends a film. You watch it. Do we therefore conclude that Netflix had no influence on your viewing decision because it doesn’t appear in any attribution chain? Of course not. That would be absurd.
But by the same logic that SEOs are applying to GEO, no click, no attribution, no channel value, Netflix’s algorithm would be a distraction, when we know it’s one of the biggest parts of the system.
I ask my mum what my dad wants for Christmas. She suggests something. I go and search for it. Google gets the credit. My mum gets nothing. She was, obviously, the entire reason I searched.
This is what AI is doing.
I open ChatGPT. I ask what the best platforms are for buying a used car. It gives me five. I open a new browser tab (and I do open a new tab, nearly everyone does) and I search for three of them. Organic search gets the win on all three. The AI recommendation that determined which three I searched for gets nothing.
Not a click, not a session, not a conversion event. And from this, some SEOs are concluding that GEO isn’t a real channel. That’s not analysis. That’s confirmation bias wearing a lab coat.
The attribution model isn’t broken. It was just built for a world where the click was the decision. That world is receding.
What the surveys actually say – and why it matters more than the traffic data
While the analytics crowd has been hunting for a signal in direct traffic, a completely different body of evidence has been quietly accumulating. Consumer surveys, multiple of them, across different markets, different methodologies, different sample sizes, are telling a consistent story. And it’s not a subtle one.
Klaviyo surveyed nearly 8,000 consumers across eight countries in late 2025. 41% reported purchasing a product that AI had recommended in the previous six months. Not considered it. Not researched it further. Purchased it.
More than one in five said AI is now their first stop when they’re making a decision, learning something new, or solving a problem. This is not a niche behaviour. This is mainstream consumer behaviour at scale, recorded across the US, UK, Australia, Germany, France, Spain, Italy, and Singapore.
Attentive went further and got closer to the question that matters. They surveyed 3,054 US adults who had shopped in the previous three months, 68% of whom had used an AI chatbot to help them shop. Among that group of AI-assisted shoppers, 71% said AI meaningfully affected which products or brands they seriously considered or chose, on at least half the occasions they used it to shop. 62% bought a different product or brand from their usual choice because of what an AI told them.
That’s brand switching caused by AI recommendations. Not theoretical, not hypothetical, not “I might consider it”. Done.
And then, and this is the bit that should make every attribution sceptic pause, 93% of those same shoppers sought information or proof outside of AI before completing the purchase. They went to Google. They went to a brand site. They went to a review platform. The AI contribution evaporated from every attribution model the moment they did.
Semrush and Exploding Topics surveyed 2,338 US adults in July 2026. Around 58% of AI users said they’d made a purchase based on an AI chatbot recommendation. Nearly 59% had discovered a new brand or product through AI.
And, this one cuts both ways, 57.5% had decided against a purchase because of what an AI told them. AI isn’t just a recommendation engine. It’s a veto machine too. If your brand is being described inaccurately, or your product claims don’t hold up to the way a language model summarises them, you’re losing buyers before they’ve typed your name into a search bar. That loss will never appear in your analytics.
CloudNine PR and TLF Research ran the same exercise in the UK, with 2,564 consumers surveyed in April 2026. Among AI users, 52% said AI had helped them discover brands they would otherwise never have found. 48% said they would consider buying from an unfamiliar brand on the basis of an AI recommendation. 79% wanted to verify elsewhere first, which takes us straight back to the Attentive finding. The recommendation happens in AI.
The transaction happens somewhere that gets the credit. The two events are completely separated in your data.
Similarweb spent six months tracking what actually happens in the days after ChatGPT names a brand. Users were 2.5 times more likely to visit a recommended brand’s site than a direct competitor within the following week.
When ChatGPT recommended Capital One, 14.2% of users visited Capital One that week, against 3.8% for American Express. When it recommended Kayak, 12% visited Kayak, against 3.4% for Skyscanner. When it recommended Sephora, 7.9% visited Sephora, against 3.3% for Ulta. Across all three categories studied, finance, travel, and beauty, the recommended brand consistently pulled more visits than its rivals. And 56% of those visits arrived through search. Not through an AI referral link. Through the search bar. Branded search got the credit. AI got nothing.
The messy middle was always messy
Underlying demand doesn’t change because AI has arrived.
You can’t get me to buy a Christmas tree in January. You cannot get people to buy things they don’t want, in markets they’re not in, at times they’re not looking.
The total volume of people buying used cars this month is roughly what it was. The total volume of people looking for a software platform, a hotel, a marketing agency: roughly unchanged.
AI hasn’t created new demand. What it’s doing is changing how that existing demand gets allocated between brands.
And this is what the aggregate data can’t show you.
Traffic has to land somewhere. It always did. AI just changes which brands it lands with.
The more famous, the more well-known, the larger the digital footprint, the more likely you are to be recommended. That’s just preferential attachment.
The recommended become recommended. The visible become more visible. Big brands will mostly hold their ground because they already dominate the training data and the recall patterns.
The real competition is happening in the middle of the pack, where two brands with similar footprints are swapping in and out of the bottom two slots of a five-brand recommendation set. Highly unstable. Highly winnable. Almost entirely invisible in aggregate data.
What being recommended more could mean… and what it might not do.
Going from a 1-in-100 chance of being recommended to a 2-in-100 chance is a 100% relative increase.
I want to be careful here, because this number can be made to sound like anything depending on what you assume underneath it.
So let’s be precise. That increase in recommendation probability only converts to commercial value if three things are true: there are enough real buyers behind the prompts you track, they behave in response to the recommendation, and the economics of your market make incremental wins significant.
If you’re selling a service at £100,000 per contract and you’re operating in a category where buyers use AI to build their shortlists (and they do, increasingly) then an additional twenty recommendations a year is not a rounding error. It could be significant. If you’re selling fast-moving consumer goods to people who make purchase decisions in-aisle, AI search probably isn’t your lever. Not every market is equally exposed to this. Be honest with yourself about yours.
The commercial question isn’t “does GEO produce trackable traffic?” The commercial question is: “Are people in my market using AI to make decisions, and does being recommended or not change who they buy from?” If the answer to both is yes, the economics are in your favour.
GEO is a direction, not a distraction
Let me state this plainly: anyone telling you that GEO is a distraction because you can’t find it in your attribution data is, knowingly or not, asking you to ignore a channel that sits at the top of nearly every Google search result and inside the default behaviour of hundreds of millions of people’s research journeys.
That’s not scepticism. It’s either ignorance or self-interest.
GEO, done well, is just targeted good brand marketing.
It’s strengthening brand co-occurrence. It’s making sure your product performance attributes are clearly articulated and easily extractable, not buried behind SEO waffle or the kind of brand copy that made sense in the Mad Men era and makes no sense to a language model in 2026.
It’s making your brand famous for the right things, in the right contexts, in ways that AI can retrieve and relay accurately. I call it push marketing. We used to pull people off search engines. Now we’re pushing people towards the consideration set before they’ve even opened a search bar.
The ROI question, the £30,000 to £150,000 question, is really just the same question you’ve always asked about brand marketing. What’s the return on being famous? What’s the return on being the first brand someone thinks of when they’re in your category?
You know the answer to that question. You’ve always known it. You just didn’t have to ask it about AI before, because AI wasn’t sitting on top of the buyer journey. Now it is.
The question isn’t whether attribution models validate GEO. They can’t. They weren’t built for it. The question is simpler than that: do people in your market use AI? And do you care what it says about you?
If yes to both, there’s nothing to debate.
If you’re not bothered what AI says about you because you’re all-in on paid search and the numbers are good, fine. I disagree with you, but that’s a defensible position.
But if you care, if you believe AI is touching the consideration sets of buyers in your market, then anyone telling you to ignore it is trying to distract you from the channel with a measurement problem.
And measurement problems are not the same as marketing problems. Brand marketing has always had a measurement problem. The answer was never to stop doing brand marketing.
See how your brand is being recommended. Understand whether it’s being recommended well. Work to improve it. And then ignore it if you dare.
My view is that if you do, it’s a risk. A quiet one. The kind that doesn’t show up in your spreadsheet until someone else’s brand does. And over a long enough time, ignoring AI search might make you go broke.
Andrew Holland