The attribution crisis is the real I/O story

Every piece written about Google I/O this month has fixated on the consumer demos. Universal Cart. Agentic booking. Information agents that monitor listings while you sleep. The framing is almost always experiential — look at this slick new way to shop, look at how Google is closing the gap between intent and action.

That framing misses the actual story. The consumer experience is fine. The consumer experience is, in some cases, genuinely better than what came before. The thing that broke at I/O wasn't the user journey. It was the attribution model that ran underneath it.

For twenty-five years, the implicit deal between Google and the businesses indexed in it has been: we'll send you traffic, and you'll figure out what to do with it. That deal had a measurement layer attached. You could see the click, log the session, run the funnel, count the conversion, attribute the revenue. Imperfect, contested, constantly gamed — but legible.

The agentic search Google demonstrated at I/O removes that legibility. Not as a side effect. As the design.

What Universal Cart actually does to your data

The mechanics are worth slowing down on, because the marketing language hides them.

Universal Cart lets a shopper add products from multiple merchants into a single Google-mediated cart that persists across sessions and surfaces. Agentic booking unifies pricing and availability from local service providers and lets the AI complete bookings on the user's behalf. Information agents monitor listings in the background and notify the user when conditions are met.

In every one of those flows, the user describes what they want. The agent does the comparison. The agent narrates the options. The agent closes the loop. Your site, your product page, your reviews, your brand story — all of it gets read, summarised, and reformatted by something the user never directly interacts with.

Google has been clear that the merchant remains the merchant of record. You still process the payment. You still own the customer relationship in the legal sense. What you don't own — what nobody outside Google owns — is the intent signal that led to the purchase.

Armando Roggio of Practical Ecommerce put it as cleanly as I've seen it framed: *"In Google's model, merchants still own the transaction, but not the purchase intent or product discovery."*

That is the entire game.

Why "you still get the sale" isn't the reassurance Google thinks it is

The transaction has never been the valuable part of the funnel. The valuable part is knowing *why* the transaction happened — which content earned the consideration, which review tipped the decision, which competitor was in the final two, which keyword the buyer started with, which time of day they were ready to commit. That data is what feeds every downstream decision a business makes. It shapes content investment, ad spend, product roadmap, pricing, retention.

Owning the transaction without owning the intent is owning the cheapest part of the funnel.

Strip it out and you're left with revenue you can count but can't reproduce. You know the sale happened. You don't know what to do tomorrow to make another one happen.

Owning the transaction without owning the intent is owning the cheapest part of the funnel.

The optimisation problem that follows is genuinely hard. How do you weight signals in agent-mediated flows when Google won't tell you which signals matter? How do you brief a content team on what to write when the audience for the content is a model whose ranking logic is opaque even to Google's own employees? How do you justify SEO investment to a finance director who, reasonably, asks how you'll measure return?

The honest answer right now is: you can't, fully. You can measure proxies. You can track brand search lift, citation share in AI Overviews, share of voice in AI Mode responses, referral traffic from the increasingly thin slice of journeys that still send a click. None of these are the conversion data you used to have. They're inferences about it.

![]()

The measurement vendors are racing to fill a hole they can't fill

Predictably, a tooling market has emerged. AI visibility platforms, citation trackers, brand mention monitors, GEO dashboards. Peec AI's recent finding that AI Overviews appear in 87% of commercial-intent queries is the kind of number these tools exist to generate. It is, to be fair to Peec, a useful number. It tells you the surface is dominant in the part of search that matters most commercially.

What it doesn't tell you — what none of these tools can tell you — is whether your appearance in any given AI Overview actually moved a buyer. The citation is measurable. The downstream behaviour mostly isn't. The tools are tracking the front door. The transaction is happening in a room they can't see into.

I've written before about the AI search measurement gap being a definition problem, not a tooling problem. Agentic commerce makes that gap worse, not better. The tools that show you AI Overview citations are measuring the discovery layer. They can't see the agent layer above it, where the actual buying decision is being made on the user's behalf.

This is the part the GEO tool category will struggle to admit publicly. Their entire commercial pitch is that visibility in AI surfaces equals influence on AI-mediated outcomes. In a click-based world that assumption was at least testable. In an agent-mediated world, the link between visibility and outcome runs through a black box owned by Google.

The asymmetry that gives the game away

There's a useful tell in Google's own behaviour. Earlier this year, Google added AI visibility metrics to Merchant Center for retailers — actual data on how products surface in AI shopping experiences. Search Console, the tool used by everyone else, still has nothing equivalent. I wrote about that contradiction when it landed, but the I/O announcements give it new weight.

Google is happy to give attribution data to advertisers transacting inside its commerce stack. It is conspicuously unwilling to give it to anyone else. That asymmetry is the entire economic shape of agentic search compressed into a product decision. The data flows to the surface that pays. Everyone else gets a citation, maybe, if they're useful enough to be summarised.

If you needed a single piece of evidence that the agentic shift is structured around Google's commercial interests rather than the user's intelligence, the asymmetry between Merchant Center insights and Search Console insights is it.

What this means if you actually run a business

The honest implications are uncomfortable, so I'll be direct.

The brands that get cited are the brands the model has heard of. That's the whole optimisation problem in a sentence.

If your business has historically depended on Google sending intent-loaded traffic to a website you control, the next two years will compress the economics of that model harder than the previous fifteen combined. The click was the unit. The unit is shrinking. You can't optimise your way out of a shrinking unit by getting better at the thing that's shrinking.

What you can do is shift the centre of gravity. Owned channels — email lists, communities, direct relationships, customer data platforms that capture intent before Google's agents intercept it — become the only places where intent data is still legible to you. This is not a new argument. I've made versions of it in the death-of-the-web piece and the Rand Fishkin response. What's new is the speed. Universal Cart isn't a 2030 problem. It's a 2026 problem.

The second move is brand. Not "brand awareness" in the soft sense — measurable brand search demand, the kind that shows up in Google Trends and that, as Tom Capper demonstrated again recently, increasingly out-predicts domain authority as a ranking signal in both classic search and AI surfaces. Brand is the only input that survives the agent layer. Agents cite what they recognise. Recognition is a cumulative asset, and unlike SEO rankings, nobody can take it away from you in a core update.

The brands that get cited are the brands the model has heard of. That's the whole optimisation problem in a sentence.

The third move — and this is the one most businesses will resist — is accepting that some portion of your funnel data is now permanently lost. Not delayed, not obscured by a privacy update, not waiting for a clever workaround. Gone. The agent layer is opaque by design and there is no commercial incentive for Google to make it transparent. Pretending otherwise leads to bad budget decisions.

The honest limits

A few things this argument doesn't cover, in fairness.

Agentic search is still early. The Peec data on commercial AI Overview prevalence is real, but actual user behaviour inside Universal Cart and agentic booking is, by Google's own admission, still small relative to traditional search. The economic risk is forward-leaning, not present-tense. A business that ignores this for another twelve months will not be wiped out tomorrow.

Regulatory pressure could change the trajectory. The EU's AI Act, ongoing antitrust action in the US, and the steady stream of publisher lawsuits all create real possibility that Google is forced to expose more data, share more attribution, or unbundle agentic commerce from search. I'm not optimistic about that, but it's not impossible.

And reasonable people disagree about how much intent data businesses ever truly "owned" in the click-based era. Google has been the gatekeeper for two decades. The attribution was always partial, always mediated, always somewhat fictional. The agentic shift is a degradation of an already imperfect signal, not a fall from a paradise that never existed.

Those caveats matter. They don't change the direction of travel.

The pivot that the smart teams are already making

The teams I see adapting fastest aren't the ones building elaborate GEO dashboards. They're the ones quietly de-emphasising last-click measurement entirely and rebuilding their analytics around leading indicators that don't depend on the click. Branded search trend. Direct traffic. Email list growth. Community engagement. Earned media mentions. Citation share across AI surfaces, tracked qualitatively rather than as a vanity metric.

This is uncomfortable for marketing leaders who came up in the era of attributable, defensible, spreadsheet-perfect ROI calculations. The honest move is to admit those calculations were always more confident than the data warranted, and that the new measurement reality is closer to how marketing actually worked before digital — directional, lagging, partly intuitive. Brand investment with an analyst's discipline rather than a performance marketer's.

The era in which a business could outsource its understanding of its own customers to Google Analytics is ending. What replaces it isn't worse, necessarily. It's just harder, and it requires marketing teams to develop muscles most of them let atrophy somewhere around 2012.

The businesses that take this seriously now have a window. The businesses that wait for a tool vendor to solve the attribution crisis on their behalf are going to be waiting a long time, because the tool vendors cannot solve it. The data they need is on the other side of a wall that Google has every commercial reason to keep up.

That's the loop. And the only way out of it is to stop optimising for the surface and start investing in the things the agent layer cannot mediate — relationships, recognition, and demand that exists before anyone types a query into anything.