AI shopping and agentic commerce statistics: measured or projected?

Every figure, and whether it was measured or projected

No chart appears on this page, deliberately. Plotting a projection next to a measurement on the same axis is the single fastest way to make a forecast look like data, and in this category that confusion is the main error in circulation. The grades follow the scheme used across the wider statistics collection, with the tracing method in the provenance audit and the population, window and status rules in the AI visibility measurement standard, the parent method behind this page.

Why a 15× multiple tells you almost nothing

Multiples are the most quoted and least informative numbers in an emerging category. A 15× increase can describe a channel going from a rounding error to a small but real share, or one going from a small share to a dominant one. Those are completely different businesses, and the same multiple covers both.

The direction is still worth knowing: something is happening fast enough to be visible above noise, and that is a real observation. But a multiple without a base cannot tell you how much revenue is at stake, whether the channel deserves headcount, or when it might matter. Anyone quoting the multiple as a reason to act is asking you to supply the missing number from imagination. Both "AI shopping is exploding" and "AI shopping is still tiny in absolute terms" draw honestly on exactly the same figure.

Fact The base figure has not been disclosed in the coverage we reviewed. That absence is the most important property of these statistics and the part almost never mentioned when they are repeated. For rough scale, the surrounding audience numbers are at least public — conventional search share, ChatGPT's monthly users and generative-AI session counts, collected in the market share hub. None is the missing denominator; they only bound the plausible range.

How to read a forecast like the 40% figure

A forecast is a claim about the future built from assumptions about the present. It is not a measurement, and treating it as one is how planning goes wrong. Four questions make one legible: what base does it grow from; what growth rate does it assume and does that rate stay constant; what definition of the outcome is being counted; and who benefits if it is believed. The last is not cynicism — forecasts in emerging categories are usually produced by parties with a position in them, which does not make them wrong but does mean the error bars are rarely symmetric.

The definitional question matters most here. "Influence a transaction" is a very wide net. It could count a purchase an agent completed end to end. It could equally count a purchase where a person asked a chatbot a question at some point beforehand. Those two readings could differ by an order of magnitude, and the headline does not say which is meant. By contrast, the general AI-search work at least states its populations — prevalence panels and click-impact studies do, which is more than most forecasts manage.

Recommendation versus agent-completed purchase

Most "AI shopping" statistics describe the left column while being read as though they describe the right. Coverage of agentic shopping and preparation guides for AI browsers both note the blur. Read any platform claim against this distinction first.

What an agent actually has to do to buy something

A recommendation is one step. A completed purchase is seven, and each is a place the chain can break.

Most current activity stops at step four. That is not a criticism — discovery is genuinely valuable — but it means a statistic about AI-influenced shopping and one about agent-completed purchases measure different points in this chain. The last row is the one merchants underestimate: an agent-placed order that goes wrong still produces a human with a problem, and whose queue that lands in is unsettled with real cost attached.

What consideration-set inclusion plausibly requires

No rigorous, disclosed methodology exists for what gets a product into an agent's shortlist, so these are plausible factors rather than validated ones: accurate and complete structured product data so an agent can compare at all; current pricing and availability, since a wrong-on-stock recommendation erodes the agent's own reliability; and being reachable by the specific crawlers each platform uses to build product understanding.

Those crawlers are documented by OpenAI, Perplexity, Anthropic and Google, catalogued in the bot registry, with the access decision covered in the blocking census. Whether any of them execute JavaScript to see your prices is an open question, and whether structured markup helps at all is separately contested — observational work on schema and AI citations found little, which is what our schema study is designed to test properly. None of these has been isolated and tested for a causal effect on inclusion.

How to measure agent traffic on your own store

Nobody is going to hand you this measurement, and your own crude number is more decision-relevant than any industry figure because it describes your actual business.

Two rules make the difference. Behaviour beats identity — user-agent strings are self-reported and easy to imitate, so the fetch pattern (product pages, feed endpoints, structured data, whether scripts and assets loaded) tells you more than the name, which is the conclusion 30-day agentic crawler log studies also reach. And write down your classification rules before you start. The temptation to reclassify ambiguous traffic once you can see whether it helps the story is strong, and it is exactly how internal numbers become fiction.

Confounds in any agentic-commerce measurement

Bot contamination is the one that does the most damage to a headline number. Scrapers, price monitors and uptime checks all hit product pages, and Cloudflare's network-scale crawler breakdown and crawl-to-click work show how lopsided the mix is. Lumping them in with shopping agents inflates the figure dramatically. Referrer stripping pushes in the other direction. Neither error has a known size, which is why the trend is usable and the level is not.

What does not transfer between platforms

Fragmentation is not a temporary inconvenience here; it is the defining fact. Checkout protocols differ, so an integration built for one flow does not work with another — a direct engineering cost, repeated per platform. Product data requirements differ in field names, required attributes and freshness expectations, so a feed that satisfies one platform may be ignored by another. Crawler access differs, so a robots.txt decision can silently exclude you from one channel while leaving another open, which publisher-blocking studies suggest is common. The wider protocol picture — MCP, A2A, WebMCP and the rest of the stack — is moving fast and has not converged.

Hypothesis We expect consolidation eventually, because merchants will not maintain many parallel integrations indefinitely. When, and around whose standard, is unknown, and betting on the answer this early is a risk rather than a strategy.

What to do now, and what to defer

Split the work in two, because presenting it as one initiative is how a hedge turns into a gamble. Do now: clean structured product data — accurate price, real-time availability, honest specifications, in a format an agent can parse. That pays off under classic SEO and GEO too, close to what the one controlled study in generative visibility found and what the tactic scoreboard grades highest, so it is not a bet on agentic commerce alone. Defer: checkout integration, which is a bet on a specific protocol and should be sized and timed like one.

The costs of early integration are usually omitted from the pitch. Engineering time on a protocol that may not survive. Margin risk: intermediated channels historically extract a fee, and the fee is set once the channel matters, not while it is being adopted — a merchant who builds a dependency early has less negotiating room later.

Data risk: handing product, price and availability to an intermediary in machine-readable form makes competitive price monitoring trivial, which may be fine but should be a decision. And support risk: deciding in advance who handles a return on an order the customer did not personally place is cheaper than discovering it during the first incident. Against all that sits a real cost of waiting — if the channel matures and you have no structured product data at all, catching up takes months.

Who this applies to matters as much as the timing. It matters most to merchants selling comparable, specifiable goods — electronics, tools, components, commodity items — because attributes are what an agent can evaluate. It matters less where the decision is aesthetic or personal. It works differently for regulated and high-consideration purchases and for local businesses, where eligibility checks and licensed intermediaries impose constraints that have nothing to do with technology. And it barely applies to services sold through a conversation with a human, whatever the headline growth figures do.

Common misreadings

"AI is already 40% of ecommerce." It is a forecast about influence, not a measurement of transactions. Two different claims, with an enormous gap between them.

"Growth is 15×, so we are late." Late to what size? Without the base, urgency is manufactured. A channel can grow fifteenfold and still not be worth an engineer's quarter.

"Schema markup is agent optimisation." Structured data is necessary for comprehension — one of seven steps in the chain above. Doing it well does not get you through the other six.

"Agentic commerce is just SEO with a new name." The discovery half rhymes with GEO. The transaction half does not: payment, authorisation and aftercare are operational problems no amount of content work solves.

Null results we would publish

• Agent-attributed revenue remains negligible in absolute terms despite large multiples. Arguably the most likely outcome and the least reported one.
• Structured product data quality shows no measurable relationship with inclusion in agent recommendations — which would undercut the main piece of practical advice on this page, including our own version of it.
• Agent traffic cannot be reliably distinguished from other automated traffic at all. If the measurement is not possible with available signals, every figure in this category inherits that problem.
• A forecast retrospective. We would revisit the 40% projection at its target date and state plainly whether it held. Almost nobody goes back to check a forecast once it has done its promotional work.

All of those would land in the null results registry, alongside every other method in the studies index. What would change this page fastest is disclosure of the absolute base behind the growth multiples; second, convergence on a shared checkout protocol; third, an independent, method-disclosed measurement of agent-completed transactions across multiple merchants. We have not found one, and its absence is why this page reads the way it does.

Open questions

Open question What actually determines inclusion in an agent's shortlist? No disclosed-method study isolates any factor.

Open question Do agent-placed orders differ from human-placed ones in return rate, basket size or lifetime value? Merchants may know. Nobody has published it.

Open question Who owns the customer relationship after an agent-completed purchase, in practice rather than in protocol documentation?

Open question Does blocking shopping crawlers measurably reduce a merchant's presence in agent recommendations, or do alternative data paths fill the gap?

Verification status

The Shopify growth figures are Partial : real and specific, industry-reported, and published without an independently disclosed method or an absolute base. The 40% figure is Hypothesis by nature — a projection, not a measurement — and its outcome definition is undefined in the coverage we reviewed. No figure on this page has been independently verified.