AI Shoppers Are Already Buying: Your Analytics Just Can't See Them

There's a familiar way this conversation usually goes. Someone brings up AI shopping, everyone nods, and then somebody says "yes, but that's a few years out." The room relaxes. It's a future problem, and future problems don't need a budget line.

That instinct was right about eighteen months ago. It is now wrong, and the gap between when it stopped being true and when most teams noticed is where the money is going.

The numbers moved faster than the conversation

Three data points, each from a different angle, all pointing the same way.

Volume. Generative-AI referral traffic to US ecommerce sites was up roughly 4,700% year over year as of July 2025, according to Adobe Analytics. Through late 2025 it was still running +830% year over year, and Adobe's own framing is that it has been roughly doubling every two months since September 2024. Compounding curves are easy to dismiss early, because early on the absolute numbers look trivial. They stop looking trivial very suddenly.

Quality. This is the part that should change how you feel about it. A year ago, shoppers arriving from AI were substantially less likely to buy than shoppers from traditional search, somewhere between 43% and 80% worse, depending on the month. That gap didn't just close. It inverted. Over the 2025 holiday period, shoppers arriving from AI were 30% more likely to convert than shoppers arriving from traditional search.

Read that twice. The channel growing fastest is also, per visitor, your best one.

Behaviour. Around 63% of surveyed consumers now use an AI assistant weekly. Roughly 15% already use agentic AI (AI that actually goes and does the task), and about 31% expect to within a year. The people who will shop this way are not a future cohort you'll need to acquire. They are your current customers, using a new front door.

So why does your dashboard look normal?

Because the traffic arrives wearing a disguise, and every layer of your measurement stack was designed for a different kind of visitor.

No referrer, or a useless one. A shopper who asks an assistant for a recommendation and then lands on your product page frequently arrives with no referrer at all, or one that gets bucketed into direct traffic. Direct traffic is the drawer where analytics puts things it can't identify. It has always been a little inflated. It's now inflated with your highest-intent segment.

The research happened somewhere you have no visibility into. In the old model, a shopper's comparison happened partly on your site: category page, filters, two product pages side by side. You could see the consideration. Now the comparison happens inside a conversation with an assistant. By the time anyone reaches you, the decision is largely made. You see the last step of a journey whose interesting parts happened off-property.

The agents themselves don't look like sessions. An assistant reading your product page to answer a question isn't a session in any sense your analytics recognises. No scroll depth, no dwell time, no engagement signal. In many setups it's filtered out as bot traffic, which is technically correct and commercially catastrophic, because that "bot" was shopping on behalf of a human with a credit card.

And the failures are invisible by construction. This is the worst part. When an assistant considers your store and decides not to mention you, nothing happens. Maybe it couldn't confirm a spec, or couldn't parse your price, or found a competitor easier to describe. There's no bounce. There's no abandoned cart. There's no 404 in your logs. A customer existed, wanted what you sell, and was routed elsewhere, and the event that would have told you left no trace anywhere in your systems.

You cannot fix what generates no signal. That's the entire problem in one sentence.

The gap that turns this into money

Now put two facts next to each other.

Shoppers from AI convert 30% better. And revenue per visit from AI traffic is still running about 18% lower than it should be.

Those two things cannot both be true unless something is being lost in between. The intent is higher. The outcome is worse. The difference is friction that happens after the assistant decides to send someone your way, or before it decides not to.

Adobe put a number on the most likely cause: the average retail product page scores 66 out of 100 on machine readability. Their conclusion, stated plainly: for retailers with thousands of SKUs, a significant share of the catalog is currently invisible to AI.

Not badly ranked. Invisible.

What "invisible" actually looks like

It's rarely dramatic. It's almost always mundane.

A shopper asks for waterproof hiking boots under $150. You sell exactly that, in stock, at $139. Your product page describes it beautifully: a paragraph about weekend trails, a photograph of the boot in a stream, a customer review mentioning a rainy hike. Every human reading that page understands the boot is waterproof.

Nothing on the page states it in a form a machine can rely on. So the assistant, which will not guess about a product claim, recommends three competitors whose pages say the word plainly. You were never in the running. You'll never know the query happened.

Multiply by every attribute a shopper might filter on, across every product you sell.

The uncomfortable framing

The instinct in most teams is to treat this as an SEO problem with a new name: get mentioned, rank in the answer, optimise for the engine. That framing is comfortable because it maps to an existing budget and an existing agency relationship.

It's also incomplete, and the incomplete part is where the money is. Being mentioned is not the transaction. A shopper who gets a great recommendation and then hits a product page where the agent can't resolve which variant is which, or a checkout step it can't get through, is a sale you were awarded and then lost. Those failures are less visible than a ranking miss and considerably more expensive, because they happen after the hard part is already won.

What to do about it this week

You don't need a platform decision to get started. You need evidence.

1. Grep your server logs for known assistant and agent signatures. Count them. Whatever the number is, compare it to what your analytics reported for the same period. The delta is your blind spot, quantified.
2. Write down the twenty questions a real shopper would ask before buying your best-selling category. Not keywords, but actual sentences with budgets and constraints in them.
3. Ask those questions of the assistants your customers use. Record what comes back verbatim. Note every time you're skipped, described wrong, priced wrong, or a competitor is named instead.
4. Then send an agent through your buying flow and write down the exact step where it stops.

That's an afternoon of work and it will tell you more than another quarter of debating whether this is real.

If you'd rather not do it by hand, that's what we built Product Gale to do: the same four steps, run continuously, with a dollar figure attached to each failure. But the manual version works, and doing it once yourself is worth more than reading another article about it.

Including this one.

---

Traffic and conversion figures: Adobe Analytics, reported via Forbes (July 2025, November 2025) and Marketing Week (March 2026). Consumer adoption figures: Marketing Week, March 2026.