Agentic Commerce: When the Buyer Is a Bot

Most ecommerce optimisation assumes a human on the other end. Photography, copy, social proof, urgency, trust signals — an accumulated craft aimed at persuading a person.

An increasing share of product discovery now runs through an intermediary that is not persuadable. Assistants are being asked to shortlist products, compare options against stated constraints, and increasingly to complete the transaction. Amazon’s shopping assistant work is one visible instance of a broader direction across every major assistant.

The filter that intermediary applies is different in kind from the one a shopper applies, and the things that survive it are not the things most teams have been investing in.

What an agent cannot do

Start with the constraints, because they determine everything downstream.

It cannot be persuaded. Copy tuned for desire does not work on a system matching stated constraints. “Loved by 10,000 teams” is not a constraint satisfier. “Supports SSO, £40/user/month, deploys in EU regions” is.

It cannot see your design. Product photography, layout quality and interaction polish — the bulk of ecommerce investment — are close to invisible. An agent reads text and structured fields. A beautiful page whose specifications live only inside images is, functionally, a blank page.

It cannot forgive ambiguity. A human seeing “from £29” investigates. An agent matching “under £50/month” either has a number to compare or does not, and a missing field is usually resolved by dropping the candidate rather than by asking.

It cannot verify most claims. “Best-in-class,” “enterprise-grade,” “industry-leading” are unverifiable and get discarded. Certifications, standards compliance, published integrations and concrete numbers are checkable and survive.

And it will not tell you it rejected you. No bounce, no session, no cart abandonment. A shortlist you did not make generates no data on your side at all — which is why this shift is so easy to miss until the aggregate effect shows up in a quarterly number with no obvious cause.

That last constraint deserves more weight than it usually gets. Every other channel in ecommerce leaves a trace when it fails: an abandoned basket, a bounced session, a search impression with no click. Agent-mediated rejection is the first failure mode in the discipline that is completely silent on your side, which means the only way to observe it is to go and ask the assistants the questions your buyers ask, and read what comes back.

What survives the filter

Structured, accurate product data. Price, currency, availability, SKU, specifications, shipping — machine-readable and current. Product markup is the standard route, documented in Google’s product structured data reference. This is unglamorous infrastructure and it is now table stakes rather than a nice-to-have.

Accuracy above all. An agent that recommends a product at a price that turns out to be wrong produces a bad user experience the platform will attribute to you. Stale pricing and phantom availability are worse than absence, because they burn trust in a way that is hard to observe and slow to recover.

Explicit constraint satisfaction. State plainly what you support: regions, integrations, compliance regimes, team sizes, contract terms. The unglamorous specification table is now a discovery asset. Most brands moved that content into images or a PDF datasheet years ago for design reasons.

Third-party corroboration. Agents weight independent sources heavily, precisely because your own claims are unverifiable. Review platforms, comparison sites, documentation and press are what an agent can check. Review sites and AI visibility covers the landscape.

Being in the consideration set at all. Before any comparison, the agent assembles candidates. If you are not in that pool, none of your product data matters — which puts entity presence and category association upstream of every optimisation above. Your brand is an entity before it’s a website covers the precondition.

The comparison is now explicit

The most consequential change is structural rather than technical.

A human comparing three products does it loosely — some tab-switching, an impression, a decision partly driven by which site felt more trustworthy. An agent does it as a table. Your product and three competitors, side by side, on the dimensions the user actually specified.

That is brutal for anyone whose advantage was presentation rather than specification, and it rewards two things: being genuinely better on a dimension buyers state, and saying so in a checkable form. It also means competitive weaknesses that were previously buried in a comparison nobody completed are now surfaced automatically in every comparison.

The strategic response is not to hide the weakness. It is to be unambiguous about the segment where you win, because an agent matching constraints is unusually good at routing a well-specified fit — better than a human skimming, in fact. Specialists tend to do well in this environment for exactly that reason. How AI recommends products covers the selection behaviour.

Where this is genuinely uncertain

It would be easy to write this as an inevitability. It is not, and three things are genuinely unsettled.

Adoption is real but early. People are using assistants for research far more than for completed purchases. The research half is happening now and is worth acting on; the transaction half is a direction of travel with an unclear timeline, and anyone giving you a confident date is guessing.

The commercial model is unresolved. Whether agent-mediated commerce runs through open web data, closed marketplace integrations, or paid placement is not settled, and the answer determines how much of this is optimisable at all. If it consolidates into a handful of proprietary integrations, “optimise your public product data” becomes a much weaker lever than it looks today.

Trust and liability are unresolved. Who is responsible when an agent buys the wrong thing? The answer shapes how much autonomy these systems are actually granted, and cautious answers mean human confirmation stays in the loop far longer.

So the honest framing: prepare for the research half, which is already happening, using work that is defensible regardless — accurate structured data and clear specifications are not a bet on agents, they are overdue hygiene. Treat the transaction half as a plausible direction rather than a plan.

What to actually do

1. Audit your product data for accuracy first, coverage second. Wrong beats missing in the worst way. Start with price and availability.
2. Move specifications out of images and PDFs into indexable text. This is frequently the single highest-impact change and it is usually resisted on design grounds.
3. State your constraints explicitly — regions, integrations, team sizes, compliance. Write the boring table.
4. Ask the assistants what they recommend for the constrained questions your buyers actually ask, and read what they say about you. Nothing else reveals how you are being represented.
5. Watch the AI referral trend as a leading indicator, remembering it is a floor rather than a total. Tracking AI referral traffic covers what can be concluded.

AEO for ecommerce covers the category specifics, and optimizing product pages for AI the page-level work.

The uncomfortable part

Much of what makes this work is the least interesting work available: keeping a feed accurate, publishing a specification table, marking up a price correctly. It is the kind of thing that gets deprioritised every quarter in favour of a campaign.

The shift worth internalising is that this material is no longer back-office plumbing. When the shortlist is assembled by something that reads fields and cannot be charmed, the fields are the storefront — and the brand work that used to carry the sale is now doing its job one step earlier, getting you into the candidate pool rather than closing from it.