For about fifteen years, the correct answer to automated traffic was to stop it. Scrapers stole pricing, bots hoarded inventory, credential stuffers tried stolen passwords against your login. An entire industry grew up around detecting non-human visitors and turning them away, and it got very good at its job.
Then the buyer became non-human. And the stack you built to keep software out cannot tell the difference between a scraper and a customer with a wallet.
The mismatch
A shopping agent looks, to a mitigation system, almost exactly like the thing that system was built to block. It is automated. It moves faster than a person. It does not have the mouse-movement entropy or the dwell time that behavioural detection keys on. It may arrive from a data center IP range rather than a residential one. Every heuristic that made bot management work is a heuristic that flags your prospective customer.
Most merchants have not noticed this yet, because agent traffic is still small relative to human traffic and because the failure is silent. Nobody files a support ticket saying "your challenge page defeated my AI assistant." The agent simply gives up and routes the request to a competitor whose checkout it can complete. You lose the sale and never see the attempt.
Where agent checkouts actually break
Bot mitigation is the failure that surprises people most, but it is one of six, and the others are just as fixable.
| Stage | Common failure | Fix |
|---|---|---|
| Discovery | Product data absent or unparseable | Structured feed, Product markup |
| Evaluation | Price, stock, or variant ambiguity | Machine-readable availability per variant |
| Cart | Bot mitigation blocks the agent | Allowlist verified agents instead of blanket blocking |
| Authorisation | No mandate the merchant accepts | Support an agent payment standard end to end |
| Settlement | Payment method unsupported for agents | Agent-capable processor configuration |
| Confirmation | No machine-readable receipt | Structured order confirmation |
We keep the longer version of this in our research on agentic checkout conversion. The pattern across all six is the same: nothing here is a persuasion problem.
Agents do not abandon carts. They fail integration tests.
This is the reframe that makes the whole problem tractable.
A human abandons a cart because they got distracted, had second thoughts, found a better price in another tab, or balked at the shipping cost. Those are psychological failures, and the entire discipline of conversion rate optimisation exists to address them. Urgency copy, social proof, trust badges, exit-intent offers.
An agent has none of those states. It does not get distracted, it does not reconsider, and it is not moved by a countdown timer. When an agent abandons, something was technically unresolvable: a variant it could not disambiguate, a field it could not fill, a challenge it could not pass. That makes agent abandonment far more diagnosable than human abandonment, and far more fixable.
The practical consequence
Agent conversion is closer to a pass/fail integration test than a funnel. Copy does not move it. Machine-readability does. Which means the team that owns it is probably engineering, not growth, and in most organisations nobody has told them.
Ignore the headline transaction numbers
If you go looking for evidence on how big agentic commerce already is, you will find very large transaction counts. Treat them carefully. Analysis of x402 volume has found that roughly half of transactions look like testing rather than genuine commerce, and the tracked real daily volume is a small fraction of the settled transaction count.
Enormous transaction counts against small total value is also just what a market made of micropayments looks like, since many agent transactions are fractions of a cent for an API call or a single inference request. Both effects push the same way: counting transactions overstates economic activity. We track the breakdown in the x402 adoption tracker.
The honest position is that agent-mediated commerce is small today and the infrastructure is arriving faster than the volume. That is an argument for readiness, not for urgency theatre. The cost of being ready is a few engineering tickets. The cost of not being ready is invisible, which is what makes it easy to defer indefinitely.
What to instrument
Most merchants cannot answer basic questions about agent traffic, because their analytics were explicitly built to filter automated visitors out. Four things worth logging separately:
Agent-identified sessions, by user agent and by declared agent identity, kept apart from both human traffic and generic bot traffic.
The stage at which agent sessions terminate. This is your funnel, and unlike the human funnel each drop-off has a specific technical cause you can go and fix.
Bot-mitigation challenge rates served to known agent identities. If this number is high, you are turning away buyers and calling it security.
Structured-data parse failures on product pages. A page an agent cannot parse is a product it cannot recommend, well before it is a product it cannot buy.
Discovery and conversion are the same problem here
This is the part that makes agent readiness more than a checkout concern. An agent that cannot parse your product data will not shortlist you in the first place. One that cannot complete a purchase will route the next similar request somewhere that works. The failures compound, because each failed attempt is also a signal against you in whatever the agent carries forward.
The order of operations we described in the money rails field guide holds here too: discover, trust, price, transact. Bot protection sits across all four. A mitigation rule written in 2019 to stop a scraper is now, silently, a discovery filter.
The fix is not to turn protection off. It is to stop treating "automated" as a synonym for "hostile" and start distinguishing verified agents acting for a real buyer from the traffic you actually want to block. That distinction is what verified agent identity programmes exist to provide, and it is going to become a gate on discoverability rather than just a security control.
Presenc tracks every agent and AI browser hitting your pages, separates them from human traffic and from scrapers, and shows where they extract cleanly against where they give up. If you have never looked at your site the way an agent sees it, the first look is usually uncomfortable, and usually fixable in an afternoon.
