Thirty years of craft on this web went into being looked at. Visual hierarchy, the hero image, the two seconds you get before a human bounces — all of it tuned for a single reader, and that reader had eyes. Something else reads the page now, sometimes by crawling it, sometimes by being handed a surface built for the purpose, and it doesn't squint at your layout for intent. It takes the fields and believes them.
A page implies. A declared field commits.
That gap was free for as long as nothing depended on it. It isn't free anymore. Schema.org, the shared vocabulary sites use to label their pages for machines, counts its price property across more than ten million domains. Google already treats disagreement between your product feed, your page, your markup, and what actually happens at checkout as an enforcement matter: drift far enough and you lose automatic price updates, then you lose the listings.
So: the error rate. Nobody publishes it. Not the platforms, not the retailers, and as far as I can find, not the researchers either. The surface that is quietly becoming the primary reader of your business has never been measured for fidelity, because nothing depended on the answer.
I've been watching this failure mode since I was seventeen and on call; it just has a new outfit. The declared price is a description of your system. Checkout is your system. Descriptions and systems come apart continuously, silently, and always in the direction that makes you look like a liar, and I have never once seen an organization staff the job of keeping them honest until something expensive made it somebody's fault. Something expensive is now reading, at scale, in public.
Then there's the part the machine doesn't carry across at all. In April, Britain's competition regulator ordered the AA and BSM driving schools to refund more than 80,000 learners and fined them £4.2 million. The offense: lesson prices advertised without a mandatory £3 booking fee, which showed up at checkout, after the learner had chosen a slot and typed in their details. The reasoning behind the order is what I keep coming back to. Put the honest all-in number up front and you get fewer clicks, because people anchor hard on the headline and meet the rest of it late. Vagueness pays. It pays because of the order in which human attention arrives.
Agents inherit plenty of our biases; they're assembled out of us. They don't inherit that one. Unbundled fees, call-us-for-a-quote pricing, tiers cut so that no two vendors ever line up side by side — those are instruments for managing the sequence in which a person discovers things. A parser asks for the total on the first request and gets back a shape engineered not to have one.
Which drops two very different companies into the same hole. The one whose declarations don't survive contact with its own checkout. And the one whose value genuinely doesn't fit into fields: the consultative sale, the it-depends, the thing that becomes obvious twenty minutes into a conversation and disappears inside a comparison table. To something reading structure, they look identical. Thin, or absent.
Nora Kaplan argued here that the moat becomes machine-readable reliability. Agreed, with one addition: a moat like that leaks from the operational end, and it will be owned the way operational problems always get owned, which is badly. Every company can tell you who owns checkout and who owns the brand. Ask who owns the feed and you get a shrug, a contractor's first name, and a date somewhere around 2019.
That used to be a perfectly good answer. It described a file nobody read.
-
The open web question: W3C and GS1 have scheduled a September workshop on e-commerce for humans and agents that asks, without answering, whether agent-readable commerce keeps the web open or funnels visibility through a handful of protocol gateways.
-
Scraping versus surfaces: A W3C community group spent May discussing structured agent-to-web tools as an alternative to agents mechanically driving a user interface — early incubation work, not a standard, but the clearest sign that sites exposing deliberate machine surfaces is becoming a design decision rather than an accident.
-
Who authorized the purchase: The IMF's April note on agentic payments separates the part of an agent that forms buying intent from the parts that authorize and settle — useful if you're trying to work out what your checkout is actually agreeing to when no human is present for the click.
-
Records that don't exist yet: A May survey of 32 competing IETF drafts on agent authorization found three unrelated proposals using the same acronym, which tells you roughly how settled the question of "who is this agent buying for" currently is.