Eurostat classifies a sale by where the order arrived. Websites and apps in one column, EDI-type messages in the other. Nothing asks who composed the order, so a purchase placed by an agent is filed as an ordinary web sale. Europe’s fastest-moving channel is counted accurately, as something else.

The definition is precise about the wrong thing. E-commerce, in Eurostat’s glossary, means orders sent by methods built to receive them, with payment and delivery free to happen offline. One exclusion follows: orders sent by manually typed e-mail do not count. The only clause that inspects how an order was written exists to disqualify the human hand.

The count is not small. In 2024, 23.6% of EU enterprises with ten or more employees or self-employed persons made e-sales. Websites or apps alone accounted for 18.0% of enterprises, EDI-type messages alone for 2.9%, and both for 2.7%. E-sales reached 19.5% of total turnover, of which EDI carried 11.1 points against 8.4 for the web. The larger share of Europe’s e-commerce already arrives with nobody typing.

An agent’s order must go in one of those two columns. Fill a checkout form and the sale reaches the seller through the website, exactly as a shopper’s would. The other column will not take it: EDI-type sales are structured messages passing between enterprises, and an agent buys for a household. So the machine’s order is recorded where the human’s is, and the questionnaire offers no way to say otherwise.

Eurostat can tell you whether the shop uses a model; it cannot tell you whether the shopper did.

The household survey is no sharper. 78% of EU internet users aged 16 to 74 bought or ordered goods or services in 2025, counted across any device: desktop, portable, tablet, mobile phone. A phone is a phone whether the customer tapped it or an assistant did. Europe knows what its shoppers bought on, and not what did the buying.

The trade has settled for private analytics instead. Adobe Analytics puts AI-referred traffic to US retail sites up 138% year on year in May 2026, converting 54% better than everything else, as Digital Commerce 360 reported in June. Those are shoppers arriving from a chat window, not software buying on their behalf, and only on sites Adobe can see. It is a referrer log, not a census.

Adding a question in Brussels is not an editorial decision. The variables the 2026 survey collects were fixed by Commission Implementing Regulation (EU) 2025/1310, adopted on 3 July 2025. The regulation sits inside the Union’s Digital Decade programme, which tracks 2030 targets including the take-up of artificial intelligence. The instrument built to watch AI spread through European business cannot see it arrive in the orders it counts.

Suppose the question were added tomorrow. Enterprises would have to answer it, and most cannot: Perplexity’s shopping agent presented itself to Amazon as an ordinary Chromium browser, and it took a federal injunction to establish that Amazon could refuse it. A survey collects only what its respondents already know. The gap will outlast the form that could close it.

None of this is a failure of arithmetic. The survey does what it was designed to do, and it was designed when the only machine placing orders worked in procurement. The first year in which software chooses a material share of European purchases will be published as a year in which websites did slightly better than before. Brussels will legislate against that sentence. It will be true, and it will describe nothing.