The Agent Revenue Leak in DAX Commerce

Discovery works. Completion does not. Across the Commerce-lane measurements that built the DAX 40 Agent-Success Index [S1], the finding repeats with different brand-shapes: an AI agent can name the brand, route the buyer, sometimes read the catalog. Then the funnel kills the transaction one click later.

adidas is the cleanest case. We asked five agent classes to add an Ultraboost 5 to a cart on adidas.de. Every one of them hit the same Akamai 403 page at homepage entry, in German: Leider können wir im Moment keinen Zugriff auf unsere Seite geben [S2]. No product detail, no cart, no checkout form ever loaded. The agent never met the catalog.

Zalando lets the agent further in, then stops. The session-warmed reference lane reached the Nike Air Max 90 size 43 cart at €134.95, but only because the session was pre-warmed. A plain HTTP fetch hit an Akamai shell. Standard browser runs failed three of three on the cold-start variant. Direct /checkout returned a This page has gone out of style dead-end without a live bag [S3]. The buyer is findable. The path to close is not public.

The brand is the answer. The brand is not the close.

The pattern is consistent across sectors. That gap, sector-after-sector, is what this article names.

There is a single reason the leak keeps surprising brands. Most measurement programs collapse agent success into one number, usually called AI Visibility: how often does an AI mention you. Hyperize measures two [S6]. AI Visibility is whether the brand surfaces in the answer at all. AI Usability is whether an agent arriving at the surface can actually complete the job.

The two move independently. BMW scores 37.64 on AI Visibility for an unbranded mid-class sedan query, which is honest, not low. But the 318i configurator hides behind Shadow DOM web components: a coding agent loads the page, finds 227 interactive elements, and reads nothing — document.body.innerText returns an empty string. The €47,000 price is configurable in a browser one time in three [S4]. The composite lands at 30.13 because Usability does the work the headline number cannot.

Mercedes-Benz inverts the same axes. AI Visibility is higher at 43.61. A browser agent builds the C 180 configuration cleanly, reads the exact price of €42,427.31, reaches the test-drive booking from the configurator. Then the booking turns out to be built only for humans, and the simplest agent classes are blocked before the page even loads [S5]. Composite: 39.52. Findable. Not callable.

A high AI Visibility score and a blocked checkout are the same brand. They are not the same success.

Some brands cut the other way. Qiagen scores 80 on AI Usability and 12 on AI Visibility, a reference-grade buying surface AI agents almost never reach [S7]. Deutsche Telekom holds 80 on Usability with discovery as the open gap [S8]. DHL reaches 68 on Usability with breadth as the open edge [S9]. The brands that close best are not the brands that get found best.

A composite score on its own is theater. Cited together, the two axes locate the leak precisely: it is upstream of the answer, on the path to it, or at the close. The fix that closes each leak is different.

Across 21 measured brands, the same three shapes recur. Each one stalls a different point on the agent path.

Entry blocks. A WAF rule that turns the simplest agent classes away at the front door. The reason is rational: anti-bot, anti-scraping, anti-account- takeover. The effect is structural: the agent never sees the catalog. adidas is the published example [S2]; the Akamai 403 sits before search, before any product, before any cart interaction.

Shadow surfaces. Custom web components and SPA shells that render product data inside a DOM no standard query reaches. The plain reader gets a JavaScript bootstrap. The coding agent gets an empty string. Only the most capable browser, with the right accessibility surface exposed, passes through. BMW's 318i configurator is the worked example [S4]; one in three browser runs reach the price the human sees instantly.

Session walls. Checkout and booking paths that depend on a pre-warmed session no fresh agent has. Zalando's cart [S3] and Mercedes-Benz's test-drive booking [S5] live on the same wall: a fresh agent reaches the front of the path, then hits a state ritual built for a human who already had cookies.

The cure is not unwinding the human experience. None of the brands measured here built a configurator a buyer would call wrong; humans configure them every day. The cure is shipping a parallel one a buyer's agent can use too.

Open the door. A WAF policy that identifies and admits declared agent classes onto the same surface humans use. Not unconditional. Auditable, rate-limited, citeable: an answer engine learns the brand accepts agent traffic and starts recommending it at all.

Surface the catalog. Reference Pages that carry product facts, price, and the buy-step in plain HTML alongside the existing web app. Answer Pages for the buying questions a customer asks before the configurator opens. An Evidence Layer that lets a buyer's agent confirm a configuration without rendering the Shadow DOM at all.

Authenticate the agent. A delegated-session pattern that lets a buyer's agent transact for a human without inheriting the human session ritual. The booking completes, the cart closes, the receipt routes back to a buyer who never had to open the page.

The brand is not the problem. The brand's surface is built for the last generation of customer. The next one is the agent.

The leak is not a bug in any one brand's stack. It is the structural mismatch between a buying generation that sends an agent and a Commerce surface engineered for a person clicking a browser. Closing it is the work the DAX 40 Index keeps measuring, brand by brand, wave by wave.