What the real freight web costs an agent.

The old vision baseline was not measured. It was assumed: page height divided by viewport height, on the premise that an agent screenshots a whole page before it can find a link. Our own side was charged for one targeted query. That is not a comparison, and we had already written the same criticism about someone else’s number.

So we measured it. Ten independent vision agents were given the actual captured screenshots of ten of these sites, one at a time in scroll order, and asked to name the visible element for each of four intents. They read 67 screens — a median of 7 per site — and answered 17 of 40.

The real ratio is 49.5×, not 148×. The assumption was doing roughly fortyfold of the work, and it was ours.

There is a second, harsher number. If you insist the whole structured read must pass through the model — it does not, the transport builds it and answers from it — the ratio is 3.7×. Both are in the table below, because which one is right depends on an architectural boundary and not on us.

The result we did not expect is the one we would now lead with anyway: structure answered 22 of 40 against vision’s 17, and there was not a single question vision could answer that structure could not.

Every sweep before this one dismissed consent walls, on the reasoning that otherwise you are measuring a cookie wall rather than a site. We never checked whether that reasoning was true, and it sat directly against something we say on this same page about the entitlement arm: that accepting terms on someone’s behalf is not ours to do. Both could not be doctrine.

An audit also broke the scoping. “Inside a recognised consent container” was implemented as [id*="consent"] and three more substring wildcards, so it would click “Continue” on a recipe page about cookies, “I agree” on a form posting to an account-deletion endpoint, and an Accept all that was display:none in preference to the visible Reject beside it.

10 of 30 sites had a banner this code would have accepted. Dismissing them changed the answer on 1 site, which is inside run-to-run variance.

So the sweep accepts nothing now. The comparison arm still runs, because “it buys nothing” is a claim and claims should be checkable.

The first sweep ran a headless browser advertising itself as headless, and did not dismiss cookie walls. On that basis ten of 35 sites gave us nothing, and we published that as a finding about the industry.

Re-running with a normal Chrome identity and consent handling, 6 of those sites served us. Maersk, MSC, DHL, FedEx, UPS and Forto were not closed to agents. They were closed to the specific way we were knocking.

One qualifier on that sentence, because a reviewer caught us overstating it: FedEx served a country-selector interstitial rather than its homepage. It let us in; it did not let us in to anything. Five of the six is the honest claim.

Of 35 rows, 30 were read. 2 are excluded by us — a duplicate host after an acquisition, and a site whose robots.txt says not to knock. 3 refused a datacentre crawler outright, and every one of those later served an entitled browser. So the number of sites that are closed to an agent, as opposed to closed to an anonymous crawler, is zero. That is a much less dramatic number than the one we first published and it is the correct one.

We did not add stealth. No navigator.webdriver masking, no anti-detection, no evasion — that is a losing arms race and we are not playing it. We stopped announcing ourselves as a robot and started handling cookie banners like any browser does. Both identities were run so the difference is published rather than quietly absorbed.

These observations do not establish why an edge refused access or whether a caller was authorised. Browser access and permission to automate are separate questions.