Study: agents judge brands by whether they can read the site

A single number does the arguing here: across 37,927 agent journeys covering 1,056 real businesses, the paper finds that agent experience, whether a browsing agent can actually fetch and parse a business's own site, predicts purchase-relevant outcomes independent of fame, prior model knowledge, and two separate proxies for AEO. Businesses were matched on those variables and then split by AX level. The AEO-heavy tactics that consultancies have been selling since 2023, seeding forum threads, stacking listicles, chasing off-site citations, washed out as a control. Whether the site itself was legible to a machine did not.

This inverts the working assumption behind most AEO advice, which treats being surfaced as the finish line. It isn't. The paper's own framing is precise: an agent opens the results and reads them before deciding, and a single buyer question can trigger several rounds of search and fetch. Getting mentioned in a Reddit thread or a "best providers" listicle earns a citation. It does not earn a fetch that resolves into usable content once the agent lands on the site itself. If the page requires JavaScript rendering the crawler skips, sits behind a cookie wall, or buries the material fact (pricing, eligibility, coverage terms) three clicks deep in a PDF, the agent's drill-down step fails regardless of how many off-site breadcrumbs pointed there.

The mechanical failure point

This is a tooling story before it is a content story. Four independent agent harnesses were run in the study, which matters because it rules out a single vendor's quirks as the explanation. The consistency across harnesses suggests a shared bottleneck: most business websites were built for human eyeballs and search-engine crawlers optimised for HTML, not for autonomous agents that need to parse structured facts on the first pass and often will not retry a failed fetch. A site that renders fine in Chrome can still be functionally invisible to an agent that gives up after one malformed request.

For the sectors that read this newsletter, that is not an abstraction. A multilateral's public-facing site, a bank's product pages, an industrial group's technical documentation: these are exactly the dense, PDF-heavy, legally reviewed properties most likely to fail an agent's fetch. Financial services firms in particular tend to gate rate information and terms behind interactive calculators or authenticated portals, which look like due diligence to a compliance officer and look like a dead end to a crawler. UN agencies and development banks publish core facts inside long-form PDFs with no machine-readable summary layer. Industrial groups bury certification and specification data in downloadable spec sheets rather than parseable page text. Every one of these is an AX failure waiting to be measured, and per the study's design, no amount of off-site citation-building will offset it.

The implication for brand teams is a reallocation of budget, not an addition to it. Communications functions have spent the past two years pouring effort into the AEO playbook: getting cited in industry roundups, seeding Q&A content on third-party forums, building the kind of scattered digital footprint search engines reward. None of that is worthless. But this study puts a number behind a suspicion many technical SEO people already had: off-site visibility gets you into the agent's shortlist, and shortlist inclusion is no longer the scarce resource. Fetchability is. A financial institution that has spent a year optimising its press mentions while its product pages sit behind a client-side rendering framework has optimised the wrong half of the funnel.