How often AI browser agents fail to complete a checkout on real commerce websites, measured by Scovant's own agent simulations. This page is the permanent source for the statistic cited on the Scovant homepage: the number, its population, its window, and its limitations, all in one place.
What counts as a run. One simulation = one AI agent driving a real headless browser through a scripted or autonomous purchase flow on a live commerce website (find a product, add it to cart, reach checkout). Scovant never enters payment details or creates accounts — reaching checkout with the cart intact counts as success.
Population. Every terminal simulation run on commerce-type sites in Scovant's production database: runs that reached a definite success-or-failure verdict. Runs that failed because of Scovant itself — an anti-bot block classified as our platform's failure, infrastructure timeouts, runs swept by our own janitors — are excluded from both the numerator and the denominator: our downtime must not be published as sites failing agents.
Corpus. Both customer-initiated runs and runs Scovant executed over its own research corpus of public commerce sites are included. Corpus sites did not opt in and are not affiliated with Scovant; a site can request removal at any time, and removed sites leave the benchmark. Agent mix spans the current Scovant model catalog; each run is a single attempt, not a best-of-N.
Limitations. This is a measurement of agent-automation compatibility, not of human conversion. The corpus skews toward popular public sites; a run tests the configuration a site served at run time (WAF rules, consent walls, and A/B variants included); and a failed run means that agent, that day, on that flow — not that the site can never serve agents. The benchmark is recomputed live; the run window shown above is the exact span of the underlying data.
Reproducibility. The scoring model is public at /scoring, the wider research program at /research, and the machine-readable benchmark at checkout-benchmark.json. Data is licensed CC BY 4.0 — reuse freely with attribution to Scovant.
Scovant is an AI agent readiness testing platform that scans, simulates, and continuously monitors websites to determine how reliably AI agents can discover, understand, and interact with them.