Static readiness scores are comfortable. They tell you your structured data is present, your sitemap is reachable, and your robots.txt says the right things. What they don't tell you is whether an AI agent can actually complete a purchase on your site. That gap is exactly what agent simulations exist to close.
The data
As of September 2026, the Scovant Checkout Benchmark — every terminal agent checkout run on commerce sites in Scovant's production corpus, excluding failures caused by Scovant's own platform — stood at 86% of runs failing across 214 runs on 88 commerce sites; the page recomputes live, so read the current figure, its run and site counts and its window there, alongside the methodology.
Read that carefully: on real stores, most agent checkout attempts do not finish. Not because the markup was missing — because something between "find the product" and "reach checkout" broke.
What a real simulation actually does
When Scovant runs an agent simulation, a real AI agent attempts an actual purchase flow in a real browser. Not a crawl. Not a schema parse. An agent that navigates, makes decisions, and either reaches checkout with the cart intact or fails at a specific, replayable step. (Simulations never enter payment details and never create accounts — reaching checkout counts as success.)
The failures cluster in a short list, visible step by step in the replays:
- Bot checks firing mid-flow — the WAF admits the product page, then challenges the cart.
- Cart state with no machine-readable confirmation — the agent adds an item and cannot tell whether anything happened.
- Consent and newsletter overlays that trap focus or cover the one control that matters.
- Ambiguous, icon-only controls a human parses instantly and a model guesses at.
- Multi-step forms that lose state on the agent's navigation pattern.
Why one model isn't enough
Model capability varies. A flow a premium model recovers gracefully can still trip up a leaner one — and the model your customers' tools use may not be the one you tested. Scovant runs the same flow against every LLM agent in its catalog — Claude Sonnet 4.6 is the default — so readiness reflects a distribution, not one lucky (or unlucky) data point.
What to do with this
- Run one real simulation on your own checkout before trusting any static score — the replay tells you the exact failing step.
- Check your bot manager against your declared robots policy; the benchmark's failure mix shows edge enforcement is a leading cause.
- Make cart state machine-visible: a cart-count element and a confirmation signal are cheap fixes with outsized effect.
- Re-test on deploys — a checkout that works today is one renamed field away from an agent regression.