Why AI pilots underperform on websites that look perfectly good
Because a human reads around ambiguity and a system cannot. Prices, availability, eligibility and service definitions usually sit inside sentences rather than in structured fields, so an agent either guesses or refuses. The failure is in the input layer, not the model.
The symptom
A business runs an AI pilot on its own content. The site is well written, the pages are clear, and a person browsing it would have no trouble. The agent returns answers that are roughly right and specifically wrong — the service category correct, the eligibility criteria invented.
The instinct is to blame the model, then to try a better one. That rarely helps, because the constraint is upstream.
What is actually happening
Human readers resolve ambiguity continuously and without noticing. A sentence like "we generally see patients within a week, though urgent cases are prioritised" carries a clear meaning to a person and almost none to a system. There is no field to read. There is a hedge, a condition and an implied exception, and a model will resolve them by guessing.
The same page rendered as structured data — a service entity with a stated lead time and a defined urgency path — produces a correct answer every time, from the same model.
The fix, in order
- Define the ten entities your business actually deals in — services, locations, availability, prices, eligibility.
- Expose them as structured data on the page, or as a read-only JSON feed alongside it.
- Generate the structured layer from the same source as the rendered page, never maintained separately.
- Version and timestamp it, so a consumer can detect staleness rather than serving it confidently.
The failure mode after you fix it
Divergence. The structured layer and the human page drift apart, and an agent serves the stale one with complete confidence. A build-time check that fails when they disagree costs an afternoon and prevents the whole class of problem.