It crawls the whole thing, walks the journeys your customers actually take in a real browser, and files what it found with the screenshot it found it in.
Not a score out of a hundred. Specific things that are wrong, where they are, and the proof.
A preview of the working surface. The product is in development.
The whole site, including what sits behind the pages a sitemap never reaches, harvested in bulk rather than one page at a time.
It models the people your site actually serves — the first-time buyer, the returning customer, the one hunting for a refund policy — and gives each of them goals worth attempting.
A real browser, following the paths those people would take. Before anything that changes state — a submit, a purchase, a delete — a deterministic gate stops and asks.
Findings are drawn from what the browser actually saw and recorded, not from a model's impression of a page it was told about.
Every finding carries the screenshot it came from. Click it and see exactly what the reviewer saw, at the step it saw it.
A finding can be re-run to verify it is actually fixed, and a monitor can keep watching so it does not quietly come back.
What may be clicked is decided by an explicit engine, not by a model's judgement in the moment. A destructive action is gated by rules you can read, which is the only version of this that is safe to point at a live site.
Findings are persisted against the screenshots and page captures that produced them. A claim you cannot check is worth nothing when you have to act on it.
External services sit behind an owned interface, and each one's replacement path is written down before it is adopted. No vendor gets to become load-bearing by accident.
Structured output is checked against a frozen schema before any consumer sees it, repair is bounded to one round, and a sensitive synthetic persona is refused before a provider call is made.
Crawl, catalogue, personas, journeys, evaluation and findings are stages of a single investigation with real state, so a run can be followed while it happens and explained afterwards.
Every row carries the workspace that owns it, and row level security refuses the rest. One customer's investigation cannot be read from another's session even if the query asks.
The investigation runtime is implemented end to endOne pipeline crawls a site, harvests its catalogue, models who it serves, drives journeys in a real browser, evaluates what it captured, and persists evidence-backed findings the API serves with clickable screenshots.
It reads a site thoroughly on purposeIt is built to reach what ordinary crawling misses, because the problems worth finding are rarely on the pages that are easy to reach.
Only sites you are entitled to testPoint it at your own site, or one you have written permission to investigate. That is a condition of use, not a suggestion.