Field report · not a study
We pointed an agent at our own API and asked for nothing. It called it “a goldmine.”
Two Claude Code sessions, one day, two unrelated projects. Neither was asked what it thought of UXXRAY. Both reached for the same word, on their own.
Ben was starting two unrelated design projects on 2026-07-16, in two separate Claude Code sessions. Both times he gave the agent the same rough instruction: point yourself at uxxray.com, see if anything sparks before we begin designing. He asked for nothing about UXXRAY itself.
Both agents found the API from the homepage HTML on their own. Neither was told it existed.
Session one
The agent pulled /patterns, /league and three exemplar bundles, called the dataset “a goldmine”, and produced six concrete, data-backed design decisions from it: an honest logo-wall substitute, a comparison table as the centerpiece (a measured +32 score lift, present on only 8% of the pages it checked), and a gap in FAQ coverage worth +27. Its sixth idea was unprompted and is the harder one to dismiss. It proposed that UXXRAY’s own /scan endpoint could score its own concepts, 0 to 100 against the corpus, before shipping them: a validation loop it invented for itself.
Session two, same day, a different project
Ben repeated the same instruction in an unrelated session later that day. This agent picked ryzesuperfoods.com as its “closest comp,” pulled its bundle, and produced a ranked steal-list ordered by the score lift each tactic measured (a review wall at +36, a comparison table at +32, above-fold proof at +29). It reversed one of its own earlier design calls after seeing the data, and called the dataset “a goldmine”: the same word, independently, in a session that shared no context with the first.
Why this is on the page at all
UXXRAY’s own rule is that a claim nobody can check does not ship. This page is the check. Both quotes are real, both sessions are described above in full, and nothing here is inferred beyond what is written. It is not evidence of customer demand. It is evidence that, given only the API and no prompting, two independent agent sessions reached for the same word and then used the data as intended.
See what it finds for you