I used Claude to rebuild SearchSeal. Here's what changed
A dated product-evolution case study on replacing synthetic visibility scores with analytics and separate crawler evidence.
The first SearchSeal measured selected prompts and model answers. The rebuild exposed why that was a research experiment rather than a universal view of brand visibility. As of July 24, 2026, SearchSeal combines privacy-friendly website analytics with a separate AI crawler evidence module.
The first product measured a sample
The earlier system sent selected prompts to selected models and summarized mentions. Every result depended on prompt choice, model version, timing, location, and sampling. Those observations can support a defined research program, but a clean dashboard cannot turn them into a universal score.
The crawler-only rebuild was an intermediate boundary
The next version focused on requests recorded by infrastructure the site owner controls. A request can include a user agent, host, path, time, source, and sometimes an HTTP status. That is useful evidence for access and delivery, but it does not measure human visits or business outcomes.
The current product keeps both views separate
SearchSeal now provides pageviews, privacy-safe daily visitors and sessions, acquisition sources, goals, revenue, and observed human AI referrals. It also provides separate crawler evidence: which identified crawlers requested which pages and the HTTP outcome when available.
A human AI referral is evidence of a visit. It does not prove that crawling or a citation caused the visit. A crawler request is evidence from infrastructure. It does not prove indexing, citation, recommendation, or conversion. Unknown remains unknown instead of becoming zero.
Where Claude helped
Claude was useful for implementing explicit contracts, tests, parsers, and repeated interface work after the evidence rules were written down. It was less useful when the product decision itself was vague. The agent could help express a boundary, but it could not choose a truthful metric without that boundary.
- Request-only evidence cannot invent an HTTP status.
- Overlapping sources cannot double-count totals.
- Policy tokens cannot appear as observed user agents.
- Detected fixes and confirmed fixes remain separate states.
- Analytics and crawler evidence cannot be collapsed into one score.
What I would repeat
I would write the evidence contract before the interface, test public outputs, keep source coverage beside every aggregate, and delete obsolete terminology when the product boundary changes.
The lasting lesson is not that AI can build a SaaS product. It is that precise evidence boundaries give the product and the coding agent something defensible to build.