Townwick phase 1: three demo sites deployed, and the SEO plan the research killed

What shipped

All three seed clients are live demos: townwick-cjs-barbershop, townwick-untamed-salon and townwick-yolks-on-us, one Vercel project each, created by script over the REST API. Each one has an ignored-build-step so editing a client rebuilds only that client. That’s proven, not assumed. A cjs-only commit left the other two CANCELED in the deployments API.

The pipeline itself grew a lot. Sharp photo processing strips EXIF and GPS, corrects orientation, and emits AVIF, WebP and JPEG at four widths each plus a 150-byte WebP LQIP. satori renders OG images at about 10KB. LocalBusiness JSON-LD. Demo and live mode genuinely diverge now. The contact form is Resend-backed with a honeypot and a minimum fill time. And there are 13 Pexels placeholders sitting under a sha256 provenance ledger.

Fixtures use verified public data only. Untamed’s hours are empty on purpose because they genuinely haven’t published any, and Yolks’ hours came from the Chamber because their own website still carries Square’s template defaults, including a map pin in San Francisco.

Decisions

Two of the five are corrections to our own design docs, which is the best possible outcome for a research round.

First: no aggregateRating or review markup in JSON-LD, ever. Google’s review-snippet policy bans self-served review markup outright for “LocalBusiness or any other type of Organization”, and the penalty isn’t a shrug. It escalates to a manual action that strips every rich result the site has. docs/04 had a snapshot-feeds-aggregateRating plan written down. It got killed before a line of code existed.

Second: demo robots.txt allows crawling and the noindex meta does the work. The original combo, disallow everything plus noindex, is self-defeating. Block the crawl and the crawler never reads the noindex, and the URL can still get indexed off an external link.

The rest is plumbing with teeth. Vercel’s own workspace-graph skipping can’t help here because clients/<slug>/ isn’t a workspace package, so Vercel treats any client edit as a global change. The ignored-build-step script fixes it, with exit codes inverted from every instinct you have. satori won’t take WOFF2 and Astro’s Fonts API only emits content-hashed WOFF2, so satori gets its own .woff devDependencies and sharp rasterizes the path-only SVG in 13ms, which dropped the resvg native dependency entirely.

Stock photos get committed as binaries with a sha256 and a license snapshot. The Unsplash and Pexels licenses survive the photo being deleted, but the URL does not, so a fetch-on-build manifest is a demo that breaks for reasons outside the repo.

What broke

An API outage killed the first implementer twice and ate both transcripts. The replacement inherited the working tree and, to its credit, didn’t trust it. It found two real defects: a CLIENT env typecheck failure that CI had been masking, and two stock photos whose ledger entries falsely claimed no visible trademarks. One had a WAHL wordmark, the other a sticker. Both cropped and re-hashed.

Offboarding the scratch client I’d used for the timing test then exposed a latent CI bug, where a deleted client’s slug still entered the Lighthouse matrix and reddened a run. Fixed by filtering to clients that still exist, which is a sentence I’d have bet money was already true.

Numbers

Review re-ran the battery independently: 124 tests, gate 107/107. Then it negative-proofed the gate by injecting an aggregateRating, and 8 checks failed exactly as designed.

Live-artifact probes did the rest. noindex confirmed on the deployed demos, 405s where they belong, honeypot submissions returning a silent 200 with no email created, and three [DEMO]-tagged deliveries verified through the Resend API rather than by squinting at an inbox.

Lighthouse medians of 5 came in at perf 1.00 and a11y 1.00 on 236KB of a 512KB budget. SEO reads 0.69, which is the demo noindex and nothing else. The scaffold-to-deployed-URL run was 2m47s against a 30 minute target, though that’s a floor, since it skips the human part of gathering a real business’s data.

Next

Phase 2 is the live pipeline: the custom-domain checklist, production Resend with a verified sending domain, Google reviews sync with compliance rendering, analytics with two events, and one seed client promoted to a real test domain. Open dependency: actually acquiring townwick.com.