Nov 14, 2026 • 10 min read

Auditing and Fixing Our Own Site: A Week-by-Week Log

A living case study: how we audited and fixed our own GEO agency site week by week, with real metrics and the mistakes we made along the way.

Third DoorAuthor
Auditing and Fixing Our Own Site: A Week-by-Week Log

Why we audited ourselves first

Before asking any client to trust our process, we ran it on our own site. This post is the running log of what we found, what we fixed, and what is still broken. It is written as a living case study, so expect updates as we ship changes.

Every number below is internal data from our own builds, CI runs, and Lighthouse checks. No external benchmarks are claimed.

Week 1: the build pipeline was the real problem

Our first surprise was that the site did not actually prerender. The prerender script launched Chrome by calling `puppeteer.executablePath()` synchronously, but that method returns a Promise in Puppeteer v25. On the Ubuntu runner the Windows Chrome paths did not exist, so the function returned a Promise object instead of a path, and `puppeteer.launch()` threw "Browser was not found at the configured executablePath ([object Promise])". Prerendering silently skipped every route.

What we fixed: made `getBrowserExecutable()` async and awaited the call. Added Linux Chrome candidate paths. Added a CI verification step that prints the resolved Chrome path.

Result: 17 routes now prerender successfully in CI (internal data: prerender run on commit `fee99b0`).

Week 2: Vercel build and deploy mismatch

The GitHub Actions workflow pulled production settings, ran `vercel build`, then deployed with `--prebuilt --prod`. But `vercel build` was invoked without `--prod`, so it produced a preview-environment build. The production deploy then failed with "prebuilt output was built with target environment preview."

What we fixed: added a conditional `--prod` flag to the build step when running on `main`.

Result: production deploys now succeed. CI run `42382658387` concluded success (internal data).

Week 3: the contact form posted to localhost

The contact form in `ContactSection.jsx` posted to `http://localhost:3000/submit-form`. In production that fails silently. We moved the endpoint into the frontend project as a Vercel serverless function at `/api/submit-form`, so it deploys with the site and runs same-origin.

What we fixed: created `api/submit-form.js` (ESM, since the project uses `"type": "module"`); pointed the form at `/api/submit-form`; added Supabase as the free storage backend (500 MB free tier).

Result: form submissions now return 201 and land in the Supabase `contacts` table (internal data: tested 2026-11-07).

Week 4: content and SEO hygiene

We found 38 lint errors (unused React imports, unused variables) and fixed them. We moved pricing into `src/data/pricing.js` so the pricing page and its JSON-LD offers share one source. We converted 12 blog posts from a JS array into individual Markdown files with frontmatter, generated routes and 301 redirects from the old numeric URLs, and added a content quality gate to CI.

Quality gate rules (internal data): title <= 60 chars; description 120-160 chars; exactly one H1 per page; no stub strings; banned words (`guarantee`, `#1`, `leading`) unless allow-listed; a percentage needs a source link or "internal data" tag; no images without alt; no broken internal links; no duplicate slugs.

Result: content check passes on all 12 posts; CI green.

What is still broken (honest list)

Limitations

This is a single-site case study with a small team and a small sample of CI runs. Nothing here is a controlled experiment. The fixes are engineering hygiene, not proof of SEO impact. We have not yet measured citation lift.

What we are measuring next

CTA

Want the same audit on your site? [Get a free snapshot](#contact) or [see pricing](/pricing).

Contact

Get In touch
with us!

Have questions or ideas? We'd love to hear from you. Reach out anytime and let's connect.

Contact Us