Wellness · Continuous SEO and GEO
A website that measures itself every month, explains its numbers and knows what to fix next
For a small retreat center: custom connectors for services that offer no API, one analytics snapshot from every source, and AI that turns it into evidence-backed recommendations that an AI coding agent then ships.
≈ 9 : 10
Visitors from AI assistants for every ten from all organic search combined, in the first weeks measured
Daily
Prices and dates synced from a booking platform that has no API, validated before anything reaches the site
100%
Recommendations that cite the data behind them. The report format rejects any that do not
< 24 h
From publishing a new guide to seeing it indexed by Google
01
The problem
A small retreat center sells its programs to international guests through its own website and through third-party booking platforms. Prices and dates lived in two places and were kept in sync by hand. Nobody could say which channel brought in inquiries, because nothing was being measured.
At the same time, future guests were changing how they search. More of them ask ChatGPT which retreat to choose, and assistants cross-check what they read: when a price or a group size differs between pages, they trust the business less, or quote the wrong number. The center needed a site that AI could quote with confidence, and a way to know, every month, what to improve next.
02
Why it was hard
- The booking platform that owns prices and dates has no API. Its public pages were the only machine-readable source, and some programs publish no structured data at all.
- A legally sensitive category: a wrong claim is worse than no claim. Being quoted by AI is worthless if what it quotes is false.
- A small site means small numbers and noisy tools. Ad blockers hide visitors from one analytics source, and the same page can score very differently on two consecutive speed tests.
- Business decisions and technical fixes look alike. An agent that “fixes” the wrong thing can change how the business sells.
03
What we built
Two systems around one source of truth. Custom connectors pull data from services that offer no API and validate it before anything reaches the site. A monthly loop merges every measurement source into one snapshot, has AI turn it into recommendations backed by evidence, and hands the approved ones to an AI coding agent.
04
Where the innovation is
01
Connectors for services that never built an API
The booking platform publishes its events as structured data on public pages. A daily job reads them and checks that the result makes sense before writing anything: enough events, the core programs present, prices where they should be. Then it emails a plain list of what changed. When the platform changed its markup in production, the guard fired, the site kept its last good data and the alert arrived the same morning.
02
One snapshot from every source
Search Console with page-by-page indexing checks, analytics with a dedicated channel for AI assistants, the hosting platform’s own visitor counts and repeated speed tests merge into one validated snapshot every month. A source that fails is reported as missing, never as zero.
03
No evidence, no recommendation
AI writes the monthly analysis, but into a strict format: every finding and every recommendation must cite the numbers behind it, with a priority, an owner and an estimated effort. The report the owner reads in the CMS and the PDF that reaches their inbox come from the same component, so they can never disagree.
04
From recommendation to shipped change, then measured
Once the owner approves a recommendation, an AI coding agent implements it under an engineer’s review. Each one is tracked as pending, shipped or dropped, and the next report checks whether it worked. Continuous improvement as a process, not a promise.
05
Written to be quoted by AI
Answer-first copy, an “at a glance” block where every line stands on its own, structured data for the business and its events, and an llms.txt rebuilt from live data, so AI assistants read the same price the booking page shows.
05
What the first reports showed
With every source in one place the picture was clear within weeks. AI assistants were already a real channel, almost as large as all organic search, and they were landing visitors on the pages that answer the hardest questions.
What a recommendation looks like
A real one from the third report, generalized. The format rejects any recommendation without evidence.
Recommendation · Speed
Lighten the homepage video
- Priority
- High
- Owner
- Engineering
- Effort
- 1 to 2 hours
- State
- Pending
Evidence
- Homepage: about 6.5 MB on mobile
- Retreats page: about 1.3 MB
Where visitors came from
Share of sessions by channel, first half of the first month measured. Almost all AI traffic came from ChatGPT.
06
Results
- Prices and dates on the site, in its structured data and in llms.txt now come from the booking platform every day. The first time the platform changed its markup, the guard caught it, the site kept its last good data and the alert arrived that morning.
- Three monthly reports delivered, every recommendation with its evidence, a priority and an owner. Five already shipped; the rest are assigned and tracked.
- New guides and articles indexed by Google within 24 hours of publishing.
- Desktop speed scores between 93 and 95, and a perfect search score on every test run.
- In the first weeks measured, AI assistants sent almost as many visitors as all organic search combined, and an inquiry arrived that named an AI recommendation as its source.
- Google rankings are still building. That is exactly what the loop is for: every change is measured in the next report, so nothing rests on assumptions.
Recommendations shipped, cumulative
Each report reviews the ones before it: shipped, pending or dropped.
07
What we learned along the way
We audited our own launch first
The first version of the site, which we built, moved to a new platform without redirecting the old addresses. Search Console made the gap obvious. Redirect rules now catch more than forty old URLs, and every launch since starts with a redirect map.
Ambiguous is acceptable, false never
A guide on the category’s legal status, drafted with AI, read well and overstated the facts. Checked against primary sources, it was withdrawn the same evening and rebuilt on the law and the health authority’s own advisory. The same verified wording now appears everywhere the topic comes up.
Not every fix is ours to make
An agent “fixed” how dates were generated. It turned out to be a deliberate business choice, and it was reverted within the hour. Since then, anything that touches prices, dates or policy goes to the owner first.
Stack
- Next.js
- React
- TypeScript
- Sanity
- Vercel (cron + analytics)
- Search Console API
- GA4 Data API
- Lighthouse
- Zod
- Resend
- schema.org
- llms.txt
- IndexNow
- Claude
- AI coding agents
Want a website that tells you what to fix next?
We build the same loop for your business: one source of truth for your data, a monthly report in plain language, and every recommendation backed by your own numbers.
Book a discovery sessionClient identity, location and commercial terms are withheld. Small figures are shown as ratios or shares; every one comes from the center’s own reports.