The problem
Staycation already had 10 city pages targeting “hotel spa + city”. He writes that “they had to create new templates by hand. It took them about half a day per page.” The old pages had a hero, a title, a list of hotels and a short generic text block. “Only the city name changed from one page to the next.” The head term was out of reach. He says “hotel spa” meant fighting Booking, Tripadvisor and “every OTA on earth.”
The blueprint (as he describes it)
- Pick the pattern your data supports. Staycation had proprietary data on hotels (descriptions, prices, reviews) across many locations, so he went all in on the location pattern.
- Ignore search volume. He builds pages for variations at 10 searches a month or less. He says keyword tool volume “is an estimate” and competitors will skip queries that show no volume. “At Staycation, the list of pages came from the hotel database, not from Semrush.”
- Set a supply floor. He says a page with 2 hotels is weak and “drags the whole silo down.” The rule: “Below 3 bookable hotels in a city, no page.”
- Sort the silo into categories. Staycation had three: cities, departments and regions. Each category gets its own template and decides the internal linking. Cities link to nearby cities, and regions link to their departments.
- Bring an angle competitors lack. On “hotel spa Bordeaux” they added geographic detail, such as a spa hotel five minutes from the center. They also gave advice by mood or season, such as a couple mood or a gloomy winter.
- Build the template from fixed blocks.
- Hero that repeats the query, with a breadcrumb and a “Mis à jour le” (updated on) date.
- Listing of real hotels with neighborhood, rating, review count, spa amenities and current discount. Each page has an average, minimum and maximum price.
- A “why choose you” block, which he says most pSEO templates skip.
- Internal linking in 3 tabs. “All three lists were generated deterministically with a script. Nothing was generated with AI or by hand, which can lead to mistakes.”
- An editorial section. “This is AI-written but data-fed.”
- An FAQ of around 9 questions per page. A script selects them and matches them to the database, so a coastal page gets a sea view question.
- A hub page that targets the head keyword and mostly distributes traffic to the child pages.
- FAQ and breadcrumb structured data, with everything rendered server side.
- Produce from one database. “One page equals one row in Supabase.” Each page section had a specific prompt. He writes: “At the time, I built the logic on n8n, with different automations to fetch the data from Supabase, process it, run the prompts, and then send it straight to their CMS.” He says today he would rather do it in Claude Code.
- Iterate on a first batch before generating. He creates at least 1 variation per use case (one page with a lot of supply, one thin) and at least 1 per sub-category. To check for hallucinations, he “removed the AI’s access to outside research, so it relies only on the database info.” He compares the pages with competitors’ pages for keywords and intent.
- Expand the offer. They started with spa, then added pool, sea view and jacuzzi.
- Measure in order. First indexation rate, then impressions by level (cities vs departments vs regions), then clicks last. He says clicks on sub-10-volume pages arrive months after indexation.
Stated results
- “In one year, excluding brand searches, we grew non-brand transactional traffic four times, reached 2,500 impressions a day in AI search, and increased purchases attributed to search by 90%.”
- Pages went from 10 hand-made city pages to 300.
- The Search Console chart caption shows weekly clicks rising from about 1,000 to a peak near 9,000 between May 2025 and mid-2026. It shows daily AI search impressions rising from about 300 to over 2,000 between July and September 2026.
- Indexation rate, time per page after the change and page-level conversion data: none stated.
Caveats
- This is a guide on Web Anatomy’s own blog, and Web Anatomy is his company. The page promotes a free offer of “3 free opportunities” across SEO, AEO, conversion and UX, with paid expert help as an option.
- The Staycation figures are client results reported by his own company. They are not independently audited.
- The source names n8n, Supabase, prompts and the client’s CMS. It does not name the model or the CMS. Claude Code is what he says he would use today, not what he used.
- This is an automated content workflow that runs prompts. The source does not describe AI agents.
- He points to a separate post on his “content machine strategy” for the real detail on production. The article is adapted from his Web Anatomy newsletter of October 2, 2026.
- Handles: X @gabrielamzallag (verified 2026-11-03: profile links webanatomy.ai and his Substack). Substack gabrielamzallag815268 (verified 2026-11-03: linked from the source article). LinkedIn /in/gabriel-amzallag (verified 2026-11-03: linked from the source article). No Bluesky, Threads, Instagram or TikTok verified.
FAQ
What tools did Gabriel Amzallag use to build Staycation’s 300 pages?
He stored one row per page in Supabase and wrote a specific prompt for each page section. He built the logic on n8n to fetch the data, run the prompts and send it to Staycation’s CMS. The model and the CMS are not named.
Was the whole page written by AI?
No. The internal linking tabs were generated by a script, and a script selects the FAQ questions. The editorial section is “AI-written but data-fed.” The source does not say how the FAQ answers were written.
How did he keep the AI from making things up?
He removed the AI’s access to outside research so it relies only on the database. He also tested thick and thin variations per use case and per sub-category before generating all 300.
How much time did the new setup save?
The source does not say. It only states that the old hand-built pages took about half a day each.
Sources
- https://www.webanatomy.ai/blog/programmatic-seo (primary: first-person guide by Gabriel Amzallag on the Web Anatomy blog, published Oct 6, 2026, adapted from his Web Anatomy newsletter of October 2, 2026)
Want agents like these working your distribution? Get It Built. Get new teardowns weekly: subscribe (double opt-in).
Get one teardown a week.
Same format, a new operator every issue. Double opt-in.