Why every site you build must be genuinely different (the uniqueness gate)

What you're doing and why

Once you own more than one site in the same trade, the AI will happily write the same page twice. Ask any model for septic FAQs and you get "How often should a septic tank be pumped?" every single time. Duplicate pages across sites you own is exactly the pattern that gets a network of sites devalued — so the builder has a uniqueness gate that compares every new page against everything you've already published and refuses to publish a near-copy. This lesson explains what it's protecting you from, and what to do when it stops your build.

The exact steps

1. Let the gate do its job. When a build halts with a uniqueness failure, that's the system working. You'll get the exact path, the similarity score and the text that clashed:

   UNIQUENESS $.services[0].faqs[1].q 1 vs septic-pumping-ozark-mo — your text: "How often does a 1,000-gallon septic tank need to be pumped out?" A score of 1 means identical. Anything above 0.3 is rejected.

2. Don't lower the threshold to make it pass. That's turning off the smoke alarm. Fix the content instead.
3. Feed it the real local questions. The research step already collected genuine "People also ask" questions for your town, and those differ market to market. Tell Claude Code:

   "Rewrite the FAQs for <my site> using the real People-Also-Ask questions from the research pack. They must not repeat any question already used on my other sites."

4. Differentiate on angle, not wording. Rephrasing "How often should a tank be pumped?" into "How frequently should a tank be pumped?" still scores as a duplicate. Change the question — local rules, local soil, seasonal demand, a different customer situation.
5. Re-run the build. It picks up from where it stopped.

What it looked like for us

Building our 8th site (septic, in a small North Carolina town) the gate stopped the build cold:

• faqs[1].q scored 1.0 — word-for-word identical to our Ozark site's question.
• faqs[0].q scored 0.778 — "How much does it cost to pump a septic tank in NC?" against Ozark's "…in Missouri?" Same question, different state.

The build retried three times, failed all three, and stopped rather than publishing. Nothing duplicated ever went live — which is the entire point of the gate.

Then we found why, and it was our bug, not the model's: the AI was being shown only 9 sample strings from existing sites (business blurbs, not FAQs), while the gate was scoring against every string on every site. The model was writing blind and being marked against an answer key it had never seen. We fixed it so the AI now receives all 160 existing FAQ questions up front, marked as taken, and is pointed at the local research questions instead.

Gotchas

• Similar towns are the risk, not similar trades. Two septic sites in different states will produce near-identical cost and frequency questions unless you force local specifics into them.
• The most canonical questions are the most dangerous. The ones everybody asks are the ones the model reaches for first. Use them on your first site in a trade; find different angles after that.
• A failed build is cheaper than a bad one. Our halted build cost about 20 cents and an hour. A network of sites quietly sharing duplicate content costs you the rankings you were building.
• If you rewrite by hand, check the answers too, not just the questions. The gate reads every string, so a fresh question with a recycled answer will still trip it.