Most published numbers about AI-generated websites quietly exclude the parts that dominate the bill. Here is ours, including the parts that are inconvenient.
The line items
- Model inference for copy and structure — scales linearly with site count.
- Similarity checking and regeneration — scales superlinearly with batch size.
- Asset processing — flat per site, but sensitive to image count.
- Build and deploy — cheap, parallel, largely fixed.
- Human review — the real cost, and the one nobody publishes.
Why per-site cost falls then rises
Cost per site drops steeply up to roughly a hundred sites as fixed overhead amortises, then creeps back up as similarity checking becomes quadratic-ish against the growing batch. We cap effective comparison windows to keep this manageable, at some cost to diversity guarantees.
The honest conclusion
Generation is cheap. Deciding what is good enough to ship is not. Any vendor quoting you a per-site price without describing their review process is quoting you the easy half.