The problem
Programmatic SEO usually forces a choice: publish slowly and keep quality, or publish at volume and ship slop. The chain from keyword to published page is four distinct jobs, and handing all four to one model in one prompt is exactly how the slop gets made.
The call that made it work
Put the LLM in the loop rather than in charge of the loop. Each step has a narrow job and a checkable output, so volume comes from automating the handoffs between steps — not from asking a model to do the whole thing at once.
Select a step for the reasoning
Keyword research
The automated chain begins at keyword research, so the target set is generated rather than hand-assembled.
- pages generated in ~5 hours
- 100+
- organic traffic
- +40%
Programmatic SEO with an LLM in the loop, without shipping slop
The volume number and the quality bar are usually treated as a trade-off. They are not, if the chain is decomposed properly: research, intent, authoring, and publishing are separate jobs with separate outputs, and only one of the four is a writing task.
Most programmatic SEO fails because it collapses all four into a single generation step and then judges the result on word count. Splitting them means each step can be checked on its own terms, and the model is never asked to invent the strategy and execute it in the same breath.