Publishing at scale without publishing anything wrong
Programmatic SEO means building a large number of pages that each answer one specific question a buyer actually types. Done carelessly it produces a content farm. We run it in our own business, and it is the clearest illustration of how agents and deterministic code divide the work.
The failure mode
Point a model at a product catalogue, ask for a page per variant, and publish the results. Hundreds of near-identical pages, confidently wrong about prices and specifications, quietly damaging the brand they were meant to promote. The mistake is not automation. The mistake is trusting the draft.
Agents do the part that needs judgement
Researching the topic, gathering source material, deciding what a page should cover, and producing a draft. This is reading, weighing and writing — genuinely the work of a model, and slow, expensive work for a person.
Code does the part that needs to be right
Nothing an agent drafts is trusted. Every draft is handed to a deterministic pipeline that checks it against a database before anything is allowed to publish:
- Do the product facts match the system of record — price, specification, availability — or has the model invented a detail?
- Does this page already exist, or duplicate one that does?
- Are the required fields present and within limits?
- Do the internal links resolve, and is the structured data well formed?
- Does the page clear the minimum bar to be worth publishing at all?
A draft that fails is not published. The failure is recorded with its reason, and the page waits for a person. Nothing reaches the live site on the strength of a model sounding confident.
Validators are code, never prompts. A prompt that enforces a rule today will quietly stop enforcing it after the next model update.
The agent proposes, the database disposes
That is the whole architecture. The database — not the model — remains the authority on what is true. The model never writes to the record; it only ever suggests something that code then has to accept.
On one run, the generator refused to emit malformed output and caught a live defect before it reached production. That refusal is the return on the entire approach.
Where else this shape applies
Anywhere volume meets a correctness requirement: quotations generated from a price list, reports assembled from operational data, responses drafted against a policy document. The pattern does not change. Agents draft, code verifies against a system of record, and nothing ships unverified.