Step 1 (week 1): confirm the Brand Brain, the sources, and the baseline
Review the Brand Brain the Content & SEO Engineer Digital FTE built from your documentation, guidelines, and existing pages: product facts, voice, vocabulary, the claims the brand never makes. Set the approved sources for the fact gate per property: documentation, policy pages, named public sources. Record the baseline: pieces published per month, revision rounds per piece, and the console and citation record for the questions you want to be visible for.
Write the fan-out list for the first three primary questions with the FTE, so the first topics have somewhere to go.
Step 2 (week 1–2): choose the publishing route with the engineers
If the site is a repository, connect it with a token that can open pull requests but not merge, scoped to the property. If it is not, install the open-source endpoint package in the site once, with the engineers reading it first. Either way, the engineers' involvement afterwards is the review they already do for any change, and nothing publishes without their merge or the endpoint's confirmation.
Agree the pull request convention: content in the repository's format, sources in the description, the fact-gate result stated.
Step 3 (weeks 2–6): run the pipeline gated and read the reasons
Topics arrive with worth-it verdicts; approve the few yeses. Drafts arrive having passed the fact gate, with sources attached and the gate's result stated; review for judgement and tone, send back with reasons. Approved pieces open as pull requests; engineers merge. Read the send-back reasons weekly and correct the Brand Brain; by week six the reasons should be about tone, not facts.
Hold the first monthly review at day 30: pieces approved and merged, approval rate with reasons, fact-gate blocks and what they revealed, usage. Promote nothing.
- Topic: worth-it verdict, approved by you
- Draft: through the fact gate, sources attached
- Approval: judgement and tone
- Pull request: reviewed and merged by engineers
- Measure: sampled, dated, honest
Step 4 (weeks 7–10): let the cadence emerge and build the cluster
By the second cycle the send-backs are few, the pull requests merge quickly because the convention holds, and the FTE's topic proposals draw on the measurement of the first pieces. Build the cluster: the sub-queries still uncovered become the next verdicts, each new piece links three related ones and names the same entities the same way, and voice-ready FAQs go on every piece. The cadence is whatever this loop sustains; it is not set in advance.
Where a real video exists for a piece, plan chapters to the H2s and publish the transcript with it.
Step 5 (weeks 11–13): the quarterly comparison
At the quarterly review, compare against the baseline: pieces published per month that passed the fact gate, revision rounds per piece against the old freelance or in-house number, console and citation movement for the target questions, approval rate, usage, and cost per published piece against what a piece cost before. Label what is unmeasured. Decide the three things, and keep publishing gated; a merge by an engineer is the right gate for a site.
Write the review note and set the second quarter's fan-out list.
What does done look like at day 90?
Done is a pipeline that runs every week with a topic verdict, a sourced draft, a human approval, and a merged pull request; a cluster growing from evidenced gaps; a Brand Brain that produces drafts right in substance; engineers who see a well-formed pull request rather than a request for their time; and a quarterly comparison read from the record. The content lead spent minutes a day on verdicts and drafts and an hour a month on the review.
It is not a content factory; every piece carries a source on every claim and something the index lacked, which is why the engine is worth running.
What should you refuse to do in the first quarter?
Refuse a fixed calendar before the loop has shown what it sustains. Refuse to let the FTE merge; the engineers' gate is part of the design. Refuse to approve a draft the fact gate blocked by adding a source afterwards. And refuse topics whose verdict was no because someone wanted the page; the verdict's reasons are usually right, and overturning it is logged for the review to read.
The engine's quality is the sum of these refusals.