In one sentence: spend the first fortnight establishing a baseline and clearing technical access, the next six weeks on extraction and evidence, and the last four on corroboration — and change one class of thing at a time so the rerun means something.
Weeks 1–2: baseline and access
Nothing here produces visible output, which is why it gets skipped, and skipping it is why most programmes cannot prove anything at the end.
- Write the prompt set. 20–40 buyer questions, no brand names, grouped by intent, versioned. This is the instrument; everything later is measured with it.
- Declare the competitor set and confirm detection is symmetric.
- Run the baseline and store it. Preserve the raw answers, not just the scores.
- Clear gates one and two — fetch as each named agent, and fetch with JavaScript off. These are hours of work and they invalidate everything downstream if left closed.
Preserve the baseline before you change anything. A programme without a before is a programme that ends in an argument.
Weeks 3–8: extraction and first-party evidence
Now the writing, but structural first. Rewrite the pages that should already be answering your highest-intent questions so that a self-contained answer sits in the first two sentences under each heading. Then fill the gaps the baseline exposed — the questions where nobody, including competitors, gave a good answer, which are the cheapest wins available.
Rerun at week 6. Expect movement in citations before movement in recommendations; citation responds to extraction, recommendation responds to corroboration, and you have not started the second one yet.
Weeks 9–12: corroboration
The slow layer. Complete directory and marketplace listings with every field filled. Get the entity into structured knowledge bases. Answer questions in the communities your citation data shows the engines actually read — your own source mix tells you which those are, so this is not guesswork.
Change one class at a time
| Phase | What changes | What you learn from the rerun |
|---|---|---|
| 1–2 | Access only | Whether absence was a fetch problem |
| 3–8 | Page structure and content | Whether extraction was the bottleneck |
| 9–12 | Off-domain evidence | Whether corroboration closes the mention–recommendation gap |
Changing access, content and off-domain evidence in the same month is the most common way to end a quarter with a better number and no idea which action produced it — which means you cannot decide what to do in the next quarter.
What to expect, honestly
Access fixes show up fast, sometimes within a crawl cycle. Extraction work shows up in citations within weeks. Corroboration is the slow one and often shows nothing inside 90 days, which is worth saying at the start rather than explaining at the end.
The end-of-programme artefact
Not a score. A before-and-after on the same question set, with the failed runs visible, a list of what changed in each phase, and the source mix showing where the next quarter's effort should go. That document is reusable; a dashboard screenshot is not.