How does the research work, stage by stage?

1. Find keywords

Starting from the profile's seed topics, the research pulls candidates from three places:

• Keyword databases. Related searches, suggestions and ideas for each topic, with monthly volume, difficulty and cost per click.
• Comparable competitors. Sites like yours are found automatically, filtered so that giant publishers and marketplaces are not treated as peers, and the keywords they rank for that you do not are added.
• Customer language. Public discussions where people describe the situation that made them search. Their phrases are extracted with a verbatim quote and checked against search volume. Phrases without measured volume stay in the table marked unverified rather than being dropped or inflated.

For a live site, everything you already rank for in the top 30 is added too, with your position and page.

2. Keep the ones that could bring customers

Every candidate is judged against the confirmed profile: could someone typing this become a customer, which offering serves them, and what is the intent, buy, compare, learn, fix or navigational. Brand searches and excluded terms are set aside. Kept and dropped keywords both keep a one-line reason.

3. Group by what Google shows

Live results are fetched for the strongest keywords. Keywords whose top results overlap are one group, because one page can win them all. Each group gets a short name and a sentence on the underlying need. Groups are ranked by demand, tempered by difficulty, with places reserved for groups that came from real customer language.

4. Read the pages that rank

For each inspected group, the top pages are read in full. Each gets a review: its format, what it delivers to the searcher, where it fails them, an overall quality, and one to three verbatim quotes. Then the group gets a gap analysis: the dominant format, what the search features answer before a click, the specific gaps no top result fills, and a verdict.

5. Check whether you can win

Your domain's strength against the domains in the top ten, the keyword difficulty, whether a site of comparable strength already ranks, and whether you already have a page for the need. The result is a plain statement of the obstacle, not a score.

6. Decide

One action per group: build, improve or skip, with the page type, the working title, what it must contain to beat the current results, the work size, the commercial reason and the confidence. Two groups that would improve the same page are folded into one.

7. Brief

Every page worth building gets a brief: title, meta description, target keywords, audience, the angle, an outline with notes per section, questions to answer, how to beat each competing page, internal links and a word count.

Where the judgment comes from

Large language models do the reading and judging, from the stored evidence only. Nothing in the plan is invented; see Is the AI making this up?.