Keyword Research for Content Sites - Find Words You Can Actually Win

Why does most keyword research produce nothing?

Because it stops at the spreadsheet. The common workflow — dump seed words into a tool, sort by volume, chase the big ones — skips the two steps that decide outcomes: does the live search results page actually answer what searchers mean, and can your page realistically beat what is there. Volume without those two checks is a lottery ticket with extra steps.

The fix is treating research as a decision pipeline: mine broadly, validate harshly, score only what survives, and record the verdict with a review date so nothing rots in a tab.

Where do keyword candidates actually come from?

Nine channels, in two tiers — the ones you can check yourself and the ones that need an account or an app:

• Search signals: Google Trends rising queries and related searches at the bottom of results pages; autocomplete walks; “people also ask” chains.
• Community anomalies: niche subreddits and forums where the same question keeps resurfacing with thin answers — a repeated question is an unclaimed page.
• Competitor gaps: sites covering your topic where a sub-section is obviously thin or abandoned; tool-marketplace reviews where users complain about what a product cannot do.
• Adjacent winners: whatever content won in a neighboring niche translates — “best X for Y” formats and beginner-guide structures port across topics with new research.

If you run this mining as a repeatable workflow — channels in order, candidates logged with source and date — the whole loop can be automated. The AnvilWiki template ships an agent skill that runs this exact pipeline and writes candidates into a decision table for you.

How do you validate a candidate in two minutes?

Open the search results page for the keyword and answer three questions:

1. Intent match: is the top ten answering what a searcher actually wants, or just sharing the words? A search for “how to start a recipe site” that returns domain-name marketplaces and app builders has failed the searcher — an intent mismatch is an open seat.
2. Weak positions: forums, PDFs, auto-translated pages, or sites with obviously dated content in the top ten are seats your page can take.
3. Search sprouting: do related searches and autocomplete sprout longer forms of the keyword (“… for beginners”, “… without coding”)? Sprouting means demand is growing, not decaying.

How do you score and decide?

Score only candidates that passed validation, and score honestly — every unmeasured factor is marked as untested, never counted as zero. The framework: keyword volume tier, difficulty estimate, weak-position bonus, trend direction, and intent match — with a written verdict into a decision table that records the candidate, the evidence, the verdict, and a review date three, seven, fourteen, and twenty-eight days out. Words with no search-data evidence yet stay in the review lane — “untested” and “zero” are different sentences.

The site-type decision itself is the output of this process: each direction on this site won its seat through the same validation, which is why the recipe-site tutorial and the coupon-site tutorial exist — both survived the table.

What does this look like on a real site?

The mining workflow above runs against this very site’s tutorial line: every tutorial in the hub entered through a scored decision row, and the rows carry review dates so stale verdicts get re-validated instead of quietly rotting. Start your own table with three to five candidates — one round, fully validated, beats fifty rows of unvalidated spreadsheet.