AI-Augmented Keyword Research: A Workflow That Beats Ahrefs Alone

Ahrefs, Semrush, and Moz all derive their keyword databases from clickstream data, third-party search panels, and SERP scraping. These methods reliably capture queries with meaningful search volume — typically queries with 50+ monthly searches in major markets. They systematically under-represent queries with under 10 monthly searches, conversational multi-clause queries, and queries that have shifted to AI assistants where the data isn't visible to clickstream panels at all.

The buyer-research reality is that most informational and evaluative queries now happen in the long tail and increasingly in conversational form. A buyer doesn't search 'dental SEO Toronto cost'; they ask ChatGPT 'how much should a small dental clinic in Toronto budget per month for SEO if I want to rank in the map pack within 6 months and I'm currently invisible'. Traditional tools see neither that query nor the dozens of variations of it.

The first phase uses LLMs to generate a comprehensive question set for a topic. The prompt pattern that works in our workflow: 'You are [target buyer persona]. List 100 specific questions you would actually ask Google or ChatGPT when researching [topic]. Include questions at every stage — initial awareness, comparison, objection-handling, post-purchase. Phrase them in the actual language a buyer would use, not marketing language.'

Run the same prompt across ChatGPT, Claude, and Perplexity (or Gemini) and de-duplicate. The three models produce overlapping but non-identical query sets — the union is typically 200–400 unique questions, of which 60–80% would never appear in traditional keyword tools. This is the breadth pool.

Take the breadth pool from Step 1 and bulk-import into Ahrefs Keywords Explorer (or Semrush Keyword Magic) for volume, difficulty, and SERP feature data. Most of the queries will return 'no data' or volumes under 10 — that's expected and not a reason to discard them. The pattern that matters is which queries return any meaningful volume (typically 30–40% of the pool) and which have low keyword difficulty relative to your domain authority.

The queries that survive Step 2 are the ones with both demonstrated demand (volume > 10) and a credible chance of ranking (KD below your domain's typical winning threshold). Typically 40–80 queries from a 200–400 pool make it through. These become the prioritised content brief candidates.

For each candidate query from Step 2, manually review the current top-5 organic results. The diagnostic question: are the existing top-ranked pages weak in ways our page could exceed? Specific patterns we look for — outdated information (last modified 2+ years ago), thin content (under 1,200 words on a topic that warrants depth), missing structured data, missing author signals, dominated by Reddit/Quora threads (signal that no purpose-built page has won the query yet).

Queries where the top-5 are dominated by strong, recent, well-authored, well-linked content from established competitors get deprioritised — winning takes too much investment relative to the query's payoff. Queries where the top-5 has visible weakness get prioritised. This SERP audit is the highest-leverage step in the workflow and is where most automated keyword tools fail because the audit requires judgment.

For the top 10–20 candidates from Step 3, run the actual queries through ChatGPT, Perplexity, and Google AI Mode and record which sources are currently cited. This produces two pieces of intelligence: (1) the competitive set in AI search (often different from the organic competitive set), and (2) the gap — queries where AI assistants are citing weak sources or struggling to find a clear authority give you a credible AI-citation opportunity.

The AI citation gap is often the strongest signal of opportunity in the entire workflow. Queries where Google ranks competitor X strongly but AI Mode is citing thin Reddit threads or struggling to cite anyone authoritatively are the queries most likely to be won quickly with a well-built page. These get the top brief priority.

The final step is to cluster the prioritised queries into pages — typically 3–8 related queries per page — based on shared search intent and user goal. The clustering matters because both Google ranking and AI citation increasingly favour pages that comprehensively answer a query cluster over pages that target individual queries narrowly.

The clustering test we use: would a buyer asking query A naturally also ask queries B and C in the same research session? If yes, they belong on one page. If no, they're separate pages even when the keywords look superficially related. This produces the actual content brief set — typically 10–25 briefs from the original 200–400 query pool, each targeting a cluster the page can credibly win.