Stop chasing keyword difficulty: a query-cluster framework for 2026

I've been doing keyword research for twelve years and I have never trusted a Keyword Difficulty score less than I do today. The scoring models were built when the top ten was reliably a stable list of ten blue links. The top of a 2026 SERP is a paragraph from an AI summary, a video carousel, two product packs, a People-Also-Ask block, and four organic results that often aren't the ones a human will read.

Single-keyword KD treats those SERPs as if they were comparable. They aren't. A 'low difficulty' keyword whose SERP is dominated by an AI summary and a video is harder to win than a 'high difficulty' keyword with ten conventional blue links.

Cluster-level competitiveness

The unit of planning isn't a keyword. It's a cluster — a tight group of queries that share intent and would be served by a single page or a small page family. For each cluster, score three things and ignore single-keyword KD:

• Format dominance: does the SERP show the same content shape across the cluster (long-form guide vs. comparison table vs. video vs. tool)? If yes, that's the format you have to ship. If no, the cluster is fragmented and the win is to be the answer for the most fragmented variant.
• Authority floor: pull the ten pages most often appearing across the cluster. What's the median referring-domains count of the publishing domain? That's your real opening bid, not a 0–100 difficulty score.
• AI surface penetration: how often does an AI-generated answer appear above the organic results in this cluster, and is the answer sourcing the pages that rank, or sourcing pages that don't? This is the single biggest deciding factor and it isn't in any keyword tool.

A worked example

Last quarter a client wanted to attack a cluster around 'how to set up SSO for a small team.' The top single-keyword KD numbers were comfortable. The cluster told a different story: the SERP for nine of the top twelve queries opened with an AI summary that cited the same two pages — Okta's docs and a Vercel guide — and a video. The lift available from a tenth blue link was tiny.

We changed the brief. Instead of writing the eleventh long-form guide, we built a configuration tool that generated the exact identity-provider config snippets the user needed. That tool earned the citation on six of the nine AI summaries within ten weeks. Single-keyword traffic was unimpressive; cluster-level branded follow-on search and direct visits to the tool were not.

Keyword difficulty isn't dead. It's a screen, not a strategy. The strategy is choosing clusters where the SERP rewards the format you can ship — and admitting when it doesn't.