Why Keyword Clustering by SERP Overlap Beats Semantic Similarity Tools

Most keyword clustering tools group by semantic similarity — embedding two keyword strings, measuring how close they are in vector space, and calling them "the same topic" past some threshold. That's a reasonable proxy. It's also not the thing that actually determines whether two keywords compete for the same page.

Similar wording isn't the same as the same intent

"Best running shoes for flat feet" and "running shoes flat feet reviews" look nearly identical as strings, and most embedding models will score them as highly similar. But wording similarity isn't what Google uses to decide whether one page can rank for both — actual ranking behavior is.

The far more reliable signal is: do the two keywords currently pull up mostly the same pages in the top 10? If eight of the same URLs show up for both queries, Google has already decided, empirically, that a single page can satisfy both. That's not an inference from the words — it's an observed fact about the current SERP.

Where semantic similarity gets it wrong

Two failure modes show up constantly with embedding-based clustering:

• False merges. Two keywords can read as semantically close but pull

entirely different result sets — one skews informational, the other commercial — because of a subtlety the wording alone doesn't capture ("proxy pricing" vs. "how proxies work," for instance).

• Missed merges. Two keywords can look unrelated as strings but

actually rank nearly the same pages, because searchers use different vocabulary for the same underlying need. Semantic models trained on general language don't always catch this; overlapping rankings do.

What SERP-overlap clustering actually catches

Checking live top-10 results for every keyword and grouping the ones that share enough ranking URLs directly answers the question that matters: "will one page, done well, satisfy both queries — or am I about to publish two thin pages that cannibalize each other?"

That's the mechanism behind Intent Clusters: not a similarity score, but the actual, current search results for each keyword, compared against each other. A minimum-overlap threshold you control (how many shared ranking pages counts as "same intent" for your niche) turns a keyword list into a set of clusters — one content brief per cluster, not one per keyword, ready to hand to the writer instead of drafting each variation separately. For a site with hundreds of near-duplicate keyword variations, that's the difference between a coherent content plan and a stack of pages quietly competing with each other in the same search results.