Updated 2026-08-02

10,000 keywords from one Cerebro export: how to shrink the list without losing the market

You cannot read a 10,000-keyword export row by row, and you must not filter it down by volume — the honest move is to collapse phrasings of the same demand into families: normalize case and punctuation, ignore word order, merge plurals, fold numbers and sizes into classes. On real corpora this removes 23–37% of the rows and, more importantly, changes which demands look big. The unit of analysis stops being the keyword and becomes the family.

Why is the raw list misleading, not just long?

Buyers type one demand a dozen ways. In a real 11,547-keyword corpus, "car phone holder" exists as 29 phrasings — "phone holder for car", "car phone holders", "phone holders for your car"… — totalling 544,816 monthly searches. Read separately, each row looks mid-sized; only the family reveals a top-3 demand of the corpus. Meanwhile a consistently-spelled generic term sails to the top of your volume sort untouched. Fragmentation is not noise; it is a systematic bias against exactly the multi-phrase demands where niches live.

What are the rules of a trustworthy collapse?

Five deterministic rules — no AI, no similarity scores, every merge explainable:

1. Normalize the surface: lowercase, strip invisible characters (real exports contain them), punctuation becomes a space — "anti-slip" is two tokens, never "antislip".
2. Ignore word order: compare keywords as bags of words. "holder for car" and "car holder" meet; function words ("for", "the") drop out of the comparison.
3. Merge plurals with bounded language rules — "holders" joins "holder". Rules you can enumerate and justify, not a statistical stemmer that fuses what it should not.
4. Fold numbers into classes: "8.5 x 11" and "8-1/2 x 11" become the same dimension marker. Sizes fuse with each other — but a phrase WITH a dimension never fuses with the same phrase WITHOUT it: a buyer specifying a size is a different search.
5. Keep duplicates within the bag: "car car holder" is not "car holder".

Why under-merge on purpose?

Every grouping system errs in one direction. Ours is chosen: when in doubt, keep phrasings separate. A missed merge costs you a duplicated row. A wrong merge is worse in kind, not degree: it fuses two distinct markets, adds their volumes, mixes their Title Densities — and can make an already-served territory look virgin. The cheap error is tolerable, the expensive one is forbidden.

What changes once you read families?

• Volume re-ranks. Fragmented demands reassemble; "car phone holder" jumps past generic terms that looked bigger row by row.
• Metrics get honest aggregation rules. Demand adds up across phrasings (searches for variants are the same wallets); Title Density takes the family maximum (one titled variant disqualifies virginity); cost-of-entry metrics take the minimum (you enter through the cheapest door). One rule per column, each with a reason.
• The display name stays human. A family is identified by its normalized bag of words, but shown as its most-searched real phrasing — you read "car phone holder", never "car holder phone".

Where the collapse sits in the full method — and what to do with the families next: You ran Cerebro. Now what?