Entity-mention share metrics: pooled denominators and dictionary phrases both get hijacked by unrelated viral usage

When ranking named entities (brands, products) by growth in social-media captions, two silent failure modes inflate or deflate results. (1) A pooled denominator ("captions naming any entity per week") is contaminated when one entity name is reused as a meme/spam hashtag: one name produced 1.2M captions in a single summer on unrelated funny videos, inflating the pool and making most other entities look like they grew. (2) Phrases that pass dictionary filters still mean something else in the scoring window: "on cloud" was 65% "on cloud 9" (a concert tour), "caesars" 78% a pizza chain, a surname became a politician in the news, a brand name was Honduran slang, a town name. Aggregate tests (unit tests, matcher diffs) do not catch these.

1) Validate the pooled denominator with a median-entity index: pick a stable panel (entities with >= 20 mentions in >= 95% of weeks), compute each panel member's weekly share / its own mean share, take the weekly median. If the index ratio between comparison windows is ~1.0 the pool is fine at window level. Weekly dips mark meme weeks. 2) Check the median growth of entities under different selection rules: selecting on the current-period level inflates the median (selection effect), so compare with selecting on the base period. 3) Before publishing any ranked list, sample 10–25 raw texts per surviving phrase from the scoring window itself, not all-time top items, which hide current usage. Exclude rows whose sample is >= ~30% another meaning, list the exclusions with counts in the output, and qualify the phrases (e.g. "caesars palace", "goodyear tires") for the next run.

Hashtag ecosystems recycle short names for SEO and memes, and natural-language phrases collide with idioms, places and people whose popularity changes over time. A share metric then measures the collision's popularity, not the entity's. All-time "top receipts" sorted by views hide this because the collision often dominates only in the recent window.

Weekly counts of entity mentions from a large caption dump, compared year over year to find rising entities.