Cosine similarity measures how closely two pieces of text point in the same semantic direction by comparing the angle between their embedding vectors. A value near 1 means the two are talking about nearly the same thing; a value near 0 means they are largely unrelated in meaning.
| Term | Cosine Similarity |
|---|---|
| Category | Statistics and Evidence |
| Also known as | Embedding Similarity |
| Where it appears | Semantic enrichment evaluator |
What it means in RankGear
In RankGear, cosine similarity is the arithmetic behind the semantic enrichment evaluator. Each passage — a sentence, a heading, a candidate addition to your draft — is first turned into an embedding, a long list of numbers that encodes its meaning. Cosine similarity then compares two of those embeddings by the angle between them rather than their length, so it judges whether two passages are about the same subject independent of how long either one is. That is what lets the evaluator ask a focused question: does this sentence actually reinforce the topic of the pages that rank, or does it merely share a few words with them?
How to interpret it
The score runs from 0 to 1 on this scale, and it is the direction that carries the signal, not the raw magnitude of the vectors. Higher means closer in meaning, but “high” is relative to the embedding model and the material — read a score against the other scores in the same run rather than against a fixed cutoff you carried in from elsewhere. Two passages can share vocabulary and still sit far apart in meaning, and two that use different words can sit close, which is precisely the case where surface keyword counting misleads and cosine similarity does not. Before trusting any single number, confirm it was actually computed on real embeddings rather than defaulted, and remember that a similarity drawn from a handful of competitor pages is a shakier basis than one drawn from many.
| Approximate value | What it tells you |
|---|---|
| Close to 1 | The two passages express nearly the same meaning; strong semantic overlap. |
| Middle of the range | Related but distinct — some shared topic, some divergence. |
| Close to 0 | Little meaningful relationship, whatever words happen to coincide. |
Example
Suppose you are enriching a page on “commercial roof coatings” and RankGear proposes two sentences to add. One discusses silicone versus acrylic coating chemistry; the other mentions roofing only in passing while describing office lease terms. Against the embeddings of the pages already ranking, the first sentence scores a cosine similarity around 0.82 and the second around 0.19. Both contain the word “roof,” but the evaluator surfaces the first as genuinely on-topic and flags the second as a near-miss that would dilute the page — a distinction a simple keyword match would have missed entirely.
Important considerations
- Cosine similarity captures semantic direction, not truth or quality: two sentences can be highly similar and both be wrong, or dissimilar and both correct.
- The numbers are relative to the embedding model that produced them — compare scores within a single run, not across different models or tools, and treat them as comparative indicators rather than Google’s own scores.
- Confirm a low value is a real measured similarity and not a missing or defaulted embedding before acting on it.
- Small SERP samples make any derived target unstable, and semantic closeness to ranking pages is a correlation: matching it is an indicator, not a cause of ranking. Statistical association does not establish causation.
Related terms
Part of the RankGear glossary · how RankGear measures · the 870 factors.