Cosine Similarity in SEO
Cosine similarity is the single most-used filter in Synthexeo's semantic SEO process — it decides what belongs in a SILO and what doesn't.
What you'll learn
- Understand what cosine similarity measures and how it's calculated from embeddings
- Learn the practical thresholds used to include or exclude a topic from a SILO
- Know the two standard Google Sheets templates used for scoring
- Apply cosine similarity to decide "one article or two?"
What Cosine Similarity Measures
Every piece of text can be converted into an embedding — a high-dimensional vector representing its meaning. Cosine similarity measures the angle between two such vectors, producing a score from 0 to 1: the closer to 1, the more semantically similar the two texts are, regardless of exact wording.
Formula: cosine(A,B) = (A·B) / (|A|·|B|). In practice, you never compute this by hand — a Google Sheets template with an embedded API script does it automatically once you paste in the topics being compared.
Practical Thresholds
| Score | Meaning | Action |
|---|---|---|
| 0.85 + | Very strong connection — near-duplicate meaning | Likely the same article; don't split |
| 0.75 – 0.85 | Strong, legitimate connection to the core topic | Separate article within the same SILO |
| 0.65 – 0.75 | Moderate connection | Judgment call — usually include, but flag for review |
| below 0.65 | Weak or unrelated | Exclude from this SILO / this content plan iteration |
"One Article or Two?"
Before splitting or merging planned content, run the candidate topics through a cosine check. Score above 0.85 with your core topic strongly suggests they should be one article, not two competing (cannibalizing) pages. Score between 0.75-0.85 means separate articles are justified, but they belong in the same SILO with mutual internal links.
Two Standard Templates
Synthexeo maintains two ready-made Google Sheets templates: one with the central topic in a single cell and a list of comparison topics in a column (returning a relevance score per row), and a second, more detailed version that also classifies each comparison as strongly/moderately/weakly related. Both must be copied (not edited in place) — the embedding API script only works in a personal copy.
Where Cosine Similarity Fits in the Bigger Process
Cosine similarity is the filter step between raw topic collection (competitor Top Pages, PAA, Wikipedia categories) and building the final SILO graph. Skipping this step is how content plans end up bloated with tangential topics that dilute topical focus instead of reinforcing it.
Key Takeaways
- Cosine similarity scores 0-1 how semantically close two pieces of text are, based on embeddings
- Practical thresholds: 0.85+ = one article, 0.75-0.85 = separate article in same SILO, below 0.65 = exclude
- Use the score to resolve "one article or two?" decisions objectively instead of guessing
- Synthexeo's two Sheets templates must be copied (not edited directly) — they carry an embedded API script
- Cosine similarity is the filter step between raw topic collection and the final SILO structure