What it means
An embedding similarity threshold is the cutoff score that decides whether a chunk is close enough in meaning to a query to be retrieved or cited. The engine measures the distance between your chunk's embedding and the query's embedding, usually with cosine similarity, and only chunks above the threshold move forward. Everything below the line is invisible to the answer, no matter how good the writing is. Kevin's own rule of thumb for page-level overlap shows how these scores behave: A cosine similarity higher than 0.7 classifies as "high," while it's "low" below a value of 0.5.
Why it matters
Retrieval is a filter before it is a “ranking.” ChatGPT retrieves about 6x more pages than it cites, and a similarity threshold is the first gate that decides what even enters that pool. If your content sits just below the cutoff for a query, you are not competing for the citation, you are not in the running at all. This is why Kevin's Perplexity guidance is to meet embedding similarity thresholds by using semantically aligned language.
Say a B2B analytics brand publishes a deep guide on "data activation" but phrases everything in internal jargon. For the query "how to use customer data," its best chunk scores 0.48 similarity, lands under the threshold, and never surfaces, while a thinner competitor page written in the searcher's words clears 0.71 and gets cited.
How to use this knowledge
Mirror the query's language in the chunk. Phrase the answer using the words your audience and the model use, so the embedding clears the threshold.
Put the answer near the query terms. Keep the question and its answer in the same chunk so they vectorize together rather than drifting apart.
Cut stilted phrasing. Overly optimized or unnatural copy pushes your embedding away from how real queries are written.
Measure your own overlap. Use cosine similarity to spot pages above 0.7 that compete with each other and consolidate them.
Growth Memo guidance
Ensure your content meets embedding similarity thresholds by using semantically aligned language. Avoid overly optimized or stilted phrasing and provide unique insights that differentiate your content.
Though not an exact science, in my experience, a cosine similarity higher than 0.7 classifies as "high", while it's "low" below a value of 0.5.
ChatGPT retrieves about 6x more pages than it cites.
Semantically aligned language — the phrasing tactic that lifts a chunk above the threshold.
Cosine similarity — the score the threshold is applied to.
Vector (embedding) — the numeric representation being measured against the cutoff.
Retrieval — the filtering step where the threshold is enforced.
Citation Rate — the downstream metric a cleared threshold feeds into.