Semantic similarity
Updated 2026-09-15
Semantic similarity is the measure of how close two pieces of text are in meaning rather than in wording, usually computed from vector representations. Retrieval uses it to find passages that answer a question even when they share no keywords with it.
Why it matters for prompt sets
Because retrieval matches meaning, two phrasings of the same question can pull different passages and produce different answers — similar, but not identical. That is why a prompt set tracks several phrasings of the important questions rather than one canonical wording, and why content that states a claim in the reader's vocabulary is retrieved more reliably than content that states it in internal product language.
Illustrative pair (example)
"Tools to track brand mentions in ChatGPT" and "how to measure AI visibility" are near-synonyms to a human. Retrieval treats them as different neighbourhoods and returns overlapping but distinct passages — so a brand can be present in answers for one and absent for the other.
Common mistakes
- Tracking one phrasing per question and concluding the category is covered.
- Writing only in product vocabulary, which sits far from how buyers phrase the problem.
- Stuffing synonyms into a page instead of writing the claim clearly once — retrieval matches meaning, not repetition.
Frequently asked questions
- How many phrasings should a prompt set include?
- Two or three for the questions that matter commercially, one for the rest. The cost is linear in samples, so breadth of phrasing competes directly with sample size.
- Does keyword research still help?
- As a source of real phrasings, yes — the value is in learning how people say it, not in placing the exact string on the page.
