If you search for content or coverage, you need to choose a method suited to finding an exact name or a broader subject. Keyword matching and vector similarity identify documents about a subject using different methods. Knowing which method produced a result helps you interpret it.
Keyword matching
The document is checked for the words you asked about. Present means match, absent means no match.
Its strengths are precision and predictability. If you search for a product name, an error code or a person, keyword matching finds exactly the pages containing it and you can explain every result.
Its failure is vocabulary. A page discussing your subject in other words is invisible to it.
Vector similarity
Text is converted into a list of numbers by a model, and that list is a position. Text about similar subjects produces nearby positions; text about unrelated subjects produces distant ones. In this product a piece of writing becomes 768 numbers. Google's embeddings overview explains the general idea.
Comparison is then between positions rather than between words, so a document can match without sharing any vocabulary with your query.
Its failure is the opposite one. It returns things that are near, and near is a matter of degree. A result can be topically adjacent and useless for your actual purpose, and the method cannot tell the difference between close and correct.
A worked example where each one fails
Suppose you sell software that helps restaurants manage staff rotas, and you want pages discussing the problem you solve.
Keyword search for "restaurant scheduling software" returns pages using that phrase. It misses an article titled "Why your kitchen is always short-staffed on Fridays" that discusses exactly your problem for three thousand words without once using your phrase. That article is probably the best prospect on the web for you, and a keyword search will never show it.
Similarity search for the same idea finds that article. It also returns pages about hospital shift patterns, retail workforce planning and restaurant point-of-sale systems, because all of them sit nearby. Some are useful adjacent markets and some are irrelevant, and the ranking will not separate them for you.
Neither method is wrong. They are answering different questions: which pages say this, and which pages are about this.
Which to use for which job
| Task | Better method | Why |
|---|---|---|
| Finding mentions of your brand | Keyword | The brand name is the thing itself, so exactness is what you want. |
| Finding coverage of your market | Similarity | Writers use their own vocabulary for a subject, not yours. |
| Finding a specific product or error | Keyword | Precision matters and near misses are noise. |
| Finding placements for a topic | Similarity | A relevant resource page may describe the field in unfamiliar terms. |
| Checking a draft against an intended subject | Similarity | The question is what the writing is about, not which words it contains. |
| Auditing whether a term is used | Keyword | You are asking a literal question about words. |
What a similarity score does not mean
Four misreadings are common enough to be worth naming, and all follow from treating a distance as a judgment.
- Not accuracy. A confident, well-written, factually wrong passage can closely match its subject. Similarity has no view on truth.
- Not quality. A short focused piece can score higher for similarity than a longer, better one that covers several subjects.
- Not a ranking signal. No search engine or assistant is involved in producing the number, so it says nothing about how one would treat the page.
- Not keyword coverage. A high similarity does not mean particular terms are present, and adding terms does not reliably move it.
One further condition applies to any similarity comparison. Both sides must have been produced by the same model. Vectors from different models are numbers of the same shape describing different spaces, and comparing them returns a value that looks like a score and means nothing.
Using both together
Use similarity search to find related material, then check exact keywords for details that matter. Check whether a page names your competitor, accepts contributions or mentions the standard you comply with.
Doing it in the other order limits you to what you already knew how to ask for.
Choose the search method when defining what you need to find. Check the returned documents so the research reflects their actual content.
See how semantic similarity supports content editing, meaning-based search for industry research, or how a draft is scored against a category.