Is it worth keeping if we have meaning-based search?

Yes, and the reason is identifiers. Codes, references, error strings and surnames all need exact matching, and a system reasoning about meaning will treat two similar codes as similar things, which is the correct behaviour for prose and the wrong one here. Almost every serious product now runs both.

Why do people say it is obsolete?

Because its weakness is vivid and its strength is invisible. Everybody has experienced searching for a phrase and finding nothing while the document sat there under different words, and nobody notices the times an exact match was silently correct. The complaint is memorable and the success is not.

What should we tell people who get no results?

That an empty page usually means a vocabulary mismatch rather than a missing document, and that trying the words the author would have used is the fastest fix. It is a small piece of guidance that recovers a surprising amount, because the material genuinely is there most of the time.

Glossary

Keyword search

Keyword search matches the words as they were typed, which makes it exact on names, codes and part numbers and useless the moment somebody phrases the same question differently.

In plain terms

Looking for the words you actually typed. If you search for a part number it finds that part number, precisely, every time, which is exactly what you wanted. If you search for how to fix a leak and the document says repairing water ingress, it finds nothing at all and reports that with complete confidence.

Why it matters

Because it is routinely described as obsolete by people selling the alternative, and it is the half of the arrangement that handles anything with an identifier in it. Product codes, error strings, invoice numbers, surnames and legal references all have to match exactly, and a system that only understands meaning will confidently return something adjacent to what you asked for.

How it works

It matches the characters rather than the sense, which is what makes it both exact and brittle in the same breath. A document containing the string you typed is found and one expressing the identical idea in different words is not, so the quality of a result depends entirely on the searcher happening to use the vocabulary the author used. That is a strong assumption in any organisation where several teams write about the same things.

Its precision is genuinely unmatched on identifiers. Nothing else reliably distinguishes one part number from a very similar part number, because meaning-based approaches treat near-identical strings as near-identical things, which is correct for prose and disastrous for codes. This is the reason the technique never went away.

The failure is silent and total rather than partial. A vocabulary mismatch produces no results at all, not worse results, and an empty page reads as an absence of documents rather than an absence of matches, so people conclude the material does not exist and stop looking. That misreading is expensive in organisations where the answer usually does exist somewhere.

Modern systems keep it and run it alongside something that understands meaning, which is the arrangement almost everything in this guide now ships. One half catches the exact string and the other catches the rephrasing, and combining them is a great deal more useful than either alone.

Two searches, two kinds of wrong

There is a useful asymmetry buried here that rarely gets stated. When exact matching fails it returns nothing, which is unhelpful and completely honest: the searcher knows immediately that something has gone wrong and can try again with different words. When meaning-based matching fails it returns a document about a very similar thing, presented with exactly the same confidence as a correct result, and the searcher has no signal at all that they are looking at the wrong product's specification. In prose that near-miss is usually a feature, because a document about the adjacent topic is often what somebody actually needed. In anything with an identifier it is the expensive failure, and it is expensive precisely because it looks like success. That is the whole case for keeping the older technique in the arrangement, and it is why the products in this guide combine the two rather than picking a side.

Seen in the wild

Enterprise search across an organisation's applications, where a part number and a vaguely worded question have to be served by the same box.
Glean
A curated knowledge layer whose agents answer with references, sitting on material where exact product names matter as much as phrasing.
Guru
A workspace assistant answering from pages and databases, where somebody searching for a project code expects that code and nothing adjacent to it.
Notion AI

Common misconceptions

People assume

Meaning-based search replaced it.

In fact

It joined it. Nothing else matches an identifier exactly, so systems that dropped this half started returning documents about a similar product rather than the one whose code was typed, and the arrangement everything now ships runs both and combines the results.

People assume

An empty result means the document is not there.

In fact

It means no document used those words. The material is frequently present and described in vocabulary the searcher did not think of, and because an empty page looks like an absence rather than a mismatch, people stop rather than rephrasing. That is the characteristic and expensive failure.

Telling them apart

Keyword search vs Semantic search

Keyword search

Finds the words you typed, exactly.

Semantic search

Finds the meaning, whatever words were used.

Ask whether a near-miss is useful. For a part number it is worse than nothing; for a question about a process it is often exactly right.

Questions

Is it worth keeping if we have meaning-based search?: Yes, and the reason is identifiers. Codes, references, error strings and surnames all need exact matching, and a system reasoning about meaning will treat two similar codes as similar things, which is the correct behaviour for prose and the wrong one here. Almost every serious product now runs both.
Why do people say it is obsolete?: Because its weakness is vivid and its strength is invisible. Everybody has experienced searching for a phrase and finding nothing while the document sat there under different words, and nobody notices the times an exact match was silently correct. The complaint is memorable and the success is not.
What should we tell people who get no results?: That an empty page usually means a vocabulary mismatch rather than a missing document, and that trying the words the author would have used is the fastest fix. It is a small piece of guidance that recovers a surprising amount, because the material genuinely is there most of the time.