When we talk about search, it can mean a bunch of different things, sort of spread along a spectrum of topicality/relevance of results: Unsolicited Recommendation → Solicited Recommendation → Find an Answer to a Question → Query-Defined-Category Sampling (Top-K) → Query-Defined-Category Retrieval → Pre-Defined-Category Retrieval (see attached chart image)

And as a user, you’re often met with unwelcome surprises, because sometimes even if you want one kind of search, you may be served results from another. Add on top of that how a search tool might move across categories over time (e.g. Google introducing paid-for promoted results), and it becomes pretty easy to tell how the user experience for search in 2026 is pretty bad.

Search Type	Description	Result Type
Unsolicited Recommendation	Ads Social media feed optimization	Few results Not what you asked for
Solicited Recommendation	Searching for “nearby restaurants” Similar product recommendations when online shopping Online dating Scam/Threat detection	Few results What you asked for
Synthesize an Answer to a Question	Search with AI	One natural language answer to your question, possibly listing or linking to other sources
Find an Answer to a Question	Google circa 2007-2016	Top result is likely the answer (for most queries) Following results are highly relevant
Query-Defined-Category Sampling	Get the top relevant results that match a user-defined category (i.e. a Query-Defined-Category) E.g. The first page of results for the query “treatments for cat allergy”	Few results Sorted by relevance to your query
Query-Defined-Category Retrieval	Get ALL results that match a query-defined-category E.g. every single article that discusses “treatments for cat allergy”	All results matching your query Sorted by relevance to your query
Pre-Defined-Category Retrieval	Old school information management systems, where documents are pre-tagged and you can click into categories to narrow down the list of options until it is small enough that you can scan it manually. While this is another version of full category retrieval, it has a very limited set of “queries” (e.g. category tags or labels) that are chosen and curated upfront, like in a physical library.	Many results Not sorted Filtered to categorical choices

Here’s what a better search system would look like: It would start with Query-Defined-Category Retrieval at the bottom, with relevance-sorted sampling (top-k results) built on top of the full retrieval, and then with answers synthesized from the sampled top-k results. Recommendations would be built on a blend of top-k (sampling), full retrieval, and synthesis. The important parts (for me as a user) are: no pollution of the results with unrelated content, optimize for minimal manual validation by the user of the results that were ultimately served. And most importantly, let users choose what kind of search they’re getting. Seriously, if I’m going for full category retrieval, you better believe I don’t want vector-similarity recommendations anywhere near that.

The important part here (and this is where today’s search technologies have diverged a lot) is that in this ideal search system, recommendation systems aren’t used to retrieve results, they’re used to synthesize recommendations from the results retrieved by full and sampled retrieval components. If you try to build full retrieval on top of recommendations, you’re going to get bad results. And that’s where we are today: most search tools today are built upside-down: fuzzy recommendation powers retrieval, which is why basic document search and retrieval is so bad in 2026.

Search Taxonomy

Experience Report: Vector Embeddings for Semantic Search