Query Intent and Taxonomies

"Apple" typed into a search box could be a request for the company's stock, the fruit's nutritional content, a nearby store, or the trailer for a film. Ranking cannot resolve this, because the disagreement is not about which document is most relevant but about what kind of thing the answer should be. Query understanding is the layer that decides the question before retrieval tries to answer it.

Broder's taxonomy, which still organises the field, splits web queries three ways (Broder, 2002, A Taxonomy of Web Search, SIGIR Forum 36(2)).

Navigational: the user wants a specific site and has one correct answer in mind. "Facebook login". Success is the right result at rank 1; there is no partial credit and no benefit from diversity.

Informational: the user wants to learn something, and several documents may each be useful. "How does BM25 work". Success is a good set, and diversity across interpretations has real value.

Transactional: the user wants to do something, buy, download, book. Success is reaching a page where the action is possible.

Modern systems extend this with domain-specific verticals and with an entity dimension: whether the query names a known entity, and if so what type, which routes it to a structured answer rather than to document retrieval.

Why classification changes the pipeline, not just the ranking

A navigational query should trigger a tight, high-precision path: strong preference for domain matches, no diversification, and no reformulation. An informational query benefits from expansion, from diversification across senses, and often from a generated summary over several sources. A transactional query should surface structured commerce data with price and availability rather than prose.

These are different systems, not different weightings, which is why intent classification usually sits before retrieval rather than being folded into a ranker's features. In retrieval-augmented generation the same split reappears: a question with one factual answer wants precise retrieval and a short grounded response, while an open-ended one wants broader retrieval and synthesis, and treating both identically is a common design failure in RAG systems.

Signals that work

Query-only classification from the string alone is weak, since the string is short and ambiguous by construction. The signals that carry information are behavioural and contextual.

Click entropy over historical results for the query is the strongest single feature: a navigational query concentrates nearly all clicks on one URL, and an informational one spreads them. Query length and syntactic form correlate with intent, since questions signal informational intent and brand-only strings signal navigational. Session context disambiguates: the same query means different things after "python tutorial" than after "snake species". Entity linking against a knowledge base identifies which known entities appear and what types they have.

When it breaks

Mixed and ambiguous intent is the normal case. A large fraction of queries carry more than one plausible intent with non-trivial probability. The design response is a probability distribution over intents driving a blended result page, not a hard classification, and systems that commit to a single label produce confidently wrong pages for the second-most-likely reading.

Head and tail behave differently. Behavioural features require history, and the tail is where most distinct queries live and where no history exists. A classifier excellent on head traffic can be useless on exactly the queries that most need understanding, and evaluation weighted by traffic hides this completely.

Intent shifts on events. A brand name gains a news intent overnight when something happens; a film title changes meaning at release. Any model trained on historical clicks lags these shifts, and the lag is longest for the queries where being wrong is most visible.

Personalisation and intent are entangled. The same query from two users can carry genuinely different intent, so a global classifier is estimating an average over a population that does not share the property. This is where per-user context helps and where it starts trading against privacy and predictability.