AI Search User Intent Taxonomy: How Users Query Generative Engines

TL;DR

Classical 3-class search intent (informational, navigational, transactional) under-describes how people actually query ChatGPT, Perplexity, and AI Overviews. Add four AI-era classes — clarification, synthesis, multi-step task, verification — for a 7-class taxonomy. Map each class to a content_type, a structural template, and a citation pattern, then audit your corpus for coverage gaps.

Why a new taxonomy is needed

Andrei Broder's 2002 web-search taxonomy split queries into three classes that still hold for ten-blue-link search. Generative engines change the picture in three ways:

1. Conversational follow-ups turn one query into a chain. Engines preserve context across turns, so users ask shorter, less keyword-shaped questions.
2. Synthesis is offered for free. Users no longer need to read three pages and combine them; they ask the engine to combine, then probe the synthesis.
3. Verification is a first-class behavior. Users paste claims and ask "is this true?" — a class barely visible in classical taxonomies.

These behaviors don't fit into informational/navigational/transactional cleanly. We extend the original three classes rather than replace them so that GEO content plans remain compatible with existing SEO programs.

The 7 intent classes

1. Informational

Definition. The user wants to learn a fact, definition, or explanation.

Example queries.

• "What is dense passage retrieval?"
• "Explain how llms.txt works."

Recommended content_type. definition or reference.

Page structure. H1 → AI summary → short definition → mechanics → misconceptions → FAQ.

Citation pattern. Engines pull a single sentence or short paragraph as the answer; multiple sources are co-cited when definitions disagree.

2. Navigational

Definition. The user wants a specific page, product, or brand resource.

Example queries.

• "Stripe Connect docs"
• "Perplexity API pricing page"

Recommended content_type. landing or hub.

Page structure. Strong title, canonical URL, and clear entity binding so the engine routes the user to the official destination.

Citation pattern. Engines link out without summarizing. Brand and entity authority drive ranking; thin pages still win if they are the canonical destination.

3. Transactional

Definition. The user wants to perform an action: buy, sign up, install, configure.

Example queries.

• "Sign up for Notion AI"
• "Install the GitHub CLI on macOS"

Recommended content_type. tutorial or landing.

Page structure. Step list with verification clauses, prominent CTA, prerequisites block.

Citation pattern. Engines either link to the action page or extract the procedure into the answer. Pages with deterministic step lists and stable anchors win retrieval.

4. Clarification (AI-era)

Definition. A follow-up that narrows or refines a previous answer. Lives almost exclusively in conversational contexts.

Example queries.

• "Show me the version that uses Python instead."
• "What does that look like for a multi-tenant setup?"

Recommended content_type. comparison or modular tutorial with variant blocks.

Page structure. Variant-aware sections ("Python", "Node.js", "Go") under predictable H3 anchors so the engine can pull the right variant on demand.

Citation pattern. Engines pull only the variant section, not the whole page. Sub-page anchor stability is decisive.

5. Synthesis (AI-era)

Definition. The user asks the engine to combine multiple concepts or sources into a single recommendation.

Example queries.

• "Compare AI Overviews and AI Mode and tell me which to optimize for."
• "Summarize the latest research on prompt injection across the top 5 papers."

Recommended content_type. comparison, framework, or guide with explicit cross-referencing.

Page structure. Side-by-side tables, summary bullets, an explicit "verdict" section the engine can lift verbatim.

Citation pattern. Engines co-cite multiple sources and may quote the verdict directly. Pages with cleanly extractable verdict blocks dominate.

6. Multi-step task (AI-era)

Definition. A task that requires several actions across pages, tools, or systems. Often delegated to agents.

Example queries.

• "Migrate this repo from Heroku to Fly.io."
• "Set up GA4 + Looker + a weekly Slack digest."

Recommended content_type. framework or chained tutorial series.

Page structure. Clearly numbered phases, agent-friendly tool contracts (see Agent-Friendly Documentation), and explicit success criteria per phase.

Citation pattern. Engines (and agents) pull individual phase chunks as sub-tasks; the citation network is fan-out across multiple pages.

7. Verification (AI-era)

Definition. The user pastes a claim, output, or document and asks the engine whether it is correct, current, or trustworthy.

Example queries.

• "Is this Stripe error code still valid in 2026?"
• "Is this prompt-injection mitigation considered best practice?"

Recommended content_type. reference with strong source-of-truth signals (canonical_url, lastmod, citation_readiness=reviewed).

Page structure. Answer-first verification verdict, evidence with dated sources, change-log block.

Citation pattern. Engines anchor the verdict to a single source they can defend. Freshness, citation hygiene, and explicit dated provenance dominate.

Mapping table

How to use this taxonomy

1. Tag every query in your demand model with one of the seven classes.
2. Pick the matching content_type when planning a page; do not write a definition for a synthesis query.
3. Audit existing pages for class drift — a page meant to satisfy clarification queries that lacks variant H3 sections is misaligned.
4. Track citations per class. LLM citation benchmarks should bucket wins by intent class so you can see which classes you under-serve.
5. Coordinate with internal-link strategy — multi-step task pages need bidirectional links between phases.

Misconceptions

• "AI queries are just longer informational queries." Synthesis and verification have distinct success criteria; treating them as informational under-serves both.
• "Navigational doesn't matter in AI search." It still matters; engines route users to canonical destinations and reward strong entity binding.
• "Transactional is dead because users buy through agents." Agents still consume transactional pages — they just consume them as deterministic step lists rather than reading prose.

FAQ

Q: Should I publish a separate page per intent class?

Yes for high-volume topics. A topic that draws informational, comparison, and verification traffic deserves three pages (definition, comparison, reference) cross-linked under one hub, not one bloated guide that under-serves all three.

Q: How do I detect a verification query in analytics?

Look for queries that contain a quoted claim, a paste of an error code or excerpt, or phrasing like "is X still true" / "is X correct". They are often invisible in keyword tools because they are conversational, but they show up in LLM session transcripts and chat-mode logs.

Q: Do these classes apply to enterprise internal AI search?

Yes, with a fourth-from-the-bottom emphasis: internal queries skew toward multi-step task and verification. Enterprise docs benefit from explicit phase-numbered procedures and explicit dated verdicts.

Q: How does this relate to reader_modes in our frontmatter?

reader_modes are the publisher's declaration of how a page is meant to be read; intent classes are the user's declaration of how they query. The mapping table above shows the natural pairs, but a single page can serve more than one intent if its structure is modular.

Q: Does Google's classical 3-class taxonomy still work for ten-blue-link search?

Yes. The 7-class taxonomy is a superset, not a replacement. Pages that win classical informational queries usually also win the corresponding AI informational queries, provided they are chunkable.