Definition

Intent Detection

Intent detection is the runtime operation that reads ambient signals plus strategically placed questions to classify each arriving user against the intent categories theorized upstream — a best guess at why this particular user showed up.

Categories come from upstream

Detection happens at runtime, but the categories it sorts against are built beforehand. The framework theorizes intent categories from user-value research, and the move from "user wants a ride to a wedding" to "user is attending a formal event" is the craft of transition scoping that turns qualitative data into a usable model. You can't detect what hasn't been theorized, and you can't theorize what hasn't been investigated. Detection is the runtime half of a two-part operation: the multi-intent model gets built upstream, and individual users get classified against it as they arrive.

Three functions at once

A single detection sequence does three things simultaneously:

1. Classification — sorts the arriving user against the theorized categories.
2. Qualification — the questions themselves tell the user what the offering commits to delivering. Users self-route out of branches that don't match what they came for, which turns the sequence into a qualification filter rather than a pitch funnel.
3. Coherence test — the answers accumulate into evidence about whether the offering's self-conception matches what users are actually arriving for.

The third is the central one. Over time, the distribution of classifications becomes a continuous check on whether the theorized categories still describe the real user base — one of the mechanisms that keeps the framework honest about its own premises.

The signals it reads

Detection integrates whatever signals are available before the user is routed into a path: search terms, referral context, campaign attribution, and answers to questions placed at the entry to the transition. The sorting hat — the in-product question-branching instrument — is one signal stream among these, and usually the most controllable, because the team builds and owns it directly rather than depending on marketing or attribution tooling. No single signal has to carry the classification alone; detection forms its best guess from the combination, then routes the user to a matched path.

Audit signal as a byproduct

Once detection is running, the audit signal falls out of routing's normal operation, so there's no separate measurement function to build. Two axes carry most of it: the OSFA-routing rate (the share of users whose intent doesn't classify into any built category and who get the one-size-fits-all default) and the sequence-retention rate (the share who complete the sequence rather than dropping out partway). Movement in these over time is the intent-mix fit signal — drift between the theorized categories and the users actually arriving. The signal is in how the axes move, not where they sit at any given moment.

Example

A design tool asks one question at signup: "What are you here to make?" The options — a logo, a social post, a presentation, a website — route new users into different first-run flows. Detection reads that answer alongside the search term that brought the user in and the referral channel, then classifies the arrival: someone who searched "free logo maker," clicked a logo ad, and picked "a logo" is a high-confidence logo-intent classification, routed to the logo path. A user who picks "something else" lands in OSFA. Months later, "something else" has climbed from 8% to 22% of signups while sequence completion held steady — the categories built at launch are drifting from who's now arriving, and the coherence test is what surfaced it.

See also

• Approximate intent
• Discriminating conditions
• Sought state