Prompt Structure as a Proxy for Decision Stage

A keyword tells an advertiser what term a person typed. A prompt tells an advertiser what that person is trying to do, who they're doing it for, what they've already ruled out, and what's stopping them from deciding. That gap, between a label and a fully-formed expression of intent, is the subject of this piece: prompt structure, read correctly, functions as a reliable proxy for where a person sits in a purchase decision, and it does so in ways keyword matching was never built to capture.

Why prompt structure carries intent information that keywords cannot

A keyword is what's left over after a person compresses an actual need into the handful of terms they believe a search box will understand. The compression is lossy by design: the budget gets dropped, the constraint gets dropped, the fact that they already tried and rejected a competitor gets dropped, because none of that fits into a search query shaped by two decades of interface convention. What survives is a vocabulary shaped by the box, not by the mind that typed into it.

A prompt doesn't undergo that compression. A person writing to a conversational system writes to be understood, so the budget, the constraint, the prior rejection, and the stage of the decision tend to appear in the sentence rather than get stripped out of it. Contextual targeting in ChatGPT ads is built around this difference: it operates on conversation-level understanding rather than page-level content matching, because the "content" being evaluated is generated in real time through dialogue, and relevance has to be reassessed continuously as that dialogue moves forward. A user planning team logistics might talk for several turns about deadline tracking and remote coordination problems without ever typing "project management software," yet a system reading the pattern of that conversation can treat it as a high-intent signal without needing the category label stated.

The targeting mechanism built for this environment reflects the shift directly. There's no keyword bidding layer standing between the advertiser and the user: the primary lever is a 280-character natural-language context hint, matched against the live conversation through a relevance-weighted second-price auction. Because relevance weighting sits alongside bid price in that auction, a more relevant ad can beat a higher bid, which turns the advertiser's ability to read conversational intent precisely into a structural advantage rather than a stylistic preference in ad copy. The rest of this piece is concerned with how to read that intent, stage by stage, from the structure of the prompt itself.

How intent evolves across a multi-turn conversation

Intent inside a conversation is not a fixed property set at the first message. It moves, and it moves in a direction that can usually be tracked, which is what makes conversational structure a proxy for decision stage rather than simply a richer version of a keyword.

A targeting system built for this environment has to interpret the whole trajectory of a conversation, not just the line typed most recently: what the user asked earlier, how the assistant answered, which follow-up questions emerged from that answer, and where the thread appears to be heading. This isn't a speculative design goal. Transformer models underlying systems like ChatGPT already maintain conversational state, the property linguists call discourse coherence, in order to generate responses that make sense in context, and advertising systems built on top of these models draw on the same contextual representations to evaluate whether an ad fits the moment.

The directionality matters as much as the state-tracking. A conversation that opens with a simple restaurant recommendation can, within the same session, evolve into trip planning, then into a budget conversation, then into transportation logistics, with each phase representing a distinct commercial opportunity that a page-level contextual system built for static web content could never have separated out. What matters for reading a conversation isn't only what a person asked but the direction the thread is moving and how quickly it's narrowing. A handful of exchanges, by the fifth or sixth message, can establish a detailed profile: a role, a stage of operation, an active shortlist, a level of purchase readiness, all built by the conversation itself rather than by a third-party data layer sitting outside it. Microsoft's Copilot targeting model is built around the same premise: ads are triggered by considering the whole conversation within a session, not only the last prompt, and Microsoft's separate "ad voice" feature, which explains to the user how a given ad relates to their conversation before presenting it, is the user-facing expression of that same session-level read.

Open-ended exploration: what early-stage prompts signal

The earliest stage of a conversation reads as structurally broad. The person is using category-level vocabulary, imposing few or no constraints, and asking the assistant to orient them rather than to help them decide. A representative prompt at this stage might read something like "what should I know about project management tools for a small team," which names a category but no product, no price range, and no prior history with any competitor.

A handful of structural markers mark this stage out. The framing tends toward open questions: "what are my options for," "how do people usually," "what should I know about." Budget, timeline, and specification constraints are typically absent. The vocabulary stays at the category level rather than dropping to a named product or a specific integration requirement, and requests often ask for a comparison across a broad field rather than between two or three named alternatives.

What this tells an advertiser is that the person is building a mental map of a category rather than narrowing toward a shortlist, and the correct job at this stage is to be present inside that map rather than to attempt a close. The scale of this stage shouldn't be underestimated: in sportswear, eMarketer found that 39% of prompts reflected exactly this kind of upper-funnel, informational activity. Because relevance is derived from the live conversation rather than from historical browsing behavior, this kind of targeting doesn't require persistent user identifiers or cross-site tracking. The conversation itself carries the signal. The privacy model looks different from the cookie-based targeting that search and display advertising have relied on for years.

Active evaluation: the structural shift that marks the middle of the funnel

The middle stage of a conversation looks structurally different from the exploratory stage, sharply enough to be the single most useful signal in the whole framework. Evaluation-stage prompts are narrower and comparatively framed: the person has already formed a shortlist, has introduced personal constraints, and is asking the assistant to help them choose between known options rather than to orient them toward unknown ones.

Several structural markers distinguish this stage from the exploratory one that precedes it. Named alternatives start to appear in the prompt itself, in constructions like "between X and Y, which". Constraint vocabulary enters the conversation: budget ranges, team size, integration requirements, geographic limits, specific timelines. Rejection signals appear too, in phrases like "I already tried" or "the problem with X is," which carry information a keyword could never transmit because a keyword has no room for a rejected alternative. Comparative questions start assuming prior research has already happened, in the form of asks like "does it integrate with Slack" or "is it better for small teams", and follow-up questions start making sense only in light of an earlier exchange, which is precisely the discourse coherence that transformer models are built to track.

Vocabulary specificity is the clearest single marker in the entire framework. The shift from category terms to product-specific or feature-specific language signals that a person's mental model has narrowed from an entire category down to a shortlist of two or three real candidates. It's the conversational equivalent of moving from a head term to a long-tail keyword in search, except the signal carries more information here, because the constraint and the rejection travel alongside the narrowed vocabulary rather than getting stripped out of it.

For an advertiser, this is the moment to show up with differentiation rather than general awareness, since the person already knows the category exists and the job of the ad is to resolve a comparison already underway rather than to introduce a new one. Ad copy that names the constraint the person has already expressed and answers it directly, something like "built for teams under 20 people, no Salesforce required," fits the register of the conversation in a way no keyword-triggered ad could manage, because the keyword never carried the constraint in the first place.

Near-purchase prompts: the structural signals of a constrained, imminent decision

The final stage in the framework is structurally constrained and action-oriented. The person has resolved both the category and the shortlist, and the conversation has shifted from helping them choose to helping them act.

The markers here are distinct from both earlier stages. Action verbs replace question verbs: "help me set up," "walk me through signing up for," "what do I need to get started with." Implementation detail enters the conversation directly, in the form of specific pricing tiers, trial terms, onboarding steps, and compatibility checks. The conversation narrows to a single option, and the questions shift from "which" to "how." Urgency and timing language often appears alongside this, in phrases built around a deadline or a specific window, such as "before," "this week," or "by the time I."

This is the stage where stated intent maps most directly onto purchase probability, because the evaluation work has already happened inside the conversation and what remains is execution rather than comparison. A conversation about financial tools for a small business can, by the fifth or sixth message, establish a role, an operational stage, an active shortlist, and a high level of purchase readiness, a profile built entirely by the conversation rather than supplied by an external data layer. Amazon's Rufus assistant demonstrates the commercial weight of this moment in practice: shoppers who engage with Rufus during a session are substantially more likely to complete a purchase, which shows that conversational specificity tracks with transaction probability. For an advertiser, the job at this stage is almost entirely about removing friction rather than adding a new message: offers, trial access, direct product links, or a checkout path that removes a step rather than introducing one.

This stage is the highest-intent moment in the whole framework, yet it is the one most likely to go uncredited by the measurement systems most advertisers already have in place, a problem the next two sections return to directly.

Translating the stage framework into context hints and campaign structure

The practical instrument for all of this is the context hint, and writing one well means describing the conversational situation each stage produces rather than listing the product features an advertiser wants to promote.

An exploratory-stage hint should describe a situation rather than an offer: "someone asking what options exist for [category], hasn't mentioned a budget or a specific product yet" matches the register of a person who is still building a mental map, not closing in on a decision. An evaluation-stage hint should describe a comparison already underway: "someone weighing [named constraint] between two or three [category] options, has mentioned team size or integration requirements," with the constraint vocabulary in the hint deliberately echoing the constraint vocabulary that appears in the prompt itself. A near-purchase hint should describe an action: "someone asking how to get started with [specific product type], requesting pricing tiers or onboarding steps," with action verbs in the hint mirroring the action verbs the person is actually typing.

Teams that write context hints as comma-separated topic lists consistently get poor matching, because a topic list labels a category the way a keyword does, while the actual skill required is describing a conversational situation in the same register the conversation itself uses. The auction mechanic rewards exactly this kind of precision: because a more relevant ad can beat a higher bid, a smaller advertiser with a precisely stage-matched hint can outcompete a larger advertiser running a vague one, an unusual property for paid media that is worth exploiting while most of the field is still writing hints the way it used to write keyword lists.

The campaign structure that follows from this is straightforward: build stage-segmented ad groups, one per stage, each carrying its own hint, its own creative register, and its own conversion action, rather than a single campaign aimed at a category broadly. The creative register has to shift with the stage: informational in tone for the exploratory group, focused on differentiation for the evaluation group, focused on removing friction for the near-purchase group. The conversion action has to shift too: a content asset or a newsletter signup fits the exploratory stage, a comparison page or a demo request fits evaluation, and a trial signup or a direct checkout path fits the near-purchase stage, where every extra step costs conversions that the conversation has already earned.

Where the stage signal breaks down

Prompt structure is a reliable proxy for decision stage, not a perfect one, and the conditions that degrade it matter as much to an advertiser as the framework itself.

The clearest limit is structural rather than behavioral. If a buyer's employer pays for a paid AI tier, every paid tier is ad-free by design. The reachable audience for ads is, by construction, limited to the segment that hasn't paid to remove them. For B2B advertisers specifically, this means the highest-intent near-purchase signals, the ones coming from senior decision-makers like CISOs or VPs of engineering who are far more likely to be on a paid seat, may simply be invisible to the ad system, because the most structurally specific prompts are happening on tiers that carry no advertising at all. The framework holds up best for consumer categories and for broad-audience B2B products where the free tier still captures a meaningful share of the people who actually buy.

The second limit is that a person can leave the conversation and convert later through a different surface entirely, and a last-click model has no mechanism for crediting the ad impression that landed at the moment of peak intent. The practical response is to stop treating the stage signal as a direct conversion trigger and start treating it as a reach-and-frequency input instead, measured through creative-level tagging and holdout testing rather than through a last-click report that was never built to see a conversation happening on a different surface, at a different time, days after the questions that actually decided the outcome.