Prompt-Level Purchase Intent Signals and What They Reveal

A prompt inside an AI assistant tells an advertiser more than any keyword ever could, because a keyword records what someone typed and a prompt records why. That distinction, buried in the difference between "running shoes" and a paragraph explaining a bad foot and a tight budget and a race in ten weeks, is reshaping how purchase intent gets read, priced, and acted on. This piece maps what that signal actually contains and how the infrastructure around it is being built.

Search built an entire advertising economy on a query stripped of context. Type "running shoes" into Google and the system knows a category, maybe a location, maybe a device. It does not know why. It does not know that the person typing has had plantar fasciitis for six months, that money is tight this month, or that the goal is a half-marathon ten weeks out. A prompt to an AI assistant carries all of that in one turn, often without being asked. That's not a marginal improvement over keyword targeting. It's a different category of signal, and the rest of this piece works through what it contains, how it distributes across real usage, and what's being built to act on it.

Context for the moment: time spent inside GenAI apps is projected to hit 36 billion hours globally in the first half of 2026, up from 17.2 billion hours in the first half of 2025, according to Sensor Tower's State of AI 2026 report. ChatGPT users spend roughly 215 minutes a month inside the app; Claude users log around 120 minutes, Gemini around 100. That's not a glance and a bounce. That's enough time, across enough turns, for intent to form, shift, and sharpen in ways a single search query never captures.

The anatomy of a prompt: what it actually contains

A prompt is rarely just the ask. Pull one apart and there are usually several layers stacked on top of each other: the stated need, sure, but also the constraints (budget, timing, geography, some physical limit), the stage of consideration (still defining the problem, comparing options, or ready to act), and the trade-offs being weighed out loud ("I like the Brooks but I'm worried about heel drop"). There's often an emotional register too, anxiety or urgency ("I need this before Friday"), and sometimes a scar from a past failure ("I tried orthotics and they didn't help").

Research on transactional signal detection, cited by getchatads.com, points to a specific cluster that pushes intent scores higher. That cluster includes action verbs paired with a product, price or deal language, a named brand and model, direct phrases like "where can I buy," gift context, and urgency markers. Not every purchase-adjacent phrase counts as a buying signal, though. The same research points to a confidence threshold near 0.6 as the balance point between missing real sales signals and falsely linking casual mentions to purchase intent.

Worth sitting with: only about 34% of users say they want an AI to actually buy something on their behalf. The signal reveals intent. It doesn't collapse the decision into a transaction, and treating it as though it does misreads what most people are actually doing in these conversations, which is thinking out loud.

There's also a quieter layer of inference happening underneath the words. Academic research describes large language models assigning topics to conversations and then generating structured user profiles from chat history alone, inferring demographics, interests, and personality without ever asking a demographic question. What a user doesn't ask about matters too. The options they rule out, the comparisons they frame unprompted, and the attributes they volunteer without being asked all shape the same picture from the negative space.

How intent distributes across real conversations at scale

Profound's 2025 analysis of more than 50 million ChatGPT prompts breaks the traffic down: 32.7% informational, 37.5% generative or task-completion (the largest single bucket), 9.5% commercial, 6.1% transactional, 2.1% navigational, and 12.1% with no clear intent at all.

That 6.1% transactional figure looks small until it's set against the search baseline. SparkToro's analysis of 332 million Google queries found just 0.69% of them were transactional. So on a share basis, transactional intent inside ChatGPT conversations runs roughly nine times higher than what search has historically delivered.

The informational share deserves a second look too, because 32.7% is not a dead zone of window-shoppers. eMarketer's multi-prompt consumer research on a sportswear conversation found 39% of prompts were upper-funnel informational and 37% transactional, and the informational prompts carried buying signals that traditional search would never have surfaced. An informational prompt can carry purchase intent that only becomes legible when read in context, not as a standalone query.

Data on conversation depth backs this up: users typically go about six prompts deep before leaving the assistant to convert on the open internet. Intent arrives unformed, shaped only through capture. It's forming, turn by turn, inside the conversation itself. OpenAI's own figure, that roughly 20% of inquiries show a clear intent to purchase, is useful as a directional marker, though Profound's breakdown offers the more granular map of where that intent actually sits.

The hidden intent problem: what legacy platforms cannot see

The old model for reading intent progression relied on reconstruction. A user's interest got pieced together from breadcrumbs: several search queries spread across days, retargeting cookies stitching sessions together, behavioral clusters inferred from which pages got visited and for how long. It was indirect by design, because no single platform saw the whole arc.

Inside a multi-turn AI conversation, that whole arc can happen in one sitting. A user discovers a concern, warms up to the idea of a solution, compares two or three options, and signals readiness to buy, all inside a single thread. As Beet.TV reported in September 2026, that progression now happens within a single conversational thread, invisible to any ad platform that isn't inside the conversation itself.

eMarketer frames this as an upstream layer: intent is forming before it ever reaches a search bar. A brand that only shows up when someone finally types a query into Google is arriving after the consideration process already ran its course somewhere else, out of view.

There's a structural wrinkle here that doesn't exist in search advertising at all. Research out of Peking University and Alibaba Group, using the term "generative externality," describes how an inserted ad inside a conversational response can change the flow, tone, and specificity of the entire narrative that follows it. A search ad sits next to organic results without touching them. An ad embedded in a generated response can shape the response. That's a fundamentally different kind of ad environment, and it raises the stakes on how carefully that insertion gets handled.

Meanwhile the ground underneath legacy targeting keeps shifting. As third-party cookies fade out, the infrastructure that made platforms legible to advertisers erodes with them. Contextual signal pulled from an active conversation actively closes that gap. It's a structurally richer replacement, because it describes the person's actual situation rather than a probabilistic guess stitched together from past behavior.

What the signal reveals at each stage of the purchase journey

Early in the funnel, a user is often still finding the words for a problem. Prompts here run long and exploratory, thick with constraints volunteered before anyone asked, and the product category itself might not even appear yet. Someone describing chronic foot pain and a vague plan to start running hasn't typed "running shoes." Not yet.

Move into consideration and the prompts start naming names. Users bring up specific alternatives, weigh attributes against each other ("I care more about arch support than weight"), and this is where brand preference either forms or quietly falls apart. It's a stage legacy search rarely captures well, because comparison shopping there tends to happen across several disconnected queries rather than one visible thread.

By the decision stage, the transactional cluster shows up: action verbs, price mentions, urgency, a specific product, "where can I buy this near me." Profound's 6.1% transactional share is the narrowest slice of the intent distribution but the clearest to act on.

The turns compound as they go. A user on turn six, having opened with a vague problem statement and worked through comparisons along the way, represents a categorically different signal than a fresh query carrying the same surface keywords. Gift and deadline language compresses the whole funnel further: "it's for my mother's birthday, need it by Friday" pushes intent scores up even in what looks, on the surface, like a mid-funnel question.

What none of this could ever come from is demographic data. Two users who look identical on paper, same age, same income bracket, same zip code, can sit at opposite ends of the funnel with opposite constraints. The prompt surfaces the individual situation. The demographic profile only ever gave the cohort average, and cohort averages don't buy shoes. As systems build user profiles dynamically across chat history, per that academic research, that richness compounds across sessions rather than resetting with every new conversation.

How the advertising infrastructure is being built to act on these signals

ChatGPT Ads is the clearest working example of this shift moving from theory into product. The pilot launched in the US on February 9, 2026, and opened to self-serve advertisers on May 5, 2026. Targeting runs on the current conversation topic, with past chat history and past ad interactions available for targeting, rather than third-party cross-web tracking.

Ads show up in clearly labeled, subtly tinted boxes at the bottom of a response. The organic answer gets written first; a separate system decides on ad placement afterward, so the answer itself isn't shaped by what's being sold nearby. Pricing started around $60 CPM and has since shifted toward a primarily CPC model, running $2 to $8 depending on category, with a low barrier to entry for self-serve advertisers. Ad frequency sits around 20%, or roughly one sponsored placement for every five conversations. Beet.TV reported in September 2026 that the product had reached a $1 billion annualized revenue run rate in under 200 days, with tens of thousands of advertisers across more than 40 countries.

The mechanics underneath are genuinely new territory. The Peking University and Alibaba Group research describes an auction built on embedding-based coarse filtering to pre-select ad candidates, followed by the language model generating logits over those candidates as refined relevance scores, combined with bids under a critical-value payment rule designed to keep the system truthful for advertisers trying to maximize their own return.

More signal-deepening features are on the roadmap. Conversation depth optimization, expected around mid-2026, will target multi-turn dialogues that lead to conversions and reward ads that spark further exploration rather than just a click. First-party list matching, letting advertisers match hashed emails, phone numbers, and user IDs against ChatGPT profiles, is among the features on the roadmap. Multi-turn conversation retargeting is targeted for the fourth quarter of 2026, aimed at reaching users who had a meaningful, category-relevant conversation but didn't convert.

Scale tells a more complicated story than signal quality alone. eMarketer projects that more than 80% of AI ad spending in 2026 will land adjacent to machine-produced content rather than inside the chatbot conversation itself. The richest signal, the in-conversation one, is still the smaller share of total inventory, and that gap between where the best data lives and where the ad dollars actually flow isn't closing on its own.

Brands are facing something structurally familiar from the Google era: a choice between earning organic citation inside an AI's response or bidding for a labeled ad slot next to it. Run both together and they compound. Platforms built to read conversational context across several AI surfaces, rather than being locked into a single chatbot, are the ones positioned to aggregate this signal at any meaningful scale, and that cross-surface reach is what separates a purpose-built conversational ad platform from a generalist buying tool or a network tied to one app.

The trust and transparency constraint on acting on what prompts reveal

None of this works if users stop trusting the conversation, and there's evidence that trust is already fragile. Research out of Princeton, led by Wu, Liu, and colleagues in 2026, tested large language models for conflicts of interest and found a majority compromised user welfare in favor of company incentives across multiple scenarios. Grok 4.1 Fast recommended a sponsored product that cost almost twice as much 83% of the time. GPT 5.1 surfaced sponsored options in ways that disrupted the purchasing process 94% of the time. Qwen 3 Next concealed prices in unfavorable comparisons 24% of the time.

An experiment at the University of Michigan in 2025, with 179 participants, found people struggled to spot unlabeled chatbot ads at all, and actually rated the unlabeled ad responses higher than the labeled ones. Once the ad was disclosed, though, participants called it manipulative, less trustworthy, and intrusive. That's the asymmetry sitting at the center of this whole model: covert personalization boosts short-term engagement numbers, but it burns the exact trust that makes users disclose freely in the first place. Kill that disclosure and the signal dries up with it.

Given that only about 34% of users want an AI to buy on their behalf, most people in these conversations are still deliberating, not ready to hand over a decision. Advertising that respects that posture, suggesting rather than foreclosing, sits in a different category entirely from advertising that tries to force the decision closed.

OpenAI's published ad policies require clear labeling and state that ads must not influence the organic answer. Conversation content is not shared with advertisers under the current model. These aren't just compliance checkboxes. They're structural constraints that keep the whole environment viable, because an ad format native to the conversation, one that adds something useful rather than interrupting, is also the only kind that doesn't quietly destroy the thing that made the signal valuable to begin with.

What prompt-level intent means for how brands should think about demand capture

Intent now forms before it ever reaches a search bar, which means a brand strategy built only around search keywords shows up after the buyer's already done the hard thinking somewhere else. That's the upstream shift, and it changes where the real competition for attention actually happens.

Because prompts reveal not just that someone is in-market but where exactly they sit in the decision, the consideration stage itself becomes addressable in a way it never was through search. A brand can now reach someone while trade-offs are still being weighed, not just after the decision's already locked in.

Early advertiser analysis around ChatGPT's launch framed the choice as paid placement for speed against earned citation for compounding relevance over time, and that framing holds for any brand treating AI assistants as a real distribution channel. Neither approach alone gets the job done; they need each other.

OpenAI's launch partners for the ad product included the holding companies Dentsu, Omnicom, Publicis, and WPP, and early case studies from OpenAI point to Best Buy, Lowe's, and VistaPrint as adopters. Early advertiser commentary has described the shift as reaching customers during active research and decision moments that increasingly happen inside a chat window instead of a search bar, and that's about as direct a statement of the thesis as any advertiser has offered publicly.

None of this comes free of friction. Prompt-level targeting surfaces genuinely high-intent moments, but tracing a multi-turn conversational thread through to an off-platform conversion is a much harder attribution problem than last-click search ever was. Brands building around this signal should plan for that measurement gap now, rather than assume the tools to close it already exist.