Intent Category Mapping for AI Ad Targeting

What intent category mapping is

Intent category mapping decides whether an AI assistant shows an ad, and if so, which one. It replaces keyword matching and page context, the two signals that built the search and social ad businesses, with something closer to reading comprehension: classifying what a person is actually trying to figure out, from the words they just typed. That shift matters because the substrate it rests on changed completely. Search ads worked off a query string and a results page. Social ads worked off a graph of past behavior. An AI assistant offers neither: no URL, no ten-blue-links page, no feed to scroll. What it offers instead is a live sentence, written by a person describing a real problem in their own words, and US AI ad spending is projected to hit $68.25 billion by 2030 on the strength of systems that can read that sentence correctly.

Intent category mapping takes a user's prompt, sorts it into a structured commercial category, and hands that category to the ad-matching system as its input. That's a different job than the one legacy ad tech knows how to do. Keyword matching works on strings: it looks for "best running shoes for flat feet" and matches literal terms or close variants. Intent mapping works on meaning, so the same purchase intent appears in a dozen phrasings that share almost no words. "My arches hurt after long runs, what should I switch to" carries the same commercial intent as that keyword query, but a string matcher would never connect the two.

It also isn't audience segmentation, though people mix the two up constantly. Segmentation describes a person based on history: past purchases, browsing patterns, demographic guesses. Intent mapping describes what that person wants in this specific turn of this specific conversation, and that can override or flatly contradict whatever the segment profile says. It isn't sentiment analysis either. Sentiment tells you whether someone's mood reads positive or negative. Intent mapping tells you where someone sits in a decision process and what kind of commercial need they have right now. Treating those as the same axis is the design mistake that appears most often in early builds.

How a prompt reveals purchase intent beyond what keywords could

A prompt is not a query, and that gap is the entire reason this discipline exists. A search query is a compressed fragment, three or four words stripped of context because the person typing knew a search box wouldn't reward more. A prompt runs the other direction. It often carries a full description of the problem, the constraints, the trade-offs already on the table, and its phrasing produces a clear signal of the exact stage of decision the person has reached.

More than a fifth of pre-purchase digital journeys now start inside an AI chat interface instead of a search bar. Travel leads that shift: 37% of travel queries begin in an LLM, and a single travel prompt can carry an unusual amount of signal. Someone might write, "planning a trip for four adults and two kids in October, want to avoid the worst crowds and stay under budget." A keyword system sees fragments to match. An intent classifier sees a party size, a season, a crowd-sensitivity constraint, and a budget ceiling, all in one breath.

Early-stage prompts also tend to arrive unbranded, because people explore a category before they've settled on a brand name. That's the moment when the consideration set is still wide open, and it's exactly the moment a branded keyword search is structurally unable to reach. Most teams underrate how much this matters. The unbranded, exploratory prompt isn't a weaker signal than a branded search, it's a signal keyword systems were structurally incapable of ever catching, which makes it the bigger opportunity, not the consolation prize.

The anatomy of an intent category: audience, intent stage, and topic

An intent category only works for targeting once three things are locked down together. An intent category only works for targeting once three things are locked down together; dropping one makes the category unusable.

Audience describes who's asking, but by behavior rather than demographics. Not "women 25-34." Instead: a first-time buyer, a professional doing comparison research, someone already three products deep into a side-by-side spreadsheet. Intent stage places the prompt somewhere along a journey, awareness, consideration, evaluation, or near-conversion. Topic narrows the commercial category down to something specific enough that one ad group can address it coherently, instead of something broad enough to mean almost anything.

OpenAI's own guidance for ChatGPT Ads makes this concrete. Advertisers write what it calls "context hints," short descriptions that pin down one buyer, in one mindset, about one narrow category, and each ad group typically runs somewhere around 5 to 15 hint variants built around a single audience-intent-topic theme. That range isn't arbitrary. Going broader blurs the match, pulling in prompts the ad group shouldn't touch. Going narrower starves the system for volume, because there aren't enough matching prompts to spend a budget against. The 5-to-15 window is OpenAI's answer to that trade-off, and it's the same trade-off every targeting system built before it has run into eventually.

Of the three components, intent stage is the one advertisers skip most often, and it's the one that costs them the most. Teams coming out of keyword-logic backgrounds default to topic, because topic is the closest cousin to a keyword, and they treat stage as an afterthought bolted on later. That's backwards: treating stage as an afterthought bolted on later costs teams the most, precisely because stage, not topic, is what should come first. Upper-funnel conversational moments, the exploring-and-comparing turns rather than the ready-to-buy ones, end up served worse ads relative to how much volume actually passes through them, and that's the gap most worth closing first, not last.

How intent classification runs at auction speed

None of this works if it runs slow. The ad stack an AI assistant runs on has four layers: demand and auction, context and targeting, creative generation, and measurement and attribution. Intent mapping sits inside context and targeting, but it has to hand off cleanly to the other three or the whole chain breaks.

The first checkpoint is a trigger layer: before any ad request goes out, a classifier, usually a fast, cheap model call or a simpler heuristic, decides whether the current turn is even commercially relevant. Most turns in an average conversation aren't. Someone asking for help rewriting an email needs no ad next to the answer, and holding back there is the system working as intended. It's the system doing its job correctly.

If a turn clears that bar, a fetch layer sends an ad request in parallel with the model's own response generation. Parallel is the operative word here: the ad call cannot block the answer, because the answer is the product and the ad is a guest riding along with it. Latency budgets run tight, on the order of 200 to 300 milliseconds for the ad request to resolve. Missing that window means the interface commits to showing no ad for that turn rather than delay the response. Blocking the actual answer on a slow ad fetch is, by a wide margin, the most common mistake in early production integrations, and it's the one mistake with no acceptable version.

The source of the intent signal when you are not the LLM

Not everyone buying this inventory gets to see the prompt directly. Direct, prompt-level access stays reserved for the platforms that own the interface itself: advertisers buying inventory inside ChatGPT go through OpenAI's own ads manager or through integrations OpenAI has explicitly approved.

Everyone else in the open programmatic market has to source the signal from somewhere else. As of March 2026, Verve Group became the first ad tech provider to activate conversational intent signals sourced from major LLM ecosystems for use in programmatic targeting outside those platforms directly. The architecture draws signal from users who've opted in and shared their AI chat activity through apps they already have installed, rather than leaning on a direct data-sharing deal with any single AI company. No raw message content gets stored or passed along anywhere in that pipeline.

That intent signal then feeds a shared data spine: zero-party data a user volunteered, search intent pulled from existing sources, and pseudonymized AI chat activity, combined into one scoring layer. The intent category derived from a chat prompt becomes one input among several, sitting next to the traditional signals programmatic buyers already know how to use, supplementing them.

Mapping intent categories to ad placement surfaces inside AI interfaces

By 2026, four ad surfaces exist across production AI chat products, and each one carries a different relationship between intent and format.

The after-answer inline card sits below the assistant's response, labeled "Sponsored," with a title, short copy, and a link. OpenAI picked this format for ChatGPT's initial commercial rollout, and it delivers the highest yield with the least disruption to the surrounding experience, commanding CPMs in the $25 to $60 range. The sidebar panel, desktop-only, sits outside the conversation column entirely and pulls a much lower CPM, in the low single digits, for the plain reason that it's easy to ignore. The sponsored follow-up chip offers small pill-shaped suggestions below an answer that, when clicked, submit a brand-favorable prompt or send the user to an advertiser. Perplexity shipped this format in November 2024, then walked away from advertising altogether in February 2026, citing user-trust concerns, which says something the industry should sit with: this exact format is the one most likely to blur a genuine suggestion into a paid one, and Perplexity's own retreat from it stands as evidence. Response-grounded brand mentions, where the model's own answer states a brand directly, carry the highest revenue ceiling and the highest cost to trust, and remain confined to early experiments as of 2026.

Intent category should decide which surface gets used, not just which ad wins the match. High commercial intent, someone deep in comparison and asking about a specific product, belongs on the inline card, because the user already expects commercial content at that point. Upper-funnel intent, someone still exploring or comparing broadly, calls for a lighter touch, a chip or a sidebar slot, so the moment doesn't turn transactional before the user is ready for that. Professional or research-flavored prompts do best on surfaces that read closer to editorial than to advertising, since a loud inline card there costs more trust than it earns in clicks.

ChatGPT's open beta currently runs three placements side by side: Sponsored Answers inside the chat response itself, Sponsored Follow-ups, and Sponsored Shopping cards. Different intent categories route to different placements even within that one platform, which is itself evidence that surface selection has become a targeting decision, not a fixed template applied uniformly.

Across every one of these formats, one line holds without exception: the ad stays separate from the answer. Classifying intent accurately enough to serve a relevant ad is not license to shape what the model actually says.

Intent category mapping in practice across verticals

Travel is the clearest proof of what this discipline does well. With 37% of travel queries starting inside an LLM, the prompts arrive dense with constraints, budget, party size, timing, crowd tolerance, and an intent category built around the single word "travel" would throw most of that signal away. The category needs to capture trip type, timing, and the specific constraint set the user laid out. Travel also splits by urgency in a way few other categories do: some purchases close within 48 hours, a one-way flight booked on impulse, while multi-destination trip planning can stretch two weeks or more before it converts. An intent category that tracks topic and ignores urgency is throwing away information the prompt handed it for free.

Auto runs a different playbook. The average consideration set is 1.96 brands: most car shoppers are choosing between two names, not ten. That narrow window makes early-exploration intent mapping worth disproportionately more here than elsewhere, because a user still weighing two options is far more persuadable than one already searching a specific model by name, which is the keyword-era equivalent of a door already closed.

Health and wellness is the fastest-growing category in AI-related advertising as of the first quarter of 2026, up 165% year over year. That growth comes at a cost: sensitivity in this category runs high enough that intent mapping can't stop at commercial relevance. It has to carry a brand-suitability classification right alongside it. A prompt about symptom research and a prompt about product discovery both read as "health" by topic, but they demand completely different handling, and this is the one vertical where sub-category granularity, prevention versus treatment versus product discovery, isn't optional the way it might be elsewhere.

Privacy architecture and the limits of what intent mapping can use

The raw conversation itself never touches the ad auction. What gets classified is the commercial intent category the prompt implies, not the personal content of the exchange, and that distinction is the entire privacy model this system rests on.

Verve's architecture rests on three commitments. The underlying data comes from users who opted in. The signals passed downstream are aggregated and pseudonymized. No raw message content or individual conversation record gets stored or disclosed anywhere in the chain. OpenAI's own ad policies run on a parallel commitment, clear labeling of sponsored content and strict independence between the ad and the organic answer. The ad can't shape what the model says, and the model's answer can't bend because an advertiser happens to be in the mix.

That separation is the load-bearing wall the entire business stands on. It's the load-bearing wall the entire business stands on. If users start suspecting an answer got quietly shaped by who paid for placement, trust in the assistant collapses, and once that trust is gone, the intent signal stops being worth anything to anyone, because its entire value as a targeting signal depends on it being an honest, unprompted statement of what someone actually wants.

Measurement and attribution when the conversion path runs through a conversation

Attribution gets harder once the funnel's first touch is a paragraph typed into a chat window instead of a click on a search result. A search ad click carries a URL, a session ID, a fairly clean attribution path that decades of ad tech were built to track. A conversational turn doesn't hand over that same artifact by default, and a user might work through a decision across several turns, spread over different days, before ever reaching the page where the purchase happens.

Connecting a prompt's intent category cleanly to a downstream conversion is the frontier this discipline is still working through. The intent category assigned to a prompt can tell an advertiser what kind of need showed up and roughly where in the journey it sat. Tying that specific moment to a conversion that happens days later, on a different platform, is a harder measurement problem than anything keyword-based search attribution ever had to solve. The systems built to handle it are still maturing, and calling it settled would overstate what's actually been proven out in production so far.