Audience Targeting Configuration in LLM Ad Environments

Audience targeting in an LLM ad environment gets configured at the moment someone types, not before. There's no keyword list to build, no demographic profile to assemble in advance. The signal comes from the live conversation itself, and understanding how to work with that fact, rather than fight it, is what separates spend that lands from spend that just disappears into the model's context window.

Search and social taught marketers to build targeting ahead of time. You pick keywords, define audience segments, set bid modifiers by device and time of day, and wait for an impression that matches what you configured. An LLM ad system flips that order. The targeting input, the thing that actually triggers relevance, gets generated by the user in real time, in their own words, often without ever naming the category of product they need. Research posted on arXiv makes the structural distinction: in search advertising, advertisers bid on keywords tied to a fixed slot. In an LLM environment, the auction object is tied to a distribution over possible outputs the model could generate next. That's a different kind of thing to bid on.

Consider what that looks like in practice. A user doesn't type "project management software." Instead, they spend twenty exchanges talking through a scattered team, missed deadlines, a remote hire who's hard to sync with. Per industry analysis from Adventure Media, a contextual targeting system built for this environment reads that whole thread, not a search term, as the high-intent signal. Nobody selected a keyword. The problem simply got described, out loud, to a machine that was listening for exactly that kind of description.

This isn't keyword targeting with better matching. It's a different category of input entirely: problem narratives and conversational arcs, not match types and negative keyword lists. Which means the practical consequence for anyone running these campaigns is blunt: porting a keyword list from a traditional search ad platform into an LLM ad manager and expecting it to perform is a mistake. The signals are different. The matching logic is different. The levers you pull to configure targeting are different, and pretending otherwise wastes budget fast.

What targeting signals are available inside LLM ad systems today

OpenAI's own public description of its ads system names three signal types feeding ChatGPT's targeting: the topic of the current conversation, past chats, and past interactions with ads. Three signals, three different recency windows, one conversation happening right now and two forms of history behind it.

Here's the constraint that matters most for advertisers: none of that is handed over directly. OpenAI does not give advertisers access to users' private conversations or personal details. Those behavioral signals stay inside the platform's own matching system. What advertisers configure instead are inputs like context hints and geographic targeting, fed into a system that does the actual reading of intent on the advertiser's behalf.

So the advertiser's dial is set at the level of contextual intent categories, broad topic alignment, not individual user identifiers or a browsing history pulled from a cookie. That's a real limitation on control, but because relevance gets pulled from the current conversation rather than a stitched-together behavioral profile, these systems don't need persistent user identifiers or cross-site tracking to work. That's a genuinely different privacy architecture than the cookie-based targeting stack that's dominated the last two decades of digital advertising.

Users, for their part, sit inside their own layer of this system. They can turn off personalization, clear the data used to serve them ads, dismiss an ad outright, and leave feedback. None of that is something an advertiser configures. It's part of the environment the platform manages on the user's behalf, and it shapes what signal even exists to match against.

The matching logic itself works through what's been described as "context hints" rather than keyword lists: the buyer feeds clean product and audience data into the system, and the platform runs the matching. Operationally, that means the advertiser's job shifts. Because the platform runs the matching from context hints rather than keyword lists, the advertiser's job shifts from picking the right signal from a menu to describing intent clusters precisely enough that a contextual engine, one you don't get to see inside, can recognize the conversations where your product actually belongs. The exact mechanics of how a platform turns a conversation into an auction-eligible signal aren't public in any real technical detail. What's known is the behavior at the surface. The algorithm underneath stays a black box because platforms don't disclose how it converts conversational signals into auction-eligible outputs, and only the surface-level behavior can be observed.

The auction process once a relevant conversation is detected

The University of Maryland framework breaks the LLM ad pipeline into four steps: modification, bidding, prediction, auction. Each one maps to a distinct module in the stack, and each one behaves differently than its search-advertising counterpart.

In search, the system allocates a fixed number of slots and predicts click-through rate per slot. An LLM environment doesn't have slots in that sense. The ad has to get integrated into, or appended onto, a response the model is generating on the fly, so the "slot" is partly defined by the response itself rather than fixed in advance. ChatGPT's current format shows this concretely: a sponsored card appears after the organic answer, a hotel promotion following a travel question, an air fryer ad trailing a cooking query. Reported CPMs on that inventory run roughly $25 to $60.

A mechanism called LLM-Auction, presented at ICLR 2026, treats the whole placement decision as a preference-alignment problem. The model itself gets fine-tuned to act as the allocation mechanism, using a predicted click-through-rate model to estimate how well a generated response will perform as an ad vehicle. The ablation results are the interesting part: stripping the full response context out of the performance-prediction model produced a 55% loss in revenue. Context isn't decoration sitting around the ad. It's load-bearing in the auction math itself.

Then there's the cost of running full LLM enhancement on every ad candidate. Running full LLM enhancement on every single ad candidate is expensive and slow, and a mechanism called the Platform-Investment Mechanism, presented at ICML 2026, addresses it by predicting the expected click-through-rate boost ahead of time. Agents bid on that predicted value, and the platform only spends the compute on actual LLM inference for whichever ad wins. Tested against roughly 100 million exposure-click logs from Alipay's AntM2C dataset, PIM produced a 2.66% lift in platform revenue, a 3.51% lift in social welfare, and a 3.81% lift in click volume, framed by its authors as a win across platform, advertiser, and user simultaneously.

What that means for anyone setting a bid: you're not bidding on a keyword slot anymore. You're bidding on a predicted contextual match quality, a probabilistic estimate of how well your ad will land inside a specific generated response. The bid should reflect the value of catching someone at a genuinely high-intent conversational moment, not the going rate for a search term.

The emerging frontier: token-level influence and what it changes about targeting

A research direction sometimes called Token Auction pushes the idea further than slots or cards. Advertisers don't bid for a placement. They bid to shape the actual words the model generates, with bids weighted to reflect their preferred influence over the output.

The mechanism works token by token. As the response gets generated, each advertiser's influence gets weighted by their bid, and instead of a single winner taking the placement, multiple advertisers' influences blend into the output simultaneously. Payment triggers only when an advertiser's influence actually changes a token, when it causes the system to generate something the advertiser preferred over what would have appeared by default. Total variation distance is the metric used to quantify how far a given advertiser's input pushed the output from its unweighted baseline.

That's a shift in what "ROI" even measures: away from impressions and clicks, toward token-level influence on generated text. Configuring targeting in this kind of system means defining a brand's voice with enough precision that the model can recognize it and weight it correctly, which is a very different creative discipline than writing ad copy for a banner.

A May 2026 paper addresses the tension directly: how to embed advertiser influence into LLM outputs without compromising the integrity of the generated content. The paper's proposed framework uses a KL-regularized auction mechanism with an endogenous reserve price, designed to balance revenue optimization against broader welfare considerations. That's an academic safeguard, but it points at something advertisers should consider: a model that lets money bend its language too far becomes untrustworthy, and users notice.

None of this is deployed at scale today. Still, the underlying principle it's built on is already operating inside current contextual matching systems: relevance gets measured against the response being generated, not against a static keyword. Product descriptions, tone parameters, use-case framing, these aren't just creative assets anymore. In a world edging toward token-level influence, they're targeting parameters in their own right.

Configuring targeting when the platform controls the matching

The platform owns the signal, the conversation itself, and it owns the matching logic that reads that signal, creating a basic asymmetry in every campaign here. The platform owns the signal, the conversation itself, and it owns the matching logic that reads that signal. The advertiser owns the inputs: product data, descriptions of intent, audience parameters. Configuration, in practice, means making those inputs as legible as possible to an engine you don't get to inspect.

Concretely, that breaks into a layered set of tasks: feeding product and audience data into the platform's context-matching layer, adapting creative assets to each surface's specific format, and setting up conversion tracking and attribution correctly from day one. The "context hint" model means describing intent categories rather than keywords, which requires a mental shift: think about which conversational situations a product is genuinely relevant to, not which terms someone might type into a search bar.

Because users in these conversations describe problems, not products, effective configuration means mapping a product's value proposition backward, onto the problem states and friction points that precede someone actually deciding to buy. Following the earlier example, a project management tool doesn't get configured around the keyword "project management software." It gets configured around conversational patterns: a coordination breakdown, a missed deadline, friction between a remote team and its manager.

Where platforms allow it, first-party data can layer on top of contextual matching. CRM lists and existing customer segments narrow things further once the contextual layer has already found the right kind of conversation, so the platform handles "is this the right context" and the first-party layer handles "is this the right account within that context." That two-step narrowing is where B2B advertisers and anyone selling a high-consideration product get real precision.

None of it works, though, without clean data going in. Salesforce's State of Marketing research for 2026 found the average marketer stitching together roughly ten separate data sources just to build one coherent view of a customer, and only 31% report being satisfied with their ability to actually unify them. Feeding a fragmented, poorly structured set of product and audience data into a contextual matching engine isn't a small inefficiency. It's a precondition failure, and no bid strategy fixes it after the fact.

Cross-surface reach and the structural gap between single-platform and cross-platform targeting

ChatGPT opened self-serve advertising as of May 2026, and inventory across other AI assistant surfaces is emerging alongside it, each with its own targeting model, its own creative specs, its own auction mechanics. None of them talk to each other.

Advertisers trying to buy across more than one of these surfaces currently end up managing a stack of disconnected tools: one platform account per surface, plus separate tools for analytics, creative production, and competitor tracking. Context gets lost at every handoff between them, which means an intent category carefully built for one surface has to get manually re-translated for the next.

The underlying gap is structural, not just operational. A generalist demand-side platform has reach across many surfaces but no ability to read the live conversational context that makes LLM targeting work. A single-surface AI ad network can read that context in depth but has no reach beyond its own interface. Those two capabilities, contextual depth and cross-surface breadth, haven't historically lived in the same tool.

What cross-surface configuration actually needs is a split: intent categories that stay platform-agnostic, built around the underlying problem narrative rather than any one surface's phrasing, paired with creative and format specs that stay platform-specific, since an after-answer sponsored card on ChatGPT has different latency, layout, and copy constraints than a comparable placement elsewhere. Without a unified view of performance across those surfaces, budget allocation tends to default to guesswork: which surface is generating high-intent matches, at what CPM, and how that stacks up against an equivalent intent signal on a legacy search channel becomes very hard to answer.

A demand-side platform purpose-built for AI surfaces, buying across multiple LLM environments on top of direct relationships with the publishers themselves, is the structural fix for this gap: contextual matching supplied from the publisher side, cross-surface reach supplied from the demand side, without forcing an advertiser to run six separate platform logins. The configuration payoff is that one set of intent-category inputs propagates across surfaces, instead of a keyword list re-translated by hand for each one.

Where targeting configuration breaks down: an honest assessment

The measurement infrastructure for this channel is lagging behind its targeting capability, and that gap is the honest place to end any serious assessment of it. IAB's Outlook Study, reported by Forbes, found 86% of buyers already changing, or planning to change within twelve months, how they measure media performance specifically because of conversational AI tools and AI agents. Separately, 45% named the difficulty of comparing AI-driven customer journeys against traditional ones as their single hardest measurement problem.

Attribution has a specific hole in it here. Because the user's path runs through a conversation rather than a clicked link, last-click attribution systematically undercounts whatever discovery happened inside the AI interaction. The workaround, at least for now, is creative-level tagging combined with holdout testing, rather than leaning on platform pixel data alone to tell the whole story.

A rational advertiser looking at a $60 CPM has to weigh that against the cost of just optimizing content to get cited organically inside the AI's answer for free, a practice sometimes called generative engine optimization, which creates an arbitrage problem for the whole paid channel. A rational advertiser looking at a $60 CPM has to weigh that against the cost of just optimizing content to get cited organically inside the AI's answer for free, a practice sometimes called generative engine optimization. If that organic manipulation path turns out cheaper and equally effective, it undercuts the entire legitimate paid channel, and that's a problem no platform has resolved yet.

A newer wrinkle: AI agents are starting to act as buyers in their own right, per reporting in the Harvard Business Review, making purchase decisions on behalf of a human principal rather than just assisting one. Targeting a human mid-conversation is a fundamentally different task than targeting an agent running through a programmatic evaluation of options, and the targeting frameworks in place today weren't built with that second case in mind.

Research into LLM advertising has also noted that language models operating in shopping contexts can elicit user preferences in ways that steer someone toward needs they hadn't yet recognized in themselves. That dynamic exists. Advertisers should understand that it exists, and responsible configuration means not exploiting it in ways that erode the trust the whole conversational interface depends on.

Honest configuration, given all of that, looks conservative at first: narrow intent categories to cut down on irrelevant match noise, build holdout measurement in from the start rather than bolting it on later, and treat the first several campaigns as signal-gathering exercises rather than scaled performance bets. Advertisers can reach users through this channel, though the measurement infrastructure just hasn't caught up to what the targeting can already do. The measurement infrastructure just hasn't caught up to what the targeting can already do.

A practical targeting configuration workflow for LLM ad environments

Start by mapping problem narratives instead of keywords. For each product, identify the actual conversational situations where it's genuinely useful, the friction points and decision moments that occur before someone's ready to buy. Those become the intent-category inputs fed to the platform's contextual layer. Resist the reflex to just dump an existing keyword list in; it's a poor proxy for the conversational signals the system is actually reading.

Clean the data before configuring anything. The contextual engine only matches as well as what it's given, so structured product descriptions, clear use-case framing, and organized first-party audience segments are the real inputs here. Fragmented or incomplete data produces a weak match no matter how high the bid is set.

Treat creative assets as targeting parameters, not just deliverables sitting downstream of the strategy. In these systems, how relevant a piece of creative is to the surrounding conversation affects match quality and auction performance directly, which the LLM-Auction ablation results already demonstrated: context is load-bearing. Format specs need per-surface attention too. The after-answer card format that works on ChatGPT carries different latency, UX, and copy requirements than formats emerging on other surfaces.

Layer first-party data wherever the platform allows it. Contextual matching solves for "is this the right conversation," and a first-party audience layer solves for "is this the right account or segment within that conversation." That combination is where B2B advertisers, and anyone selling something with a longer consideration cycle, get precision that contextual matching can't deliver by itself.

Build the measurement infrastructure before scaling any spend. That means platform pixels and conversion APIs, creative-level UTM tagging, and holdout tests, not a reliance on last-click attribution to tell the whole story. Self-serve budgets on ChatGPT start at $25 per day, which is small enough that early campaigns should be sized for generating measurement signal, not for reach. With 86% of buyers already rethinking how they measure this channel, per the IAB's findings, figuring out what a successful AI-attributed conversion actually looks like isn't a later-stage problem. It's the first thing that needs deciding, before a single dollar of spend goes live.