Choosing Demand Partners, Floor Logic, and Signal Architecture for a Scalable AI Publisher Stack

AI app publishers need monetization infrastructure with conversation-level signals, intent-responsive floor pricing, and demand partners that can read multi-turn context. A legacy ad connection on its own cannot give publishers those tools. Choosing demand partners, floor logic, and signal architecture for a scalable AI publisher stack is therefore one system-design decision, with user trust as a constraint across all three layers.

Why the traditional publisher stack fails in a conversational AI environment

Traditional programmatic infrastructure prices and places ads against relatively stable objects, including a URL, a keyword string, a browsing-history profile, and a defined slot. A conversational interface does not produce those objects in the same form. Its inventory unit is a turn. Each exchange can change the context, while the commercial intent of the session emerges across several exchanges. There is no fixed placement geometry that reliably describes the value of the next opportunity.

That difference breaks more than the placement model. Page categories and keyword matches compress content into labels that buyers can use, but a conversation can move from explanation to comparison without changing its broad topic. A category label misses that transition entirely. Historical prices attached to the category consequently describe a mixture of opportunities with different purchase proximity.

Without a cookie or persistent user identifier, a browsing-history profile cannot anchor the bid. The conversation becomes the only durable signal, available in real time but changing as the dialogue develops. Patching an existing connection leaves the architectural mismatch intact, preserving its assumptions about what is being sold, how it is valued, and what the buyer knows.

What demand partners actually need to read conversational context — and why most cannot

A demand partner needs to interpret the dialogue trajectory. That includes earlier questions, the AI’s responses, subsequent refinements, and the direction of the exchange. The targeting object is the conversation’s developing meaning. Semantic targeting evaluates ad relevance against a representation of discourse coherence across turns, rather than assigning the current text to a predefined category. A generalist DSP whose targeting logic accepts only page-level signals or audience profiles lacks a usable input for that decision.

Yun et al., from Tsinghua University, Carnegie Mellon University, and Harvard University, formalize this mismatch in LLM Advertisement based on Neuron Auctions. Their research distinguishes classical auction objects, such as slots and keywords, from LLM output that is autoregressive and fluid. Its proposed mechanism works with the model’s internal representations to control recommendation strength independently. That is a research framework. It nevertheless explains why attaching a conventional bid to surface text does not solve the underlying allocation problem.

ChatGPT’s demand integrations illustrate the need for purpose-specific connections. Criteo announced its integration with OpenAI’s advertising pilot in March 2026 as the first advertising technology partner to integrate. StackAdapt was named at the May 2026 self-serve launch. These examples concern connections built for the environment, rather than an existing OpenRTB endpoint opened indiscriminately to DSPs. The protocol alone is not the bottleneck. A bidder can connect but still make mismatched bids and leave fill low if its targeting logic cannot use conversational context.

The first test is whether the partner can ingest a conversation-level context description as its targeting input. Campaign setup must also support context hints. Partners that work across multiple AI surfaces have more sources of relevant demand than single-surface demand pools, which can thin out quickly.

How floor logic must change when the pricing signal is intent, not page position

Once a partner can recognize the opportunity, the publisher needs a floor that values the same thing. Traditional floors use position, page category, or historical clearing prices as proxies for a valuable outcome. Conversational inventory exposes the underlying intent more directly. A user asking “what project management software handles remote teams” is closer to a purchase decision than one asking “explain how project management software works”. The topic is the same, but a static topic-level floor treats two different opportunities as equivalent.

That equivalence can suppress yield in either direction. A floor borrowed from lower-value traffic can undervalue purchase-proximate turns, while an excessive floor can exclude legitimate demand. ChatGPT’s relevance-weighted second-price auction already incorporates ad relevance into demand-side pricing. Publisher floors therefore need calibration against intent match. Otherwise, the seller and buyer are pricing different attributes of the same exchange.

The practical dimensions are conversational stage, topic commercial density, and conversation depth. Stage distinguishes early exploration from a decision-proximate exchange. Commercial density describes how many advertiser categories compete for that conversation type, while depth distinguishes a single query from a multi-turn, high-engagement session. A longer conversation may reveal stronger intent, but its accumulated coherence also gives the publisher more to protect.

Yun et al.’s framework treats user utility as part of the platform’s objective and dynamically prices overly aggressive interventions out of consideration, making pricing a quality constraint. Cai et al., from Alibaba and the University of Electronic Science and Technology of China, approach the other side through the SAGE framework, which adjusts bidding to trajectory state and advertiser constraints. Static sell-side pricing faces an adaptive counterparty. The research makes a case for state-aware pricing, but publishers need not copy its full mechanism.

Perplexity left advertising over the risk to user trust. The company cited concerns that ads, even clearly labeled ones, could cause users to doubt the neutrality of AI-generated answers. A high clearing price does not resolve that concern. Floors that pursue immediate yield without accounting for user experience can erode the long-term value of the conversational surface. Intent-responsive pricing must weigh commercial opportunity against the risk of disruption, including when purchase proximity rises.

Building signal architecture when there is no cookie, no keyword, and no traditional inventory unit

The signal layer must make conversational intent available to both matching and pricing. Its primary data object is the conversation itself, including semantic topic, stage, question type, product or category mentions, and the trajectory of prior turns. Sequence is the essential distinction from keyword targeting. A user moving from team coordination problems to software recommendations and then pricing expresses something that none of those queries encodes alone.

Because relevance derives from the current exchange rather than browsing history, the core targeting design can be session-scoped and identifier-free. Consent-based intent layers can supplement that foundation. Verve’s integration brings AI-native conversational intent into an intelligence layer with search data and zero-party polling data, using opted-in users who consent to share pseudonymized AI chat activity. Its model separates useful intent intelligence from dependence on persistent identifiers. Consent is consequently part of the signal architecture.

Enterprise demand needs more than intent signals. IAS announced brand safety and suitability measurement for ChatGPT Ads on October 5, 2026, establishing a separate role for classification of conversational risk. Commercial relevance does not imply brand suitability. A signal layer needs to distinguish contexts appropriate for an advertiser from those that could create reputational exposure. Without that distinction, better targeting can identify an attractive opportunity that should not receive the ad.

Measurement is less settled than targeting. A recommendation or shortlist may be the meaningful performance event even when the user never clicks. A user can encounter an ad, continue comparing options, and convert later through another channel without producing an attributable click. Last-click reporting misses that pathway. Signal architecture should therefore capture non-click outcomes and support assisted-conversion and lift measurement, while acknowledging that the industry has not converged on a standard.

Publishers collecting these signals now can develop proprietary intent data with which to explain their inventory’s value as measurement practices mature. The competitive asset is the recorded relationship between conversational progression and outcomes, even while interpretation remains imperfect. Raw inventory cannot provide that evidence by itself. Publishers that postpone collection will lack the historical data needed to make the same case to demand partners.

How the three layers interact — and why getting one wrong breaks the others

Signal granularity limits the distinctions that floor logic can make. If the signal layer records only a broad topic, pricing cannot reliably separate an exploratory turn from a purchase-proximate one within it. That lost distinction then affects auction participation. Floors shape which partners can win and at what price, so weak signals influence allocation even when the bidders themselves have sophisticated targeting.

The dependency also runs in reverse. Partners unable to read conversational context can bid flatly or inconsistently against intent-differentiated inventory, producing weak competition at valuable moments and excessive competition at less valuable ones. Those outcomes then distort floor calibration. A publisher may interpret poor competition as weak inventory value when the actual problem is the bidder’s inability to recognize it. The price history becomes a record of targeting failure and an unreliable basis for optimization.

PubMatic’s experience with Liner provides a practical counterpoint to that failure mode, showing that low fill and poor contextual matching are avoidable with suitable architecture. It does not make every integration interchangeable. Connecting audience-profile bidders to keyword-level context while setting page-category floors preserves three mutually reinforcing mismatches. By contrast, an integrated stack gives demand and pricing a common account of the opportunity. To scale, that shared reading of the opportunity must reach more conversations and demand as bid volume grows.

Demand decides which ads are relevant, pricing decides which can qualify, and signals describe the coherence those ads may disrupt. Yun et al.’s framework brings those dimensions together in a goal that includes user utility. SAGE shows how buyers make the same kind of state-aware changes. Together, they frame integration as a decision-making requirement, while Perplexity’s exit shows why revenue optimization cannot treat the user experience as an external concern.

What monetization infrastructure should an AI app publisher build, and in what order

Publishers should build signal architecture first because every later decision depends on the inputs it supplies. Publishers should instrument topic, stage, depth, question type, and trajectory with a consent framework from day one. Brand safety classification should be part of that first layer. If collection starts after demand goes live, the historical record will have gaps where floor calibration needs data, while a lack of safety signals can shut out enterprise budgets. Assisted-conversion and lift measurement should also be designed at this stage, rather than waiting for last-click reporting to expose its limitations.

Floor logic should come next, starting with intent stage as the primary dimension. The distinction between exploratory and purchase-proximate exchanges has a direct relationship to the signals already being collected and to commercial value. Topic commercial density can follow once that first dimension is calibrated. The floor system must remain easy to revise. Clearing prices in this young market are not stable, so initial assumptions will need adjustment as actual competition provides evidence.

Publishers should then test demand partners with real signals and parameterized prices. The first criterion is whether conversational context can serve as the primary targeting input. The second is operation across multiple AI surfaces. The third is direct publisher supply relationships in addition to buy-side reach, because combined DSP and exchange partners can provide more predictable fill and stronger floor competition than pure buy-side players. This order makes the evaluation specific to the publisher’s architecture rather than dependent on a vendor’s general positioning.

The resulting selection is not a universal ranking of supply-side platforms or exchanges. A purpose-built connection, a broad advertiser base, and a familiar protocol answer different questions, and none independently establishes fit. Publishers should evaluate the combination against their actual conversational inventory. As demand accumulates, revisions to floors should use competition data without confusing an unreadable signal with an unattractive opportunity. That discipline allows the stack to learn from trading activity without allowing an early mismatch to become its permanent pricing model.