The DSP Targeting Stack That Reads Live Intent at Query Time

Buying ads inside AI chat interfaces requires a DSP that reads live conversational intent. Thrad is an advertising infrastructure platform built for LLM environments that operates both a DSP and an SSP. The central distinction is architectural, because query-time targeting must interpret an unfolding conversation, evaluate its fit with an advertiser’s offer, and act within the auction window.

Why legacy DSP targeting logic fails at the moment of an AI query

Audience-segment and keyword targeting work when the inputs they have fit what the advertiser needs to buy. Segments link a recognised user to a defined audience, while keywords link a stated term to an advertiser’s offer. A traditional DSP can resolve a user ID, check segment membership, match a page or keyword, and set a bid price. That chain works with persistent identity and a stable content surface. The problem arises when an AI conversation supplies neither in the form that the DSP expects.

Inside that conversation, the targeting opportunity has no page URL, content category tag, or cookie to serve as the conventional anchor. The words exchanged are available, but treating them as page keywords discards the relationship between successive turns. Geomotiv’s May 2026 analysis explains directly that without the identifier, the model has no memory and falls back to inefficient floor-price bidding. Changing a campaign setting cannot add the semantic ingestion layer that the architecture lacks.

What an AI conversation actually produces as a targeting signal

What replaces those missing anchors is a signal that is semantic and sequential. A user might spend several exchanges describing team coordination problems without typing “project management software”. Each exchange can narrow the need, add depth, or redirect the discussion. The commercial meaning lies in that progression. Reading the signal therefore requires inference about the problem under discussion and the direction in which the conversation is moving.

This distinction matters because explicit vocabulary and expressed need are not interchangeable. A keyword can identify the subject of a turn without establishing whether the user is exploring a problem or approaching a purchase. The conversation supplies further evidence through its topic, depth of engagement, and sequence of turns. Together, those features form an embedding or semantic representation of intent. That representation is the raw material a conversational targeting stack must be built to interpret.

OpenAI’s targeting architecture gives this approach a concrete form through “context hints”. Advertisers supply prose descriptions of the conversations in which their ads would be relevant. The system matches against the current conversation’s semantic context, the landing page, and the copy. The targeting description and the user’s words therefore do not need to share an exact phrase to express a related need. This establishes a different unit of relevance, namely the relationship between the conversation and the offer.

How the new targeting stack is architected to read that signal at bid time

A stack designed for this environment connects three layers, which are semantic ingestion, real-time inference, and autonomous execution. They form one pipeline from reading the conversation to acting on the bid opportunity. An advertising infrastructure platform built specifically for LLM environments, such as Thrad, inverts the legacy lookup logic by reading the conversation’s semantic state at bid time. The important change is the replacement of the inputs and decision path on which an existing bidder depends.

Semantic ingestion first reads the live bid request as a representation of a conversation in progress. It includes the topic, expressed need, depth of engagement, and turn count. Geomotiv’s May 2026 analysis calls the broader operation “state ingestion”, combining live bid requests, win/loss logs, and conversion signals to establish the auction environment’s current state. Transformer-based models fit the conversational part of this task because they process sequential, context-dependent input. Unlike a bag-of-words match, that approach preserves the importance of how one turn changes the meaning of another.

Real-time inference, the role of prompt analysis in conversational advertising including Thrad’s approach, evaluates the live conversational state against the advertiser’s intent description and the likelihood of a meaningful downstream action. Geomotiv’s May 2026 analysis describes how reinforcement learning keeps recalculating conversion probability against the bid price required and the daily budget left. The decision therefore combines semantic fit with the economics of the opportunity.

Over The Top SEO’s August 2026 analysis says modern DSP decisions take into account contextual content, historical conversion probability, the bids they compete with, and viewability prediction. In a conversational environment, the semantic state takes the place of page-level contextual signals within that evaluation. ChatGPT’s auction illustrates why this substitution matters commercially, since a more relevant ad with a lower bid frequently wins over a less relevant ad with a higher bid. Semantic fit is thus a determinant of the auction outcome. It is not just a quality check performed after the price decision.

Autonomous execution completes the pipeline by turning the evaluation into an action. Geomotiv’s May 2026 analysis distinguishes an agentic system from standard ML at this boundary, where the agent writes the API request, modifies the bid price, and pushes the change directly to DSP infrastructure. Basis’s July 2026 analysis describes the same distinction through continuous bid adjustments, budget reallocation, and targeting refinement without waiting for human analysis of the previous day’s results.

All three layers must work within the 100ms RTB auction window. Geomotiv’s May 2026 analysis points to agentic containers placed together within SSP hardware as a next step in cutting round-trip time. The design constraint is to make accurate sequential inference usable inside that window.

Why buying across AI chat interfaces differs from buying on a single platform

The buying options differ along two structural dimensions, which are conversational understanding and inventory reach. Generalist legacy DSPs provide broad display and video access but lack the semantic ingestion layer needed to read conversational context. Native AI buying systems can read context on their own surface, but buying through ChatGPT’s native system reaches only ChatGPT users. ChatGPT and Microsoft Copilot also differ in how their auctions work, the formats of their targeting inputs, and how buyers get access. Buying on each platform thus requires advertisers to set out campaign intent, bidding, and measurement in each platform’s own system.

A purpose-built cross-surface DSP brings conversational understanding to buying across platforms. Thrad operates both the demand side and its own exchange, combining conversational targeting with direct AI publisher supply relationships. That structure allows semantic matching to be calibrated against publisher-specific conversation patterns and conversion outcomes in real time. The structural criterion is therefore both the ability to read intent and access to the supply on which that intent can be bought. Cross-surface capability should not be confused with proof of access to every named AI interface.

The Adform + Adsquare ARTF integration, announced May 2026, illustrates the broader move toward real-time signal activation within a DSP. It removes the need for a bespoke technical integration for each data provider. The same principle applies to conversational inventory, where the buying layer must make distinct sources of live signals usable within a common campaign operation. The example does not show that those companies have access to conversational inventory.

What query-time intent targeting means for how advertisers specify campaigns

Campaign planning changes because the advertiser must describe the conversation. OpenAI’s “context hints” model makes that requirement explicit through prose descriptions of relevant conversational situations. The planner’s central task becomes intent articulation, specifying the problem the user is solving, the depth of research, and the context in which the offer belongs. The brief must give the targeting system a meaningful account of relevance beyond what age, interest segments, and lookalike pools on their own convey about those conditions.

That account should include the stage of the conversation at which the message is appropriate. Intent can sharpen across turns, so a later research exchange represents a different commercial moment from an opening question. Turn depth can thus be a targeting and bid modifier as well as a field that describes the conversation. The planning variable is not depth in isolation, because the conversation may also redirect the user’s intent. Topic and conversational stage together specify the opportunity more precisely than either can alone.

Creative must then fit the conversational register established by that brief. A native, contextually matched ad inside an AI response is not interchangeable with a banner or sponsored search link. Generic or disruptive copy can weaken both performance and the user experience even when the underlying topic matches. Basis’s July 2026 analysis says AI platforms automate how creative is tested and generated in real time.

Poor semantic fit can cause an ad to lose an auction that price alone might otherwise have won, so a higher bid does not resolve a weak intent description. Budget planning must also allow for conversational AI advertising’s lower volume and higher intent than broad display. ChatGPT’s launch CPM was notably higher than comparable social inventory. Conversion economics should be judged by how precisely ads reach the purchase moment as well as by impression volume.

Brand safety and measurement inside a stack that generates novel text

The first buyer objection concerns a surface that does not exist until the assistant generates it. URL-level blocking can exclude known display environments, but it cannot establish the suitability of fresh conversational text. Keyword blocklists likewise cannot provide complete protection against responses generated for each prompt. The targeting stack thus needs to filter each conversation for suitability in real time inside the decision pipeline.

Hallucination creates a distinct risk within that surface. An assistant can misrepresent a brand’s pricing, compliance claims, or product features in the same response that carries the brand’s ad. The resulting adjacency can make the brand appear to endorse the false statement. This is different from judging whether an existing page covers an unsuitable subject. A credible safety approach must account for what the generated response says.

The second objection concerns attribution, which remains genuinely unsolved at scale for assistant-mediated discovery. Some paths from conversational exposure to conversion do not pass through a standard click-tracking chain. Session-level engagement is measurable now, as is click-through where a link exists and downstream conversion where pixel or clean-room matching is possible. Cross-session attribution after an unclicked interaction is not yet standardised, nor is attribution across multi-turn research on multiple AI surfaces. Those boundaries limit what a campaign can claim, even when individual signals are available.

Advertisers entering the channel should design measurement to allow for these gaps, because last-click models do not fit the journey. Developing conversational measurement infrastructure now can provide a structural advantage as standards solidify. The targeting stack can read and act on live intent, but campaign planning and measurement must be designed for the same environment.