DSPs That Parse Live Query Intent Versus Those Still Relying on Audience Segments

Real-time prompt intent analysis means reading the semantics of a live conversation, the actual words a user types into an AI assistant, and using that as the targeting signal. Platforms built for this, one model among them, match ads to what a person is asking right now rather than to who that person has historically resembled. The vendors that do this well are structurally different from traditional DSPs: they connect to conversational auction infrastructure, parse context hints, and treat the DMP layer as optional. The rest of this piece lays out why that distinction exists, how to evaluate it, and which platforms handle each side of it.

Targeting architecture divides DSPs into two structurally different tools

The DSP market looks, from a distance, like a single category with varying degrees of polish: some platforms have more inventory, some have sharper creative tools, some bolt on more AI. That appearance is misleading. The real divide sits one layer deeper, in what each platform is willing to count as a targeting signal. A segment-based DSP answers one question: who is this user, based on what the platform already knows about them. A query-intent DSP answers a different question entirely: what is this user trying to do, right now, in this conversation. These are not two points on a maturity curve, and a vendor doesn't graduate from one to the other by adding a machine learning layer on top of its existing audience stack. They require different data inputs, different auction mechanics, and different inventory surfaces, because the signal itself is a different kind of object. AI Digital's "Programmatic Ad Targeting: Best Strategies, Tools, and Tactics in 2026" documents that AI-driven predictive optimization is now standard across enterprise DSP environments, but it also makes clear that the signals those optimization engines run on vary fundamentally by platform type. A buyer choosing a DSP is, whether they frame it this way or not, choosing which class of signal they believe actually predicts conversion. That choice determines which inventory they can reach at all, not just how well they perform on the inventory they already buy.

How audience segments are built

Audience segments are built from demographics, purchase history, CRM records, and modeled behavioral inferences, and every one of those inputs describes a past state. Viant's overview of programmatic targeting lays out the standard DSP data stack in three tiers: first-party data drawn from CRM systems, purchase history, and site interactions; second-party data brought in through partnerships; and third-party aggregated behavioral data purchased from outside vendors. Each tier, regardless of its source, describes what a user has done. None of them describe what a user is doing at the moment the ad is served. That's not a flaw in execution; it's the basic design of the segment model, and for a wide range of marketing goals, it works. Lifecycle marketing depends on knowing who a customer has been over time. Suppression lists depend on past purchase records. Retargeting depends, by definition, on a prior behavioral trace. Platforms like 6sense, Demandbase, and ZoomInfo have built substantial B2B targeting businesses on exactly this kind of firmographic and intent-adjacent data, and Bombora's cooperative data model for B2B intent signals works because it aggregates research behavior across a network of sites over time, which is a legitimate and useful signal for a certain kind of decision. The trouble is structural, a matter of basic design that better pipelines can't fix. AI Digital's comparison of DSPs and DMPs notes that classic DMP models are under pressure for two reasons: third-party data availability has narrowed, and the segmentation job itself is increasingly handled elsewhere, by CDPs, clean rooms, warehouse-first segmentation setups, or DSP-native audience tooling. The segment layer is fragmenting even on its own terms. But fragmentation aside, there's a ceiling no amount of first-party enrichment gets past. No volume of CRM data, no depth of behavioral modeling, tells a platform that a specific person is, at this exact moment, mid-conversation with an AI assistant, asking a question that reveals live purchase intent. That moment simply does not exist inside a segment, no matter how recent or well-built.

Live query intent as a signal

A live conversational query is a declared, real-time statement of intent, made in natural language, in the present tense. When someone asks an AI assistant a question, they're stating a need directly, as it forms. Valasys frames this distinction sharply: an interest stack is a snapshot, while a conversational model is a live feed. The difference goes beyond timing, though, into depth. A single keyword captures a moment. A conversation captures a developing decision. Consider someone who opens by asking how to start running, mentions they've had trouble finding comfortable shoes in the past, specifies that they run mostly on pavement, names a budget, and then asks for a specific recommendation. No single keyword in that exchange captures the purchase intent on its own. The full thread does, and it does so with a level of specificity that no segment label, however granular, could match. This kind of interpretation doesn't require access to a user's identity or their long-term browsing history. The intent lives inside the conversation itself, so this signal type has a genuine privacy advantage that segment-based targeting, built on accumulated personal history, structurally cannot offer. AI Digital draws a useful line between contextual targeting, which looks at what a user is consuming right now, and behavioral targeting, which looks at what a user has done in the past. Live query intent occupies a third category that fits neither description cleanly: what the user is actively constructing, in real time, through dialogue. That's why a finance brand chasing refinance-ready homeowners gets nothing useful out of buying a "homeowners 35-54" segment. The signal that actually identifies a refinance-ready person is the sequence of questions they're asking right now, not a demographic label applied weeks or months earlier by a modeling vendor.

AI assistant interfaces make this divide a practical operational problem

Conversational AI interfaces have turned this distinction from an architectural curiosity into a live operational constraint, because they opened a new advertising channel and the structural gap between DSP types maps directly onto which part of that channel a buyer can reach. ChatGPT launched a limited, invite-only advertising pilot on February 9, 2026. Open self-serve access followed in May 2026. Criteo was announced as the first named technology partner in March 2026, and programmatic access through StackAdapt opened that same May. Buying into this inventory requires reading conversational context. Advertisers submit context hints, prose descriptions of the conversations and topics where a given ad makes sense, rather than exact-match keyword rules, though OpenAI's API also accepts keyword-style inputs for advertisers who want them. OpenAI has been explicit that these context hints are not guarantees of delivery in any specific conversation. Matching on ChatGPT draws on the current conversation's context and intent, the advertiser's landing page and copy, and, only when a user has enabled ads personalization, select signals from that user's broader ChatGPT activity. A segment-based DSP has no route into that auction on comparable terms, because the auction isn't asking the question a segment is built to answer. The mechanics compound the problem: OpenAI runs a relevance-weighted, second-price auction, where relevance functions as a multiplier on the bid rather than a tiebreaker, so a well-matched ad with a lower raw bid can outrank a poorly matched ad with a higher one. Contextual signal quality has become a direct cost lever in that auction, not a nice-to-have. AI Digital's research confirms that AI-driven predictive optimization is standard across enterprise DSPs in 2026, but optimizing against conversational intent signals is a distinct capability from optimizing against behavioral or demographic ones, and the two don't transfer. Publisher-side supply is bifurcating along the same line: AI-powered widgets now live on publisher sites, generating conversational interactions with readers that become monetizable through contextual chat ads. That inventory sits outside the traditional display and video pipes that segment-based DSPs were built to access. The gap is a description of where a meaningful share of inventory already sits today, reachable only by platforms built to read it.

DSP architecture required to operate on conversational surfaces

Operating effectively on conversational AI inventory requires a DSP built around semantic context matching from the ground up, not one retrofitted from legacy display infrastructure with a conversational layer bolted on afterward. The first requirement is a context hint architecture that accepts prose descriptions of conversational settings and matches them against live query semantics, fundamentally a natural language understanding problem. The second is auction-layer integration with LLM SSP infrastructure, the supply-side layer that identifies monetizable moments inside AI chat sessions, evaluates conversational context, and runs the relevance-weighted auctions those moments depend on. A standard display SSP connection doesn't provide any of that, no matter how well it performs on banner or video inventory. AI Digital's research on AI in programmatic advertising describes DSPs optimizing across several axes at once, including predictive bidding, budget allocation, frequency modeling, and creative optimization, and identifies data quality and signal clarity as the real bottleneck to scaling. On conversational surfaces, signal quality means the richness of conversational context captured, not the depth of an audience segment file. Publisher controls form a third requirement, and they matter at the architectural level: a responsible LLM SSP has to let publishers block specific ad categories from appearing in specific conversational contexts, the way a financial news publisher might block gambling ads or a health site might restrict pharmaceutical advertising, and the DSP has to honor those controls at the bid-request level. Brand safety enforcement on these surfaces works differently than it does in display, too, because the exclusions are tied to the content of the conversation itself. OpenAI excludes ads from mental health and political conversations regardless of advertiser category, and personal health conversations are kept free of ads in sensitive or emotionally vulnerable contexts, though an April 2026 policy update allows non-sensitive health queries to carry ads from approved health advertisers. A DSP operating on this surface has to receive and act on those placement exclusions in real time. The brand safety logic has to live inside the bidding pipeline. Cross-surface reach is what separates a genuine conversational DSP from a single-surface ad network: a platform that can only buy on one AI assistant has context-reading ability but limited reach, while a platform connected to multiple AI surfaces through an exchange layer offers both. That is an architectural choice about what signal drives matching, not an added sophistication layer sitting on top of a conventional DSP. AI Digital's own analysis confirms that a DSP can run effectively without a traditional DMP when targeting relies on contextual signals and platform-native audiences, and on conversational surfaces the live query itself substitutes for the segment, so the DMP layer matters far less than the context-reading layer sitting next to it.

Industry-specific cases where the segment-versus-intent divide produces different outcomes

The stakes of this divide play out differently by vertical, but the underlying pattern holds across all of them: categories where users begin their purchase journey inside an AI conversation are the categories where segment-based DSPs are most exposed. In e-commerce, a brand with a synced product catalog can surface real products and live pricing the moment a user asks a shopping-related question, whether that's a product comparison or a request for gift ideas within a set budget. The match happens against the live question, not against a browsing session from three weeks prior, and that timing difference is the entire value proposition. In finance, OpenAI has opened selected financial services advertising in the United States, though only under manual approval, licensing requirements, and strict category limits. The targeting advantage in this vertical is concrete: a refinance-ready homeowner is identifiable through the specific questions they're asking in the moment, not through a demographic label like age or homeownership status applied weeks earlier by a third-party data vendor. Legal services face a similar structure, permitted in the United States for licensed advertisers subject to manual approval, but prohibited outside the US. Health and healthcare sit on the restricted side of the line for now. OpenAI's current ad policies keep healthcare advertisers, medical services, prescription products, and health claims about prevention, diagnosis, or treatment out of ChatGPT's ad inventory during the test period, though certain U.S. subcategories, including dental services, hospitals, health insurance, and medical testing, may qualify for case-by-case manual approval. Hospitals, private practices, telehealth brands, and most consumer-facing health businesses can't currently buy sponsored placement inside ChatGPT conversations. For this category, the workable path runs through organic citation visibility via generative engine optimization rather than paid placement, at least under current policy. In B2B and SaaS, education and training providers who reach users mid-conversation about career decisions or skill gaps are reaching those people at the exact moment they're weighing options, which is a fundamentally different entry point than a display ad served against a "professionals interested in training" segment built from inferred job titles and content consumption. Platforms such as Intentsify and Captify have built real businesses around sharpening intent signals for B2B and contextual buying respectively, and Quantcast and Inuvo offer their own approaches to audience and contextual reach. Each reflects a genuine, defensible point on the spectrum this piece has been describing. Buyers should not ask which of these platforms is more advanced than the others; they should ask which class of signal, segment or live query, actually describes the moment where their category's purchase decisions get made, and which platform architecture was actually built to read that signal when it happens.