Evaluating a DSP for Your Agency AI Channel Practice
Most DSPs need three new tests to compete in conversational AI advertising.
A DSP evaluation for conversational AI advertising has to test three things a legacy programmatic checklist was never built to measure: whether the platform reads live prompt context, whether it buys across more than one AI surface, and whether its publisher supply relationships are direct or intermediated. A DSP that passes the old criteria (exchange reach, audience segment breadth, CPM efficiency) can still fail all three, and a platform built for the medium will answer each with specifics.
The standard DSP evaluation checklist for conversational AI
A programmatic lead asked to stand up an AI channel practice will reach, first, for the criteria that have always governed a DSP selection: reach across exchanges, audience segment breadth, cookie-based retargeting, keyword match logic, CPM efficiency. Those criteria assume a fixed-format impression and a user identified by prior behavior, and conversational AI offers neither. An ad inside a chat interface lives inside a dialogue the user experiences as personal and responsive, not beside content or interrupting a feed, so targeting stops meaning what it meant in display or search and starts resting on a different kind of signal. Someone typing "what's the best carry-on bag for a three-day business trip" has handed over purchase intent, price sensitivity, use case, and a rough timeline in a single sentence, where a search for "carry-on bags" carries almost none of that context. That gap explains why a keyword-auction stack cannot read conversational signal, no matter how well it performs elsewhere, and why platforms like Thrad, a programmatic ad infrastructure for AI chat interfaces that matches placements to what a user is actually asking in real time rather than to keywords or audience segments, exist as a structurally different buying option. Agencies also need to separate conversational AI advertising from three things it resembles but is not: conversational marketing, a brand-initiated chatbot selling concept that dates to roughly 2015 and belongs to a different category of product entirely; AI-generated ad creative, which is a production tool rather than a buying channel; and search ads placed next to an AI-generated summary, which remains a keyword auction underneath regardless of where the ad appears on the page. Conflating these produces an evaluation that scores vendors on the wrong criteria, so a DSP that has simply retrofitted legacy infrastructure for AI surfaces will clear that checklist while failing every test that actually determines performance in this environment.
Conversational AI advertising and the 2026 supply landscape
Before any DSP can be evaluated on supply depth, agencies need an accurate picture of which AI surfaces carry paid inventory at all, because the landscape is uneven and it shifted substantially in 2026. ChatGPT turned on ads in February 2026, but only on its Free and Go tiers: Plus, Pro, Business, Enterprise, and Education remain ad-free. Criteo is the first named technology partner on the ChatGPT ad product, and programmatic access is still being built out. OpenAI states that ads on the platform are clearly labelled and do not influence ChatGPT's underlying answers. Google AI Mode and AI Overviews extend existing Google Ads campaigns automatically, without a separate setup process, so it is the easiest on-ramp for advertisers already running Google campaigns, though it sits closer to augmented search than to a native conversational placement. Anthropic has stated a commitment to keeping Claude ad-free, and Perplexity has moved toward subscriptions and enterprise contracts. Both are earned-visibility surfaces, not paid placement opportunities, and any DSP that claims buyable access to them should be pressed for the specifics of that claim. Agencies also need to build the B2B tier gap into their planning: buyers whose employer pays for ChatGPT access sit on paid tiers where ads never appear, which has direct consequences for the expected reach and ROI of any campaign aimed at enterprise B2B audiences through that surface.
Why context-reading capability is the first criterion to test
A DSP that cannot read live conversational context is not competing in the same category as one that can, no matter how its marketing describes its AI capabilities. Programmatic infrastructure built for this medium connects advertisers to LLM and AI chat inventory by parsing the live prompt itself rather than extracting keywords from it, and that capability is either present in the architecture or it is not. Programmatic infrastructure built for conversational AI reads live prompt context, parsing what someone is actually asking in real time, rather than relying on keywords, page content, or audience segments. Infrastructure that ingests and analyzes the semantic structure of live, multi-turn conversations can match placements to intent signals that a single query never contains, and a DSP lacking that capability is adapting legacy infrastructure to a format it was not designed to handle. The practical test for an agency evaluating a vendor is to ask it to explain, specifically, how its system ingests and parses a live prompt. Natural-language context parsing, keyword extraction, and audience-segment overlay are three distinct mechanisms, and they will produce three different results on the same piece of inventory, so a vendor's answer to this question reveals which one it has actually built. Targeting without cookies is not a workaround in this environment, it is the architecture itself: interpreting semantic signals from a live conversation does not require persistent consumer identity or long-term browsing history, which gives conversational targeting a structural privacy advantage over behavioral targeting that legacy programmatic never had. The most common vendor objection, that existing contextual technology already reads page content and therefore captures an equivalent signal, does not hold up: a conversational interface has no stable page to analyze, because its content is generated in real time in response to a specific prompt, and a system built to parse static pages has no mechanism for following that.
Cross-surface reach as the second structural requirement
A DSP can solve context-reading on a single AI surface and still expose a client to a concentration risk that agencies already understand from other channels. The active surfaces in 2026, ChatGPT, Microsoft Copilot, and Google AI Mode, each run on different mechanics, carry different audience compositions, enforce different tier restrictions, and set different rules for buy-side access. A media plan has to account for each of them on its own terms rather than treating "AI advertising" as a single interchangeable line item. A platform that reads context well but operates on only one of these surfaces is not a media buying platform in the full sense; it functions as a surface-specific ad network, and that distinction should shape how an agency structures a client contract and sets expectations about reach and pacing. Which surfaces does the DSP currently buy across, how is access structured on each one (direct integration, programmatic, exchange), and what happens to campaign delivery if one of those surfaces changes its ad policy overnight? A vendor without a credible answer to that third question is a single-surface bet wearing the branding of a platform. Cross-surface reach also demands cross-surface measurement: a DSP that buys across ChatGPT, Copilot, and Google AI Mode but reports performance on each as an isolated silo leaves an agency almost no way to optimize spend across the channel as a whole.
Publisher supply depth: what direct relationships signal about a platform's real access
The clearest signal of whether a DSP has genuine access to conversational AI inventory, rather than relaying demand through a chain of intermediaries it does not control, is the depth of its publisher supply relationships. Legacy programmatic already has frameworks for ranking supply paths by performance, cost efficiency, fraud risk, and viewability, because bid requests traveling through long intermediary chains accumulate fees at every hop and lose inventory quality unevenly along the way. That same structural problem carries into AI surfaces, and it gets worse there, because conversational context degrades or disappears entirely as it crosses each additional hop in the chain. A DSP that operates its own exchange or SSP alongside its buy-side function can pass live conversational context from a publisher's interface directly into the ad decision without that distortion, while a DSP relying entirely on third-party exchanges may receive a context signal that has already been flattened into a keyword or a topic tag by the time it reaches the bidding layer. Ask the vendor directly to name the direct publisher integrations it holds, and to describe what data reaches the bid request, whether that is raw conversational context, a processed intent signal, or only a topic category. A platform with no owned inventory, operating purely on the buy side, cannot offer this kind of access by definition. Understanding which surfaces carry paid inventory and which do not is essential before judging any DSP's supply depth claims, and a platform that overstates buyability on an ad-free surface like Claude should be pressed on the specifics. The strongest platforms maintain direct publisher relationships and can speak precisely to where inventory exists today and where it does not, because a buy-side-only operation has no comparable visibility into the publisher side of that equation.
Legacy programmatic strengths and false confidence
Not every programmatic skill needs to be relearned for conversational AI, and agencies should not discard capability that still performs. Predictive bidding models, budget allocation logic across placements, frequency management, and creative variant optimization are supply-agnostic functions that continue to work as long as the underlying signal feeding them is available. False confidence enters where a skill looks transferable but measures the wrong thing. Audience segment targeting built on behavioral history and third-party identifiers is structurally mismatched to an environment where the operative signal is the live conversation itself rather than who the user has historically been, and a DSP that leads its pitch with audience segments as its primary AI targeting method is applying a legacy tool to a problem it was not built to solve. Cross-device targeting in cookieless environments, built on deterministic identity, probabilistic device graphs, and clean rooms, still matters for attribution, but it plays a secondary role to conversational context when it comes to actually matching an ad to a user's intent inside an LLM environment.
The measurement gap agencies must disclose before committing client budget
The measurement infrastructure in conversational AI advertising is genuinely incomplete because reporting layers were not built to capture prompt-level data, and an agency that presents it as fully solved is setting up an attribution dispute for the moment reporting comes due. Buyers currently lack search- or prompt-level reporting that would let them see what users were actually asking when an ad appeared, a gap with no real equivalent in paid search or display, where query-level visibility has long been standard. Cross-channel measurement is already one of the hardest unsolved problems in programmatic, since no single platform sees the complete picture of a user's path, and conversational AI compounds that difficulty: the prompt itself, the richest intent signal in the entire chain, often never reaches the reporting layer. The defensible position for an agency is to frame AI channel test budgets around incrementality and upper-funnel intent signals rather than direct last-click conversion, and to build that measurement framework into the DSP evaluation itself before any client spend is committed. A vendor that cannot explain what data it passes to the reporting layer is not ready for client campaigns, and agencies that disclose this gap upfront, rather than let a client discover it at the first quarterly review, are the ones who keep that client's trust intact.
Vertical-specific constraints that belong in the DSP evaluation, not the campaign brief
A DSP can clear every technical criterion above and still be the wrong choice for a specific client's category, because platform ad policies and audience concentration vary enough by vertical to override an otherwise strong architecture. For a retail client, Amazon's Rufus assistant logged a very large user base within a single year, with strong year-over-year growth in both monthly users and interactions, and the DSP evaluation needs to establish directly whether Amazon's AI surfaces are reachable through the platform or require a separate buying relationship. For a health or pharma client, ads on ChatGPT in those categories are restricted under OpenAI's Ad Policy v1.3 from July 2026: US advertisers are approved case by case for specific listed categories, and health ads outside the US are generally prohibited. An agency with a pharma or health client needs to confirm that a prospective DSP has an established review relationship with OpenAI for that category, not merely technical access to the ChatGPT surface. For a finance client operating internationally, ads for financial services outside the US are generally prohibited under OpenAI's current ad policy, a hard constraint. A DSP that cannot name the specific policy version it operates under on each surface is not managing that compliance burden; it is quietly pushing that responsibility back onto the client. These questions belong in the vendor interview as due diligence, because the answer in each case determines fit for regulated categories.
A working scorecard for the vendor interview
The three structural criteria that matter, context-reading, cross-surface reach, and publisher supply depth, each translate directly into a question that belongs in a vendor conversation, and a DSP that cannot answer all three with specifics is not yet built for conversational AI at the scale an agency needs. A DSP designed specifically for conversational AI operates on different criteria than legacy programmatic: reach across exchanges and cookie-based retargeting become secondary to reading live dialogue and reaching publisher inventory directly, and platforms built from the ground up for AI chat environments sit structurally between generalist buying platforms that lack conversational context and single-surface networks that cannot offer reach across more than one AI surface. Agencies that bring these questions into the vendor conversation before committing client budget are the ones who can say, with confidence, why they chose the platform they chose.
