Evaluating DSP Prompt-Targeting Depth When the Signal Is a Live User Query

Criteo is the named adtech partner in OpenAI’s ChatGPT ads pilot, while Google AI Mode and Microsoft Copilot receive ads through existing campaign types. Buying ads inside AI conversations therefore requires distinguishing purpose-built conversational targeting from existing search integrations. A DSP evaluation should test how deeply the platform reads and acts on a live conversation.

Why a live user query is a fundamentally different targeting signal than a keyword

A three-word search query leaves the advertiser reconstructing intent from a fragment. A multi-turn conversation can state the user’s budget, use case, shortlist, and rejected options before an ad is considered. The targeting system receives a different kind of input. Instead of estimating what someone probably wants from a few words, it must interpret what that person has explained.

Match types, negative keywords, quality score, and keyword-level bid adjustments do not describe ChatGPT’s conversational targeting system. Its primary control is a natural-language context hint written at the ad group level. The auction assesses that hint with the ad title, copy, and landing page against the live conversation.

Google AI Mode and Microsoft Copilot attach placements to qualifying campaign types automatically, without a separate opt-in or opt-out. Evaluate how deeply a DSP can read, parse, and act on conversational context in real time.

What prompt-level signals actually contain and why they are richer than any prior intent proxy

A live prompt can function as a declared intent brief. It can reveal research stage, budget, constraints, prior product rejections, and a specific use case together. A prompt expressing interest in starting to invest but feeling completely overwhelmed signals exploration, while “compare CRM platforms for small businesses” signals near-decision purchase intent. Those are different advertising contexts, even before any audience data is considered. A system that extracts only the category discards much of the useful signal.

The conversation then adds a trajectory to that brief. Early prompts can create awareness-level opportunities, while later turns accumulate refinements and exclusions that make a conversion-focused offer more appropriate. Users often go through several prompts before leaving for the open web to convert. Reading only the latest topic can miss that progression.

Contextual targeting reads what a page is about, while behavioral targeting reads past activity. Prompt targeting reads what someone is trying to accomplish now. Its standard is therefore not simply an improvement over cookies, but the ability to interpret a brief without losing its constraints.

The four dimensions of prompt-targeting depth that a DSP evaluation must cover

The first test is semantic understanding of conversation content. A buyer should present two travel conversations, one about planning a family trip on a budget and another about comparing business-class upgrades. The matched ads should reflect their different constraints. Recognizing “travel” in both is not enough, because domain recognition without product and use-case fit remains shallow targeting.

The second test is intent-stage classification within a category. The platform should distinguish research, comparison, and decision contexts, then demonstrate how creative and campaign objectives follow those distinctions. Objective choice is part of this test. A Reach/CPM approach belongs with exploratory conversations, while a Conversions/oCPC approach belongs with high-specificity, near-decision contexts rather than every conversation about the same product category.

The third test is continuity across the conversation thread. By prompt six of a travel-booking discussion, a user may have narrowed the choice a great deal since prompt one. Earlier rejections must remain relevant to later matching. The buyer should ask whether the auction considers prior turns in the session or evaluates each prompt independently, because a reset loses the accumulated specificity that can signal proximity to purchase.

The fourth test checks how the relevance-weighted auction works and how much of that buyers can see. In ChatGPT’s relevance-weighted, second-price auction, a more relevant ad can beat a higher bid. Precision can therefore change the placement outcome. Buyers should ask whether relevance scores appear at the ad group or creative level and whether the platform explains wins and losses in terms of relevance alongside the bid and delivery result.

How the context hint — the primary targeting control — works and what good versus poor hint-writing produces

The context hint describes the conversations in which an ad belongs. It is written at the ad group level in natural language, with 280 characters used as a practitioner convention without an officially documented limit. A useful writing interface should support precise descriptions of intent.

Good hints describe what a person is trying to accomplish and why the offer is useful at that moment. Lists of topics separated by commas make the input more like keyword targeting and lead to poor matching. The landing page also contributes to relevance. OpenAI’s ad system crawls it during review, so a precise hint cannot be evaluated separately from the title, copy, and destination that support the offer.

For buyers, the operational test is whether the DSP helps advertisers write, compare, and refine those inputs. Useful tools include vertical-specific guidance, warnings when hints are too broad, and A/B testing with relevance-level reporting. Automation also deserves a close look. Criteo’s June 2026 investor press release describes Prompt Smart Ads as combining catalog data, commerce intelligence, and prompt-level insights to generate creatives and product descriptions tailored to the consumer’s query.

Ad matching runs in a separate service while the model processes the query and generates its response. Advertisers cannot pay to influence the model’s answer. Buyers should treat claims about “influencing AI responses” as a red flag or as a description of something other than paid conversational placement, and assess whether hint tooling adapts to different AI surfaces rather than assuming every conversation style is interchangeable.

What the auction structure reveals about a DSP's ability to act on conversational context at bid time

A precise hint has value only if the buying system can use it at auction time. In ChatGPT’s auction, building the relevance signal is a key part of running a campaign as well as a writing task before launch. Bid controls alone are insufficient for optimization. A DSP evaluation should establish which signal the auction used and how the platform explains the resulting placement decision, rather than treating contextual understanding as a feature demonstrated only in the interface.

Campaign objectives expose another part of that execution logic. The available choices are Views (CPM), Clicks (CPC), and Conversions (oCPC or oCPM), with different implications for exploratory and near-decision conversations. Optimization and billing are distinct. OpenAI’s billing documentation states that eligible advertisers using Conversions can select click billing through oCPC or impression billing through oCPM while optimizing toward a selected standard conversion event. OpenAI also states that it does not yet have performance benchmarks across advertisers, industries, or campaign types.

Cross-surface buying adds a further test because the underlying auctions differ. Google AI Mode uses existing campaign types, including broad match, AI Max, Performance Max, and Shopping, while Microsoft Copilot uses Search, Shopping, Performance Max, and Audience. Neither offers an opt-out for these qualifying placements. Those systems use existing search auctions for conversational contexts, while ChatGPT’s system was built for conversations and uses the context hint as its primary targeting control.

A buyer should check whether a DSP supports conversational inputs or gains access through existing campaign integrations, then ask how it brings both into one buying strategy. Surface coverage does not establish semantic capability. Conversely, strong context reading on one surface does not establish the ability to translate that precision across different auction rules and conversation styles.

How measurement and attribution work — and do not yet work — in conversational ad environments

Multi-turn conversations complicate even the definition of an impression. Surfaces do not share a standardized answer about whether counting attaches to the turn displaying the ad, session start, or user engagement. Conversion introduces a second break in visibility. A user may continue through several prompts and then move to the open web, leaving a less direct path between exposure and purchase than a conventional search click.

OpenAI’s October 5, 2026 Ads Blog post describes expanded measurement infrastructure through Pixel and Conversions API support, an Event Quality Score, and incrementality partnerships with Haus, Measured, and WorkMagic. The one-day view-through window had shipped in late August 2026. These additions improve the available measurement tools. They also show why DSP evaluations must examine current measurement capabilities rather than assume the channel already offers mature attribution.

Performance evidence requires similar care about what is being measured. Adobe Analytics data shows that AI-referred visits to US retail sites convert at a meaningfully higher rate than non-AI traffic, but referral performance is not itself proof of paid-ad effectiveness. Microsoft Advertising’s August 2025 data shows higher click-through and conversion rates for Copilot than for traditional search, along with shorter customer journeys. Those results come from the vendor and should be described that way. Lower raw click volume across conversational environments should not take away from conversion quality, but conversion quality still needs to be backed by campaign-level evidence.

Paid placement and organic AI citation also need to be measured separately. Buying an adjacent ad and receiving the model’s recommendation in its answer are separate outcomes with separate explanations. Buyers should therefore ask whether the DSP distinguishes view-through from click-through attribution, supports incrementality testing, and exposes relevance measures beyond engagement metrics. A credible answer also states what the platform cannot yet measure, rather than presenting incomplete attribution as full coverage.

How to run a structured prompt-targeting depth evaluation across DSP candidates

The comparison should run through five tests in order. Begin with paired conversations in the same category but with different constraints, then assess whether matched creatives differ in message, stage, and product fit. Next, ask to see how research-versus-decision classification works within that category. The choice of objective should follow that difference in intent. This separates a platform that recognizes the topic from one that can use its meaning to shape a campaign.

For the third test, ask whether the auction takes earlier turns in the session into account or assesses each prompt on its own, and assume single-prompt evaluation if the vendor cannot give a clear answer. The fourth test examines relevance-signal transparency. Request ad group and creative-level reporting, explanations of wins and losses, and clarity about the signals available for optimization. Impression and click reporting alone does not show how relevance affects the auction.

Assess vertical-specific guidance, broad-hint warnings, and support for testing variants with relevance-level reporting rather than accepting a general claim of AI-assisted targeting. Supply access also needs direct scrutiny. Ask how publisher relationships make conversational signals available. Ask what a platform relying on third-party auction data can actually read. Cross-surface reach belongs in this same evaluation, with evidence that context handling transfers across ChatGPT, Google AI Mode, and Microsoft Copilot.

The evaluation should conclude with a measurement audit. Ask about incrementality experiments, view-through windows that fit conversational journeys, separate organic-citation tracking, and what current attribution can and cannot measure. Honest uncertainty is useful information for a buyer. Working through these tests forces the team to understand the signal, the auction, and the measurement gaps before selecting a platform, producing a stronger grasp of the channel regardless of which DSP wins the evaluation.