Rebuilding Your Intent Signal Architecture Around Live Prompts Instead of Cookie Pools
Live prompts reveal evolving decisions that cookies never capture, reshaping how ads target buyers.
Real-time prompt intent analysis for programmatic advertising uses the problem, constraints, and decision stage in a live AI conversation to decide how relevant an ad is. Rebuilding intent architecture around that signal requires changes to targeting, bidding, and measurement. Cookie pools show what someone did in the past, while prompts show the decision they are making now.
Why cookie pools are the wrong unit of measurement for how people now make decisions
Cookie pools organize evidence around a user’s previous actions, but a conversational decision develops through exchanges that those pools cannot observe. Page visits and product views provide traces of activity on surfaces advertisers can tag. A conversation can contain the comparison, rejection, and refinement that precede a purchase without producing those traces. The intent signal has moved beyond the observable page.
A prompt such as “help me find running shoes for flat feet under $150” combines use case, physical constraint, and budget. Even when browsing history shows an interest in running shoes, it cannot show which requirements matter for the choice being made now. In a conversational environment, the relevant unit is therefore the developing decision. Retargeting can still reach someone who visited a page, but it cannot substitute for understanding the problem being resolved before that visit.
What a live prompt actually contains that a cookie never could
A live prompt shows intent in the user’s own words, including what they need and the tradeoffs they are weighing. The thread adds how those requirements developed, which options were ruled out, and what the assistant suggested before the next question. These are different signals from retrospective behavior. A keyword can indicate that someone is in the market, while a conversation can explain what would satisfy that person.
The dialogue arc also reveals intent without a category name. Someone who describes a team struggling to coordinate, missing deadlines, and finding it hard to work together remotely may show stronger intent for project management software than someone who types that keyword once. The problem supplies the category connection. Conversational targeting reads earlier exchanges, assistant responses, and follow-up questions as a trajectory. Because relevance can derive from this self-contained context, the targeting mechanism does not require cross-site tracking or persistent user identifiers. That privacy distinction follows from the signal’s structure, rather than from making a behavioral profile less detailed.
Why the existing ad stack fails at the structural level when applied to conversational surfaces
Keyword matching, audience segments, and bid modifiers based on past behavior assume that relevance can be judged on a surface that stays much the same. A conversation changes the surface with every exchange. An initial request for general information can become a comparison and then a decision under specific constraints. A relevance judgment made at the beginning can therefore become inappropriate later. Adding a conversational placement to an existing campaign does not, by itself, give the targeting system a way to reassess that progression.
Reach and contextual understanding are also separate capabilities. Generalist DSPs built for display and social provide distribution but do not read conversational context, while single-surface AI ad networks can understand context within one assistant. Neither on its own allows contextual buying across multiple assistants. The architecture needs to bring those functions together at the prompt level.
Google AI Overviews, AI Mode, and Microsoft Copilot automatically include eligible campaigns, with no opt-out for those placements. Advertisers can therefore end up on conversational surfaces without choosing to rebuild their targeting or bidding inputs. The campaign may look familiar even though the context in which its ads appear has changed. That disconnect makes an inventory audit more urgent than selecting another campaign type.
How can live prompts become programmatic ad signals in practice?
A prompt-native targeting input describes the situation in which an offer fits, including what the person is trying to resolve. Context hints make that distinction concrete. A 280-character natural-language description of the relevant conversational situation is a different artifact from a comma-separated list of topics. Precise problem narratives support better matching, while broad topic lists discard the relationships between the user’s need, constraints, and desired outcome.
Intent stage then qualifies the topic match. A research conversation needs different creative and landing page logic from a comparison or decision conversation, and conversation depth provides a readable signal of that progression. The landing page also participates before the click. OpenAI’s ad system crawls it as a relevance signal during review, so a page that continues the specific problem expressed in the conversation contributes to match quality. Treating that page only as a conversion destination leaves part of the targeting system disconnected.
The implementation has five distinct jobs, which are buying the placement, targeting the conversation, producing creative for each angle, matching a landing page to each ad, and counting conversions once across surfaces. These jobs cannot simply inherit their legacy equivalents unchanged. They must also operate within the conversation’s latency constraint. Decisions need to happen in milliseconds rather than interrupt the exchange, which favors pre-computed intent clusters over real-time user-profile lookups.
How bidding logic has to change when conversation state replaces the keyword auction
A relevance-weighted second-price auction against conversational context rewards specificity as well as price. A precise context hint at a moderate bid can outperform a vague hint attached to a higher bid. Spend alone does not settle the match. Because the auction evaluates the context hint, ad title and copy, and landing page together, bid strategy and creative strategy operate as a unit.
Conversation stage also changes the value being purchased. Someone six exchanges into a decision about a specific product presents a different opportunity from someone beginning an informational exploration. A bid strategy that treats both as equivalent risks overpaying for early research and underpaying for developed purchase intent. Bids therefore need to use intent-stage inputs as well as how keywords have performed in the past. This is especially urgent for campaigns already running in Google and Microsoft conversational placements with bidding logic that was not built for them.
Competitive analysis changes with the auction’s inputs. There is no equivalent keyword layer through which a competitor analysis tool reveals rivals’ context hints. Contextual specificity and copy quality therefore become more important sources of advantage than reverse-engineering another advertiser’s keyword bids. The task is less about identifying an overlooked string and more about expressing the offer’s fit to a particular decision.
What measurement architecture has to handle that attribution models built for cookies cannot
The intent event and conversion event can occur on different surfaces, days apart, with no cookie connecting them. Last-observable-click attribution credits what remains visible before conversion, even when an earlier dialogue shaped the choice. Some of this gap is currently unsolvable with available tooling, and a credible measurement design must acknowledge that limit.
Click-through rate alone introduces another distortion. ChatGPT ad CTR benchmarks at roughly 0.9%, below Google search ad CTR on a like-for-like basis, but a conversational referral can arrive after a recommendation. Lower click volume can therefore coexist with stronger downstream conversion economics. First Page Sage’s 2026 projections show ChatGPT ads with higher conversion rates than Google Ads in most of the industries studied.
Incrementality experiments offer the most defensible way to assess causal contribution despite incomplete journey visibility. OpenAI states that it is partnering with Haus, Measured, and WorkMagic to explore geo-based experiments for ChatGPT Ads. This approach requires test-and-holdout design that many teams have not built. In parallel, one attribution rule applied across ChatGPT, Google AI surfaces, and other channels provides a more consistent accounting basis than separate platform-native reports. Consistency avoids assembling a channel comparison from systems that each overcredit their own inventory.
Where prompt-level signals are most commercially potent, by category
High-consideration decisions make the value of contextual detail especially clear. In retail, a prompt can combine material, budget, use case, and body constraints at the product-fit stage. Those requirements describe a choice more precisely than a retargeting segment does. Macy’s “Ask Macy’s,” built on Google Gemini, went fully live in March 2026 and showed that shoppers using the assistant spent about 4.75 times more than those who did not. That comparison illustrates the commercial importance of assistant-supported evaluation, though it does not by itself isolate an advertising effect.
BCG found that a majority of US adults who use AI during the purchase journey research or plan travel purchases with AI help. Trip.com’s official newsroom reports that TripGenie users show “twice the order conversion rate” and are “30-40% more likely to return to the platform.” The opportunity lies in developed travel intent. As with retail, assistant usage outcomes indicate the value of the decision context without establishing that an ad caused the difference.
Financial services and technology show different gaps between what audiences do and what advertisers do: eMarketer’s 2026 analysis shows that consumers use AI in financial services about as much as in retail, while total search spending is roughly half that of retail even though search takes a large share of category budgets. That suggests an earlier transition with room to establish conversational presence. Technology and electronics have the highest consumer AI usage among the categories studied, but advertiser AI spending lags. In both cases, the commercial opportunity depends on aligning placement with where evaluation is happening, rather than extending historical audience categories unchanged.
ChatGPT ads reach free and Go tier users, while Plus, Pro, Business, and Enterprise plans are ad-free. Employer-paid accounts therefore exclude a managed-account buyer segment that many B2B advertisers prioritize. Category prioritization must consider both the quality of the conversational signal and whether the desired buyer can actually receive the placement.
The strategic sequence for rebuilding intent architecture without abandoning what still works
The practical sequence is selective rebuilding. The first move is to audit eligible campaigns already serving inside AI Overviews, AI Mode, and Copilot before adding new spend. That establishes where conversational placements already depend on legacy inputs. The next move is to build an intent category map around the problems a product solves, then write context hints from those problem narratives.
Landing pages come next because they participate in relevance as well as conversion. Each page should address the needs and constraints expressed by its matched ad, rather than hand the visitor to a generic category destination. Measurement must then put more weight on conversion economics than on click volume. A single attribution rule across surfaces establishes consistent accounting, while incrementality experiments, where possible, test contribution beyond that rule. Neither step requires pretending that every conversational influence can be assigned to an individual conversion.
At the brand layer, organic presence should precede paid reinforcement. A citation sits inside the answer, while an ad sits beside it, so buying placement without organic visibility can mean paying to accompany a competitor’s recommendation. This does not make other useful signals obsolete. Mastercard Commerce Media uses a different base of permissioned transactional insights, drawing on billions of annual transactions to match offers at the right moments for hundreds of millions of enrolled consumers. The distinction is between signals suited to the decision and inferred histories used as substitutes for it.
Platform selection follows the same architectural logic. Buying across multiple AI surfaces and reading conversational context at each one is a different capability from broad distribution alone or contextual access to one assistant. As more conversational surfaces open up, that combination gives teams more contexts to learn from. Teams building now can learn which intent clusters convert and which problem narratives match the right conversation stages while competition is still thin. That learning does not transfer from a cookie pool and cannot be bought retroactively.
