Why Cookies Were Never Capturing the Moment Buyers Actually Decided

Real-time prompt intent analysis for programmatic advertising reads the meaning and progression of a live AI conversation to assess ad relevance, rather than reconstructing intent from browsing history. Cookies were never capturing the moment buyers actually decided because they recorded only actions. Reaching that decision moment requires conversational inference.

What cookies were actually designed to do, and the decision moment that design skips

Cookie-based tracking was built around observable page behavior. A page visit, a click or an exposure becomes a record after the action happens, while the reasoning that led to it remains outside the record. That distinction is the architectural limit. A pixel can count an action without knowing whether the person is curious, comparing options or ready to buy.

An action count also lacks the commercial meaning advertisers need. By itself, it cannot distinguish a modest purchase from a large one, a new customer from a returning one, or a completed sale from an abandoned cart. Attribution windows also limit what the feedback loop can track. B2B buying cycles span weeks or months, while purchases of furniture, electronics and subscriptions can outlast the permitted window. When conversion falls outside that window, the algorithm loses the connection between the earlier signal and the result.

Cross-device journeys make the discontinuity concrete. A person researches on a phone in the morning and purchases on a laptop in the evening, but cookie logic can represent that journey as two different people. The buyer is continuous. The record is not. The resulting match-rate gap can be severe enough that a site with meaningful traffic can retarget only a fraction of those who showed purchase intent.

Rules on cookies add to these limits, which were already built into the system. CNIL fined Google hundreds of millions of euros in September 2025 for making cookie refusal harder than acceptance. But even a fully compliant cookie system would still record behavior without recording the conversational progression through which a buyer settled on a choice.

Where intent actually crystallizes — inside the conversation, not on the page

AI conversations increasingly contain that progression before a buyer reaches a brand’s site. Someone can describe prior experience, specify constraints, narrow a budget and request a recommendation within one thread. The resulting signal has a different depth from a keyword. It links a proposed purchase to the conditions under which the person would consider it, rather than leaving the advertiser to infer those conditions from separate visits.

A request for vegetarian options illustrates the distinction. On its own, the phrase says little about whether someone is choosing a restaurant or considering activities during a trip. In a conversation, earlier exchanges establish the concern, and later exchanges can show whether a recommendation has resolved it. A user can acknowledge a concern, signal openness to new products and reach readiness to buy without leaving the thread.

Travel makes the commercial consequence particularly clear. A meaningful share of travel queries now begin in an LLM. Discussions of an itinerary give tour operators and activity providers signals about what a user has in mind, while comparisons of places to stay can show when users are close to booking. These are different stages within a developing plan. A page visit might show interest in a destination, but the conversation explains what would make a given option a good fit.

That creates a gap even for well-run campaigns. A brand can run ads with great technical skill on Google and Meta yet be absent while an assistant helps a buyer narrow the options. Later visits may record what came out of that discussion without showing the discussion itself. The missing layer is the place where the buyer’s requirements become a choice.

Why reading conversational intent requires a different mechanism than reading page content

Conversational targeting must read how turns in a conversation fit together as well as match words in the latest prompt. Earlier questions, the assistant’s responses and the follow-ups all help establish what the user is trying to accomplish. Meaning changes as the exchange develops. A request that appears broad in isolation can be highly specific when read after several rounds of comparison, while a product mention may express a concern.

The transformer architecture underlying LLM assistants maintains discourse coherence to generate responses. Advertising systems built on that infrastructure can use the same contextual representations to evaluate relevance. The distinction from keyword matching is mechanical. The system evaluates an advertiser against an evolving understanding of the discussion, rather than treating each new input as an independent string to match.

Static contextual advertising, by contrast, analyzes page content as a fixed context. A conversation that begins with restaurant recommendations can move into trip planning, then budget discussion and transportation logistics. Each transition changes which advertising could be useful. Reading only the opening topic would miss those shifts, while reading only the last prompt would discard the constraints already established. Conversational relevance therefore requires reassessment as the dialogue progresses.

As the system reassesses the conversation, it can use the intent stage to target ads. Research, comparison and decision can be distinguished through the depth and direction of the discussion. This also changes the privacy dependency. Ads can be relevant to the conversation at hand without persistent user identifiers or cross-site browsing history. Each conversation becomes its own context for targeting, though this design still leaves questions about privacy and measurement.

How do major platforms turn live prompts into ad placements as of late 2026?

ChatGPT, Google and Microsoft have made different choices about the relationship between the answer and the advertisement. ChatGPT Ads appear as labeled sponsored cards below the assistant’s answer. The commercial unit remains visibly separate from the response. Placement uses a relevance-weighted, second-price auction, so a more relevant ad can beat a higher bid. Precision can therefore give an advertiser with a modest budget an advantage over a less relevant advertiser spending more.

OpenAI announced an Ads Manager plugin on September 16, 2026, enabling campaign creation and updates through natural-language prompts. The announcement included AI creative suggestions with opt-in text customization and separate Shopify and HubSpot integrations, its first ecommerce and CRM partners. It also introduced a test of Sponsored Agents. In that test, users start a conversation with a brand’s agent after the click, extending the commercial interaction beyond the sponsored card.

OpenAI published a broader measurement update on October 5, 2026, that covered Pixel and Conversions API integrations, an Event Quality Score and a one-day view-through window. It also described early-stage geo-based incrementality experiments with Haus, Measured and WorkMagic, together with 10+ attribution partners and a new visual ad format. Not every feature originated with that update. The view-through window and Event Quality Score had appeared earlier. These are vendor-reported capabilities.

Google monetizes AI Mode and AI Overviews through existing Google Ads infrastructure. Search campaigns using broad match or AI Max, Performance Max and Shopping campaigns are eligible, with smart bidding required. There is no opt-in or opt-out control. Qualifying campaigns can already serve inside AI answers, even though most advertisers have not audited that participation.

Microsoft Copilot also automatically opts eligible campaigns into serving, with no advertiser opt-out and no guaranteed display. Ads are assembled from assets already in the advertiser’s account and appear below the organic response. Microsoft calls the connecting summary “ad voice.” It introduces how the ad section relates to the discussion, while session-level awareness of the whole conversation shapes relevance. The entire arc, rather than only the last prompt, can determine which advertisers fit.

Perplexity’s exit complicates any assumption that answer engines will converge on advertising. Fewer than a fraction of a percent of brands that applied were ever admitted. Access was exceptionally narrow before the exit. That decision makes conversational intent unavailable to advertisers without establishing that the intent itself has little commercial value.

Anthropic sells no ad placement in Claude, while the Gemini app and Meta AI have no ads within their chat interfaces as of October 2026. Meta instead uses Meta AI conversations to personalize ads across its other platforms. Those distinctions prevent a simple ranking by conversational sophistication. For advertisers, including B2B teams, platform availability and who can actually see a placement constrain reach before relevance can improve performance.

What the budget reality looks like and where most AI ad spend is actually flowing

Most AI ad spending in 2026 goes to traditional paid search listings that appear alongside Google’s AI Overviews, with less going directly into chatbot conversations. That distribution separates the market’s label from its operating mechanism. Much of the market is extending existing search infrastructure into an AI-shaped experience.

eMarketer’s 2026 forecast projects US AI search ad spend rising from a small base in 2025 to a much larger total by 2029, reaching a meaningful share of all search ad spend. The projected growth establishes the commercial direction. It does not establish that the underlying attribution gap has closed. The majority of near-term spending still runs through legacy search infrastructure, where the feedback system is behavioral and post-hoc.

Advertisers who enter AI Overviews through Google’s existing auction therefore face the same cookie-era attribution problems, even when their ads appear on an AI surface. Meanwhile, informational queries are increasingly resolved without a click, while action-oriented queries remain largely unimpacted. Informational and educational searches consequently become less efficient from a paid search perspective, as the answer can satisfy the user before a visit occurs.

This creates two distinct changes within the same market. AI answers alter whether a user needs to click, while conversational inference alters what an advertiser can understand before a click. Neither change guarantees the other. A conversational intent layer built for the task must read what the user is trying to do and the limits they face as the conversation develops, even when an existing search placement now appears near generated content.

What brands and agencies need to rethink about intent measurement now that the decision moment has moved

The practical implication for brands and agencies is a different definition of an intent signal. A behavioral record describes where someone has been, while a semantic reading describes what that person is trying to accomplish now. Those signals answer different questions. Treating conversational targeting as another way to construct an audience segment would miss the progression within the conversation that makes the signal distinctive.

That progression spans the full funnel. A broad, exploratory conversation can support an awareness message, while clarified needs make more targeted advertising useful during consideration. Threads at the decision stage can support placements aimed at conversion. The appropriate commercial response changes with the user’s reasoning, without requiring a cookie-stitched history to infer every transition. Intent is not limited to the final expression of purchase readiness.

The distinction is especially relevant in B2B software, financial services, education, travel, and health and wellness. In these categories, users seek detailed information and recommendations as they work toward a settled choice. For practitioners assessing targeting capabilities, the meaningful comparison is therefore whether a system reads the whole dialogue, recognizes changing constraints and distinguishes intent stages.

View-through windows, event quality measures and incrementality experiments are still being established within conversational interfaces. Brands that treat them as a solved attribution system risk misreading their results. A richer signal about current intent is not, by itself, proof that a placement caused a sale.

Conversational targeting can remove the dependency on persistent browsing identity for relevance, but it still needs credible ways to connect placements with outcomes. Its value lies in capturing reasoning that page-level tracking could not observe.