Demand Capture vs Demand Generation in AI Advertising Channels

Conversational AI advertising does something search and display never could separately: it catches ready buyers and shapes category awareness for curious browsers, in the same interface, sometimes in the same conversation. That dual capacity breaks the budget logic advertisers have used for two decades. I've watched three separate ad tech cycles promise to unify funnel stages and fail, and this one might actually pull it off, mostly because the mechanics finally match the claim.

What a user's prompt actually tells an advertiser about where they are in the funnel

A prompt carries hedging language, context from earlier turns, and a degree of uncertainty a search query rarely holds. Someone typing "best running shoes" into a search bar hands an advertiser almost nothing beyond the words themselves. That same question, asked of a chatbot after three prior turns about foot pain and marathon training, hands over a much fuller picture of where the person actually stands.

eMarketer looked at over a billion daily signals and found that more than 20% of pre-purchase digital journeys now start in AI chat instead of on a search results page, and in travel, that figure climbs to 37%. The brand-consideration numbers say something too: a user who begins in chat considers an average of 1.96 brands in the auto category, while a user who begins in search considers 4.21. The chat user shows up narrower and more focused, and that changes what an advertiser should say to them, and when.

An unbranded opening prompt, something like "what should I look for in a running shoe," signals category curiosity rather than brand intent, which puts it squarely in the upper funnel. A prompt naming a brand or asking for a side-by-side sits mid-to-lower funnel and needs a different kind of ad entirely. eMarketer also found users carry an average shortlist of 1.43 brands from search into a follow-up AI conversation, so that chat session often works as a closing moment rather than a discovery one. The AI answer becomes a sounding board for a decision that's already half made, a strange thing to sit with if you spent a career thinking of chat as a top-of-funnel channel by default.

Informational and transactional prompts appear in roughly even measure inside a single interface, across categories. Legacy search would have scattered those two audiences across separate channels entirely; conversational AI holds them together in one thread. Read correctly and in real time, the prompt itself is the funnel signal now.

How the mechanics of conversational AI advertising make simultaneous capture and generation technically possible

Search ads match against a fixed keyword list decided in advance. LLM ads match against the live content and trajectory of a conversation as it unfolds, which is a different targeting problem entirely. Platforms look at the current prompt, the turns before it, and the broader topic thread, then pull ad inventory from that, with no cookie required and no demographic proxy standing in for intent.

ChatGPT's rollout, launched in February 2026 across the U.S. Free and Go tiers, shows the model in practice. Ads appear in clearly labeled, subtly tinted boxes beneath AI responses, and targeting draws on conversational context, chat history, and how a user has responded to ads before. The underlying AI answer itself stays untouched. That separation is the whole trust proposition here, and it's worth sitting on for a second: mess that up, and the entire channel loses the thing that makes it different from a banner ad.

The bidding side closed the loop shortly after. OpenAI introduced cost-per-action bidding for select advertisers in late May 2026, so advertisers pay only when someone clicks through, signs up, or buys. That puts the channel on measurement footing search advertisers have leaned on for twenty years. A lower-funnel prompt can now trigger a performance ad tied to an actual conversion goal, not an impression logged and forgotten. Contextual matching handles the other half: an upper-funnel, unbranded prompt can trigger a brand-awareness message aimed at someone who arrived with no brand intent at all. Demand generation happens mid-conversation, while attention is already high, instead of during a scroll past a banner.

Researchers studying auction frameworks built on retrieval-augmented generation have found that bid value and contextual relevance score both determine placement, and a strong relevance score can offset a weaker bid. That rewards ads that actually fit the question asked, not just the ones paying the most. Speed matters here too, and not as a nicety, because a placement that lags breaks the conversational flow and reads as irrelevant the second it arrives late. Latency is a quality signal in this channel.

What early performance data shows about AI channels versus legacy search

Search still sets the baseline everyone gets measured against, and that baseline keeps climbing. Thrad, a programmatic ad marketplace built for AI chat interfaces, operates in this gap between legacy benchmarks and conversational inventory. 2026 Google Ads benchmarks across more than 13,000 campaigns show a 6.64% average click-through rate, an 8.18% conversion rate, and a cost per lead of $66.69. Fair or not, that's the ceiling AI channels get compared to.

The ground under search is moving, though. Average search cost-per-click rose 12% year-over-year to $2.96 in the first quarter of 2026, and CPCs climbed across 87% of industries in the most recent benchmark period. Paid click-through rate on AI-Overview queries fell from roughly 19.7% to 6.34% per independent measurement, and advertisers are paying more to chase a shrinking pool of clicks, a trend line that gets harder to defend the longer it continues.

Against that backdrop, the early AI channel numbers look strong, maybe suspiciously so. Microsoft Advertising reported, in data published August 2025, that Copilot delivered 73% higher click-through rates and 16% stronger conversion rates than traditional search, with customer journeys running 33% shorter. Criteo's aggregated data across 500 U.S. retailers in February 2026 found users referred from LLM platforms like ChatGPT convert at roughly 1.5 times the rate of other referral channels.

That lift tracks with the funnel clarity a prompt provides. A user who arrives through a high-intent prompt has already done a chunk of the work an advertiser used to pay for separately. Worth noting, though: most of these figures come from the platform operators themselves or from early-cohort advertisers who volunteered to be case studies, so they deserve the same grain of salt you'd apply to any vendor's first-year numbers. The generation side is harder to pin down still, since brand lift from an upper-funnel placement doesn't show up on a CPA dashboard, and the measurement infrastructure built to capture that value is playing catch-up to the ad product itself.

Why the same creative strategy doesn't work for both funnel positions inside a conversation

The common mistake is treating an LLM ad placement as just a new slot for old search copy: short, transactional, loaded with a call to action, regardless of where the triggering prompt sits in the funnel. That approach throws away what the prompt already told the advertiser.

An upper-funnel prompt means the user is still building a mental model of the problem in front of them. An ad opening with a brand name and a price tag shows up too early and reads as noise. What earns attention there is a message that helps the user think about the category differently, something that widens the frame instead of closing the sale. A lower-funnel prompt flips that: the user has already done the exploring, and any friction in the ad (too much brand story, no clear next step) costs the advertiser the exact moment they were waiting for.

There's a tonal constraint running underneath both cases, too. Ads inside AI responses live somewhere users experience as personal, closer to a dialogue than a page of results. An ad that reads like a banner or a keyword-stuffed headline feels wrong there no matter which funnel position it's aimed at. eMarketer's data suggests brand visibility inside LLM outputs has become its own strategic variable: a brand absent from upper-funnel chat responses loses its shot at the consideration window before the user ever reaches a purchase-ready prompt.

The practical answer is running at least two creative tracks, one built for exploratory and informational prompts, one built for evaluative and transactional prompts, with the platform's contextual matching deciding which fires. Early participants in ChatGPT's ad pilot have described the channel as a way to reach users at a moment no prior ad format could touch.

How to rethink budget allocation when one channel covers what previously required two

The 70/30 split between capture and generation budgets, per eMarketer, was built on an assumption: two separate populations, reachable only through two separate sets of channels, search and retail media for the ready buyer, CTV and display for the person who doesn't yet know they have a need. That made sense when search and display genuinely couldn't do each other's job.

That assumption gets harder to defend once one channel does both jobs at once. The real question shifts from which bucket gets funded to how much of the conversational AI spend is doing capture work versus generation work, and whether the creative and bidding strategy reflects that split or just defaults to old habits out of comfort.

Lean too hard into capture and an advertiser misses the earlier moment when the consideration set is still forming. A chat user who starts unbranded ends up considering only 1.96 brands on average, so early presence narrows the competitive field dramatically in the advertiser's favor. Lean too hard into generation, though, and an advertiser ignores the CPA infrastructure now live on ChatGPT, leaving measurable, attributable conversion volume on the table. Neither imbalance holds up given what's technically possible now.

The economics argue for shifting weight toward conversational AI generally. Legacy search CPCs rose 12% year-over-year while early AI channel benchmarks show real lift on both click-through and conversion. A balanced approach favors dynamic allocation, where the platform's own intent-matching infrastructure does the routing, over a fixed percentage split drawn up months in advance in a media plan. The measurement gap needs planning around directly, too. Upper-funnel AI placements need brand-lift or assisted-conversion attribution to show their real value, and any advertiser measuring the channel purely on last-click CPA will undercount the generation motion and pull budget away from it too soon.

What publishers inside conversational AI interfaces need to understand about monetizing both intent types

A chatbot session is a continuous, multi-turn dialogue where every turn carries a fresh intent signal, and the strength of that signal shifts turn by turn. Publishers haven't occupied a spot quite like this on the open web before.

High-intent, transactional prompts are premium inventory here. They draw performance advertisers willing to pay CPA rates, because the conversion probability is demonstrably higher, and Criteo's 1.5x conversion figure across 500 U.S. retailers is the kind of proof point that lets a publisher justify a higher CPM for that inventory specifically. Upper-funnel, exploratory prompts represent a different asset, one that brand advertisers in finance, travel, health, and e-commerce will pay for even without an immediate conversion attached, just to be present the moment a user starts forming a consideration set.

A flat, single-tier CPM floor treats all of that as identical inventory, and that leaves money sitting on the table that a smarter pricing model would catch. Publishers who can classify inventory by intent signal, transactional versus exploratory, can set differentiated floors and capture more of what advertisers actually want to pay. The catch: most conversational AI publishers don't have ad tech built for this yet. Legacy SSP infrastructure, designed around banner slots and pre-roll video, doesn't parse prompt-level intent and can't run the kind of relevance-weighted auction this inventory needs. OpenAI was processing 2.5 billion queries a day as of July 2025, and that scale, repeated across a growing set of AI applications, adds up to billions of daily conversational moments as inventory. Turning that volume into real revenue depends on publisher infrastructure built specifically for this environment, not retrofitted from the banner-ad era.

The remaining friction points that will define how fast this channel matures

Measurement is the biggest one. The generation motion, upper-funnel brand impact, doesn't fit cleanly into last-click or even assisted-conversion models built for search. Until someone solves that, finance teams will keep undercounting what the channel actually delivers, and budget will keep flowing toward whatever's easiest to measure rather than whatever's working.

Creative production comes right behind it. Serving both funnel motions well means building creative that responds to prompt signals in real time, and that's a production workflow most brand teams and their agencies haven't built. It's a genuinely new discipline, and a lot of agencies are going to find that out the hard way, probably around the third quarter they miss a performance target and can't explain why.

Trust matters more here than it ever did in search. Ads inside an interface users treat like a personal assistant carry a heavier disclosure obligation than an ad next to a search result, and the norms around labeling and transparency are still getting worked out. Industry observers broadly treat this as an open question, not a solved one. Layered on top is the interpretation problem itself: reading funnel position from a prompt is a guess dressed up in a probability score. "Best running shoes" could be idle curiosity, a gift search, or someone one click from checkout, and getting that classification wrong hurts both campaign performance and the user's experience of the product.

None of that changes the calculus for advertisers and publishers deciding whether to move now. The consideration set is still narrow (1.96 brands in chat-first journeys is a small field to compete in), the competitive landscape inside AI placements is still early, and CPA bidding just went live. Those conditions favor whoever shows up first, and that head start compounds before the channel settles into whatever normal looks like a few years out.

The 70/30 split between capture and generation was a workaround, built because search and display genuinely couldn't do each other's job. Conversational AI advertising removes that constraint. The purchase journey was always continuous, even when the tools available forced advertisers to treat it as two audiences waiting on two separate channels, and the tools just hadn't caught up yet. Now that one channel holds both motions at once, the old line drawn between them doesn't hold up anymore.