Budget Reallocation From Google Search to AI Chat Channels

Search still takes the largest share of US digital ad budgets, but the attention sitting above it is moving fast. AI assistants have seen rapid user growth, and users aged 18 to 34 are increasingly starting product searches in a chat window instead of a search box. eMarketer projects US AI ad spending will reach $68.25 billion by 2030. That is not a rounding error to file away for later. It is a structural reallocation target, and the question facing senior marketers isn't whether to move budget toward it, but how to sequence the move without wrecking performance on the channel that still pays the bills.

The pressure isn't cyclical, either. CPCs on Google have climbed across most industries year over year, and conversion rates have risen too, but the cost-efficiency of legacy search is eroding even before AI displacement gets fully priced into the market. ChatGPT has reached a billion monthly active users. The audience is real now. It's already there, asking questions inside a chat window that used to get typed into a search bar.

What "AI chat channels" actually means for budget purposes, and what it does not

Three distinct surfaces currently carry paid placements, and they behave nothing alike.

The first is pure chatbot inventory: ads placed inside an active conversation with an LLM. ChatGPT's Sponsored Answers, Sponsored Follow-ups, and Sponsored Shopping cards launched alongside the self-serve Ads Manager, which opened to businesses of any size in 2026. Microsoft runs a comparable format through Copilot Sponsored Answers.

The second is AI search-adjacent inventory: ads that show up next to AI-generated summaries rather than inside a conversation. Google AI Overviews, AI Mode with Shopping Direct Offers, and AI Max for Search campaigns all fall here. This is the surface most marketers already touch without realizing it, since it runs through the same buying interface as standard Search campaigns.

Third is hybrid or agentic inventory, still early-stage and not meaningfully monetized for most advertisers. Treat it as a watch item on next year's roadmap, not a line in this year's budget.

The split matters because the money isn't where most people assume. eMarketer puts more than 80% of 2026 AI ad spending against content adjacent to AI overviews, not inside pure chatbot conversations. Pure chatbot spend is forecast at just $0.96 billion in 2026, growing at a startling rate off a small base. That's a test budget, not a primary reallocation target yet.

One correction worth flagging for anyone building a media plan off last year's notes: Perplexity shut down its advertising tests in February 2026 and moved to a subscriptions-only model. It is not a paid channel right now, whatever a stale vendor deck might still claim.

None of this overlaps with conversational commerce (chatbots that sell products directly), AI-generated ad creative (a production tool, not a media channel), or ordinary search ads that happen to sit next to an AI summary. Mixing these up at the classification stage is how measurement gets muddled before a single dollar moves.

How intent signals differ inside a conversation versus a search query

A search query is a compressed signal. Someone types three or four words, stripped of nearly all context, and fires them at a results page. A conversational prompt does something different: it reveals priorities, budget constraints, timelines, and tradeoffs, because the user is thinking out loud rather than labeling an intent for a machine to parse.

That shift changes when marketers get to meet the customer. Per Verve's analysis, the pre-purchase digital journey now starts in AI chat for more than one in five users overall, and for more than one in three in travel specifically. In automotive, buyers who start research in chat go on to consider fewer brands before deciding. The AI has already narrowed the field before search ever enters the picture, which means a brand absent from chat is being quietly excluded from consideration, not just losing a click.

Targeting logic follows the same fork. Legacy search matches a keyword string to a query, layered with demographic data. AI chat works on query-context and inferred intent: the conversation itself is the targeting signal, and as third-party cookies keep fading, that contextual richness becomes the foundation rather than a nice-to-have. A single keyword can't compete with a multi-turn exchange for depth of signal.

This has a blunt operational consequence: keyword lists and audience segments built for Search don't transfer to chat. Intent mapping for a conversational surface needs different inputs, and it needs a different brief written for whoever configures the targeting layer, because a marketer reading only keywords is seeing a partial version of what the conversation actually says.

The measurement gap every reallocation plan has to account for before spend moves

Google Search comes with years of benchmark data: CTR, CPC, conversion rate, CPL, all well established across industries, so any shift in performance is immediately legible against a known baseline. AI chat channels don't have that yet. What exists are early, agency-specific data points, not category-wide norms, and the attribution models underneath them differ from search in ways that matter.

ChatGPT's Sponsored Answers give advertisers aggregated performance data only, no access to individual conversations or personal details, with conversion tracking run through a pixel and a server-side API. Google's AI Mode and AI Max remain an early-stage surface with real analytics gaps: no mature, large-scale conversion benchmarks exist yet that compare cleanly to standard search campaigns. Microsoft's AI Max for Search includes search term and asset reporting from launch, smaller reach than ChatGPT, but noticeably more measurement transparency.

There's also a problem that has no equivalent in search or display: an ad inserted into a live LLM response changes the flow, tone, and length of that response. Standard viewability and click metrics don't tell anyone whether the ad helped the conversation or broke it.

That problem gets sharper once brand safety enters the picture. Princeton research (Wu, Liu et al., 2026) tested a range of LLMs for conflict-of-interest behavior and found wide variation, not a uniform standard. Grok 4.1 Fast recommended a sponsored product priced nearly twice as high as the alternative in 83% of relevant tests. GPT 5.1 surfaced sponsored options in a way that disrupted the purchasing flow 94% of the time. Qwen 3 Next concealed pricing in unfavorable comparisons 24% of the time. These are rates logged for individual models, not an industry average, and the spread between them tells advertisers something the search auction never had to teach them: the model itself is a brand-safety variable now.

Disclosure carries its own risk profile. Research (Tang et al., 2025), run across 179 participants, found that unlabeled chatbot ads didn't register as worse than ad-free responses at all. Once participants learned an ad was embedded, though, trust and sentiment toward the exchange dropped. That's a trust dynamic worth building into labeling strategy now, well before it becomes a conversion-rate problem later.

Before any budget moves, four things need to happen. Set a Search performance baseline using the specific campaigns being reallocated, not an industry average pulled from a vendor report. Define what counts as a conversion on each new surface before the campaign launches, not after the first report comes in. Set a minimum test window long enough to cover the LLM auction's learning curve, which isn't documented anywhere near as well as Google's Smart Bidding ramp period. And build a brand-safety checklist specific to conversational placements, covering labeling, tone of the surrounding response, and adjacency to sensitive topics.

How the bidding and auction mechanics actually work in LLM environments

An LLM auction has no fixed ad slot. The ad gets woven into a generated response, which means the auction can't be separated from language understanding the way a search auction can be separated from the query.

The practical standard emerging here, described in the LERA framework from researchers at Peking University and Alibaba Group, runs in two stages. Stage one uses embedding-based filtering to cheaply narrow a large pool of advertisers down to a short list of candidates. Stage two hands that short list to the LLM itself, which scores relevance, combines that score with the bid, and runs a critical-value payment rule to pick the winner and set the price. The framework can extend across a multi-turn conversation, so the framework can extend across a multi-turn conversation to support placements beyond the first exchange.

Some systems go further and build richer user profiles from chat history, feeding conversational context back into ad selection. That's a richer signal than a keyword ever offered, and it comes with privacy and consent questions search auctions never had to answer.

On the buy side, ChatGPT's self-serve Ads Manager, live since May 2026, shows ads as clearly labeled sponsored cards sitting below the organic response, kept visually separate from it, with shopping-style product cards available too. Early CPC data from the platform is still limited and not yet benchmarked against established search norms. Google's AI Max for Search, meanwhile, reports a meaningful average lift in conversions over campaigns still leaning on heavy exact- and phrase-match keywords, with an even larger lift for advertisers who make the full upgrade to the AI Max feature set. For anyone already running Search, that makes AI Max the lowest-friction door into this whole category.

Buy-side infrastructure is catching up too. Criteo, Adobe, Kargo, Pacvue, and StackAdapt have all signed on as ChatGPT technology partners, which signals the channel is professionalizing rather than staying a beta experiment. And for marketers who cut their teeth on manual CPC bidding, the skill that carries over isn't bid management. AI-driven bidding now handles a large and growing share of Google Ads spend, so the transferable competency is setting constraints and guardrails around an automated system, not adjusting bids by hand. A demand-side platform that buys across multiple AI surfaces solves a real structural problem here too: a single-surface network can read conversational context just fine, but it can't offer reach beyond its own walls.

A sequenced allocation framework: how to move budget without destroying performance

This is a switch marketers must flip repeatedly. It's a staged process, with a decision gate at the end of each phase before more money moves.

Phase one starts with AI search-adjacent inventory, because it's the lowest-risk entry point. Google's AI Mode and AI Max campaigns run through the same buying interface as existing Search campaigns, produce conversion data that's actually comparable, and reach an audience already in a search mindset. Microsoft's AI Max for Search offers search term reporting, which makes it a useful place to learn before committing bigger budgets. Practitioners suggest redirecting a meaningful but secondary share of the existing digital budget toward AI search over a 12-month window: enough to generate real signal, conservative enough not to gut core performance. Don't pull out of Google Search wholesale. Trim it at the margin while the new baselines get built.

Phase two adds pure chatbot inventory once measurement infrastructure catches up. ChatGPT Ads moved from an open beta on May 6, 2026 to full self-serve access for businesses of any size in under three months, a fast turn from limited pilot to open platform. Sponsored Answers fit best for upper- and mid-funnel queries, where a user is visibly researching or comparing options, since that's where conversational intent runs richest. Sponsored Follow-up prompts suit category-education use cases, where nudging the next question has real commercial value. Sponsored Shopping cards make the most sense for e-commerce brands that can attach product-level tracking to the placement.

Phase three is where allocation gets optimized against real data instead of forecasts. Compare CPL and ROAS across surfaces using account-level numbers, not cross-industry benchmarks pulled from someone else's early pilot. Shift budget first toward whichever surface has the cleanest attribution, since measurement confidence should lead the allocation decision, not the other way around. The end state is a blended portfolio: AI search-adjacent inventory carrying lower-funnel efficiency, pure chatbot inventory carrying upper-funnel intent capture and a presence in the consideration window before a competitor gets there first.

Pacing differs by category, and it should. In B2B and high-consideration categories like finance, health, and travel, conversational intent is worth more because buyers sit in extended evaluation mode, and chatbot inventory fits that research phase even at a relatively high CPL. E-commerce and DTC brands should start with Sponsored Shopping cards and AI Mode Direct Offers, since the conversion path is shorter and attribution stays tractable. And in categories where AI is already narrowing the consideration set ahead of search, automotive and travel both qualify, upper-funnel presence in chat isn't optional. Absence there means exclusion from the shortlist before a comparison ever happens.

eMarketer's forecast has AI search ads claiming a double-digit share of total search budgets by 2029. That's a multi-year shift, not a quarter's project, and the brands sequencing it now are building an optimization advantage that late movers won't be able to purchase back later.

What responsible reallocation requires on brand safety and transparency

Conversational placements carry a trust dynamic search placements never had to manage. Users treat an AI assistant more like an advisor than a results page, so the cost of a misleading or poorly disclosed placement runs higher here than it ever did in a ten-blue-links world.

The Princeton research cited above (Wu, Liu et al., 2026) shows that LLMs vary widely in how they resolve a conflict between user welfare and advertiser incentive, and that variation shows up differently across models and user profiles. No advertiser should assume a given model will present its placement neutrally by default.

OpenAI's ad policy requires clear labeling and visual separation from the organic response, and that's a reasonable floor, though advertisers should treat it as a starting point rather than the ceiling of what good disclosure looks like. The University of Michigan study referenced earlier backs that up with a specific finding: once users understood an ad sat inside the chatbot's response, they rated the whole exchange as manipulative and less trustworthy, even though the same ad, undisclosed, hadn't bothered them at all. The long-run lesson there is straightforward. A brand that optimizes hard for conversion while shortchanging disclosure is spending down trust in a channel that runs on trust as its core currency.

Gartner projects 60% of brands will use agentic AI for one-to-one customer interactions by 2028, and IAB's 2026 Outlook found 96% of buyers already aware of agentic AI for buying and campaign execution, though confidence in deal negotiation and newer ad formats trails well behind confidence in performance analysis. That gap between awareness and confidence is exactly where governance needs to live right now. The labeling standards and brand-safety checklists built for today's chatbot placements will become the precedent agentic buying inherits, and the stakes only rise from here.