The AI Attention Stack: How Advertising Is Moving into New AI Surfaces

Artificial intelligence and advertising are starting to meet in places that do not look like ad inventory at first glance: an answer box that compresses search results, a chatbot that remembers the last question, a retail assistant that helps narrow a basket before checkout. That matters because advertising has usually scaled where attention, intent, inventory, and measurement could be made to line up. AI interfaces are now testing that alignment in three different ways, and it is too early to treat any of them as a finished media channel.

BCG’s AI Attention Stack is useful because it separates the market into three layers with different adoption curves and commercial logic: search-embedded AI, assistant-native AI, and retail or commerce AI. Search-embedded AI includes Google AI Overviews and Perplexity-style answer engines; assistant-native AI includes ChatGPT, Gemini, and Meta AI; retail commerce AI includes Amazon Rufus, Walmart Sparky, and Instacart Ask. The distinction is not cosmetic. Each layer changes where the consumer expresses intent, where a brand might intervene, and what kind of trust problem the ad product creates.[1]

The product signals are already visible. OpenAI is testing ads within ChatGPT for U.S. users, on a service reported to have approximately 400 million weekly active users.[2] Perplexity has tested sponsored follow-ups, a format that fits more naturally into answer refinement than into a traditional blue-link results page.[1] Walmart has opened advertising inside Sparky, its AI shopping assistant, while Google’s Universal Commerce Protocol points toward a world where retailer data can be activated outside a retailer’s owned environment.[1] None of this proves scaled performance. It does prove that the industry is no longer discussing AI surfaces only as discovery tools or productivity software.

The Stack Starts With Search, But It Does Not Behave Like Search

Search-embedded AI is the easiest layer for marketers to understand and the easiest one to misread. It inherits query intent from search, which makes it feel familiar. A consumer asks a question, compares options, looks for a recommendation, or checks a claim. The difference is that the interface increasingly resolves the query inside the answer itself. If the page of links becomes less central, the old bargain behind search marketing changes: ranking, paid placement, and content utility still matter, but they may be mediated by an AI answer that decides what is worth summarizing.

That is why AI Overviews and Perplexity-style products deserve attention before their ad models are fully settled. Reported pressure signals already suggest that answer engines can change traffic distribution: Improvado cites figures indicating that Google AI Overviews appear for 15% of queries, reduce organic click-through rates by 18% on average, and create declines of up to 47% for informational queries; it also cites 527% year-over-year growth in AI search engine traffic.[3] Those numbers should not be converted into a universal forecast for every category. They are better read as evidence that the search surface is being reallocated, especially for information-heavy journeys where the answer can satisfy a user before a site visit.

The advertising question in this layer is not simply whether brands can buy a slot. It is whether they can remain legible to systems that summarize, compare, and recommend. In paid search, a marketer could separate bidding strategy from landing-page content more than was healthy, but the interface still exposed the brand’s destination. In AI search, the content, product data, reviews, availability, authority signals, and structured information all become part of the substrate from which the answer is assembled. Sponsored follow-ups may become an inventory type, but the preparation work starts earlier than the media buy.

That creates an uncomfortable planning problem. Search teams are used to benchmarks. They know how to argue over cost per click, impression share, query match, and incremental conversion. AI answer surfaces are not yet mature enough to support that level of channel management across categories. A sensible search-embedded AI plan should therefore separate two tracks: controlled testing of available sponsored formats, and a broader visibility program that improves the quality, consistency, and machine readability of brand and product information. The second track will not look like a media plan, but it may determine whether the first one has anything useful to amplify.

Assistant-Native AI Turns Attention Into a Conversation

Assistant-native AI is the layer with the largest imagination premium and the least settled ad experience. ChatGPT, Gemini, Meta AI, and similar assistants are not just alternative search boxes. They are environments where users may ask for planning help, draft messages, compare tradeoffs, troubleshoot purchases, or return to a thread over time. That gives the surface potential commercial value, but it also makes interruption more delicate. A bad ad in a feed can be ignored. A bad ad inside a trusted assistant can contaminate the assistant.

The budget behavior is moving faster than the operating model. BCG cites Forrester research showing that 53% of organizations already allocate budget to conversational advertising as a distinct line item, and nearly three-quarters plan significant increases.[1] That is a serious signal, especially because it reflects budget architecture rather than a one-off experiment. Once a line item exists, teams begin assigning owners, asking for pilots, drafting measurement assumptions, and expecting vendors to explain where the money can go.

Still, assistant-native advertising has to solve a harder experience problem than search. Search advertising has long trained users to understand that some results are sponsored. Social advertising has trained users to expect branded content in a feed. Conversational assistants are being adopted partly because they feel useful, direct, and personal. If a recommendation appears inside that exchange, the user needs to know whether it is sponsored, why it appeared, and whether the assistant’s answer has been distorted by commercial pressure.

This is where trust stops being an abstract brand value. BCG reports that 69% of consumers feel manipulated when AI is used without disclosure, and that 70% identify certain data categories as off-limits.[1] Those figures put a boundary around the monetization story. The more conversational advertising depends on memory, personalization, and inferred intent, the more it needs visible disclosure, consent discipline, and rules for sensitive data. Otherwise the same capabilities that make the surface valuable for marketers make it feel invasive to consumers.

For now, assistant-native AI should be planned as an emerging environment, not as a direct replacement for paid search or paid social. Marketers can test where formats exist, but they should be careful about importing old success metrics too quickly. A sponsored suggestion inside a conversation may influence consideration without producing an immediate click. A brand mention may matter even when the assistant does not route the user to a site. A referral may be preceded by several prompts that current analytics cannot see. The measurement architecture will lag the behavior.

Retail Commerce AI Has the Cleanest Path to Revenue

Retail commerce AI is the most commercially legible layer because the consumer is already closer to a transaction. Amazon Rufus, Walmart Sparky, Instacart Ask, and similar assistants sit inside environments where product data, availability, price, basket history, promotions, and conversion events can be connected more directly than in open-web discovery.[1] That does not make the ad experience simple, but it does make the monetization path easier to explain to a retailer, a marketplace seller, or a brand funding retail media.

BCG reports that shopping-related GenAI use grew 35% in 2025.[1] The important part is not that every shopper now wants an AI assistant. It is that generative AI is becoming part of shopping behavior, not only part of work behavior or entertainment. When a shopper asks for a weeknight dinner plan, a skincare routine, a replacement part, or a comparison between two products, the assistant is operating near the point where preferences become purchases.

That proximity will attract budgets. Retail media has already taught brands to pay for access to commerce intent and closed-loop reporting. AI shopping assistants add a new layer: the recommendation moment can happen before the shelf is displayed. If a consumer asks Sparky for help choosing a product, or uses a retail assistant to narrow a category, the brand wants to know whether it can be included, preferred, explained, or compared fairly. The retailer wants to monetize that moment without making the assistant feel like a sales script.

Google’s Universal Commerce Protocol points to an even broader shift because it suggests retailer data may become useful beyond the retailer’s own walls.[1] If commerce identity, product feeds, availability, and transaction signals can move into AI-mediated shopping journeys, the boundary between retail media, search, affiliate, and assistant advertising becomes less tidy. That is a planning challenge for brands that currently manage those budgets, teams, and data contracts separately.

Retail commerce AI also changes the creative unit. A sponsored banner or product tile may still exist, but the more valuable asset may be the product explanation the assistant can use: ingredients, compatibility, claims, use cases, substitutions, constraints, and proof points. A retailer assistant cannot recommend what it cannot understand. Brands that treat product content as a catalog maintenance task will be slower than brands that treat it as conversational infrastructure.

What Marketers Can Prepare Before the Buying Interfaces Stabilize

The useful work now is not to declare a winner among ChatGPT, Google, Perplexity, Amazon, Walmart, Meta, or any other platform. The useful work is to build the muscles that will matter across the stack. Formats will change. Pricing will change. Reporting will be uneven. Procurement and legal teams will ask questions before vendors have clean answers. A good preparation agenda gives the organization a way to learn without pretending the market is more mature than it is.

Structured experiments should be narrow enough to teach something. A brand might test sponsored follow-ups in an answer engine where available, compare product-content changes across retail assistant surfaces, or run controlled creative variations designed for conversational prompts. The point is not to produce a heroic case study from an immature channel. It is to learn what can be measured, what cannot, which teams need to be involved, and what vendor claims fail under basic scrutiny.

Data hygiene is less glamorous and more decisive. AI surfaces need clean inputs: structured product feeds, consistent names, current pricing, clear availability, properly governed first-party data, documented claims, and content that answers the questions people actually ask. This work often sits across ecommerce, SEO, CRM, legal, product, and media teams. If it is treated as nobody’s job, the brand enters AI advertising with weak raw material.

Creative readiness also needs to change shape. The asset library for AI surfaces should include more than finished ads. It should include modular claims, comparison language, product education, FAQs, category explainers, imagery metadata, offer rules, and approved variations by audience or use case. Teams already building governed AI creative workflows will have an advantage here because they are learning how to produce more variations without losing control of claims, tone, or compliance. For a deeper operating model, Signal & Convert’s guide to governed AI creative advertising workflows is a useful companion.

Governance has to move closer to the media plan. Conversational advertising can blur lines that older formats kept separate: ad targeting, recommendation logic, personal assistance, and product advice. Marketing leaders need rules for when AI use is disclosed, which data can inform personalization, which categories require extra review, how sponsored answers are labeled, and when a human must approve claims or exclusions. Legal review at the end of a campaign will be too late if the assistant experience is already designed around questionable data use.

Measurement assumptions should be written down before pilots begin. In search-embedded AI, a valuable outcome may be inclusion in an answer, a qualified referral, or protection against visibility loss. In assistant-native AI, it may be assisted consideration, captured preference, or downstream conversion that cannot be attributed cleanly. In retail commerce AI, it may be recommendation share, basket influence, substitution defense, or incremental sales within a retailer environment. None of those measures is a universal KPI. They are hypotheses to be tested against the surface being used.

This preparation also has to connect to existing media infrastructure rather than sit in an innovation corner. Programmatic teams will need to understand whether AI surfaces create new supply, compete for budget, or demand different brand-safety standards. Search teams will need to work more closely with content and data teams. Retail media teams will need stronger product-content and commerce-data capabilities. Teams evaluating the broader marketing tool landscape can use Signal & Convert’s role-by-role guide to AI marketing tools and its analysis of AI in programmatic advertising to keep the new surfaces tied to current operating reality.

Use the Four Futures as a Planning Tool, Not a Forecast

BCG describes four plausible futures for the AI advertising market: Search 2.0, Agentic Commerce, Ambient Promotion, and Regulated Neutrality.[1] The value of that framework is not in choosing the most dramatic scenario. It is in forcing different operating questions.

If Search 2.0 dominates, the main work is defending and buying visibility inside AI-mediated discovery. Search budgets, SEO operations, content authority, and structured data become more tightly linked. If Agentic Commerce accelerates, the purchase path shifts toward assistants that can compare, select, and transact on the user’s behalf. Brands then need product data, commerce partnerships, and rules for how agents evaluate offers. If Ambient Promotion becomes more common, advertising appears more continuously across work, messaging, browsing, and commerce contexts, which raises frequency, consent, and brand-safety questions. If Regulated Neutrality becomes the constraint, platforms may face stronger limits on personalization, sponsorship, or data use, making transparent disclosure and compliant measurement more important than aggressive targeting.

A planning team can use those futures to pressure-test decisions that are already on the table. Which data investments still make sense if targeting becomes more restricted? Which content investments help whether the winning surface is search, assistant, or retail? Which partners have enough transparency to survive a more regulated environment? Which pilots would teach something useful even if the platform changes its ad format six months later? Scenario planning earns its keep when it changes today’s backlog.

The Next 18 Months Are for Capability Building

The temptation in every new media cycle is to ask how much budget should move now. For AI advertising, that is the wrong first question. The better question is what the organization needs to be ready to buy intelligently when the surfaces become more stable. Some budget should go to disciplined pilots, especially where a brand’s category is already exposed to AI search, assistant recommendations, or retail media competition. But a hype-driven reallocation before formats, pricing, reporting, and user expectations settle is not strategy.

The competitive advantage is likely to come from readiness that looks unexciting in the quarterly deck: cleaner product data, stronger content systems, faster creative variation under governance, clearer consent rules, better measurement hypotheses, and teams that know how to evaluate an AI surface without either dismissing it or overbuying it. That work compounds. When inventory becomes easier to buy, the brands that have already prepared their inputs and operating rules will move faster than those still debating ownership.

AI attention is becoming buyable unevenly: first as tests, then as native formats, then as platforms with reporting conventions and budget gravity. The next 18 months are not the moment to pretend the stack is mature. They are the moment to build the data, governance, experimentation, and scenario-planning capabilities that determine who can act when it is.

References

1. How AI Is Reshaping Advertising for the First Time in a Decade, BCG, Jan 2026.
2. How AI is disrupting the advertising industry, CNBC.
3. AI Overviews, organic CTR, and AI search traffic benchmarks, Improvado.