AI Attention Stack: How GenAI Reshapes Discovery, Ads, and Retail

Artificial intelligence is no longer a feature add‑on to digital advertising — it is remaking the very architecture of discovery, measurement, and monetization, and the Boston Consulting Group’s recent briefing frames that shift as the most important tectonic event in advertising since paid search first rewired the web.

Background / Overview​

For the past decade advertising operated on a familiar playbook: audiences navigate the web via links, feeds, and store pages; platforms intermediate attention; advertisers buy impressions, clicks, or conversions. That model is breaking down because large language models (LLMs) and agentic AI are collapsing discovery, evaluation, and transaction into single conversational moments. The result is an emergent AI attention stack — new surfaces where intent concentrates and new inventory forms — and with it a fundamentally different set ofising will be bought, measured, and trusted. BCG’s framing is direct: shopping-related GenAI usage surged in 2025 (BCG reports a 35% increase between February and November 2025), conversational advertising is already receiving budget allocations, and retailers and platform owners are quickly turning their assistants into monetizable surfaces. Those claims are echoed in BCG’s consumer research and its separate customer-insight publications.

---

The AI Attention Stack: surfaces, formats, and economics​

What counts as the AI attention stack?​

BCG and other industry observers describe three layers where attention — and therefore ad inventory — is consolidating:

• Search‑Embedded AI: search experiences that synthesize multiple sources into a single answer (for example, Google AI Overviews and other “AI‑mode” responses).
• Assistant‑Native AI: always‑available assistants that live in workflows and apps (ChatGPT, Gemini, Microsoft Copilot, Claude, Meta AI).
• Retail & Commerce AI: retailer‑owned agents trained on first‑party data (Amazon Rufus, Walmart Sparky, Instacart Ask) that convert intent upstream.

These surfaces don’t just re-skin old ad slots — they change the shape of inventory and the value of a single “impression.” Conversational sessions carry richer intent signals (budget, constraints, preferences, sequential queries), concentrating commercial opportunity into fewer, higher‑value moments. That changes economics for platforms and for advertisers: compute‑heavy AI platforms need sustainable monetization (ads look inevitable), while advertisers face a world where being selected by an answering model matters more than ranking for a keyword.

Three emerging ad formats​

BCG outlines (and the market is already testing) three distinct ad primitives inside conversational outputs:

• In‑answer ads — sponsor messages embedded directly into a synthesized answer when commercial intent is detected.
• In‑conversation ads — suggestions, sponsored follow‑ups, or contextual prompts that appear alongside a dialogue, sometimes even for non‑transactional queries (Perplexity’s sponsored follow‑up model is a clear example).
• Agentic ads — sponsored options surfaced when an AI agent takes on planning or purchase tasks (retailer agents inserting “sponsored prompts” or “click‑to‑buy” cards inside a shopping flow). Recent pilots at Amazon and Walmart illustrate this model.

These formats differ from search and display in a crucial way: they must feel like help. If the advertisement interrupts the adviser’s role or undermines credibility, user backlash is swift and measurable.

---

Discovery, measurement, and attribution: a new operating logic​

Search used to be a marketplace of links; now it can be a generator of answers. That shift has three practical consequences for measurement:

• Intent, not keywords — models cluster meaning and intent across phrasing. Brands must optimize product data structures, availability signals, and model‑friendly content, not just keywords.
• Fewer clicks, higher value — assistants can resolve needs without sending a user to a publisher or an e‑commerce grid, compressing the funnel. Attribution models based on clicks and last‑click rules break down and must be replaced with holdout tests, server‑side event logging, and conversion‑lift measurement.
• New pricing levers — pricing may shift from CPM/ CPC to intent‑weighted outcomes (pay per assisted action, pay per agent selection, revenue share for cited publishers). Perplexity’s Publishers Program and similar revenue‑share pilots show how platforms are seeking ways to compensate creators even as referral volumes decline.

Platforms know this economics story. Google has built mechanisms to let merchants and advertisers surface within AI Overviews; Perplexity is experimenting with sponsored follow‑ups and revenue sharing; major retailers have begun pilot ad programs inside their agents. These developments are not theoretical — they’re already live in early tests.

---

Retail media: from owned destinations to distributed agentic commerce​

Retail media was already the fastest‑growing ad category because of first‑party transaction data. AI accelerates that trend but reorients where value is captured.

• Traditional retail media monetized on‑site attention (grid placement, sponsored listings). Agentic AI shifts attention upstream — agents can make selections without sending users to grid pages. For retailers, the economics change: value is in participation within distributed agentic flows, not just pageviews.
• Platforms and the industry have responded with interoperability standards. Google’s newly announced Universal Commerce Protocol (UCP) is a material development: UCP is positioned as an open standard to let agents and merchants interoperate for discovery, checkout, and post‑purchase flows — a way to keep retail commerce fluid across ecosystems rather than locked inside any single retailer.
• Practically, retailers are already testing ad insertion within their agents. Amazon’s Rufus and Walmart’s Sparky have reported pilot aage statistics that validate the approach: Amazon’s CFO/CEO statements and Walmart sales presentations indicate that agent engagement boosts conversion rates and that sponsored prompts were tested as early ad formats.

The upshot for brands: grid placement matters less than being algorithmically discoverable and selected. That means clean SKU data, up‑to‑date availability, clear fulfillment options, and content that models can parse as evidence of relevance.

---

Trust, disclosure, and the credibility battleground​

BCG stresses what should be obvious to any marketer: trust is the currency in an advice‑like interface. Consumers treat assistants as advisors; poor disclosure or heavy‑handed monetization is experienced as manipulation.

• BCG cites consumer research showing rising unease: a large proportion of consumers identify certain categories (private messages, health, precise location) as off‑limits for AI‑driven targeting, and many report feeling manipulated when AI is used in advertising without disclosure. The precise percentages vary across studies and markets, but the pattern is consistent: users demand transparency and meaningful choice.
• Platforms are moving cautiously. OpenAI has publicly stated that it will test ads in ChatGPT for U.S. Free and low‑cost Go users with clear labeling and safeguards (no ads for minors or adjacent to sensitive topics), and it emphasizes user controls for personalization. Google and others are likewise experimenting with labeled inline sponsored recommendations. These early product statements show intent to balance monetization with trust — but the exact UX, frequency caps, and measurement semantics remain fluid.
• Regulation and industry standards are already converging on transparency. The EU AI Act contains provisions that increase transparency obligations for certain AI systems, and industry groups like the IAB have released disclosure frameworks for AI usage in advertising. Those developments indicate that disclosure and provenance will soon be formal obligations in some markets. Marketers who lead with clear labeling and opt‑in models will face lower regulatory and brand risk.

Caveat: headline consumer figures (e.g., “69% feel manipulated”) can vary by sampling and question wording. Treat single survey numbers as directional; the strategic takeaway is unambiguous — misaligned advertising inside advice interfaces is reputationally risky.

---

Four plausible market futures (and what each means for advertisers)​

BCG sketches four endpoint scenarios for how AI advertising could settle; reality will likely mix elements of each:

• Search 2.0 — Ads embedded into synthesized answers; advertising optimized around intent clusters rather than keywords. This rewards structured product metadata and model‑readable creative.
• Agentic commerce — Agents manage research‑to‑purchase flows; retail media shifts to influencing agent defaults and pre‑selection logic. Brands that control SKU truth and availability win.
• Ambient promotion — Commercial influence becomes contextual andwer discrete ad units and more algorithmic signal shaping (e.g., paid prioritization baked into ranking heuristics). Measurement and governance become harder.
• Regulated neutrality — Strong legal and platform guardrails enforce separation, disclosure, and limits on conversational targeting. The playing field looks more like traditional search with explicit labels and constrained personalization.

Each scenario demands different investments. The safe recommendation: prepare for all four by building model‑ready data and transparent governance, and move fast to test early wready behaving.

---

Practical actions for marketing and product leaders​

BCG’s call to action is operational: get access, experiment, and build the operating system to win in AI‑native media. Translating that into concrete steps:

• Foundation — signal hygiene
• Audit and normalize product and service metadata into model‑friendly schemas (structured titles, attributes, availability, price, returns, shipping). Agentictraints; missing fields exclude you.
• Centralize first‑party signals (CRM, loyalty, on‑site behavior) into a single truth layer teams can access for training and testing.
• **Experimentation — rapid, measurement‑drivy access to platform betas (Google AI Overviews, Perplexity publisher programs, retailer agent ad tests, OpenAI testing programs).
• Run holdout experiments and server‑side event capture to measure assistant‑driven lift (assistant treatments vs. control cohorts). Replace brittle last‑click rules with conversion‑lift testing.
• Iterate creative that augments conversation (concise, utility‑first copy, verified product cards, labeled sponsored follow‑ups).
• Platform & partner strategy
• Build interoperable commerce feeds (support UCP and similar protocols) so your assortment is visible across agentic ecosystems; prioritize channel parity for fulfillment and returns.
• Negotiate for transparudit rights with platform partners — if a platform cannot provide verifiable lift metrics, treat spend cautiously.
• Governance, privacy & disclosure
• Define a corporate policy for conversational data: what’s permissible to use for ad targeting, what’s off‑limits, and how consent is captured and logged. Build transparency into UX and creative.
• Adopt industry disclosure best practices now — even where not legally required — to minimize consumer harm and reduce regulatory friction. The IAB’s disclosure guidance is a timely baseline.
• Organizational design
• Collapse silos between brand, performance, retail, and product data teams. The assistant era favors organizations that can iterate creative, data, and activation in hours not weeks.

---

Strengths and near‑term opportunities​

• Higher signal-to-noise: conversational inputs deliver richer intent, reducing wasted impressions. That can drive better ROI if advertisers capture intent‑rich moments.
• New direct monetization paths: agentic commerce and in‑answer placements open fresh revenue models (sponsored prompts, agent defaults, revenue share for cited publishers). Perplexity’s Publishers Program and retailer pilots are concrete proof points.
• Creative and experience premium: brands that design utility‑first, clearly‑labeled conversational creative can leverage assistants as helpers, turning interruption into assistance.

---

Risks, unknowns, and governance failings to watch​

• Trust erosion: poorly labeled or deceptive advertising in a conversational advisor will produce disproportionate consumer backlash. Survey signals and academic work consistently show authenticity and disclosure determine response.
• Measurement opacity: assistants can close funnels without clicks, creating attribution black boxes. Independent holdouts will be needed.
• Concentration risk: if discovery concentrates across a few assistant providers, those platforms gain gatekeeper power with shifting placements and terms. Diversification remains important.
• Regulatory uncertainty: the EU AI Act and emerging transparency consultations make it likely that labeling/disclosure obligations will stiffen; U.S. policy may follow in specific areas (consumer protection, children, health). Marketers should design for compliance, not reaction.

---

Verification notes and flagged claims​

• BCG’s central statistics — for example, *shopping‑related GenAI use grew by 35% from Februaryare taken from BCG’s own consumer research and its Global Consumer Radar; BCG publishes the methodology behind that figure. Treat that number as BCG’s empirically measured trend rather than a universal industry constant; other firms report rising GenAI commerce signals consistent with BCG’s direction but not identical magnitudes.
• Platform product statements are time‑sensitive. OpenAI has announced plans to test ads in ChatGPT for U.S. Free and Go users (January 2026), with clear labeling and constraints; that rollout was announced and testing is described as imminent but not yet universal. The precise ad UX, targeting semantics, and scaling timeline remain subject to platform design choices and user response.
• Retail pilots (Amazon Rufus, Walmart Sparky) have been publicly discussed by company executives and reported in multiple outlets. Amazon executives have cited strong engagement and conversion uplift from Rufus; Walmart has run early sponsored‑prompt experiments in Sparky. These are live market facts, but the business models and revenue splits remain proprietary and variable by test.

Where exact numbers or internal splits are not public, the article flags those as proprietary or pilot stage and recommends treating vendor claims as directional until independently audited.

---

Conclusion — what leaders must do now​

The transition from feeds, links, and grids to synthesized answers and agentic assistance is already under way. The next 12–18 months will determine which platforms, protocols, and measurement conventions dominate. The organizations that win will be those that:

• move faster to experiment in live product betas;
• invest in model‑ready data and interoperable commerce interfaces (UCP/agent protocols); and
• adopt governance and disclosure norms that protect trust before regulation forces the rules.

This isn’t a simple media‑shift — it’s a systems re‑design: discovery is now a service the platform provides, and advertising must either be useful to that service or be excluded. Brands that prepare — by cleaning data, redesigning creative for conversation, and baking privacy and consent into every interaction — will not only be resilient; they’ll shape the rules of a new commercial layer where decisions, not impressions, define value.

---

---

Source: Boston Consulting Group How AI Is Reshaping Advertising for the First Time in a Decade