Who will own the decision layer?

The next phase of AI search is not just about whether a brand appears in an answer.

It is about whether AI helps shape the choice.

Search engines organised information. Marketplaces organised products. Social platforms organised attention. AI systems are beginning to organise decisions.

That shift is already visible, although it is less complete than some product announcements suggest. Google has launched an open commerce protocol and says direct UCP checkout now supports hundreds of thousands of brands and retailers across its surfaces. OpenAI is testing clearly labelled Sponsored Agents with select US advertisers after a user clicks an ad. Anthropic has promised an ad-free Claude while expressing interest in user-initiated commerce. [1][2][3][5]

These are not three versions of the same model. They are competing architectures for commercial influence.

Consumer behaviour is also developing unevenly. NIQ reported that 42% of an approximately 500-person monthly US sample had used at least one AI tool to shop in the past month. Only 5% reported using a fully autonomous agent to place an order. In a separate survey of 322 US consumers, Gartner found that 31% would let AI narrow household-supply choices, while willingness to let it make purchase decisions topped out at 11%. [7][8]

The direction is clear. AI is becoming more important before the transaction. Broad delegation remains much less established.

That makes the question more precise:

<strong>Who will control the layer between a consumer's need and the eventual choice?</strong>

The answer will determine which brands are considered, how commercial influence is labelled, who controls the customer relationship and what AEO practitioners will need to optimise.

<h2>First, define the decision layer</h2>

The <strong>decision layer</strong> is the system that interprets a need, assembles the relevant options, compares them and helps determine what happens next.

It may recommend a shortlist. It may introduce an advertiser. It may connect the consumer to a merchant. Under a more delegated model, it may execute a transaction within user-defined limits.

This is wider than search ranking and narrower than the entire transaction system.

<div class='not-prose my-8 grid gap-4 sm:grid-cols-2'> <div class='rounded-2xl border border-[#CAD7D3] bg-white p-5'><p class='font-mono text-xs uppercase tracking-[0.18em] text-[#5B45D6]'>01 · User mandate</p><p class='mt-3 font-sans text-sm leading-6 text-[#15303C]'>Who sets the goal, constraints, budget and approval rules.</p></div> <div class='rounded-2xl border border-[#CAD7D3] bg-white p-5'><p class='font-mono text-xs uppercase tracking-[0.18em] text-[#5B45D6]'>02 · Platform interface</p><p class='mt-3 font-sans text-sm leading-6 text-[#15303C]'>Who controls discovery, ranking, presentation and monetisation.</p></div> <div class='rounded-2xl border border-[#CAD7D3] bg-white p-5'><p class='font-mono text-xs uppercase tracking-[0.18em] text-[#5B45D6]'>03 · Payment authority</p><p class='mt-3 font-sans text-sm leading-6 text-[#15303C]'>Who holds credentials, verifies the mandate and manages risk.</p></div> <div class='rounded-2xl border border-[#CAD7D3] bg-white p-5'><p class='font-mono text-xs uppercase tracking-[0.18em] text-[#5B45D6]'>04 · Merchant responsibility</p><p class='mt-3 font-sans text-sm leading-6 text-[#15303C]'>Who owns inventory, order acceptance, fulfilment, returns and seller-of-record obligations.</p></div> </div>

No current protocol assigns all four layers to one party. Open standards such as UCP, AP2, A2A and MCP can connect systems, but openness does not guarantee neutral ranking, portable preference history or consumer ownership. [1][10][11]

The most likely outcome is therefore not one winner. It is a contest over different parts of the stack.

<h2>Four futures, expressed as editorial scenarios</h2>

<div class='overflow-x-auto not-prose my-8'> <table class='w-full min-w-[820px] border-collapse text-left font-sans text-sm'> <thead><tr class='bg-[#15303C] text-[#F7FAF8]'><th class='px-4 py-3'>Scenario</th><th class='px-4 py-3'>Editorial weight</th><th class='px-4 py-3'>Why it is plausible</th><th class='px-4 py-3'>Critical boundary</th></tr></thead> <tbody> <tr class='border-b border-[#CAD7D3]'><td class='px-4 py-3 font-semibold'>Consideration operating system</td><td class='px-4 py-3'>35%</td><td class='px-4 py-3'>Research and shortlisting are currently more prevalent than autonomous ordering.</td><td class='px-4 py-3'>Not a measured probability or market share.</td></tr> <tr class='border-b border-[#CAD7D3] bg-[#EAF1EF]'><td class='px-4 py-3 font-semibold'>Hybrid marketplace</td><td class='px-4 py-3'>40%</td><td class='px-4 py-3'>Discovery, labelled commercial surfaces and merchant-controlled transactions already coexist.</td><td class='px-4 py-3'>The layers have different incentives, owners and rollout states.</td></tr> <tr class='border-b border-[#CAD7D3]'><td class='px-4 py-3 font-semibold'>Personal buyer agents</td><td class='px-4 py-3'>20%</td><td class='px-4 py-3'>Protocols can support bounded, user-directed execution across systems.</td><td class='px-4 py-3'>Portability, liability and trust remain unresolved.</td></tr> <tr><td class='px-4 py-3 font-semibold'>Slower, fragmented adoption</td><td class='px-4 py-3'>5%</td><td class='px-4 py-3'>Tests, betas, geography limits and platform-policy differences are already material.</td><td class='px-4 py-3'>Fragmentation is a current condition, not a negligible residual.</td></tr> </tbody> </table> </div>

The weights are AEO Updates editorial judgements as of 24 September 2026. They are not measured probabilities, market shares or outputs from a calibrated forecasting model.

<h2>Scenario one: AI becomes the consideration operating system</h2>

<p class='font-mono text-sm uppercase tracking-[0.18em] text-[#5B45D6]'>Editorial weight: 35%</p>

In this scenario, AI becomes the default place to research, compare and narrow choices, but the consumer still completes most transactions elsewhere.

This is the closest extension of current behaviour.

NIQ's same-study contrast is useful. Forty-two per cent of its monthly US sample had used at least one AI shopping tool in the prior month. Five per cent reported using a fully autonomous agent to place an order. Gartner found a similar willingness gap between narrowing choices and surrendering the final decision. [7][8]

The NIQ and Gartner results do not use the same denominator and should not be pooled. They support only a directional interpretation: consideration-stage use is more prevalent than autonomous purchase delegation in the reviewed evidence.

Under this model, AI systems do not need to own checkout to exert commercial influence. They only need to shape the shortlist.

That would give conversational systems power over which brands are surfaced, which attributes become decision criteria, which evidence is treated as credible, which alternatives are compared and which trade-offs are made visible.

A brand excluded from that shortlist may never reach the merchant site, marketplace page or retail shelf.

For marketers, the primary contest would remain <strong>eligibility and preference</strong>. The brand must first be understood as a plausible answer. It must then provide enough evidence to be preferred for the relevant need.

This scenario would make <a href='/articles/what-makes-brand-show-up-ai-evidence-environment'>Machine Positioning</a> more commercially important. It is not enough for an AI system to know what a company sells. The system must also have credible reasons to associate the brand with the right customer, use case and decision territory.

<h2>Scenario two: AI becomes a hybrid marketplace</h2>

<p class='font-mono text-sm uppercase tracking-[0.18em] text-[#5B45D6]'>Editorial weight: 40%</p>

This is the leading scenario because parts of it are already visible.

The system combines independent answers or organic product discovery, clearly labelled commercial placements, optional brand conversations, merchant-controlled offers and platform-enabled checkout infrastructure.

Google's UCP illustrates the commerce layer. Google launched the protocol in January 2026 as an open standard spanning discovery, buying and post-purchase support. By September, the company said UCP enabled direct checkout for hundreds of thousands of brands and retailers across Google. That count does not tell us how many consumers used the flow or how much transaction value passed through it. [1][2]

Google also keeps important responsibilities with retailers. Its launch announcement said retailers remain seller of record. The merchant controls product information, order handling and fulfilment while Google controls important parts of the discovery and interface environment. [1]

OpenAI's Sponsored Agents illustrate a different commercial layer. After clicking an ad, a user can choose to enter a clearly labelled conversation with a business-sponsored agent. OpenAI says this conversation is separate from the original chat and from ChatGPT's independent answers. As of 16 September, the product was a test with select US advertisers. [3]

This is not paid inclusion inside an answer. It is a paid path adjacent to the answer.

A hybrid marketplace could therefore make commercial influence more conversational without making it invisible. The platform might recommend independently, display a sponsored route, allow the brand to answer follow-up questions and then send the user to a merchant-controlled or protocol-enabled transaction.

The central risk is incentive creep.

When the same interface controls discovery, advertising and the transition to checkout, the commercial value of influencing the shortlist rises sharply. Labelling and system separation become essential. Those boundaries will need to be measured, not assumed.

For AEO practitioners, this scenario means earned and paid systems converge operationally but should remain analytically separate.

<ol><li>Inclusion in an independent answer.</li><li>Preference within the recommended set.</li><li>Eligibility for a sponsored commercial experience.</li><li>Conversational readiness inside a brand agent.</li><li>Transaction readiness through structured commerce infrastructure.</li></ol>

One dashboard score will not explain all five.

<h2>Scenario three: personal agents act for buyers</h2>

<p class='font-mono text-sm uppercase tracking-[0.18em] text-[#5B45D6]'>Editorial weight: 20%</p>

This is the most transformative scenario.

Instead of relying on one platform's assistant, consumers use persistent agents that understand their preferences, constraints and history. The agent can work across retailers and services, requesting information, comparing options and executing authorised tasks.

Open protocols make parts of this technically plausible.

AP2 documents payment mandates that can bind authority to an amount, merchant, product, time or other condition. MCP provides a common way for models to access tools and data, while calling for explicit user consent. A2A is intended to help agents built by different vendors communicate. [10][11]

But interoperability is not ownership.

A consumer-aligned agent would require portable preferences and history, revocable permissions, transparent objectives, clear payment and liability rules, independent switching between providers and confidence that recommendations serve the user's interests.

Current evidence does not show that these conditions have been solved at scale.

Anthropic's position helps illustrate the distinction. The company says Claude will remain ad-free because it wants the assistant to act unambiguously in the user's interests. It also says it is interested in agentic commerce in which Claude could handle a purchase or booking on a user's behalf. [5]

That is a product philosophy, not a broadly available autonomous checkout. Anthropic's public commerce blueprint is reference code. Its shopping agent can research, compare, prepare a cart and hand the customer to checkout, but the reference implementation does not place orders or move money. [6]

If a personal-agent model develops, the key asset may be the trusted mandate rather than the largest marketplace.

Brands would then need to become legible to many agents, not just visible inside one platform. Structured data, interoperable product information, verifiable claims and consistent third-party evidence would become the commercial equivalent of being stocked by many retailers.

<h2>Scenario four: adoption remains slower and fragmented</h2>

<p class='font-mono text-sm uppercase tracking-[0.18em] text-[#5B45D6]'>Editorial weight: 5%</p>

Fragmentation deserves more than residual status because it is already present.

Products are announced in different states: open standard, developer documentation, pilot, beta, waitlist, gradual rollout and general availability. Geography varies. Platform policy varies. Anthropic rejects advertising inside Claude conversations. OpenAI is testing sponsored business conversations. Google is developing ads, brand agents, protocol-enabled checkout and merchant analytics across multiple surfaces.

Consumer readiness also varies by task.

Routine, bounded and reversible decisions may be easier to delegate. Subjective, identity-laden, medical or morally consequential decisions are likely to retain more human oversight. Peer-reviewed algorithm-aversion research supports this distinction, but not precise category forecasts. [12][13][14]

Regulation, payment liability, merchant economics and data portability may also slow convergence.

Under this scenario, AI still shapes decisions. It simply does so through a patchwork of assistants, search experiences, merchant tools and partial checkout integrations rather than one coherent agentic-commerce system.

For marketers, fragmentation is not a reason to wait. It is a reason to avoid building strategy around one platform's current interface.

<h2>What can be delegated first?</h2>

There is not enough evidence to assign credible 2030 percentages by category. A more useful framework is to group decisions by <strong>delegation suitability</strong>.

<div class='overflow-x-auto not-prose my-8'> <table class='w-full min-w-[760px] border-collapse text-left font-sans text-sm'> <thead><tr class='bg-[#15303C] text-[#F7FAF8]'><th class='px-4 py-3'>Delegation band</th><th class='px-4 py-3'>Illustrative decisions</th><th class='px-4 py-3'>Likely AI role</th></tr></thead> <tbody> <tr class='border-b border-[#CAD7D3]'><td class='px-4 py-3 font-semibold'>More suitable for bounded delegation</td><td class='px-4 py-3'>Repeat household replenishment and low-risk recurring purchases</td><td class='px-4 py-3'>Execute within explicit price, product and timing constraints.</td></tr> <tr class='border-b border-[#CAD7D3] bg-[#EAF1EF]'><td class='px-4 py-3 font-semibold'>Assist, compare and require approval</td><td class='px-4 py-3'>Travel, fashion, beauty, durable goods and everyday services</td><td class='px-4 py-3'>Research, shortlist, configure and prepare the transaction for approval.</td></tr> <tr><td class='px-4 py-3 font-semibold'>Least suitable for unattended delegation</td><td class='px-4 py-3'>Healthcare, complex financial choices and highly personal or morally consequential decisions</td><td class='px-4 py-3'>Support research and comparison while preserving stronger human review.</td></tr> </tbody> </table> </div>

These are not permanent category assignments. A decision can move between bands when the mandate becomes narrower, spending limits are lower, reversal is easier or human approval is added.

The practical design question is not simply, ‘Will consumers delegate?’

It is: <strong>what will they delegate, under what constraints, with whose credentials and with what ability to reverse the decision?</strong>

<h2>The strategic consequence: Machine Brand Equity</h2>

If AI systems increasingly influence which brands enter a consideration set, brand equity acquires a machine-readable dimension.

AEO Updates calls this <strong>Machine Brand Equity</strong>.

<blockquote><strong>Machine Brand Equity</strong> is the strength of a brand's association with a relevant need, audience or decision territory inside machine-mediated choice.</blockquote>

This is an analytical concept, not a validated accounting measure.

It extends traditional brand equity rather than replacing it.

Human brand equity asks whether people know, trust and prefer a brand. Machine Brand Equity asks whether an AI system can identify the brand, understand what it stands for, find credible evidence for that position and include it when the relevant decision is being made.

The two can diverge.

A famous brand may have weak evidence for a new use case. A smaller specialist may have unusually strong, current and third-party-supported evidence for a narrow customer need. In a machine-mediated shortlist, the second brand may be more competitive than its unaided awareness would suggest.

That creates a direct connection between brand strategy and AEO.

The brand must decide what it wants to stand for. Then it must build a sufficiently consistent evidence environment for machines and people to reach the same conclusion.

<h2>What marketers should do now</h2>

<h3>1. Separate visibility from decision influence</h3>

Measure whether the brand is mentioned, whether it enters the shortlist, how it is characterised, whether it is preferred and whether a transaction follows. These are different stages.

<h3>2. Build machine-readable commercial truth</h3>

Product attributes, availability, compatibility, pricing conditions, service policies and evidence should be structured, current and accessible. Inaccurate data becomes a brand problem when consumers must verify every recommendation.

<h3>3. Strengthen third-party evidence</h3>

First-party content explains the brand's intended position. Independent research, reviews, expert commentary, retail data and customer evidence help make that position credible beyond the brand's own site.

<h3>4. Define a narrow decision territory</h3>

Most brands will not own an entire category. They can still build a strong association with a specific need, audience or trade-off. Narrow, defensible positioning is easier to support than broad, generic superiority.

<h3>5. Prepare for multiple commercial architectures</h3>

Do not assume one platform or protocol wins. Build portable product information, stable entity identity, clear policies and evidence that can survive across assistants, marketplaces and merchant agents.

<h3>6. Govern delegation risk</h3>

If an agent can act, define the mandate. Spending limits, product constraints, approval steps, expiry, revocation and error handling should be explicit.

<h3>7. Track the scenario triggers</h3>

The editorial weights should change when the evidence changes. Watch for broad consumer availability rather than announcements, verified transaction volume rather than merchant counts, repeat use rather than trial, cross-platform portability of preferences and permissions, enforceable liability and payment standards, and independent evidence about ranking, trust and commercial influence.

<h2>The base case is not one winner</h2>

The most defensible 2030 view is a layered market.

AI systems are likely to become more important in discovery and comparison. Some will add labelled commercial experiences. Some will preserve ad-free models. Merchants will continue to control important transaction responsibilities. Payment providers will control credentials and risk. Consumers will delegate selectively.

The strategic mistake would be to reduce this to a contest between chatbots.

The larger contest is over the decision interface, the commercial incentives around it and the evidence machines use to construct a shortlist.

The brands that prepare now will not merely try to appear in an answer.

They will build the positioning, data, proof and transaction readiness required to remain eligible for the decision.

<h2>Editorial methodology</h2>

This article separates observed product status, observed consumer research and editorial scenarios. Product claims retain their platform, geography and availability status. Survey findings retain their sample, geography and behavioural stage. Results from incompatible studies are not pooled. The 40/35/20/5 scenario weights are AEO Updates editorial judgements as of 24 September 2026, not measured probabilities, market shares or statistical forecasts.

<h3>References</h3>

[1] Google. <a href="https://blog.google/products/ads-commerce/agentic-commerce-ai-tools-protocol-retailers-platforms/" target="_blank" rel="noreferrer">New tech and tools for retailers to succeed in an agentic shopping era</a>. 11 January 2026.

[2] Google. <a href="https://blog.google/products-and-platforms/products/shopping/google-shopping-updates-holiday-shopping/" target="_blank" rel="noreferrer">Boost your holiday sales with these agentic commerce updates</a>. 16 September 2026.

[3] OpenAI. <a href="https://openai.com/index/reimagining-advertising-with-ai/" target="_blank" rel="noreferrer">Reimagining advertising with AI</a>. 16 September 2026.

[4] OpenAI. <a href="https://developers.openai.com/commerce/guides/key-concepts" target="_blank" rel="noreferrer">Key concepts: Agentic Commerce</a>. Accessed 23 September 2026.

[5] Anthropic. <a href="https://www.anthropic.com/news/claude-is-a-space-to-think" target="_blank" rel="noreferrer">Claude is a space to think</a>. 4 February 2026.

[6] Anthropic. <a href="https://platform.claude.com/docs/en/about-claude/use-case-guides/commerce-agents" target="_blank" rel="noreferrer">Commerce agent</a>. Accessed 23 September 2026.

[7] NIQ. <a href="https://nielseniq.com/global/en/news-center/2026/42-of-consumers-now-use-ai-tools-to-shop-niq-data-shows/" target="_blank" rel="noreferrer">42% of Consumers Now Use AI Tools to Shop, NIQ Data Shows</a>. 5 May 2026.

[8] Gartner. <a href="https://www.gartner.com/en/newsroom/press-releases/2026-05-27-gartner-survey-finds-consumers-want-ai-shopping-help-but-not-ai-purchase-decisions" target="_blank" rel="noreferrer">Consumers Want AI Shopping Help, But Not AI Purchase Decisions</a>. 27 May 2026.

[9] Accenture. <a href="https://www.accenture.com/content/dam/accenture/final/accenture-com/document-fy26/q4/Accenture-Consumer-Pulse-2026.pdf" target="_blank" rel="noreferrer">Consumer Pulse Research 2026</a>. January 2026.

[10] Agent Payments Protocol. <a href="https://ap2-protocol.org/" target="_blank" rel="noreferrer">Documentation</a>. Accessed 23 September 2026.

[11] Model Context Protocol. <a href="https://modelcontextprotocol.io/specification/2025-06-18" target="_blank" rel="noreferrer">Specification</a>. 18 June 2025.

[12] Castelo, N., Bos, M. W. and Lehmann, D. R. <a href="https://journals.sagepub.com/doi/abs/10.1177/0022243719851788" target="_blank" rel="noreferrer">Task-Dependent Algorithm Aversion</a>. Journal of Marketing Research, 2019.

[13] Longoni, C., Bonezzi, A. and Morewedge, C. K. <a href="https://academic.oup.com/jcr/article-abstract/46/4/629/5485292" target="_blank" rel="noreferrer">Resistance to Medical Artificial Intelligence</a>. Journal of Consumer Research, 2019.

[14] Dietvorst, B. J. and Bartels, D. M. <a href="https://myscp.onlinelibrary.wiley.com/doi/abs/10.1002/jcpy.1266" target="_blank" rel="noreferrer">Consumers Object to Algorithms Making Morally Relevant Tradeoffs</a>. Journal of Consumer Psychology, 2022.