An operating model for the agentic customer

Five connected layers—and a 90-day plan—for agencies preparing brands to be found, compared and chosen by AI

AI is changing how people find, compare and buy products. The response should not be a new dashboard sitting above the same disconnected teams.

An agent may need information from a product page, a structured data field, a merchant feed, a policy page and an independent source before it can help someone make a decision. A paid message may enter that journey, but payment does not guarantee organic selection. A technically perfect feed cannot rescue an unsupported claim. A strong campaign cannot fix a broken checkout.

The work crosses creative, media, SEO, commerce, experience design, data, legal and reputation. Agencies need an operating model that connects those disciplines without pretending they are the same job.

Here is one practical model. It has five layers: promise, proof, product truth, path to action, and performance and governance.

1. Promise

Begin with the value for the customer.

What need does the offering meet? Who is it for? Under what conditions is it useful? When is it a poor fit? What makes it meaningfully different from the alternatives?

These are familiar brand questions, and they remain essential. Agents act on human goals. If the agency starts with machine visibility instead of customer value, it may make an irrelevant product easier to find.

The promise should be clear enough to guide both creative work and product information. “Designed for small kitchens” may be a useful brand promise. It becomes actionable when the team also knows which dimensions, installation conditions and use cases support it.

The output of this layer is a short decision brief: the customer need, the audience, the circumstances of use, the reasons to choose and the reasons not to choose. It should be specific enough that another team can test whether the product and its evidence match the promise.

2. Proof

Turn important claims into evidence that can be checked.

Create a claims register. For each objective claim, record the exact wording, the supporting evidence, its limits, the markets where it applies, the owner and the next review date. In the United States, the FTC says advertisers should have a reasonable basis for objective claims before publishing them.¹

Start with claims that are likely to affect comparison: performance, safety, compatibility, durability, price, availability, delivery, returns, sustainability and support.

The aim is not to remove judgment or emotion from creative work. It is to know which statements are brand expression and which make a factual promise. “Made for everyday adventures” is different from “waterproof to 30 metres.” The second statement needs evidence with a defined test and boundary.

Proof should also be easy to retrieve. A certification hidden in a PDF, a warranty that contradicts the product page or an undated sustainability claim may exist, but it will be difficult for a person or system to evaluate confidently.

The output is not a larger legal disclaimer. It is a maintained record connecting each important claim to current evidence.

3. Product truth

Maintain one reliable product record.

Identifiers, specifications, options, price, inventory, shipping, returns, warnings, support terms and product media often live in different systems. When those systems disagree, the discrepancy becomes part of the customer experience.

Publish the same current facts on product pages, in supported structured data and through the feeds used by relevant platforms. OpenAI says its shopping system considers product and merchant information, including factors such as availability, price and quality. Google’s product guidance similarly relies on structured details such as price, availability, shipping and returns.²

This does not mean every channel must have identical copy. It means the underlying facts must agree. If the product page says an item is in stock while the feed says it is unavailable, that is a data problem, not a wording problem.

Give each field an owner and a source of record. Define how quickly changes must reach each channel. Monitor the facts that decay fastest, especially inventory, price, delivery estimates and policy terms.

The output is a product-truth map: what information matters, where it originates, where it is published, who owns it and how inconsistencies are resolved.

4. Path to action

Test whether a person—and an authorized agent—can complete the task safely.

The journey may include product selection, configuration, account creation, consent, payment, delivery, cancellation and support. Critical buttons, menus and forms need clear labels and predictable behavior. Totals, permissions, errors and cancellation paths should be visible rather than implied.³

The same principle applies to protocol-based commerce. Emerging specifications are trying to standardize how agents, merchants and payment systems exchange product and checkout information. They are still developing, so an agency should adopt one when it solves a real distribution or transaction problem, not because the acronym is fashionable.⁴

Keep the control points explicit. Who authorizes the purchase? Who is the merchant of record? Who handles payment, fraud, tax, fulfillment, returns and disputes? What record shows what the user approved?

The output is a tested journey with documented permissions, responsibilities and failure paths. A successful demonstration is not enough. The team should also know what happens when inventory changes, payment fails, a user withdraws consent or the agent cannot complete the task.

5. Performance and governance

Measure different kinds of influence separately.

At minimum, distinguish four outcomes:

Retrieval: Can the system find and understand the relevant information?
Organic inclusion: Does the brand or product appear in an unpaid answer or comparison?
Paid exposure: Was a clearly identified advertisement shown?
Business outcome: Did the interaction contribute to a lead, sale, repeat purchase or other useful result?

Do not combine them into one unexplained “AI visibility score.” A high citation rate is not the same as a high recommendation rate. A recommendation is not a sale. A paid placement should not be reported as earned preference.

Measurement is still incomplete. Some platforms provide referrals or reporting for AI-driven appearances, but an answer can influence a decision without producing a click.⁵ Use platform data where it exists, then pair it with controlled testing, referral data, customer research and business outcomes.

Save a representative set of customer questions and resulting answers. Review them over time for factual errors, missing products, improper claims, competitor inclusion and paid disclosure. Treat those examples as observations, not a permanent ranking formula.

Governance needs named owners. Assign responsibility for claims, product feeds, crawler policy, agent testing, ad disclosure, consent and incident response. Decide who can correct an error and how quickly they are expected to act.

The output is a scorecard with separate measures, a repeatable test set and a clear escalation path.

A practical 90-day plan

An agency does not need to predict the final shape of agentic commerce before beginning useful work. It needs a bounded first engagement.

Days 1–30: map and observe

Choose one category or customer decision. Audit ten to twenty important products or services rather than the entire catalog.

Map the questions a customer might ask, the facts needed to answer them and the systems that hold those facts. Capture current results across the relevant AI and search platforms. Record factual errors, omissions, competitor inclusion, citations and the distinction between organic and paid appearances.

At the same time, identify the owners of product data, claims, policies, media, analytics and checkout. Most early problems will be failures between teams rather than failures inside one tool.

Days 31–60: repair the foundation

Resolve conflicting specifications, prices and policies. Strengthen the evidence for important claims. Improve page structure, supported markup and the feeds that matter for the selected category.

Make the customer journey clearer. Fix ambiguous controls, missing totals, inaccessible forms and weak error paths. Confirm who is responsible at each transaction and consent point.

If paid placements are available, create messages around distinct, specific benefits. Keep them clearly separate from organic recommendations.

Days 61–90: test and establish ownership

Repeat the original questions and compare the results. Test retrieval, organic inclusion and paid placement as separate layers. Connect referral traffic to later customer behavior, while recognizing that some influence will occur without a click.

Publish the first scorecard. It should show what systems can find, what they can verify, what they present and what customers eventually do. It should also show open data problems, claim risks and accountable owners.

End the engagement with a prioritized backlog and review schedule. The goal is not to declare the brand “optimized for AI.” It is to establish a working system that can keep improving as products, evidence, platforms and customer needs change.

The agency’s role

The opportunity is larger than a new media product and more disciplined than a promise to “get recommended by AI.”

Agencies can connect the customer promise to the evidence, the evidence to the product record, the product record to the journey and the journey to accountable measurement. That coordination is valuable because no single dashboard or department can do it alone.

The standard should remain simple: give people a clear promise, give systems reliable information and keep paid influence visible.

Notes

Sources and further reading

Want to talk this through for your clients?

The first conversation is free.

Book a Scoping Sprint

More research

The next customer is an agent7 min
AEO for people who already do SEO8 min