Meta introduced Muse, a shopping agent that surfaces products and carries out shopping steps inside its own apps, in early October. The interesting question is not whether an agent can complete a purchase. It is what the agent reads before it decides what to show. An agent does not browse a shelf. It queries a product graph, ranks candidates against an inferred intent, and presents a short list. Whatever is not machine-readable is, for practical purposes, not in the assortment.
Key Conclusions
This shifts the unit of competition from the impression to the citation. In a feed, a brand competes for pixels. In an agent-mediated answer, a brand competes for inclusion in a set of three to five candidates, and inclusion depends on whether the product data answers the question the agent is trying to resolve. A listing with a clear title, structured attributes, and a stated use case is legible. A listing with a poetic title and a lifestyle image is not.
The measurement problem follows immediately. If an agent recommends a product and the purchase happens inside a closed app, the retailer's own analytics may see a direct visit with no referrer, or no visit at all. Attribution models built for click-based funnels will systematically undercount agent-driven demand, and the resulting misallocation will look like a creative problem rather than a measurement problem.
What an Agent Actually Reads
An agent resolving a shopping intent typically consumes four kinds of signal: categorical attributes, constraints, availability, and evidence. Categorical attributes tell it what the item is. Constraints tell it whether the item fits, in size, compatibility, or price band. Availability tells it whether the item can be delivered in time. Evidence tells it whether the claim is credible, usually through reviews or structured specifications. A catalogue that is strong on the first and weak on the last three will be ranked but rarely selected.
Constraints are where most catalogues fail
Attribute completeness is a solved problem for large retailers and an unsolved one for everyone else. The failure is rarely the primary attribute, such as colour or size. It is the constraint attribute: the maximum load, the compatible model, the voltage, the material that cannot be machine washed. Agents ask constraint questions because that is how a shopper narrows a set, and a catalogue that cannot answer them gets filtered out before it is ever ranked.
Evidence is the tiebreaker
When two candidates satisfy the same constraints, the agent needs a reason to prefer one. Structured review summaries, verified purchase counts, and specification tables serve that role. Prose marketing copy does not, because it cannot be compared across candidates. This is why the same product with a specification table and a hundred verified reviews outperforms a better product with a persuasive paragraph, in agent-mediated discovery specifically.
Best Practices
Three moves make a catalogue legible to an agent, and all three can be executed by a merchandising team without new platforms. They are a constraint field audit, a review summarisation layer, and an attribution instrument for agent traffic.The first two change what the agent can see, and the third changes what the retailer can prove.
Audit constraint fields, not attribute counts
Pick the twenty queries that generate the most revenue and write down the constraint each one implies. Then check whether the catalogue can answer it as a filterable field rather than as text in a description. Most teams find that five or six constraints carry most of the demand and that two of them are missing entirely, which is a fixable problem in a single quarter.
Instrument agent traffic explicitly
Add a parameter to links that agents are likely to follow, and treat the resulting sessions as a distinct channel rather than lumping them into direct. Without that separation, agent-driven demand looks like brand strength and receives no incremental investment. The parameter does not need to be sophisticated; consistency matters more than coverage.
Common Mistakes
The first mistake is optimizing the product page for a human reader while the entry point has moved to an answer. A page can be beautiful and still be invisible, because the agent never renders it. The work that matters happens in the feed and in the structured data, before the page is ever requested.
The second mistake is assuming that agent traffic will be measurable with existing tools. Closed-app purchases often leave no referrer, and cross-device journeys break the session. Teams that wait for clean attribution before acting will wait through the entire learning period, while competitors iterate on feed quality with imperfect data.
The Feed as the New Shelf Plan
Merchandising teams are used to planning a shelf: which products face the aisle, which sit at eye level, which are grouped by use case. A product feed supports the same decisions in a different vocabulary. Field completeness determines whether an item can be considered, and field consistency determines whether it can be compared. The closest analogue to eye level is not placement but specificity, because the agent selects the item whose data most precisely matches the stated constraint.
This has a structural consequence for assortment. Wide assortments with thin data lose to narrow assortments with rich data, because the agent can only surface what it can distinguish. Retailers that have spent a decade expanding SKU counts may find that the marginal SKU now contributes negative value, not because it does not sell, but because it dilutes the feed's ability to answer constraint questions precisely.
Summary
Meta's Muse is a reminder that agent-mediated shopping moves competition from impressions to citations. What matters is whether the product data answers the question the agent is trying to resolve, which is a feed problem rather than a creative problem.The work is unglamorous and it happens before any page is rendered.
The practical sequence is to audit constraint fields, add a review summarisation layer, and instrument agent traffic as its own channel. Evaluate on inclusion rate in agent shortlists and on conversion from agent-attributed sessions, not on impressions, which an agent does not generate.
Data Sources
3. Gen Z holiday shopping trends
FAQ
What does an agent read before recommending a product?
A: A: Categorical attributes, constraint attributes, availability, and evidence such as structured reviews or specifications.
Why do constraint attributes matter most?
A: A: Agents narrow candidate sets by constraints, so a catalogue that cannot answer them is filtered out before ranking.
Will agent traffic show up in analytics?
A: A: Often not. Closed-app purchases may leave no referrer, so agent sessions need explicit instrumentation.
Does a wider assortment help?
A: A: Not necessarily. Thin data on many SKUs dilutes the feed's ability to answer constraint questions precisely.
Is page design still relevant?
A: A: It matters after selection, not before. The entry point has moved to the answer, which is built from feed data.
