The Brand Truth Layer: Why Product Feeds Are Not Enough

The agentic commerce stack has a visibility problem that is almost solved and a recommendability problem that almost nobody is working on.

Getting products in front of AI shopping agents is fundamentally a plumbing challenge. Hard, but tractable. The Universal Commerce Protocol, backed by Google, Shopify, and a coalition of major retailers, creates a standardized agent surface for product catalogs. OpenAI and Stripe are building the Agentic Commerce Protocol in parallel. Both aim at the same outcome: agents can query product catalogs, check availability, and complete transactions without scraping storefronts or calling proprietary APIs. The infrastructure is being built. The discovery problem is being solved systematically.

The problem nobody is building toward is this: being present in an agent's catalog and being chosen by an agent are two different outcomes.

Feeds solve presence. Reasoning drives selection.

---

What agents actually need to make a recommendation

A shopping query handled by an agent involves reasoning: weighing signals, constructing a recommendation, deciding between products that share a category. Consider a simple jacket request: warm, good for travel. The work is evaluating fit against stated constraints, checking for social proof, handling a size-out scenario, and calibrating confidence in the output.

Product feeds are structured inventory containers: title, price, variants, tags, descriptions, availability state. They transmit that information reliably through the protocol. What they do not transmit is the reasoning that distinguishes two products priced comparably and categorized identically.

I spent time auditing six Shopify-hosted merchants through the UCP protocol surface, querying each with the same request. The protocol delivered cleanly. The pipes worked. But one signal was absent across every single merchant in the set: none of them exposes reviews or ratings through the agent layer. Every merchant whose own storefront pages render star ratings drops that signal entirely at the protocol boundary. Stars exist on the website; the agent session receives no trace of them.

That gap turns out to matter considerably. HBR's 2026 study, "Research: Traditional Marketing Doesn't Work on AI Shopping Agents" tested four leading language models across eight promotional mechanisms and roughly 16,000 simulated shopping rounds. The finding was pointed: star ratings were the only badge that consistently pushed agent selection upward across every model and every category. Discounts, free shipping, and endorsements were either neutral or counterproductive. Among the more capable models, GPT-5 and Gemini 2.5 Pro appeared to treat heavy promotional language as a quality signal in the wrong direction, an observed tendency rather than a universal rule, but one that suggests the direction of travel for frontier systems. Merchants are maintaining their star ratings carefully on their own storefronts and dropping them at the protocol boundary, the precise surface where they would matter most.

Star ratings are the most tractable social proof signal, the kind that maps cleanly to a structured field if a merchant decides to expose it. The structural issue runs deeper than that, though. Feed descriptions were written to rank in search results and satisfy a human who had already decided to click. A shopping session uses that same copy as its only substrate for reasoning, before any click, before any page view, to decide whether your item belongs in a recommendation at all. Same words, different reader, different job.

---

The five things your feed does not carry

Across the audit and across the merchants I have worked with building agentic catalog tooling, the gap between what feeds carry and what agents need reduces to five dimensions. Collectively, I call these brand truth.

Fit. Which customer is this item for, how does it run relative to size, and when is it the right pick over alternatives in the same catalog? A description written to rank on "women's linen blazer" gives the session nowhere to stand when the actual question is whether this blazer works for a petite frame that needs something that travels without wrinkling. Fit is authored reasoning about who a product serves and when, informed by the brand's actual experience with how customers use it.

Substitution. When the requested variant is out of stock, the agent needs a recommendation and a reason. Product feeds carry availability booleans. They carry zero substitution logic, which means the agent either abandons the session or guesses based on surface similarity. The brand's intended answer lives in neither outcome.

Policy. Returns, warranties, claim integrity, what the merchant will and will not promise. A brand that stands behind a lifetime guarantee on stitching wants that surfaced when comparing against a brand that offers a 30-day return window. Policy is a competitive differentiator in exactly the kind of comparison session agents run. The current feed schema has no field for it.

Routing. Which product belongs in front of which prompt? A brand with four styles of chore coat has authored opinions about which one leads for "casual weekend," which one for "workwear," which one for "gift for dad." Feed search does keyword matching against description text and tags. Routing is editorial judgment about product-intent pairing, and it lives only with the people who built the catalog. Agents have no access to it today because there is nowhere in the feed schema to put it.

Voice. How does the merchant explain themselves, and what do they decline to say? A merchant that built its identity around fair labor does not want to be described in the same register as one that did not. One sourcing from a single mill has a story that distinguishes it. Voice also covers restraint: claims the merchant will not make, comparisons it will not draw, language it has decided is off-brand regardless of what a session might surface from the description fields. This is the reasoning the business chooses to carry through every customer interaction, and the agent surface is a customer interaction.

None of these five dimensions appear in any product feed today. They require merchant authorship. No platform will build them, and there is a structural reason why.

---

Why the platforms leave this layer empty

The major commerce infrastructure companies are building shared pipes. UCP (Google, Shopify, Etsy, Wayfair, Target, Walmart) and ACP (OpenAI, Stripe) represent parallel protocol efforts aimed at the same outcome. The work is hard and genuinely valuable. Standardized agent-accessible catalogs will unlock a new class of shopping experiences that the current web cannot support.

Universal pipes carry standard schemas. What belongs in a standard schema is what every seller shares: title, price, inventory, variants, taxonomy. The fields that can be defined once and filled everywhere. A schema's value comes precisely from its uniformity, and that uniformity is structurally incompatible with carrying the variation that makes one brand's reasoning different from another's.

A given brand's fit logic applies to its specific products and the customers who buy them. Substitution rules for a denim company bear no resemblance to those for a home goods company, and neither resembles a cosmetics brand. Policies differ by category and by ethos. Routing differs by catalog shape. Voice differs deliberately, as an expression of identity. Shopify can ship a fit_notes field in the protocol schema; the content of that field for Gymshark is something only Gymshark can write, informed by years of product development and customer feedback that no platform has access to.

The reasoning that makes a merchant recommendable is essentially editorial work. It requires someone with genuine knowledge of the products, the customer base, and what the brand is substantively willing to promise. It accumulates across the catalog and gets applied to every query that touches it, across every protocol in operation. Infrastructure companies construct the infrastructure. Merchants write what travels through it.

That authored reasoning layer stays empty by default. Waiting for a platform to fill it means waiting for something that has no path to exist, because the platform cannot know what belongs there.

---

What the brand-truth layer actually is

The artifact: authored by the merchant, structured for machine consumption, versioned like code so it can evolve as products, policies, and positioning change.

It encodes those five dimensions in a format built for a session evaluating a query: a reasoning document shaped for a reader that arrives before any page view, before any click, and has only your catalog fields to work with. The narrative PDPs your copywriters wrote for post-click human readers are doing a different job. Conflating the two artifacts is how brands end up with protocol surfaces that carry their inventory and none of their thinking.

Consider what your current product copy was optimized for: search crawl, SERP ranking, human click, page read. That pipeline rewarded keyword density, scannability, and a CTA at the bottom. A shopping session running through UCP or ACP arrives through none of that, with no prior search, no click, no page rendering, just your catalog fields and whatever reasoning can be applied to them before the shopper sees anything. The system forms a recommendation from what it has, which right now is fields designed for a different pipeline entirely.

A brand-truth artifact gives the session a richer substrate: a reasoning document that travels through whatever protocol is in use and sits alongside the catalog data. The system can then represent your products the way you would represent them in a conversation with a knowledgeable customer.

I walk through a worked example on a real product page: what the catalog carries today, what a brand-truth artifact adds, and how the session's output shifts when the authored reasoning is present.

---

The moat is authorship

Every merchant on Shopify that enables the protocol gets a feed. It is the minimum required to appear on an agent surface at all. Reachability is the floor. Being chosen requires more.

The merchants who author their truth, who write down their fit logic, document their substitution rules, articulate their policies in machine-readable form, and decide how their voice should carry through a conversation with an AI, those merchants will be recommendable. The agent has something to reason with. The output comes shaped by the brand's own thinking rather than whatever can be inferred from a 146-character description and twelve merchandising tags. The description itself is the first lever, and the case for rewriting it as reasoning material rather than search copy is worth understanding before moving to the fuller artifact.

The merchants who ship warehouse defaults will be present in the catalog and invisible to the recommendation. They will be the results a system returns when it runs out of better options. In the catalog, but not chosen.

This window is early. Protocols are new, agent surfaces are still thin, and most brands have not yet thought about the shopping agent as a reader with specific needs. The merchants who start authoring now are building something that compounds. The reasoning gets richer. The sessions improve. The gap between authored and default becomes wider as every round of agent interaction trains users to expect the quality that authored reasoning produces.

If you want to think through what your brand-truth layer looks like in practice, I would like to talk. Sartorial is working with early partners on exactly this: authoring the reasoning layer that sits above the protocol, below the session, and between the brand and the agent. Request early access.

---

Sumit Jagdale is the founder of Sartorial.