For twenty years, merchants learned to write product data for machines that rewarded brevity. Google Shopping wanted short titles, tight attribute sets, and clean taxonomies, so catalogs were compressed into the minimum viable feed. AI agents flip the requirement: they read like attentive humans, they answer conversational questions, and they can only sell what your catalog actually says. The new playbook is enrichment, not compression.
This extends our agentic-commerce readiness check, that piece asks whether an agent can find, understand, and buy from your catalog. This one is about the middle step: making the catalog worth understanding.
Why is the old feed playbook failing?
Because the queries changed. Google shared at its 2026 I/O and Marketing Live events that the average AI Mode query is roughly three times the length of a traditional search, that more than one in six U.S. AI Mode searches now arrive as voice or image, and that planning- and decision-style queries are growing meaningfully faster than AI Mode usage overall. Shoppers aren't typing "running shoes men" anymore. They're asking "which running shoes work for a heavy runner with flat feet who mostly does treadmill?"
A terse feed can't answer that question. The agent fielding it goes looking for pairing advice, fit context, use cases, and trade-offs, and it pulls those from whoever provides them. If your catalog only carries a title, a price, and six attributes, the agent either skips you or describes you using someone else's words.
The feed era rewarded merchants for saying as little as possible. The agent era rewards them for saying everything. In a structure machines can read.
What should an enriched catalog contain?
Think of it as four layers, each one machine-readable text on (or linked from) the product page and feed. Agentic-commerce analysts. Notably Scot Wingo, whose Retailgentic newsletter calls the catalog the "gold dust" of agentic commerce. Have converged on the same direction: expand every layer of product context.
| Layer | What it includes | What it wins |
|---|---|---|
| Identity & specs | Complete attributes: materials, dimensions, sizes, identifiers, category, price, availability | Being eligible for the answer at all |
| Conversational context | Pairings ("works with…"), occasions, who-it's-for, how-to-use, care instructions | The long, specific queries agents actually field |
| Social proof as data | Review themes and counts exposed as crawlable text, not a widget | Trust signals the agent can quote |
| Objection handling | Product-level FAQs: sizing, compatibility, returns, "is this right for X?" | The follow-up questions that close or kill the sale |
The classic mistakes still apply at every layer: content locked in images, tabs that only render on click, or JavaScript an agent won't execute is invisible. Everything above should exist in server-rendered HTML with structured data labelling what it is.
Where are the protocols heading?
The commerce protocols are standardizing exactly this enrichment. As of mid-2026, Google's Universal Commerce Protocol (UCP) defines expansion attributes for conversational commerce. Pairings, occasion suitability, links to documentation, and the Agentic Commerce Protocol (ACP) from OpenAI and Stripe asks merchants for a rich product feed in its own shape. The details and names will keep moving; the direction won't. Every protocol is a consumer of the same underlying asset: a verbose, structured, truthful catalog.
That's the practical reassurance. You don't have to bet on a winner. Enrichment work done once feeds whichever surfaces matter: AI Mode, ChatGPT, agent checkouts, and ordinary AI answers alike.
Where does the enrichment content come from?
From context you already have, not from a model's imagination:
- Your search logs: the exact phrases shoppers use, including the zero-result queries your catalog vocabulary doesn't cover.
- Cart and PDP behaviour: what gets abandoned and what shoppers compare, which signals missing information (size doubts, compatibility questions).
- Review themes: what buyers praise, complain about, and ask; the raw material for honest conversational attributes and FAQs.
- Support tickets and store-staff questions: the objection list, pre-written.
One warning: don't point a general-purpose LLM at a product list and let it invent the enrichment. A model's generic priors about your product. Often absorbed from forums and video transcripts. Will write confident, wrong context into your catalog, and you'll spend the next year correcting the record. Generate from your own data (reviews, logs, spec sheets), and review before it ships. The goal is a catalog that's verbose and true.
What to do next
Pick your ten most important products and read each PDP as an agent would: can it answer a long, conversational, "is this right for me?" question from the text alone? Fill the four layers where it can't, expose them in structured form, and then watch what the engines do with it, which answers you appear in, how you're described, and whether your own pages become the cited source. That feedback loop, across every engine, is exactly what Buffy Intel is built to report.