Building a Product Feed for AI Shopping Agents

For twenty years, everyone in e-commerce built product feeds for humans. You wrote a title, set a price, uploaded an image, and hoped a person scrolling a results page saw it and clicked. The feed was plumbing. Nobody actually read it. That is changing, and it changes what a good feed even is.

I run growth and analytics for a B2B retailer, and I own the product feed end to end. Over the last year I replaced the paid software we used to manage it with an engine we build in-house, and lately I have been enriching that feed for a specific new reader: the AI agent. Not the agent I built to run my own operations, but the agent on the other side of the transaction, the one a buyer now sends to do their shopping. Here is what that shift actually looks like from inside the catalog, and what I am doing about it.

01What is agentic commerce, and why should a product catalog care?

Agentic commerce is when an AI agent, rather than a person, evaluates and recommends products on a buyer's behalf, and your catalog should care because the agent reads your product data directly instead of a human browsing your listing.

When someone asks an AI assistant for the best product for a job, the assistant does not scroll a page of results the way a shopper would. It reads the underlying product data across many catalogs, compares them, and hands back a short recommendation. The buyer often sees only the answer, not the shelf. That means the moment of truth moves upstream, from the storefront to the feed. Your product is either good enough data to make that shortlist, or for that buyer it does not exist. The catalog stops being an input to an ad and becomes the thing the decision is actually made on.

02How does an AI shopping agent read a product feed differently than a human?

A human forgives a thin listing and fills the gaps from the photo, but an agent reads only the structured data, so a product with missing attributes loses even when it is the better product.

This is the part that took me a while to internalize. A person looks at a picture of a door closer, reads a two-word title, sees a price, and their brain quietly supplies the rest: it looks heavy-duty, it is probably fine for a busy entrance. An agent does none of that. It cannot infer from a photo. It reads what is written down. If one feed says "door closer, $180" and another says "door closer, ANSI Grade 1, adjustable spring size 1 to 6, aluminum, ten-year warranty, fits doors up to 48 inches," the agent recommends the second one. Not because the product is better. Because the feed is. The richer feed answered the buyer's real question, and the thin one did not. Everything an agent knows about your product is what you told it in structured form.

03Does agentic commerce matter for B2B, or just consumer retail?

It matters for both, because the agent does not care whether the buyer behind it is a consumer or a purchasing manager, it only cares whether your data is structured enough to evaluate, though I will be honest that the B2B case is still unproven.

Almost every conversation about agentic commerce assumes a consumer, someone at home asking an assistant for the best product under a certain price. I do not run a consumer store. I run a B2B retail catalog with contract pricing, bulk SKUs, and trade buyers. And whether a B2B retailer wins in AI-driven discovery is genuinely unproven right now. Nobody has clean receipts yet. I am building for it anyway, and the reason is simple: the agent reading my feed does not care who is behind it. Machine-readable data wins the same way whether the buyer is a homeowner, a consumer's shopping assistant, or a procurement agent doing the sourcing a sales rep used to do. I am not betting on one consumer surface. I am betting that buyers of every kind increasingly arrive through an agent, and that the catalog which reads clean to a machine is the one that gets picked. The downside if I am early is that I have a richer feed, which helps everything else regardless. That is a bet worth making.

04How do you actually enrich a product feed for AI agents?

You add the structured attributes, richer descriptions, and question-and-answer content an agent needs to judge a product, and you emit it in the emerging agentic feed formats as another output of the pipeline you already run.

Concretely, enrichment means giving the agent the things a human used to infer. That is:

• Structured attributes, not prose. Material, grade, dimensions, compatibility, certifications, warranty, each as its own machine-readable field rather than buried in a paragraph.
• Descriptions written to answer questions, not to sell. An agent is trying to match a product to a need, so the description should state what the product is, what it fits, and what it is rated for, plainly.
• Product question-and-answer content, the specific things a buyer asks before purchase, answered in the data so the agent can lift them directly.
• The new conversational-attribute fields, which the major platforms began formalizing this year, designed exactly for how an assistant queries a catalog in natural language.

The important design decision is not to build a separate system for this. The enriched feed is another projection of the same nightly pipeline that already builds the feeds I run today. It rebuilds every night alongside the rest. New reader, same engine.

05Do you need a separate integration for every AI shopping channel?

No, you enrich one canonical feed and let it propagate, because the major surfaces already pull from the same source, so a single enrichment reaches more than one channel without a new integration for each.

This is the operator move that made the whole project a low-risk decision instead of a big one. The temptation with any new channel is to chase it with its own bespoke integration, and agentic surfaces are multiplying fast. But the leading feed-based surfaces I care about most already ingest from the same underlying feed. So I enrich once, at the source, and the enrichment shows up across the surfaces that read from it, no per-channel plumbing required. There are competing feed standards forming, one backed by the search-and-commerce incumbents and one pushed by the large model labs, and they are not yet identical. I am not trying to serve all of them perfectly on day one. I am leading with the feed-based surfaces I already reach, and treating the rest as fast-follows. And here is the part that de-risks it entirely: the same enrichment strengthens the ordinary paid-search feed the human buyers still use. There is no version of this where the work is wasted, proven for B2B or not.

06Is this about AI agents buying products, or something else right now?

Right now it is about discovery, not checkout, because the market pivoted from agents completing purchases end to end to agents deciding what to recommend, so the near-term goal is to be recommended, not to be bought from by a bot.

It is worth being precise here, because the hype skips it. A while back the story was that AI agents would handle the entire purchase for you, find it, add it to the cart, pay, done. That mostly has not materialized, and the industry has largely refocused on the earlier step: agents that research and recommend, with a human still completing the buy. That is the version that is real today, and it is the version I am building for. I want my catalog to be the one the agent surfaces when a buyer asks. Being recommended is the fight that is actually live, and it is upstream of everything else. If you are not in the answer, the checkout mechanics never matter.

07What is the risk of waiting until it's proven?

The risk is invisibility you never see, because if the agent cannot read your data when a buyer asks, you are simply absent from the answer, and you will never get an alert for the sale you were left out of.

Most competitive disadvantages announce themselves. Traffic drops, a ranking slips, a campaign underperforms, and you go investigate. This one does not. When an agent shortlists three products and yours is not among them because your feed could not answer the question, nothing shows up in your reports. There is no impression to explain, no click that got away. You just quietly are not considered. That is what makes waiting for proof risky in a way that feels safe: the cost of being unreadable is paid in demand you cannot measure because it never reached you. So I would rather have the enriched, machine-readable feed ready before the behavior is fully mainstream than scramble to build it after I notice the buyer already changed. The companies that win the next few years will not be the ones with the loudest ads. They will be the ones whose data an AI can actually read.