Agentic commerce

Agentic commerce just added more work to your brand. Who is responsible for doing it?

AI agents drove a sharp rise in shopper traffic last BFCM. The absolute numbers are still small. The growth rate is not. Couple that with what the models can now read, and the catalog work this creates is real. The question is who does it.

8 min read
Sticky-note reminders on a studio wall before peak trading week
Contents

The numbers are still small. The growth is not.

Most people in this industry have seen the recaps by now. AI agents drove a sharp rise in shopper traffic to ecommerce sites last BFCM. The absolute numbers are still quite small. The growth rate is not. Couple that with how fast the models themselves are improving, and it is obvious this changes how products get sold online. Not all at once, and not only in November. But the direction is not ambiguous.

Agents do not browse the way people do

The way an agent moves through the web is not a faster version of a person clicking around. Agents are much better at parsing large amounts of semi-structured data. They are worse at navigating a graphical interface, and worse at making sense of a mix of text, images and video.

Which means a lot of the work that made you findable to humans does not transfer. Merchandising by placement. Copy that leans on the photography to explain half the product. A size chart or care guide that only exists inside a PNG. Traditional SEO and online merchandising have to adapt to a sales channel that does not walk the site.

Data quality has always mattered. In the agentic channel it is critical. And it is no longer only a question of being correct.

Correctness is not the same as context

Agents can process a lot of detail, so the queries they run on a shopper's behalf tend to be more detailed and more nuanced than a search box. Waterproof jacket becomes a merino base layer that works at minus five and survives a machine wash. The catalog has to carry the context that lets an agent pick you over the next option.

You can see this in the specs already shipping. Google Merchant Center added conversational attributes so AI surfaces can answer product questions rather than guess from a title. OpenAI's product feed asks for the same kind of richness: question and answer pairs, related products, reviews. Fields that never lived in a description, and that a human browsing a product page could infer from a photo.

Agentic queries are more detailed than a search box ever was. Your catalog has to be, too.

This is a prohibitive amount of typing

For a lot of retailers, filling those fields by hand is not a project with a name on it. It is a wall. More attributes per SKU, in more languages, plus Q&A, related products, the details that currently live in a PDF or a PNG. Thousands of SKUs is thousands of rows of work. Nobody is hiring for that in October.

That is the new work the channel just added. Someone has to do it. The question is whether it is a person, SKU by SKU, or a system that can actually keep up.

The same agents can do the work

Here is the part that gets skipped. Agents are perhaps even more disruptive to how a company can operate internally than they are to the shopping journey. Managing a large product catalog is high-volume, structured work. That is exactly what they are good at.

There is a catch, and it is the same catch as on the storefront. For the same reasons agents struggle with a GUI and a mix of images and copy, they struggle with the way most brands actually run product operations: a spreadsheet out to an agency, back into Shopify, repeated for German. Traditional enterprise workflows are as unreadable to an agent as a merchandised homepage is.

A Claude seat is not how you get leverage

Most companies go about internal AI the wrong way. Handing every employee a Claude seat and asking them to use it more is not where the leverage is. You still have a human in the prompt-review loop, product by product, which is the same bottleneck with a better autocomplete.

The value is in identifying the workflows that can be fully managed by AI, and pulling people out of that loop. Enrichment, localization, feed fields, the conversational attributes the new channels are asking for. The model runs them. People steer.

That takes more governance than human-driven work, not less. You need a way to keep quality in bounds, to see what the model did, and to intervene when it drifts. If you get that right, a whole class of problems stops being a backlog. Data quality and richness just fall out of the infrastructure.

The real value is in workflows the model can run without a human in the prompt-review loop.

If the infrastructure is right, the catalog follows

Keep it in one place. One catalog where the fields, the brand rules and the languages live together, and which writes to Shopify itself. Then an agent has something it can actually operate on, instead of a pile of exports.

That is what we built Emfas to do. You decide the rules once: how the brand describes fit, tone and care, which fields each channel needs, how that reads in Swedish and in German. Descriptions, attributes, translations, metafields and alt text, generated on-brand and pushed into Shopify, plus the feeds the rest of your channels read. People set the bar and step in when something needs a steer. They do not review every SKU.

ICIW were blunt about the alternative, saying that without it they would “have to hire a new person to manage the tasks that it's automating.” Samsøe Samsøe went the other way and brought the whole thing in-house:

Emfas helped us bring product enrichment fully in-house and cut collection prep from 2.5 people to one person. No more translators or agencies.

Javier Artal Herbella, Marketing & Digital Director, Samsøe Samsøe

Get that in place before the first drop, not during it. Then the new fields, the new feeds, the new questions agents ask, are not a scramble in November. They are a byproduct of a catalog that already runs this way. We go deeper on the mechanics in your product feed is your AI storefront, and on the terminology in what is agentic commerce.

Sources

FAQ

The absolute numbers are still small. The growth rate is not. AI-referred traffic to US retail sites was up 670% year on year on Cyber Monday 2025, and the models themselves are getting better at reading product data at the same time. That combination is what changes how products get sold, even if agentic volume is not yet the majority of the basket.

More than a correct title and description. Agents run detailed, conversational queries, so they need the context that used to live in a photo, a size guide or a review: Q&A, related products, care, fit, use case. Google Merchant Center's conversational attributes and OpenAI's product feed both ask for this kind of richness, including question_and_answer, related_product and reviews.

A seat at every desk still leaves a human in the prompt-review loop, SKU by SKU. The leverage is in workflows that the model can run end to end: enrichment, localization, feed fields, with quality checks, an audit trail and a way to steer when something drifts. That is an operating model, not a chatbot rollout.

The work is cross-functional, but it should not be cross-system. Attributes, copy, translations and feeds all edit the same product. Keep one catalog that agents can actually operate on, with governance around it, and ownership stops being a standing agenda item because there is only one copy to own.

Not if you stop treating it as per-product typing. Define the workflows the agents will run, the quality bar, and how you intervene when they miss, and the catalog inherits it, including products you add during peak. ICIW backfilled their assortment to full coverage in hours rather than weeks.

Get the catalog work off your team's plate

Bring us your catalog and we'll show you which enrichment workflows can run without a human in the prompt-review loop, and how much of the gap is closed before peak week.

Book a demo