Introduction
Earlier this year we published a series of articles on agentic commerce, covering what it is and why it matters for consumer and business channels, how retailers can get ready for agents, and how generative engine optimisation (GEO) shapes the way brands appear in AI answers.
The market has moved quickly since then, with Stripe's Agentic Commerce Suite now gearing up for launch to businesses across Europe, the Middle East and Africa (EMEA), following Stripe's commitment at Stripe Tour London that UK businesses would be able to sell inside AI interfaces this year.
In October, Sierra announced the Personal Agent Protocol, an open standard developed with Meta and partners including Shopify, Stripe and Walmart for how personal agents identify themselves to businesses and act on a consumer's behalf, and has since partnered with Stripe so that its agents can take card payments during customer calls.
With retailers now entering the golden quarter, the most important trading period of their year, it feels timely to share the observations we are seeing in the market and the practical lessons we have learned from working with retailers on agent readiness.
Over recent months we have been testing premium retail catalogues for agent readiness, running real product data against agentic commerce specifications and looking at how products appear when shoppers ask AI assistants for recommendations. The same set of issues came up again and again, and almost none of them were about the payment step itself. They were about whether an agent could be certain which product it was looking at.
For most retailers, the work required is a data programme that sits upstream of every agentic channel they may choose to support, whichever assistant or payment provider they work with.
What retailers are telling us
Many retailers have explored building their own first party shopping agents, and those conversations increasingly turn to the cost of running them. Every interaction with an agent consumes tokens, and in a sector with tight margins and very high volumes of browsing, much of which never converts, the cost per conversation can be difficult to reconcile with the value of an average basket. Retailers are understandably cautious about taking on an operating cost that grows with every session before the uplift in conversion has been proven.
As a result, we are seeing far more openness to third party deployment models, in which the retailer makes its products discoverable and purchasable inside assistants such as ChatGPT and Gemini and treats those assistants as transaction surfaces. In this model the assistant provider carries the cost of running the conversation, and the retailer can focus on supplying accurate product data, a dependable checkout and a strong brand presence wherever its customers are already asking for advice.
Across these conversations it has become increasingly clear that data quality decides whether a retailer is discoverable at all, and discoverability is what most retailers want above everything else as they approach peak. An assistant can only recommend and sell a product it can identify with confidence, which brings the discussion back to the quality of the product catalogue.
How an agent shops on your behalf
An agent acting for a shopper has to be certain it has found the right product before it can do anything useful with it. A typical request might be to find a particular serum in the 50ml size, check who has it in stock and buy it at the best price with next day delivery.
To carry that out, the agent needs to work through a sequence of steps, each of which relies on the one before it.
Interpret the request and identify the exact product, size and variant the shopper means.
Find that same product across retailers and brand sites, and be confident each listing refers to the identical item.
Compare price, availability, delivery and returns on a like for like basis.
Complete the purchase through an agentic checkout and payment flow.
A person browsing a website can resolve ambiguity with their eyes, a photograph and some common sense. An agent works from structured data and text, and where that data is vague or inconsistent it will either compare the wrong things, lower its confidence and move on to a retailer whose data is clearer, or present the shopper with a recommendation that turns out to be wrong. Payment is the final step in this sequence, and it can only succeed if the earlier steps are built on reliable product data.
One description used for many products
The most frequent issue we have found is where a single product description is reused across several different products. This typically happens in ranges where a brand supplies one block of copy for a whole collection, or where a merchandising team writes once and copies the text across shades, sizes, fabrics or fits to save time.
On a product page this rarely troubles a human shopper, because the photograph, the colour swatch and the size selector make the difference obvious in most instances. However, for an agent, the description is often the main evidence it has about what a product is, so ten handbags or six moisturisers that share identical text look like the same item, or like near duplicates that cannot be told apart with any confidence. This in turn creates indecision in the agent and if it can’t complete its main objective, it will either stop and move on or surface a sub-standard response back to the customer.
These consequences are practical and measurable.
Agents can merge distinct products into one, recommend the wrong variant, struggle to answer specific questions such as whether a jacket is lined or a cream suits sensitive skin, and in some cases favour a competitor whose listing describes the exact product more precisely. If customers reinforce that they are satisfied with the response, the agent/LLM will store this for future reference and surface the recommendation again in the future. Therefore, retailers must treat every sellable product as needing its own description that states what is distinctive about it, with shared range copy kept as supporting context rather than the primary text.
No consistent identifier for the same product
The second common finding was the absence of a unique identifier that follows a product consistently wherever it is sold. Many catalogues rely on internal stock keeping unit (SKU) codes that mean nothing outside the business, and where global identifiers such as Global Trade Item Numbers (GTINs) or European Article Numbers (EANs) exist they are often missing for a proportion of the range, attached at the wrong level, or different from the codes the brand itself publishes.
This matters because comparison is one of the main reasons a shopper asks an agent to help. To check whether a fragrance is cheaper elsewhere, or whether another retailer has a dress in stock in a particular size, the agent has to be sure that the listing on your site and the listing on a brand site or competitor site refer to exactly the same item. Without a shared, reliable identifier the agent falls back on matching titles, images and descriptions, which is slow, error prone and especially weak when descriptions are shared across products.
The pattern is most visible in premium and concession led retail, where product data arrives from many brands in many formats. We therefore suggest that retailers should make it a rule that every sellable variant carries a valid global identifier at variant level, that the identifier matches what the brand publishes, and that the same code is used consistently across the website, product feeds, marketplaces and any agentic commerce catalogue.
Through our work in this space, we have found other gaps that can undermine agent confidence. Alongside descriptions and identifiers, a consistent set of further gaps appeared across the catalogues we reviewed. They are;
Attributes held in free text. Size, colour, material, ingredients and fit were often written into titles or descriptions rather than held as separate, normalised fields, so an agent cannot reliably filter or compare on them.
Unclear variant structures. Parent and child relationships between a product and its sizes or shades were inconsistent, which makes it hard for an agent to know which variant a price or stock level belongs to.
Stale price and availability. Feeds updated daily or less often meant an agent could recommend an item at a price or stock level that had already changed, which damages trust at exactly the moment of purchase.
Inconsistent taxonomy. The same type of product appeared under different categories depending on the brand that supplied it, which weakens how agents group and recommend products.
Fragmented sources. Product content, pricing, inventory and imagery were held in separate systems, including brand supplied concession data, with no single view showing which source each field came from.
Build a data layer and semantic model for agents
There are steps retailers can take to optimise their site for agentic readiness. One key step Retailers should take is to design a product data layer whose primary audience is an AI agent, with the website treated as one of several presentations of that data. Most product data today has been shaped for people, through page layouts, imagery and marketing copy, and much of the meaning sits in how a page looks rather than in the data underneath it.
An agent ready data layer brings product content, identifiers, attributes, pricing, inventory and imagery together into one governed source, with each field traced back to the system it came from. On top of that sits a semantic model, which defines what each product is, how it relates to its variants, range, brand and category, and what each attribute means in consistent, machine readable terms. Much of this would look dry or repetitive on a product page, and that is acceptable, because its job is to remove ambiguity for software rather than to persuade a person.
In practice the semantic model should cover the following.
A canonical product and variant structure, with a global identifier at variant level and clear links between parent products and their sizes, shades or fits.
Normalised attributes with controlled values and units, so that "navy", "dark blue" and "midnight" can be resolved to a common colour family where that helps comparison.
A consistent taxonomy across all brands and suppliers, mapped to the categories used by the major agentic commerce and shopping specifications.
Distinctive, factual descriptions for each product, written to state what it is and how it differs from its nearest siblings.
Price, stock and delivery data refreshed frequently enough that an agent can act on it with confidence at the point of purchase.
Reviews and customer insight about the product. Agents thrive on trust signals and customer feedback is a positive reinforcement.
Once this layer exists, publishing to a new channel becomes a mapping exercise. The same governed data can feed an agentic commerce catalogue, a ChatGPT app or Model Context Protocol (MCP) server, marketplace feeds and structured data on the website, and every channel benefits from improvements made once at source.
Where to start in practice
We’ve found the most effective first step is to measure where your catalogue stands today against the standards agents and agentic commerce platforms actually apply. We would suggest working through the following sequence.
Consolidate a single product feed. Bring together content, identifiers, pricing, inventory and imagery from each source system into one mapped feed, recording which system supplied each field.
Test the full catalogue for conformance. Validate every product against the specification of the agentic commerce platform you intend to use, ideally in its sandbox, and group the failures by error type, category, brand and source system so the largest problems are obvious. If this feels like to much of an overhead, start with a single department, brand or product type.
Score agent readiness on a sample. Assess a representative sample of products for distinctive descriptions, identifier coverage, normalised attributes, taxonomy and image quality, since these judgements need more care than automated checks provide.
Measure how agents see you today. Run realistic shopping prompts across the main AI assistants and record whether your products appear, whether they are described accurately and how you compare with competitors.
Fix at source and measure again. Prioritise the issues affecting the most products and revenue, correct them in the systems that own the data, and repeat the tests to show the improvement.
This work tends to surface organisational questions as well as technical ones, particularly around who owns product data that arrives from brand partners. Agreeing data standards with suppliers, including identifier and description requirements, is often the change that makes the improvement last.
We know this works as we have applied it across a number of our retail customers who are embarking on an agentic commerce readiness effort.
Closing thoughts
Agentic payments will make it far easier for a shopper to buy through an AI assistant, but an agent will only buy from a retailer when it is confident it has found the right product at the right price. That confidence is built from distinctive descriptions, consistent identifiers, structured attributes and fresh pricing and stock, held in a data layer that agents can read without guesswork.
Retailers who start this work now will be able to connect to each new assistant, marketplace and payment rail as it arrives, using the same governed product data each time. Those who wait are likely to find that the agents shopping on their customers' behalf recommend whichever retailer describes the product most clearly.
If you are interested in understanding how WeBuild-AI can assist your organisation in getting ready for agentic commerce then please get in touch via hello@webuild-ai.com or register your interest via the Get In Touch form and we can outline our Agentic Commerce Accelerator approach and the difference it can make for your business.
