By John Readman, Founder & Product Owner at ASK BOSCO

TLDR: AI agents are already filtering which products get recommended to your customers, before a human ever sees the results. AI chatbot traffic to retail sites grew 670% year-on-year, yet only 11% of retailers say they are ready to scale AI across their businesses. The problem is not consumer adoption, it’s that most product catalogues are built for human browsers, not machine-readable decision engines. Brands with incomplete schema markup, inconsistent product identifiers, or agent-incompatible commerce infrastructure are already invisible in AI-driven discovery. This article explains how the purchase funnel has changed, what AI agents actually check before making a recommendation, and what eCommerce teams can do about it this quarter.

AI agents are already deciding what to recommend to your customers. Most eCommerce teams have no idea what those agents can and can’t see. There is a version of your product catalogue that no human ever sees.

It’s the version an AI agent reads when a shopper types “find me the best running shoes under £100 with next day delivery” into ChatGPT, Gemini, or Perplexity.

The agent doesn’t open your website, browse your homepage, scroll your category pages, or notice your promotional banner. It queries structured data, cross-references product identifiers, checks availability signals, and makes a recommendation, in seconds, without a single human click.

If your catalogue isn’t built for that process, your products won’t appear.

This isn’t a future problem. During Cyber Week 2024, 20% of global orders were influenced by AI agents.

AI chatbot traffic to US retail sites grew 670% year-on-year over the same period. More than 20% of shoppers are already using generative AI for product discovery.

The customer journey hasn’t just changed, in many categories, it’s been restructured around a blocker your eCommerce team never had to think about before.

The funnel you built no longer describes how people buy

Most eCommerce teams still operate on a version of the same model: a customer becomes aware of a need, searches for a product, lands on your website, browses, converts, and leaves.

The job of eCommerce is to optimise each of those stages – better search visibility, better landing pages, better checkout flow.

That model hasn’t disappeared. But one has emerged alongside it and it works very differently.

In agentic commerce, the customer delegates workload. They tell an AI agent what they need; budget, brand preference, delivery requirement, any constraints. The agent then handles discovery, comparison, and often the transaction itself.

OpenAI’s Instant Checkout already enables customers to buy from Shopify merchants and Etsy sellers directly through ChatGPT, without ever visiting the retailer’s site.

Google’s Universal Commerce Protocol is being adopted by Walmart, Sam’s Club, and retailers globally to make their commerce infrastructure readable and transactable by AI agents at scale.

The consumer still makes the decision, but the agent filters the options before the consumer sees them. That filtering stage is the one most eCommerce teams aren’t optimising for, and it is increasingly where the sale is won or lost.

Here is what has changed at each stage of the funnel:

Discovery

Previously: search engine rankings, paid ads and social media.

Now: AI agent queries, which pull from structured data sources and third-party references rather than ranked web pages.

BCG found only an 8–12% overlap between traditional search results and AI-generated answers, meaning a brand with strong SEO is not automatically visible in AI-driven discovery.

Evaluation

Previously: product pages, reviews and comparison sites.

Now: Real-time agent combination of pricing, availability, specs, return policies, and cross-referenced identifiers.

If your product data is inconsistent, incomplete, or formatted only for human browsers, an agent cannot reliably evaluate it.

Purchase

Previously: your checkout flow.

Now: Agent executed transactions via tokenized payment credentials, operating within user-defined parameters. The customer might never even visit your site.

5 things an AI agent checks before recommending your product

When an AI agent is evaluating products in your category, here is what it’s actually looking for and what makes the difference between being recommended and being invisible.

1. Is your product data structured and complete?

Agents read schema markup, not marketing copy. They look for schema.org fields in JSON-LD format – specifically Product, Offer, MerchantReturnPolicy, and AggregateRating schemas.

These tell the agent what your product is, what it costs, whether it’s available, what the returns policy is, and how customers have rated it.

If those fields are missing, incomplete, or inconsistently formatted, the agent can’t accurately represent your product and will default to a product it can.

ASK BOSCO’s Recommended Action: Audit your schema markup. Prioritise product, pricing, availability, and returns policy fields.

Every product should have a consistent, unique identifier (SKU, GTIN, MPN) – agents cross-reference across sources, and inconsistencies reduce trust in your data.

2. Do your product identifiers match across every channel?

An agent querying your product doesn’t just look at your website. It cross-references your catalogue feed, your Google Merchant Centre data, third-party review sites, price comparison platforms, and any other source it can access.

If your product is listed with different SKUs, different names, or different specs across those sources, the agent either can’t confidently match them or produces an inaccurate composite.

This is how Ballantine’s Scotch whiskey ended up miscategorised as a prestige product by a major AI model – inconsistent and incomplete data across sources allowed the model to fill the gap with an inference that was simply wrong.

ASK BOSCO’s Recommended Action: Treat data consistency as a brand integrity issue, not just an ops issue. Audit how your products appear across every feed and third-party listing. Discrepancies that don’t affect human shoppers are actively harmful to agent-driven discovery.

3. Are your policies machine-readable?

Delivery timeframes, returns policies, and stock availability are among the most important signals an agent uses to filter and rank products.

A consumer who has told their agent “next day delivery only” won’t see your product in the results if your delivery information is buried in a PDF, written in ambiguous language, or not marked up in a way an agent can analyse.

ASK BOSCO’s Recommended Action: Ensure delivery options, returns policies, and stock status are exposed in structured formats, not just written on a page.

Real-time inventory accuracy matters here: an agent that recommends an out-of-stock product loses user trust, and so does the brand it recommended.

4. How does your brand appear in third-party sources?

AI agents do not rely solely on your own website and feeds. They synthesise information from across the web – review platforms, forums, Wikipedia, Reddit, and other high-authority sources. Search Engine Land shared that Reddit and Wikipedia are among the most-cited domains in ChatGPT responses.

If your brand or product is represented inaccurately, incompletely, or negatively in those sources, that representation feeds directly into agent recommendations.

ASK BOSCO’s Recommended Action: Actively monitor how your brand appears in the sources AI models are known to weigh heavily.

This is not traditional ORM (Object-Relational Mapping) – it’s about ensuring the informational ecosystem around your brand is accurate and consistent, not just positive.

5. Can an agent actually transact with you?

Discovery without transaction capability is a dead end. If a consumer’s agent identifies your product as the right recommendation but cannot complete or initiate the purchase, because your commerce infrastructure is not compatible with protocols like ACP or UCP – the agent moves to the next option.

ASK BOSCO’s Recommended Action: Understand which agentic commerce protocols are relevant to your platform (Shopify merchants already have access to OpenAI’s Instant Checkout via ACP).

Ensure your payment infrastructure, API availability, and checkout flow can support agent-initiated transactions. This does not require a full platform rebuild — but it does require knowing where the gaps are.

The conversion problem hiding inside the adoption data

Consumer adoption of AI shopping tools is accelerating fast, almost half of all shoppers used or planned to use AI for purchasing in 2024.

Among 25–44 year olds, over two-thirds say they are willing to delegate repetitive purchases to an AI agent. But only 13% of consumers have completed a purchase after an AI agent directed them to a website.

That gap, between intent and conversion, isn’t a consumer behaviour problem. It’s an infrastructure and trust problem. Customers want to use AI agents for shopping.

They hesitate at the point of purchase because they aren’t confident the agent has accurate information, that the transaction is secure, or that the brand is who the agent says it is.

For eCommerce managers, this isn’t a reason to deprioritise agentic commerce readiness, it’s a reason to prioritise it.

The brands that close the trust gap first, through accurate data, transparent policies, and agent-compatible infrastructure, are the ones that will capture the conversion uplift when adoption hits.

Shoppers directed to retail sites from AI platforms are already 30 times more likely to make a purchase than those arriving through traditional channels. The traffic quality is there, the conversion infrastructure, for most brands, is not.