When Machines Go Shopping: How AI Agents are Changing Retail
What if your next million customers never visit your website? What if an agent goes on their behalf? It scrapes what it needs, comes back with a product shortlist, and the human only ever sees the answer. In this video, Dr Greg Fletcher pulls back the curtain on what it looks like when AI agents control product discovery… And what this means for your Ecommerce catalogue in practice.
When Machines Go Shopping: How AI Agents Are Changing Retail (Webinar Transcript)
Hello everyone, and welcome to the webinar, "When Machines Go Shopping." Thank you all for joining us today. We have 20 minutes and quite a lot to get through, so I'll try to keep the pace up, but hopefully it will be useful for everybody.
To kick things off, let me start with a quick thought experiment: what if your next million customers never visited your website? Imagine they never click, never see your homepage, never click an ad, never scroll a category page. Instead, they get recommended directly by an AI agent. In this scenario, where does the thinking happen? What form does this thinking take? And crucially, what can you do about it? These are the questions we're going to explore over the next 20 minutes or so.
To introduce myself, my name is Greg. I'm the co-founder and CTO of Ocula Technologies. We use agents to build agents that generate data to be consumed by other agents. Everything that follows is based on our daily experience and the impact we're seeing with a fantastic customer base of enterprise retailers across the UK and North America.
Why Traditional Product Search Fell Short
Let's recap the old shopping journey — this slide gives me shivers every time I see it. How did we used to do things? The shopping journey was incredibly fragmented. You'd open Google, have a bunch of tabs open, read listicles, watch videos. If you were really desperate, you might resort to the family WhatsApp group. As a brand, that meant dozens of touchpoints in this journey. The problem was it was exhausting, it took a long time, and you ended up going around in circles. However, a lot of thinking happened in this process, and many decisions were being made all the while.
Fast forward to where we are now: AI can do all of that research on your behalf, in a fraction of the time. This is great news, because we get a lot of time back. But the point I really want to dig into is that thinking and reasoning about your products still takes place in this new world — it just takes a radically different form. We'll dive into that in a moment.
Agentic Commerce Is Already Happening
Before we do, I want to address anyone who might be skeptical about the urgency here. Agentic shopping isn't a prediction — it's already happening. Here's the obligatory slide with some big numbers and a McKinsey citation, but the point is that consumer behaviour has already changed, and the potential size of the prize here is astronomical. The question for everybody on the call is no longer whether to adapt, but how fast.
What makes this different from previous AI hype cycles is that the infrastructure is already live: the protocols exist, the integrations are running, and major retailers are signing up. Let me show you what I mean.
What’s in a name? ACP and UCP
Two major protocols have emerged. There's the Agentic Commerce Protocol (ACP), developed by OpenAI and Stripe, powering ChatGPT commerce. And there's the Universal Commerce Protocol (UCP), developed by Google and Shopify, powering AI Search Mode and Gemini. These are defined APIs that let agents reason about your products — they can query your catalog, check real-time availability, and initiate transactions on behalf of the user. Think of these as schema.org for the agentic era.
I was at Google Cloud Next a few weeks ago, where Wayfair spoke about how they co-developed with ACP and are now live with over a million products on the protocol. So these aren't theoretical specs… the practical reality is that most retailers will need to support both, because they unlock different ecosystems.
OpenAI vs. Perplexity: How AI Shopping Platforms Compare
Let's look at OpenAI first. OpenAI has been testing instant checkout in the US but actually scaled it back in March, which is interesting. Instead of owning the checkout themselves, they'll direct shoppers to merchants' own native checkouts, meaning the focus now is really on product discovery. Only this week, they announced OpenAI Ads, a new feed allowing your products to be served as ads on ChatGPT.
By contrast, Perplexity took a different approach. They launched Instant Buy in late 2025, letting customers check out directly in the chat, with PayPal as one of the payment options. The interesting number here: shoppers referred to merchants from Perplexity spent 57% more per order than via other AI platforms.
The point we're trying to make isn't that frictionless checkout drives higher average order value — that's a fairly obvious point. What we're saying is that different AI platforms attract different shopper profiles. Perplexity's citation-rich, research-heavy approach is currently attracting shoppers with higher purchase intent. So when you start tracking agent-referred conversions, it's important to segment by source and not lump all traffic into one bucket — it isn't homogeneous.
Four Product Data Attributes AI Agents Need to Recommend You
How AI Agents "Think"
Product discovery has never been easier with AI agents, but they lean on very different signals than humans do.
As a quick mental model: when a shopper asks ChatGPT for a product recommendation, the model rewrites the conversational query into structured searches. It pulls in parallel from web search, structured product feeds, and reviews, chunks everything into small pieces, ranks each chunk for relevance, and synthesizes the top-ranked chunks into a natural language answer. The key thing to notice: at no point is the model reading your full product detail page top to bottom. It's pulling chunks, evaluating attributes, and synthesizing. Therefore, the optimisation playbook needs to change. And here’s how we’d recommend doing that.
1. Product Highlights: Writing Standalone, Chunk-Ready Copy
As mentioned, when an agent reads a product page, it doesn't read the whole page at once — it breaks the content into small semantic chunks, as few as 50 words at a time, and evaluates each chunk in isolation. There's no narrative thread, no assumed context, so each section needs to stand on its own and make sense independently. If you use words like "it" or "this" without context, the AI agent gets confused, doesn't understand, and won't recommend your product. The key takeaway: reframe how you think about product highlights. They're not just marketing copy, they're valuable data points. Write them like you're populating a fact sheet: able to stand alone and be understood out of context, not written as flowing prose.
2. Q&A Pairs: Preventing AI Hallucination on Product Details
When a shopper asks a specific question and an agent can't find the answer, it either skips your product or starts to hallucinate. Both are bad. Creating clear sets of Q&A pairs for each product on your site is a great way to combat that risk of hallucination. By answering common customer questions, you're giving the LLM a script to feed back to people asking for recommendations or advice.
One of the most common mistakes we see retailers make is simply rebadging the specs table as Q&A. If the specs already exist in the product data feed, the LLM has access to it — rewriting them as questions and answers is, at best, a missed opportunity, and at worst, adds noise to the signal. A simple test: could a shopper work this information out from the spec table? If so, it doesn't belong in a Q&A pair. If not, it does. These are exactly the kind of long-tail questions that never make it into structured data.
3. Competitive Differentiators: Standing Out in AI Search Results
Once the model has grabbed and chunked the data, and possibly matched a question to a Q&A pair, it needs a reason to rank the different chunks and decide what to return to the user. If everything in your category looks the same in the structured data, the model will default to picking on price, availability, or randomness. Differentiators state explicitly what makes your product distinct within its category. Is it lighter than other models, or are your running shoes made specifically for wide feet, for example. This matters more than ever because LLMs are stripping the brand layer from their answers. The customer doesn't see your logo or your hero shot; they see a synthesized recommendation, and differentiators are the signal most likely to survive that synthesis.
Here's an example in practice: I searched "best running shoes under £100" in ChatGPT. It didn't give a single answer — it gave several. On Google, if someone searches "best running shoes," you basically need to rank for that phrase. When someone types the same thing into ChatGPT, it explodes the query into comfort, durability, use case, price, performance, and so on, then builds a full answer from all of that synthesized information. So now you're not optimizing for one query — your product content needs to be rich enough that AI can find you across multiple angles of that query fan-out.
It's a reasoned set of recommendations, each matched to a specific use case, doing the job that blog posts, comparison sites, and category pages used to do. Brands that appear in results like this don't get there by accident — they get there because the product data was rich enough for the AI agent to understand who would use the product and why.
4. Use Cases and Intent Tags: Matching Conversational and Voice Search Queries
As search becomes more conversational, users are asking 10-plus-word queries with full context. Here's a real example from one of our clients: "Suggest durable pants with lots of pockets for shift work." Look at that prompt — there are almost no keywords in it. The agent has to translate the intent into product attributes, but it can only do that if your data already speaks the language of intent.
That's where intent tags come in — they act as the bridge, stating exactly who your product is for and what specific problems it solves. Looking further ahead, this only gets more pronounced with voice-based queries: a 10-plus-word typed query becomes 25-plus words when spoken. That richer intent context is a great opportunity, but it also makes the need for detailed, additive attributes even more acute.
Why Freshness and User-Generated Content Build AI Trust Signals
AI rewards freshness and authenticity. A product review, a Reddit comment, or an editorial mention can matter more than brand-written copy because it's user-generated. Increasingly, models check whether your product information aligns with the broader online consensus — which is a huge opportunity to boost trust signals and improve discoverability.
What if a new product review turns into an actual product use case? What if a Reddit thread informs your product's Q&A, and an editorial mention informs your differentiators? What if all of that user-generated content integrates seamlessly into your product copy? Here's what that could look like: a static product listing becomes an always-on, always-enriched, high-conversion page that AI agents consistently prioritise. Structured use cases populate the right tags, highlights and Q&A pairs directly answer user queries, and differentiators give the model a reason to return your chunk of information over a competitor's.
AEO and GEO Fundamentals: The Data Basics You Still Need
Alongside these four data points, it would be remiss not to mention the fundamentals of AEO and GEO: GTINs and MPN identifiers, structured shipping data, real-time (not batch-updated) stock information. None of these wins you anything on their own, but missing them means you don't get cited at all.
And trust is more important than ever. In addition to aligning with broader online consensus, your data needs to be clean, clear, and reliable, or LLMs will start ignoring your site — the same principle as how Google treats sites that consistently disappoint users, except the feedback loop is even faster and harsher. For example: if your feed says "in stock" and you're actually out, the agent hits a "merchandise not available" error and stops surfacing your products. Trust degrades fast and recovers slowly, so your product data needs to be live, not lagging.
The Scale Challenge: Enriching Product Data Across Thousands of SKUs
Most product content pipelines were built to serve a Google crawler and a human reader. That was fine 15 years ago, even 5 years ago. It's not fine now. The gap between what AI agents need and what most retailers currently provide is very wide. That gap is both a problem and an opportunity.
Scale is where it gets real. Imagine a retailer with 100,000 SKUs. Each one needs an enriched title, an enriched description, around 30 structured attributes, use case tags, compatibility data, competitive differentiators — everything just discussed — plus multilingual variants and channel-specific formatting. At a conservative estimate of 15 minutes per SKU, that quickly adds up to 12 people working full-time for over a year, just to reach baseline. And that's before accounting for the fact that the data needs constant updating, since products change, pricing changes, and the conversation about your products evolves over time as UGC shifts.
This is a terrible task for humans, but a perfect task for AI agents. You might be wondering who could help with something like that… this isn't a sales call, so I'll leave you to draw your own conclusions.
You need new language and new measurement alongside organic rankings: look at citation rates — how often you're mentioned in LLM responses for your category. Don't just look at page views; look at referred conversations from other platforms too.
5 Steps to Optimize Your Product Data for AI Agents
Audit your data: Map every attribute to what an AI agent actually needs, based on your findings.
Enrich your data: Don't do this manually; use an AI agent to do it.
Ensure your product data is consistent and unified across all channels.
Start setting a baseline.
Start tracking whether LLMs can actually find your products.
Key Takeaway: Structured Product Data Is a Competitive Advantage
If you'd like to see what a couple of your products look like enriched for agentic commerce — what that audit might reveal — please take a minute to respond to the poll in the comments; there are a couple of options there. Please follow us on LinkedIn for industry updates and more content like this, and book a demo if you'd like to see what this looks like in practice.
I'll leave you with a final thought: retailers who invest in structured, enriched product information now will be compounding their advantage as agentic commerce scales, and the brands that don't will spend the next couple of years catching up.
Thank you so much for listening. Happy to take any questions.
Related Posts
