1.416.619.5349 Ext.325 
What the conversation revealed about the new rules of product discovery
The way consumers discover products is changing. Not gradually, fundamentally.
Instead of typing keywords into a search bar and scrolling through results, shoppers are increasingly asking AI agents direct questions: “What’s the best cushioned running shoe for long distances under $250?”
The AI doesn’t return a list of links. It interprets the question, searches available data sources, evaluates options against the stated criteria, and delivers a recommendation, often with reasoning attached. This is agentic commerce. And it requires brands to rethink everything they know about product data.
To understand what this shift actually looks like in practice, we ran an experiment. We gave an AI agent a realistic shopping query, watched how it navigated to a recommendation, and then asked it to explain exactly how it got there. The results were illuminating and should be a wake-up call for any brand investing in product content.
The Query
We started with a straightforward prompt:
“Recommend the best men’s long-distance running shoes available in Canada. Premium budget, prioritize cushioning.”
This is the kind of question real shoppers ask. It’s specific enough to be actionable, but open enough that the AI needs to interpret, research, and synthesize. What happened next revealed how AI agents actually approach product discovery, and what data they depend on.
How the AI Navigated to a Recommendation
Before the AI could recommend anything, it had to build context. Here’s the process it described:
- Step 1: Define the Criteria
The agent immediately parsed our query into structured requirements: men’s sizing, long-distance use (marathon training, half-marathon, long runs), cushioned feel (not minimalist), premium budget (~$200-$300 CAD range), and available in Canada.
This is taxonomy in action. The AI translated natural language into filterable attributes. If your product data doesn’t include these attributes, or uses inconsistent terminology, you’re already invisible.
- Step 2: Source Independent Validation
The AI didn’t just pull from one source. It searched multiple independent review sites to cross-reference claims:
“RunRepeat tested long-distance running shoes in and out of the lab and discovered Novablast 5 to be the best overall, citing impact protection, weightlessness, and durability.” “Runner’s World identified Brooks Glycerin Max as best overall cushioned shoe and Nike Vomero 18 as best for long runs.”
This is critical. The AI weighted third-party reviews that included measured data, such as stack heights, shock absorption scores, and energy return, over marketing copy. Brands that have rich, testable specifications documented across independent sources gained an advantage.
- Step 3: Verify Availability and Price
A recommendation isn’t useful if you can’t buy the product. So the AI searched Canadian retailers specifically:
“HOKA Bondi 9 is available at Sporting Life and Running Room at $200 CAD… Brooks Glycerin Max at Altitude Sports for $249.99 (on sale for $199.99)… ASICS Novablast 5 at $189.99-$190.”
The agent specifically noted which retailers had stock, current pricing, and even sale information. Products without clear availability signals got deprioritized.
- Step 4: Synthesize and Recommend
Only after gathering review consensus, technical specifications, and retailer availability did the AI produce its final shortlist. It weighted multiple factors: category fit (cushioned, long-distance), review validation, Canadian availability, and price alignment with “premium budget.”
What Data Actually Mattered
When we asked the AI to explain what made certain products easier to recommend, the answer was revealing:
“The easiest path was combining a retailer with visible stock and price, plus a review that confirms or contextualizes feel and performance. That mix gives both practical buy-now signal and technical validation.”
In other words, the AI needed two things working together:
- Structured merchant data, including clear pricing, stock status, and product positioning
- Rich product context, including technical specifications, use-case fit, and independent validation
Brands that had both were easy to recommend. Brands missing either created friction, and friction in an AI-driven discovery process means getting skipped.
The Roadblocks: What Made Products Hard to Recommend
Equally instructive were the challenges the AI encountered:
- Incomplete Technical Specs
“Running Room pages showed product exists but required JavaScript to display full details. Many retailers list features but not measurements (stack height, weight, drop).”
If the AI can’t parse the data, it can’t use the data. Pages that hide content behind JavaScript, load slowly, or present information in non-standard formats create barriers. - Fragmented Information
“Had to combine brand pages (positioning, tech features) + review sites (measured data) + retailers (price, stock) to get complete picture.”
The AI had to piece together a complete product profile from multiple sources. Brands that forced this extra work were at a disadvantage compared to those with comprehensive, consistent data in one place. - Missing Standardized Attributes
“No standardized way to compare cushioning levels across brands.”
This is a taxonomy problem. Without consistent attribute definitions (what exactly does “max cushion” mean?), the AI must rely on proxy signals, such as review language, measured specs from third parties, and brand positioning. Brands using proprietary, non-standard terminology make their products harder to categorize and compare.
Two Modes of AI Shopping: “Help Me Find” vs. “Buy It For Me”
One thing became clear as we analyzed this experiment: our query was a research task, not a purchase task. We asked the AI to help us find the right shoe, not to buy it for us.
That distinction matters enormously for what data the AI actually needed.
- “Help Me Find” Mode
This is where most AI shopping lives today. The agent acts as a research assistant: reading pages, synthesizing reviews, interpreting content, and surfacing recommendations. It works primarily with visible, human-readable information.
In this mode, what matters most is: page content, review language, visible specs, and stock indicators that appear as text on the page. The AI is essentially reading like a human would, just faster and across more sources.
Notably, structured schema data didn’t play a direct role in our experiment. The AI relied on what it could read and parse from page content, not on underlying schema markup or product feeds. Many merchant pages embed structured data behind the scenes, but because the AI was operating in research mode, it focused on what was reliably viewable. - “Buy It For Me” Mode
This is where AI shopping is headed. In this mode, the agent doesn’t just recommend, it executes. It adds to cart, selects size and color, applies payment, and completes checkout.
In this mode, structured data becomes essential infrastructure. The agent can’t guess at your shoe size from a paragraph of marketing copy. It needs machine-readable product data: exact SKUs, GTINs, size and color variants, real-time inventory status, and checkout integration.
This is where schema markup, product feeds, and API-ready data become critical. Without clean structured data, an AI agent literally cannot complete a purchase on your behalf.
The Strategic Implication
Brands optimizing only for the conversational layer, with good content, reviews, and visible specs, will win in today’s “help me find” environment. But brands also investing in the transactional layer, with clean schema, structured feeds, and API-ready product data, will win tomorrow.
The trajectory is clear. AI agents are moving from research assistants to purchasing agents. The question is whether your product data infrastructure is ready for both.
What This Means for Brands
The running shoe experiment illustrates a broader shift. Product discovery is becoming conversational and agent-mediated. The implications are significant:
Taxonomy is Now Discovery Infrastructure
Your product hierarchy and attribute structure don’t just help humans navigate a website anymore. They teach AI agents what your product is, who it’s for, and when to recommend it. Sloppy taxonomy means missed recommendations.
Completeness Beats Creativity
AI agents don’t care about clever marketing copy. They care about structured, parseable data: specifications, attributes, stock status, pricing. The brands that won in this experiment were the ones with complete, consistent product information across touchpoints.
Third-Party Validation Amplifies Visibility
The AI weighted independent reviews with measured data heavily. Brands with strong presence on review sites, with products that have been tested and measured, had a significant advantage. This isn’t just about earning reviews, it’s about ensuring your product has rich, testable claims documented across the web.
Availability is Part of the Recommendation
The AI specifically checked Canadian retailer stock before finalizing recommendations. Products without clear availability signals, or that appeared out of stock, got filtered out. Your inventory data is now part of your discoverability.
Structured Data is Infrastructure for Tomorrow
While schema markup didn’t drive today’s research-mode recommendations, it will be essential as AI agents evolve into purchasing agents. Clean product schema, accurate GTINs, real-time inventory feeds, and machine-readable variant data are the foundation for agentic commerce. Brands that invest now will be ready when the shift accelerates.
The Bottom Line
When we asked the AI what separated easy-to-recommend products from hard-to-recommend ones, it gave a simple answer:
“Brands whose data was complete and accessible got recommended. Brands with gaps got skipped or deprioritized.”
This is the new reality of product discovery. AI agents don’t browse. They interpret. They don’t scroll through results. They synthesize and recommend. And soon, they won’t just recommend, they’ll buy.
If your product data isn’t structured for AI agents to understand, with complete attributes, consistent taxonomy, and clear availability, you’re not just making it harder for them to recommend you. You’re making it easy for them to recommend your competitors instead.
The digital shelf just moved. The question is whether your product data is ready to move with it.
Is your product data ready for AI-driven discovery? Learn more about Agent Discovery Readiness at RetailTaxonomy.com.
