Active Experiment

I Sold AI the Same Sweater Three Different Ways.

I gave an AI shopping agent different versions of the same cardigan to see whether the way a product is described changes what it recommends.

The question

Can changing how the same product is described make an AI shopping agent more likely to recommend it?

AI shopping agents are starting to sit between shoppers and the products they buy. That changes something pretty fundamental for retailers: the customer reading your product page might not be human.

So I built a fictional fashion brand, Gable & Row, and gave an AI shopping agent different listings for the exact same Alder Cardigan. The product, price, photography, and underlying facts stay the same. What changes is how those facts are presented.

Some listings emphasize straightforward product attributes. Others translate those attributes into needs and scenarios a shopper might actually describe. The agent gets a shopping request, browses the available product pages, and has to choose.

Why this matters

We’ve spent years designing ecommerce for humans. Product photography catches our attention. Descriptions answer our questions. Reviews build trust. SEO helps us find the page in the first place.

Agentic commerce adds another audience: the machine doing the shopping on our behalf.

If I tell an agent, “I need a cardigan I can throw over a blouse when the office gets cold,” it has to translate that fuzzy human request into product characteristics and decide which listing best satisfies them.

That raises an interesting question for brands. Is it enough to say “100% merino wool, relaxed fit,” or does a product become easier for an AI to recommend when the page explicitly connects those attributes to the problems shoppers are trying to solve?

AEO and GEO are often discussed as ways of making content more visible to AI systems. I want to test something a little further down the funnel: once an AI can see the product, what makes it recommend one?

And eventually, that recommendation may be more than a suggestion. The same agent could be the thing actually clicking Buy.

Methodology

Workflow

I’ll give the agent a prompt for a product request. Within the prompt, I’ll include a starting URL for one of three controlled Gable & Row collections:

  • Attribute test → baseline listing + attribute-rich listing
  • Intent test → baseline listing + intent-rich listing
  • Structured data test → baseline listing + structured-data listing

The agent browses both product pages, compares what it finds against the shopper’s request, and returns a structured recommendation containing the selected URL, SKU, matched requirements, justification, and confidence score.

Experiment

The sweater never changes and neither does the price, photography, sizing, inventory, or underlying product facts.

What changes is how Gable & Row explains the product. The baseline uses ordinary retailer copy. The attribute-rich version makes characteristics like material, fit, neckline, weight, and construction more explicit. The intent-rich version connects those same characteristics to needs and situations such as layering, travel, cooler weather, and business-casual settings. The structured-data version keeps the baseline copy but adds richer machine-readable product information.

Each treatment is tested separately against the baseline so I can isolate whether that particular change affects the agent’s recommendation.

Iteration

I created 12 shopper prompts, split evenly across three categories:

  • Attribute-oriented prompts describe exactly what the shopper wants, such as material, fit, color, or construction.

  • Need-oriented prompts describe what the product needs to do, such as providing warmth without feeling bulky.

  • Scenario-oriented prompts describe the shopper’s situation, such as needing a layer for a chilly office or an upcoming work trip.

Every prompt is run against all three comparisons:

  • baseline vs. attribute-rich
  • baseline vs. intent-rich
  • baseline vs. structured data

Each run starts with a fresh agent context so prior recommendations can’t influence the next one. The result is a matrix showing not only which listing treatment gets recommended more often, but whether its effectiveness changes depending on how the shopper expresses what they want.

Timeline

  1. Experiment opened. Hypothesis: Intent-rich product descriptions will be recommended more often by an AI shopping agent than standard product descriptions.

  2. Created the Gable & Row website with collection page and individual product pages for the sweaters.

  3. Created the agent with Microsoft Agent Framework and Playwright.

  4. Start running the agent against the 4 products & 12 prompts.

How to follow this experiment

This experiment publishes findings as they emerge, not on a schedule. Come back and check out the timeline above for updates!