A quiet shift is happening in how products get bought online. Increasingly, the "shopper" isn't a person scrolling a product page — it's an AI agent. Someone asks ChatGPT, Claude, Gemini, or Perplexity to "find me the best X under $50 and order it," and the model researches, compares, and picks. No website visit. No human eyeballs on your hero image. No checkout funnel.
For developers and ecommerce teams, this breaks a few assumptions worth talking about.
How an AI agent actually chooses a product
A human shopper is swayed by photography, branding, social proof, and a clean PDP. An agent is not. It reasons over text and structured data:
Product titles and descriptions
Specs, dimensions, compatibility, materials
Price, availability, shipping, and return policy
Reviews and ratings (as text it can summarize)
Whatever structured/schema data it can parse
So the agent's "pick" comes down to which product is the most legible and convincing in machine-readable form for the user's specific request — not which one looks best.
Why ranking #1 on Google isn't the same as winning the agent
This is the part that catches teams off guard. You can rank well in traditional search and still lose the agent's recommendation, because the agent isn't browsing a results page and clicking — it's reading, comparing, and deciding on the user's behalf. If a competitor's listing answers the prompt more directly ("waterproof, fits a 15-inch laptop, ships to the EU in 3 days"), the agent picks them, and you never even know you were in the running.
That's the uncomfortable bit: brands currently have almost zero visibility or control over the outcome of an agent purchase.
What "agent-commerce optimization" looks like
Think of it as SEO's next layer — sometimes called GEO (generative engine optimization). Practical moves:
Write for extraction, not just persuasion. Put the facts an agent needs (use cases, constraints, specs) in plain, structured text.
Make specs machine-readable. Clean schema markup, consistent attributes, no critical info trapped in images.
Answer the job, not just the keyword. Agents match intent ("for a beginner," "for travel," "compatible with X").
Keep reviews and policies parseable. Returns, warranty, shipping windows are decision factors the agent weighs.
A workflow for it
The approach I've been working on breaks into three steps:
Simulate the agent purchase — run the same prompts a real shopper would, across the major models, and see who gets picked.
Diagnose — figure out why you win or lose that pick (missing spec? weaker comparison? unclear use-case fit?).
Rewrite — fix the listing so it wins the next time the agent runs.
Disclosure: I'm building écentic, a tool focused on exactly this — the optimization layer for ecommerce brands when AI is the buyer. So I'm clearly biased, but the underlying shift is real regardless of which tool you use.
Top comments (1)
Some comments may only be visible to logged-in visitors. Sign in to view all comments.