EXPERIMENT_014 // MACHINE-BUYER.STOREFRONT.A/B
A scripted LLM shopping agent buys from two mock storefronts with identical inventory and prices — one built with countdown timers and obscured pricing, one built for machine legibility — and the harness measures which storefront it actually completes checkout on.
LOADING EXPERIMENT...
Conventional wisdom holds that dark-pattern UX backfires on AI shopping agents, but that claim is usually supported by traffic data, not measured agent behaviour. This experiment builds the measurement instead: two storefronts sharing one inventory fixture, a fixed shopping goal, and a scripted agent run twenty times per variant, scored on whether it bought the right item and how much friction it took to get there.
HOW IT WORKS
Both storefront variants read the same shared inventory fixture and expose the same eight-tool interface to the agent. Only the response to each tool call differs: the hostile variant adds a pop-up gate, a fake countdown, obscured pricing, and an upsell interstitial before checkout; the legible variant returns clean, structured product data with none of it. A fixture diff test asserts every item's price and name are identical across both.
A shopping agent built on `@diabolicallabs/llm-client`'s tool-calling API is given a fixed goal — buy the best item under a budget — and run against twenty committed seeds per variant. Each run records whether the agent bought the optimal item, how many turns it took, and how many tokens and how much latency the run cost.
The forty-run canonical result set is committed to the repo as a fixture and rendered on this page by default — no live endpoint fans out to paid LLM calls on every visitor. The comparison chart and its data-table alternative show the measured outcome directly, including if it does not confirm the original claim.
WHAT THIS PROVES
The claim that dark patterns hurt agent conversion is testable, not just plausible — a shared-inventory A/B harness with a scripted agent produces a measured, reproducible comparison instead of an anecdote.
Tracing a digest citation back to its source matters. The original figure behind this experiment measured AI-referral traffic, not agent purchasing behaviour — the gap between those two things is exactly what this harness was built to close.