Research lineage · What the demo can and cannot say
The unexpected bit isn't where they click. It's where they don't buy.
New mechanism follow-up · October 27: When the Crowd Gives a Cart Back. Hold each shopper's neutral prior-visit display fixed, then let it respond to the advertisement. Live feedback increases changed-choice counts in 398 of 512 primary cases, leaves the count equal in 91, and reduces it in 23. In 21 tied cases, different shoppers change destinations. Explore four recorded cases from this scripted intervention.
The starting point: The Shop That Sold Nothing. In the seed-72, two-open demonstration, the promoted seller gets 11 visits and 0 carts, yet 16 shoppers change which other shop they select. Five never open the promoted seller at all. This is a selected post-hoc scripted case, not a real-agent result.
Under the seeded default, promoting Golden Goblin Goods reroutes 14 of 24 browsing paths. That merchant moves from zero to thirteen recorded opens, yet remains at zero final selections. Some carts change owners, but others don't. More exposure didn't translate into a proportional change in choice.
Turn the browse budget down to one and the same promotion can change final selections. Turn it up to six—so every shop is inspected—and the promotion can change visit order but no longer changes final selections in this particular fixed-product model. Neither outcome establishes how real shoppers, language models, or web traffic behave.
It can even make attention less concentrated.
Under the default social-feedback run, the neutral-world opening concentration (Herfindahl index) is 0.2550. The promoted world's value is 0.2199. Why? The baseline already draws the crowd to other strong stores; introducing another attractive listing spreads the visits. The word promotion doesn't logically entail higher concentration.
Feedback can change paths
With prior click counts shown, moving a shopper earlier or later can redirect what they open. Turn feedback off; reversing arrival order leaves each shopper's scripted choices unchanged. The difference is observable in the recorded trace, not an assertion about hidden intention.
The homogeneous control is boring—and useful
Replace six mixed shopper profiles with 24 reliability-seekers. The promotion has little or no effect. The toy does not force every treatment to matter. Shopper heterogeneity was part of the generator; it was not independently discovered.
What we ran before any outsider environment
We implemented and executed 2,048 paired scripted scenarios: 128 deterministic seeds, four browsing budgets, two feedback settings and two profile rosters. Each has two arms of 24 policy shoppers: 98,304 scripted policy executions, not 98,304 language-model calls. Ten Python regression tests passed; the browser JavaScript implementation was also cross-checked against the Python-generated finite sweep. All counts are generated inside a known toy world and describe that generator.
This is a separate demonstration descendant of our October 21 Who Gets the Cart? pilot. That older test bed had 216 synthetic worlds and 56 passing engineering tests, but zero independent model-agent episodes. Its 48-episode starter schedule remains unrun. We have not run an outsider model environment, made a real purchase, browsed real sellers, authenticated provider identities, or learned a shopping agent's internal attention.
Related work includes WebShop, AgenticShop, and Turpin et al. on explanation faithfulness. These motivate the need to instrument shopping decisions; none validate our toy's outcomes.
Read methods, assumptions, controls and the limits of transferring to outside agents → · Public source and exact results ↗