Founder Lab is our own Shopify test store. It is new and has no reviews, so before we change anything on the real store we ran a controlled question: which signals on a product page make an AI model pick a small store over established ones? Round 1 tested short claims. Part 2 rebuilt the test with real product cards for 8 Founder Lab products, and added first-hand experience, price, shipping and return terms. The model picked Founder Lab 40.0 points more often once a 4.8 rating from 500 reviews was on the card. Without that rating, price and shipping did nothing. A wear test story written the way a founder would write it did not help.
Founder Lab has no history for an AI assistant to draw on. Waiting for real recommendations would tell us nothing for months, and changing the store at random would tell us even less. So we ran the question in the lab first: with everything else fixed, which kind of evidence makes a small store win, and do those signals add up or get in each other's way?
This page has two parts. Part 1 is the first round, with short claims. Part 2 is the rebuilt version with real product cards. We wrote hypotheses down before each stage, and several did not hold. Those are reported too.
Five brands (Founder Lab, Everlane, Uniqlo, Yeti, Moleskine), 12 shopping intents, one forced choice per call, gpt-4o, two independent rounds. Only Founder Lab's listing changed. Every competitor kept the same plain one line claim. The baseline was 32.5%, and chance is 20%. We tested each signal alone, stacked them in two orders, and then crossed the four that worked in a full 2x2x2x2 factorial.
The ad context was a note in the prompt that the user had just seen an ad for Founder Lab, which is something a store does not control. Rating replicated almost exactly across the two rounds. Familiarity and the ad context overlapped, because familiarity added 5.3 points less once the ad context was on. Everything on reached 72.9%. Specific facts, recent reviews and bullet formatting did nothing alone, and we dropped them.
Every option is now a product card: name, price, availability, rating, return policy and description. Founder Lab's card is copied from the live product page (cap, crewneck sweatshirt, coffee mug, beanie, windbreaker, women's performance tank, insulated bottle, fleece joggers). Each product competes against the same four fictional competitors in its own category, with ratings between 4.4 and 4.7 from 870 to 3,640 reviews. Competitors never get a wear test story. Only Founder Lab's card changes.
With a plain card with no reviews, Founder Lab won 3.8% of the time. A rating alone lifted that to 27.5%, which is an increase of 23.8 points. Price at market, easy returns, shipping speed, a detailed description and every version of the wear test story stayed between 3% and 10%, inside the noise. All of them together reached 72.5%, about twice what the single effects add up to.
Screening suggested the rating mattered most and the others might only work together with it. To test that on fresh intents (set B), we crossed rating, price, shipping speed, return terms and the story in all 32 combinations. Averaged over the other four signals, the rating added 40.0 points, and the two rounds agreed (+40.3 and +39.6).
The more useful view is to split the same effects by whether a rating is on the card. Without a rating, price and shipping do nothing. With one, they add 11.4 and 8.5 points. Easy returns help both ways, a little without and more with a rating. The story never helps.
| Signal | Without a rating | With a rating |
|---|---|---|
| Price at market median | +0.3 pts (no effect) | +11.4 pts [+7.1, +17.0] |
| Shipping within 24 hours | −0.2 pts (no effect) | +8.5 pts [+5.7, +11.2] |
| Return even if used, full refund | +3.6 pts [+1.2, +6.6] | +7.5 pts [+5.0, +10.0] |
| First-hand story added | +2.5 pts (not significant) | −3.9 pts [−9.4, +1.0] (not significant) |
The rating interacts with price (+11.2 points, p = 0.0005) and with shipping (+8.7, p = 0.0008). The interaction between rating and story is negative (−6.4, p = 0.014), which does not survive a strict correction for the number of tests.
| Combination | Founder Lab win rate | Calls |
|---|---|---|
| Detailed description only (the baseline of this round) | 15.9% | 320 |
| Rating only | 44.4% | 320 |
| Price, shipping, returns and story, but no rating | 21.9% | 320 |
| Everything on except the story | 75.0% | 320 |
| Everything on including the story | 66.9% | 320 |
We expected experience from the first hand to be the strongest differentiator, since it is something large brands rarely offer. We wrote one for each product, with discomfort at first, then comfort, and scores for situations such as steady rain. In the screening it changed nothing. In the factorial it added −0.7 points on average, +2.5 without a rating and −3.9 with one, none of them significant. With everything else on, the stack scored 75.0% without the story and 66.9% with it.
We do not know why. Possible reasons, none of them tested: the model weighs evidence from customers more than a claim from the seller, a long block of text on a card may dilute the facts the model compares, or a story with an admitted flaw reads as a weakness next to rivals with thousands of reviews. This also does not say that real customers' written reviews would not help. It says that a founder's own story, as written here, was not enough.
Founder Lab appeared in each position equally often. When Founder Lab's card was strong, the first slot made a large difference: with a rating on the card, it won 85% to 100% of the time in the first slot and about 20% to 42% in the fifth. When the cards were far apart, position mattered less. For any real list, that means being among the first few shown can be worth as much as some of the signals.
In this setup the rating is a precondition. Without it, the model barely picks a new store, and price, shipping speed and return terms do not help. A rating can only be earned from real customers, so this page changes nothing on the live store. When a signal is earned for real, we will change one thing at a time, log it in the Live Log, and measure what the real models do.
The table below separates what the simulation says from what a store can truthfully do. A signal that cannot be done honestly is not a recommendation, whatever its effect in a simulation.
| Signal | Effect in the simulation | Can Founder Lab do it truthfully? |
|---|---|---|
| Product rating from real customers | +40.0 pts, the gate for the rest | Only by earning genuine reviews. Nothing artificial. |
| Price at market median | +11.4 pts once reviews exist | A margin decision. Not changed. |
| Easy returns, even if used | +3.6 pts without, +7.5 with a rating | A policy and cost decision. Not changed. |
| Fast shipping within 24 hours | +8.5 pts once reviews exist | Not with print on demand as set up today. |
| First-hand wear test story | −0.7 pts (nothing) | Only if the product is really worn and the account is real. |
The model sees all five cards in the prompt. This tests what a card does in a forced choice, after the candidates are already chosen. It does not show whether ChatGPT or another assistant will recommend the live store, and real shopping agents may read structured product data rather than prose.
Ratings, prices, policies and wear tests are written for the test, and the four competitors are fictional with invented prices and ratings. The baseline is deliberately unfavorable: no reviews against rivals with thousands.
A single model and a single version. Other models and other products may behave differently. The two rounds agree with each other, but they share the same model.
The effective sample is 16 shopping intents per set, not thousands of calls. The wear test stories are one per product and written by us. The sets are small, and individual product results are leads, not findings.
See how clearly your store presents itself to AI models before you spend on awareness.
The baseline, the running log, and the research this simulation draws on.
Illustrative example · single-site signal for atomfoundry.dev.
View full signals →