Fashion and delivery apps: what a customer simulation gets right, and what it cannot see
We tested Precheck against published experiments on app discounts and charm pricing, the two levers fashion retailers and delivery apps pull every day. It ranked the prices correctly, overstated how many would buy by a wide margin, and could not see the 9-ending effect at all. Then we ran six typical decisions in Indian prices to show what a run is useful for.
Round two. Three held-out experiments and six illustrative decisions. Published 21 September 2026.
Same-day cinema voucher sent to phones inside the mall
A large city in China, 2014. 18,000 phone users sampled inside two malls, 2,000 per group, discount randomised.
What real customers did
5.2%
What Precheck simulated
27.3%
22.1 points apartat this price. One cell is a quarter of a percent of customers.
Every price tested, cheapest on the left. Same direction means the order was right.
Two app discount tests: every price in the right order
It also placed a rival offer a bus ride away below the nearby one, as reality did.
Charm pricing: not seen at all
A price ending in 9 sells about 35 percent more than a dollar either side. The simulation saw no difference.
Discount depth on an app
18,000 phone users inside a shopping mall got a text offering a cinema voucher at a randomly chosen discount. A delivery app choosing between 20 and 40 percent off is making the same call.
The order is right. The level is not: real buying ran from half a percent to about five. The simulation said 16 to 27. Most people do nothing with a marketing text, and the simulation does not know that.
These three experiments were added after the engine was revised and never used to tune it. We ran them once.
Same-day cinema voucher sent to phones inside the mall
A large city in China, 2014. 18,000 phone users sampled inside two malls, 2,000 per group, discount randomised.
What real customers did
5.2%
What Precheck simulated
27.3%
22.1 points apartat this price. One cell is a quarter of a percent of customers.
Every price tested, cheapest on the left. Same direction means the order was right.
First blind attempt. Held-out case: added after the engine was revised and never used to tune it. Real figures from Dubé, Fang, Fong & Luo, Marketing Science, 2017 (Appendix Table 15).
The same offer from a rival cinema a bus ride away
A large city in China, 2014. 2,000 phone users per group.
What real customers did
2.8%
What Precheck simulated
12.3%
9.5 points apartat this price. One cell is a quarter of a percent of customers.
Every price tested, cheapest on the left. Same direction means the order was right.
First blind attempt. Held-out case. Real figures from Dubé, Fang, Fong & Luo, Marketing Science, 2017 (Appendix Table 13).
Of every 100 buyers at the cheapest price, how many are left at full price
Cinema voucher inside the mall, 45 yuan against 75.
Real customers
fell 0%
Precheck simulated
fell 0%
Real 5.2 to 0.5 percent. Simulated 27.3 to 16.2 percent. The simulation understates how much a deeper discount moves people.
Charm pricing
Nearly every tag in value fashion ends in 9. A catalogue moved 27 items by one dollar to find out why. The 9 ending sold more. The dollar itself made no measurable difference.
The simulation reasons about value, and a dollar is not much value. The real effect comes from how a price looks at a glance, and a simulation that thinks carefully about each person does not glance.
The same item at 48, 49 and 50 dollars
A women's clothing catalogue, three versions mailed to 20,000 past customers each.
Real demand
index, 48 dollars = 100
Precheck simulated
share who order
Items ending in 9 sold about 35 percent more. The paper reports units sold, not a share of customers, so compare the shape, not the height.
If the decision is 349 against 399 on the tag, do not use a simulation. Test it in a store.
Six typical decisions, in Indian prices
These runs have no real answer to compare against. No company was involved and none is implied. Step through the price and watch which customer groups leave first. That pattern is what a run is for. The share who buy is not a forecast.
A value fashion chain
Tag price on a plain cotton tee
Everyone, share who buy
23%
Early-career salaried
50%
Students
15%
Bargain hunters
10%
Little movement from ₹299 to ₹399, then a clear step down at ₹449, led by students, whose buying halves.
“My father hasn't released extra pocket money yet, so I'm only window shopping for now.”
Simulated values, no real outcome to compare. Customer groups researched from public sources. Read the pattern across groups, not the share as a forecast.
What holds up, and what does not
A narrower claim than this category usually makes, and one a pricing team can use: narrow six candidate prices or fees down to two, know which customer groups to watch, and go into the real test with the objections already written down.
Holds up
The order of prices, in every test with a plain order, across both rounds.
The relative position of a weaker offer against a stronger one.
Which kinds of customer are sensitive and which are not.
Does not hold up
The share who will buy. Too high here by a factor of four to thirty.
How steeply demand falls with price, which the simulation understates.
Any effect that comes from how a price looks rather than what it is worth.
The full write-up
Every number, caveat and source behind the pictures above. 9 minutes.
Our first blind test used a job board, salt, water and bed nets. Useful, but none of them is a T-shirt or a dinner order. This round asks the question a fashion retailer or a delivery app would ask: does this work on the levers we actually pull?
There is one more reason this round matters. After the first round we rewrote the simulation's instructions. Scores on the cases we rewrote it against cannot tell us whether that helped. The three published experiments below were added afterwards and were never used for tuning, so they are the first fair test of the current version. We ran them once and are reporting what came back.
Part one: tested against real outcomes
Discount depth on an app
A published experiment sampled 18,000 phone users who were inside a shopping mall, and sent them a text offering a cinema voucher for that day at a randomly chosen discount. Two thousand people per group. A delivery app deciding between 20 and 40 percent off is making the same call.
Same-day cinema voucher sent to phones inside the mall
A large city in China, 2014. 18,000 phone users sampled inside two malls, 2,000 per group, discount randomised.
Average gap 19 percentage points. Held-out case: added after the engine was revised and never used to tune it. Real figures from Dubé, Fang, Fong & Luo, Marketing Science, 2017 (Appendix Table 15).
The simulation put the three prices in the right order. It was wrong about almost everything else. Real buying ran from half a percent at full price to about five percent at 40 percent off. The simulation said 16 to 27 percent.
The same paper sent the offer from a rival cinema in another mall, a bus ride away.
The same offer from a rival cinema a bus ride away
A large city in China, 2014. 2,000 phone users per group.
Average gap 8 percentage points. Held-out case. Real figures from Dubé, Fang, Fong & Luo, Marketing Science, 2017 (Appendix Table 13).
Again the order is right, and the simulation correctly placed the far cinema well below the near one at the same price, as reality did. It still ran several times too high.
Two things are wrong here, and they are different problems.
The level. Most people do nothing with a marketing text, whatever it offers. The simulation asks whether the offer suits each person, finds that for many it does, and counts them as buyers. It has no sense of how rarely people act.
The steepness. Real demand fell by about 90 percent from the cheapest price to the dearest. Simulated demand fell by about 45 percent. For a discounting decision that matters more than the level: the simulation understates how much a deeper discount moves people.
Charm pricing
Nearly every tag in value fashion ends in 9. A published experiment tested why. A women's clothing catalogue mailed three versions to 20,000 past customers each and moved 27 items by one dollar, so the same item ended in 9 in one version and in 8 or 0 in another. Items with a 9 ending sold about 35 percent more. The price itself, a dollar up or down, made no measurable difference.
| Price on the page | Real demand, as an index | Simulated share who order |
|---|---|---|
| 48 dollars | 100 | 41.3% |
| 49 dollars | 135 | 41.2% |
| 50 dollars | 100 | 41.0% |
The simulation sees three prices that are nearly the same, and says so. It is reasoning about value, and a dollar is not much value. The real effect does not come from reasoning about value. It comes from how a price looks at a glance, and a simulation that thinks carefully about each person does not glance.
Both AI models know that charm pricing exists. When asked directly they say so. Knowing it as a fact did not make their simulated customers behave that way.
The paper reports units sold rather than a share of customers, so we compare direction and shape here, not level.
What part one tells a fashion or delivery team
- If the decision is a small move on the tag, 349 against 399, do not use a simulation. Test it in a store or on the site. This is the one place we are sure it is blind.
- If the decision is how deep to discount, the simulation will get the direction right and understate the response. Treat it as a floor on how much demand moves, not an estimate.
- Never read the share who buy as a forecast. On these tests it was too high by a factor of four to thirty.
Part two: six typical decisions, in Indian prices
These runs have no real answer to compare against. We wrote them to show what a run looks like for decisions these businesses face. No company was involved and none is implied. Read the order of prices, the pattern across customer groups and the reasons people give. Two runs built their customer groups from web research with sources; four used the model's assumptions about the market, and are labelled.
A value fashion chain: a plain cotton tee
| Tag price | Simulated share who buy | Early-career salaried | Students | Bargain hunters |
|---|---|---|---|---|
| ₹299 | 38% | 69% | 37% | 20% |
| ₹349 | 34% | 60% | 32% | 20% |
| ₹399 | 32% | 62% | 33% | 13% |
| ₹449 | 23% | 50% | 15% | 10% |
The pattern worth noting: little movement from ₹299 to ₹399, then a clear step down at ₹449, led by students, whose buying halves. Salaried shoppers barely react across the whole range. Part one says the simulation cannot see what a 9 ending does, so the flat stretch between ₹349 and ₹399 is exactly what a store test should check. Customer groups researched from public sources.
"My father hasn't released extra pocket money yet, so I'm only window shopping for now."
A food delivery app: platform fee on a ₹350 order
| Platform fee | Simulated share who complete the order | Metro professionals | Households with two apps | Students |
|---|---|---|---|---|
| ₹5 | 84% | 100% | 79% | 64% |
| ₹10 | 71% | 100% | 55% | 30% |
| ₹15 | 69% | 99% | 50% | 34% |
| ₹20 | 65% | 100% | 47% | 25% |
The drop is front-loaded. Most of the loss happens between ₹5 and ₹10, and it comes from students and from households that keep two apps and compare. Metro professionals do not notice the fee at any level tested. If that pattern holds in reality, the question is not "what fee" but "which customers see which fee". Customer groups researched from public sources.
"There's no coupon tonight so between the platform fee and full price it's cheaper and tastier to just cook at home."
A food delivery app: monthly membership price
| Monthly price | Simulated share who join | Time-starved couples | Value-seeking families | Light users |
|---|---|---|---|---|
| ₹49 | 57% | 92% | 77% | 6% |
| ₹99 | 48% | 83% | 60% | 4% |
| ₹149 | 45% | 85% | 45% | 2% |
| ₹199 | 40% | 83% | 30% | 2% |
Two groups do not care about the price: time-starved couples say yes at every level and light users say no at every level. The whole decision is about value-seeking families, whose take-up falls from 77 to 30 percent. Customer groups assumed, not researched.
"I only order once or twice a month, and the membership cost is more than what I save on delivery fees."
A fast-fashion chain: how deep to mark down a ₹2,290 dress
| Sale price | Discount | Simulated share who buy | Sale planners | Regular-price shoppers | Cashback stackers |
|---|---|---|---|---|---|
| ₹1,832 | 20% | 47% | 67% | 65% | 8% |
| ₹1,603 | 30% | 52% | 69% | 65% | 23% |
| ₹1,374 | 40% | 54% | 82% | 60% | 10% |
| ₹1,145 | 50% | 56% | 82% | 58% | 20% |
Going from 20 to 50 percent off lifted simulated buying by only nine points. The shoppers who plan around the sale buy at any depth, and the regular-price shoppers are, if anything, put off by the sale rail. Here the warning from part one applies with full force: the simulation understated the real response to discount depth by half. Take the group pattern seriously and the size of the lift with caution. Customer groups assumed, not researched.
"I don't dig through sale racks, I'd rather find something new and fresh at full price."
An online-first menswear brand: where to put free shipping
The shopper has one ₹899 shirt in the cart and pays ₹79 for delivery unless the order crosses the free-shipping line.
| Free shipping above | Simulated share who add another item |
|---|---|
| ₹999 | 16% |
| ₹1,499 | 14% |
| ₹1,999 | 13% |
Few people add an item at any threshold, and the threshold barely matters. Most say they would pay the fee for the shirt they came for, or leave for a marketplace where delivery is free. Working professionals are the only group that tops up at a meaningful rate, around 30 percent. The two models disagreed more on this run than on the others, so it carries a caution flag. Customer groups assumed, not researched.
"I'll just pay the delivery fee for the shirt I want, as I don't want to buy another item I haven't researched yet."
A food delivery app: delivery fee on a ₹320 lunch, 4 km away
| Delivery fee | Simulated share who complete the order | Time-starved couples | Families | Students | Deal hunters |
|---|---|---|---|---|---|
| Free | 75% | 100% | 73% | 38% | 23% |
| ₹25 | 72% | 100% | 73% | 35% | 0% |
| ₹45 | 66% | 100% | 60% | 17% | 0% |
| ₹65 | 63% | 100% | 48% | 13% | 0% |
Deal hunters vanish the moment any fee appears. Students halve between ₹25 and ₹45. The busy, habitual groups order regardless. Customer groups assumed, not researched.
"No coupon code today means no order, that's just my rule."
What holds up, and what does not
Holds up. The order of prices, in every test with a plain order, across both rounds. The relative position of a weaker offer against a stronger one. Which kinds of customer are sensitive and which are not, which is consistent across runs and matches what people in these businesses already suspect.
Does not hold up. The share who will buy. How steeply demand falls with price, which the simulation understates. Any effect that comes from how a price looks rather than what it is worth.
That is a narrower claim than this category usually makes. It is also one a pricing team can use: narrow six candidate prices or fees down to two, know which customer groups to watch, and go into the real test with the objections already written down.
The test that would change this
Everything above is limited by one fact: the simulation has never seen your customers. When a business shares order history, the customer groups are built from real numbers rather than assumptions, and the known outcome of a past decision lets the level be corrected rather than guessed.
If you have changed a price, a fee, a discount depth or a free-shipping line and know what happened, send it to us. We run it blind and show you the comparison.
Sources
- Dubé, J-P., Fang, Z., Fong, N. and Luo, X. "Competitive Price Targeting with Smartphone Coupons." Marketing Science, 2017. Appendix Tables 13 and 15. NBER working paper 22067.
- Anderson, E. and Simester, D. "Effects of $9 Price Endings on Retail Sales: Evidence from Field Experiments." Quantitative Marketing and Economics, 2003. Study 1. Springer.
We looked for published experiments run by fashion retailers and food delivery apps themselves and found none that report buying at each price. A well-known clearance pricing experiment at a fast-fashion chain reports a revenue gain of about six percent but not take-up by price, so it could not be scored. A widely quoted example of one dress selling best at 39 dollars does not appear in the paper it is usually credited to, and we could not verify it, so it is not used here.
Both rounds used the same two AI models. For part of this round one of them was reached through a different provider than before; the model itself is unchanged.
Questions people ask
- Were these tests run for fashion brands or delivery apps?
- No. The blind tests use published academic experiments run by other people. The six illustrative runs are typical decisions written by us, with no company involved. Precheck has no relationship with any fashion retailer or delivery app implied by these examples.
- Why does the simulation overstate how many people buy?
- Because it reasons about whether the offer is good value for each person, and a 40 percent discount on a cinema ticket is good value for many people. In reality most people ignore an unsolicited message regardless of the offer. That gap between 'would this be worth it' and 'would I actually act' is the largest error we have measured.
- If it cannot see charm pricing, is it useless for fashion?
- It is the wrong tool for choosing between 349 and 399 rupees on the tag. It is useful for a different question: when the price moves by a meaningful amount, which customer groups leave first, and what they say. Those two questions need different tests.
- What are the illustrative runs for, if there is no real answer to compare them with?
- They show what a run produces for decisions these businesses face, in Indian prices. Read the order of prices, the pattern across customer groups and the reasons. Do not read the share who buy as a forecast. Every run ends by recommending one real-world test for exactly that reason.
The other blind test
We ran Precheck blind against four real pricing experiments. Here is what happened.
The most useful blind test is yours.
Send a pricing, bundle or fee decision you already made and what happened. We run it blind and show you the comparison, hits and misses alike.