Assortment: ● 1,000,000 eBooks in 33 languages ● 65,000 audiobooks ● 5,000 new books appear in the catalog each month.
The retailer's distribution and content delivery channels: ● two options: subscription and per-unit pricing ● a few websites ● 25 iOS and Android apps ● 10,000 partners' websites and apps (for example, mobile operators' channels)
Within one country, any book price is the same on every channel. There is only one exception, the App Store channel (iOS). The reason is App Store Tiers — preset fixed pricing levels, managed by iOS:
A screen of App Store pricing instructions. Any price had to be equal to some tier set by the App Store.
How the eBook retailer's pricing worked before the pilot test launch
● To estimate the potential effect of dynamic pricing implementation.
● To hit +5% revenue growth, with the same sales volume (in pieces)
● AI-driven price optimization
/ AI - artificial intelligence /
Before the pilot test launch, the company priced eBooks and audiobooks manually, using an expert opinion based approach.
Pricing process details were as follows: ● For a book price calculation, they considered the publishing costs of this book. ● They used various rules and business constraints, such as "a eBook's price has to be less than the print book's price" and "an audiobook's price has to be higher than the eBook's price and can be approximately as high as the print book's price". ●They changed prices once a year or less frequently.
In 2021 and 2022, the eBook retailer has run two dynamic pricing pilot tests with the Imprice platform.
The company is a leading retailer in its markets. It was crucial to avoid any negative customer reaction to dynamic pricing. For this reason, the company ran two pilot tests:
1) A preliminary test with a small sample (50 books only). The main point: to start a price optimization journey without a risk of customers' negative reaction.
2) The main test (1000 books sample). The main point: to evaluate more precisely the revenue growth caused by AI-powered price optimization.
During these pilots, Imprice machine learning algorithms calculated prices for all channels, except the App Store (because of App Store Tiers reason). On the App Store, the retailer remained "old" manual prices.
Description of pilot tests
How Imprice AI-driven pricing works:
Imprice AI algorithms explore demand fluctuations and switch prices to the optimal level:
1. Set the expected optimal price for the current day or week. 2. Study the demand with that price. 3. Try to increase the price when consumers' willing to pay raises. Try to reduce the price when demand decreases.
The Machine Learning tool solves the Exploration–Exploitation Dilemma: to find the best price as soon as possible and exploit it as long as possible
The results were as follows:
● Pilot test 1: +7% revenue growth +2% sales volume (in pieces) growth
● Pilot test 2: +8% revenue growth +6% sales volume (in pieces) growth
Pilot test 1: Demand sensing. Price elasticity books groups
PILOT TEST STEPS:
1 - Sampling.
2 - Preliminary demand sensing.
3 - Price range corrections and AI-driven price optimization
As the first step, 3 groups with 50 books each were sampled: ● a pilot test group, ● a control group ● and a second control group.
Book sampling was made by Imprice time series clustering algorithm. To ensure "test-control" pairs were chosen correctly, retailer's experts manually checked and approved these book groups.
Let's go over some criteria of sampling: ● each group included both eBooks and audiobooks, ● the selected books had the most similar sales dynamics, ● these books were in thesame genre, ● these books had no cross-impact; sales of any sampled book didn't affect sales of other books from these groups. For example, if one book from a series was sampled, other books from that series couldn't be included into groups.
During the pilot test, the retailer's team switched off all marketing tools and activities (from paid ads to "bestseller" icons on websites and apps), for books from 3 test and control groups.
Step 1. Sampling
As mentioned above, the company changed prices once a year or less frequently. The sales history of any book from the pilot test group was as: "For the last 12 months, a book X had the same price, equal to Y, never changed".
The retailer's experts defined a price range permitted as (Y-10%; Y+50%), where Y is the initial, manual calculated price. As well, they set an additional constraint: Y+50% couldn't be higher than the printed book market price.
Then the Imprice AI-driven algorithm began the demand sensing. Within the set price range, (Y-10%; Y+50%), the algorithm started to raise and reduceprices, evaluating how it affected sales. Preliminary demand sensing took 4 weeks, because there was no price change history, but only "For the last 12 months, a book Xi had the same price, equal to Yi, never changed".
Prices were allowed to change one time in 3-4 days, because the retailer's sales cycle was equal to 2 days. Due to that, mostly, customers didn't see price change during decision-making. It helped to avoid any negative reactions. The Imprice AI-powered algorithm set prices fully automatically. A pricing specialist only had to monitor the results.
The second step, a preliminary demand sensing, was completed in 4 weeks.
Step 2. Preliminary demand sensing
After completing the preliminary demand sensing, the pricing team figured out the facts:
1)Some books from the test group had an elastic demand: demand grew significantly, when the price was reduced. For some books, the company gained the maximum revenue with the lowest price in the range: the initial price minus 10%. For these books, pricing experts changed the lowest limit of the price range permitted; the AI-driven algorithm was allowed to put prices lower than (Y-10%).
2) Some books from the test group had an inelastic demand: demand didn't reduce, when the price was raised. For some books, the company gained the maximum revenue with the highest price in the range: the initial price plus 50%. For these books, pricing experts changed the highest limit of the price range permitted; the AI-driven algorithm was allowed to put prices higher than (Y+50%).
Step 3. Price range corrections and AI-driven price optimization
An example of one book price change; the book had an inelastic demand. Even though the AI-driven algorithm raised the price of the book, and the final optimum price was two times higher than the initial price, the average monthly sales volume of the book has grown a bit and the revenue has doubled.
Orange plot: revenue in euros. Green plot: sales volume (in pieces). Blue plot: average price of a month.
The initial price: 8.6 euros (April 2021) The initial price range permitted: (7.8 euros; 13 euros). In June, the team began corrections; they allowed algorithms to put prices higher than 13 euros. The final optimal price: 18.6 euros (March 2022)
In September 2021, for two weeks the book was unavailable to buy because of technical issues. The sales volume decreased due to that. As we can see, the algorithm reacted to the demand fluctuation and cut the price.
The third step, price optimization, was completed in 10 weeks.
Due to the pilot test, the retailer's team found out:
● About 50% of the books in the tested group were undervalued (before the pilot). Their prices had been raised to the optimum level.
● About 30% of the books had an elastic demand. Their prices had been reduced to the optimum level.
● About 20% of the books had a chaotic demand based on some unknown external factors, not on their prices.
First 4 weeks (preliminary demand sensing period): Revenue and sales volume in the pilot group were lower than in the two control groups.
Next 10 weeks (price optimization period): +8% revenue growth, compared to the two control groups.
The entire pilot test period: +7% revenue growth +2% sales volume (in pieces) growth
iOS was the only channel with the old manually calculated prices. iOS sales figures were removed from total sales results.
The first pilot test results:
Pilot test 2: Experiment scaling
1 -Across 1000 eBooks and audiobooks
2 - Across the entire assortment, except for new arrivals
Generally, the second pilot test was similar to the first one: ● one test group and two control groups, ● each group contents 1,000 books (eBooks and audiobooks), ● the team observed subgroups of books with elastic demand, inelastic demand and chaotic demand.
The second pilot test was completed in 12 weeks. The results were as follows:
●+8% revenue growth ●+6% sales volume (in pieces) growth
As the business targets of the pilot tests were surpassed, the retailer's team started the rollout proceeding. They decided to complete the Imprice implementation across their entire product assortment, except for new arrivals.
As mentioned above, the target of the pilot test was:
● "To hit +5% revenue growth, with the same sales volume (in pieces)"