Is offering a discount always a win?

Finding the optimal discount strategy for app subscriptions

Who doesn't love a good discount? Regardless of the era or industry, discounts have long been considered a reliable strategy to boost revenue. Sure, increasing sales volume without offering discounts might sound like a better plan, but discounts hold significant meaning as a price differentiation strategy. It's the only strategy that unearths hidden revenue opportunities in pricing that other strategies simply can't find.

But you have to be careful. An improper discount strategy can actually lead to an overall drop in revenue. In fact, we recently saw lower revenue from an experiment group in Alarmy where a discount was offered. In this post, let's dive into the characteristics of discount strategies and what to watch out for.

The starting point of a discount strategy: Who gets it, and how much?

Discounts are typically used as a price differentiation strategy. If you could perfectly tailor prices—charging more to users who would buy at a higher price, and less to users who need a lower price to convert—you could generate revenue from every single user. (See the 'Perfect Price Differentiation' graph in the image above.) You can tell this results in much higher revenue than maintaining a single price point (the 'No Price Differentiation' graph). Even operating just one additional discounted price point (the 'With Price Differentiation' graph) can generate more revenue than a single price.

However, doing this requires the ability to estimate users' spending habits. You need to determine whether price elasticity is high or low using actual consumption-related behavioral data (drop-off rates by purchase funnel, average purchase price, purchase frequency, etc.) or user attribute data highly correlated with spending power (age, occupation, income level, etc.).

But in Alarmy's case, we are constrained on both types of data. Since there's no mandatory sign-up and we aren't selling physical goods, it's hard to precisely pinpoint individual user spending power. We do at least have a purchase screen for subscriptions and some data on individual purchase history. When identifying users based on limited data, the accuracy will inevitably be lower. The only way to improve that accuracy is to run fast experiment iterations and inductively accumulate the lessons learned.

Inductive approach 1: Sales volume must absolutely increase

Fundamentally, offering a discount means conceding a portion of the unit price, so you have to drive enough extra sales volume to make a profit. It's simple math. To be precise, you must achieve an incremental sales volume greater than the percentage you conceded on price. (See the table below.)

If you offer an M% discount, you need to drive sales volume up by 1/(1-M%) just to break even on revenue, and the increment 1/(1-M%)-1 will always be greater than the discount rate M%. Plus, the larger the discount rate, the bigger the required increment becomes.

Actually, even if offering a discount results in the exact same revenue as before, it's practically a loss. Making $1,000 from 10 users is very different from making $1,000 from 100 users. This is because higher maintenance costs—like server operations and customer support—are incurred. Therefore, if you offer a discount, you have to approach it with the mindset that overall sales volume absolutely has to grow.

Inductive approach 2: Zero cannibalization is a must

When offering a discount, the best-case scenario is not giving it to users who would have purchased at full price anyway. When a discount is unnecessarily offered to users who would have bought at full price, causing them to purchase at the discounted rate, it's called cannibalization. It refers to a situation where revenue actually shrinks because people bought at a discount, whereas if left alone, they would have paid full price and generated more revenue. Thus, you must strategize to prevent cannibalization when offering discounts.

The table above shows discounted purchases increasing exclusively, with zero harm to full-price purchases. However, for us—since we can't sharply target users—a phenomenon like this is highly idealistic. The picture below is probably much more realistic.

Some users who would have bought at full price shifted to the discounted price, causing cannibalization, but fortunately, the total revenue generated was similar to the case where no cannibalization occurred at all.

The key here isn't thinking, "Cannibalization is inevitable, so we just have to live with it," but rather, "We need to find a strategy that minimizes cannibalization." And the metric used to judge this is comparing the total revenue generated in each case. In other words, you don't just look at the additional subscription revenue generated by the discount; you also examine the fluctuations in subscription revenue for the full-price product to check for any inevitable cannibalization.

Inductive approach 3: Finding the optimal app subscription discount strategy while factoring in cannibalization

The variables that make up an optimal discount strategy for subscription products generally include:

• Which subscription product gets the discount? (Annual vs. monthly)

• How much of a discount should be offered on that product? (10%? 30%? 50%?)

• Which users will receive it? (A mix of attributes & behavioral data)

• When should it be offered? (Time-based schedule or behavioral trigger)

• How long should the discount grace period last? (1 hour? 1 day? Anytime?)

• …

You decide on the control variables, choose the independent variables to use in the experiment, and then run it. After the experiment, you pick a winner based on the inductively generated data and move forward with follow-up hypotheses.

A recent subscription discount experiment in Alarmy

As shown at the beginning of the post, Experiment Group 1 and Experiment Group 2 both saw lower revenue than the control group. Even though these experiment groups lost from a revenue standpoint, they helped us inductively identify which variables were meaningless, allowing us to exclude them in future experiments.

Fortunately, another group—Experiment Group 3—showed a massive revenue increase, ensuring the overall experiment also delivered on the revenue front. Experiment Group 3 differed from groups 1 & 2 by just a single variable, yet it generated a whopping 1.4x increase in revenue.

On a side note, Experiment Group 3 was dramatically(?) added right at the tail end of the planning phase. Because the sudden reason (hypothesis) for its inclusion turned out to be spot-on and won, it made the experiment all the more meaningful.

Actually, if you tear into the data for Experiment Group 3, there was quite a bit of cannibalization. The number of full-price subscriptions was much lower than what occurred in the control group (sitting at around 70%), but the sheer volume of additional discounted subscriptions far exceeded this drop, driving a massive increase in total revenue. Now, for the follow-up experiments, we just need to tweak variables in a way that minimizes that cannibalization.

As you can see, a discount strategy is not some dead-simple, foolproof win. It's a difficult strategy that requires you to look at things from all angles before applying it.