A restaurant marketplace wants to increase orders from its customers.
The obvious idea is simple:
Give customers a bigger discount.
After all, if ₦1,000 off encourages more people to order than ₦500 off, then surely the bigger discount is better.
But there is a problem with that reasoning.
More orders don't necessarily mean more contribution.
The discount itself costs money, and there are other questions too.
Should the offer go to everyone or only selected customers?
Should it be sent immediately or after a few days?
Should the message be personalized?
And do high-value customers respond differently from ordinary customers?
Instead of choosing the promotion based on intuition, I treated the problem as a decision that could be tested.
The four decisions
We looked at four elements of the promotion:
Discount: ₦500 or ₦1,000
Timing: sent immediately or after 7 days
Spread: selected restaurants or platform-wide
Message: generic or personalized
That creates 16 possible combinations.
Rather than asking which individual choice “sounds best,” the goal was to determine which combination generated the greatest incremental contribution per customer after accounting for the cost of the promotion.
That distinction matters.
A promotion that generates more orders can still be a bad promotion if the additional revenue doesn't compensate for the cost of the discount.
The result was not what you might expect
The best-performing configuration was:
₦500 off + sent after 7 days + selected restaurants + personalized message
It generated an estimated incremental contribution of:
₦247.63 per customer
And all four factors had statistically significant effects after correcting for multiple comparisons.
But the individual results are even more interesting.
The discount wasn't the biggest lever
The discount had a positive effect.
Increasing the discount did generate additional orders.
But its net contribution effect was only approximately:
+₦18.41 per customer
Why so little?
Because the additional orders generated by the larger discount were largely offset by the additional cost of giving customers that discount.
So:
A bigger incentive did produce more response.
But that didn't make it the better business decision.
This is an important distinction in promotion analysis.
Response isn't the same thing as profitability.
The biggest levers were message and timing
The two strongest main effects were:
Message: approximately +₦63.51
Timing: approximately +₦43.80
That means the way and timing of communicating the offer mattered considerably more than simply increasing the amount of the discount.
The implication is straightforward:
Before spending more money on an incentive, a business should ask whether it can get more from the incentive it already has.
And then there was the surprising result about reach
The experiment also tested whether the promotion should be offered broadly across the platform or restricted to selected restaurants.
The broader reach performed worse.
The estimated effect of platform-wide spread was approximately:
−₦41.38 per customer
In other words, reaching more customers wasn't automatically creating more value.
It was creating additional promotional cost without enough corresponding benefit.
That's an easy trap in marketing.
It's tempting to think:
More people reached = more sales = better campaign.
But a business doesn't ultimately care about reach.
It cares about what the additional reach contributes after the associated costs.
Personalization was valuable, but not equally valuable
One of the most useful findings appeared when the results were examined by customer segment.
Personalized messaging produced a much larger incremental contribution among high-value customers:
+₦141.64 per customer
For regular customers, the effect was much smaller:
+₦29.96 per customer
Personalization helped both groups.
But its value was concentrated among high-value customers.
That creates an important strategic distinction.
The question isn't necessarily:
“Should we personalize every message?”
It becomes:
“Where is personalization worth the cost and effort?”
If personalization is expensive to scale, the evidence suggests that high-value customers are the natural place to prioritize it.
Meanwhile, discount, timing and restaurant selection performed consistently enough across the two segments that there was no comparable need to create separate strategies for them.
What about the interactions?
The experiment also tested whether combinations of the factors produced effects that couldn't be explained by simply adding their individual effects.
Three interaction pairs were tested within the design:
Discount × Timing
Discount × Spread
Discount × Message
None showed statistically significant interaction effects.
That's useful information too.
It means that, within this experiment, there wasn't strong evidence that the effect of one of those decisions fundamentally changed depending on another.
Sometimes finding no interaction is itself a finding.
It means we don't need to invent a complicated explanation when the evidence doesn't support one.
Then we tested the answer
There is another important step in any experiment like this.
Finding a promising combination in the original data isn't enough.
The recommended combination was tested on a held-out combination that wasn't used to determine the recommendation.
The analysis predicted approximately:
₦181.25 per customer
The held-out result was:
₦167.39 per customer
The difference was not statistically significant at the 5% level (p = 0.09).
That's not a reason to claim the prediction was perfect.
It isn't.
But it does provide marginal support that the relationship found in the original experiment was not simply an artifact of those observations.
So what should the business actually do?
The conclusion isn't:
“Always give customers ₦500 off.”
It's more specific.
Under the conditions of this experiment, the business would be better served by:
using the smaller discount,
waiting seven days,
restricting the promotion to selected restaurants,
and personalizing the message.
The larger discount isn't automatically better.
The widest reach isn't automatically better.
And personalization doesn't have to be applied equally to every customer.
The data point toward a more targeted strategy.
The lesson goes beyond promotions
This case is really about a common business problem:
Several reasonable ideas are competing for the same budget.
Someone says:
“Give a bigger discount.”
Someone else says:
“Send it to everyone.”
Another person says:
“Personalize everything.”
And another says:
“Send it immediately.”
All of those ideas sound plausible.
But plausibility isn't the same thing as evidence.
The experiment allowed us to evaluate the decisions together and ask a more useful question:
Which combination actually creates the most value for the business?
And the answer wasn't the loudest idea.
It wasn't the biggest discount.
It wasn't the broadest campaign.
It was a relatively modest promotion, delivered at the right time, to the right places, with a message that was particularly valuable for the right customers.
The bigger principle
Businesses often measure marketing success by things like:
orders;
clicks;
reach;
conversion;
revenue.
Those numbers matter.
But ultimately, a business needs to ask:
What did the additional activity contribute after the cost of creating it?
That's why the objective in this case wasn't simply to maximize orders.
It was to maximize incremental contribution.
And that's what made the result interesting.
The bigger discount could buy more behaviour.
The wider campaign could reach more people.
But neither necessarily created more value.