Introduction
SaaS companies should be more eager to sell to bad customers.
More specifically, SaaS companies can be better off selling to customers with poor LTV/CAC economics. On its face, this surely appears a strange claim. I hope to demonstrate here some conditions under which it’s true.
The summary logic here is simple:
SaaS unit economics are generally described using historical average analysis rather than projected marginal analysis
An incremental sale is profitable so long as marginal value exceeds marginal cost
SaaS companies bear significant fixed costs even in go-to-market operations
Average costs and marginal costs frequently differ, particularly where fixed costs are a material consideration (as in SaaS companies)
Historical CAC is an average cost, not a marginal cost; it’s therefore a dangerous input to forward-looking operational decisions
We can sell many incremental units with LTV < historical CAC profitably
Preliminary thoughts on SaaS economics
Suppose there’s some enterprise software company, Saasco. In July, Saasco acquires 10 customers; each of these identically commits to pay $1,000 per year so long as the company uses Saasco’s product. We can see that Saasco’s July cohort starts with $10,000 of ARR. For simplicity, let’s suppose Saasco has 100% gross margins — there are no maintenance costs.
Without understanding the amount of time during which Saasco retains its customers, it’s ambiguous whether our cohort here is profitable. If each of these customers churns after their first year, then Saasco has only collected an aggregate of $10,000 from the cohort. If each customer stays for exactly two years, Saasco collects $20,000 from the cohort.
We also have to understand what it costs to ‘buy’ this annuity in the first place; it tends to be very expensive to get a SaaS customer. Let’s again imagine each customer churns after their first year. If acquiring the customers cost $20,000 in total, then Saasco lost $10,000. By contrast, if acquiring the customers cost $5,000 in total, then Saasco still makes some money.
It should be clear now that the profitability of a given cohort depends on a few things:
The cost of acquisition for a contracted dollar
The retention of a contracted dollar
The margins on a contracted dollar
A more formal model of SaaS economics
Setting up our assumptions
Forget everything from the last section (regarding Saasco). We’ll here develop an algebraic model to describe the business under several scenarios.
Suppose there exist two kinds of customers in Saasco’s market. Within each of these two archetypes, customers are homogenous.
Type N customers: these are good customers. They retain at healthy rates, have solid ACVs, and they have nice-looking logos on Saasco’s website.
Type M customers: these are bad customers. They churn like crazy, have meager ACVs, and Saasco would never use them as a customer reference.
Then construct some arbitrary acquisition cohort for Saasco; that is, within some fixed period of time, Saasco lands some number of new customers. Within that period, we set up the following general conditions:
Let n be the count of type N customers acquired within the period
Starting ARR is $1 for each of these N customers
Each contracted dollar retains at some annual rate g where 0 < g < 1, such that in some year t, these n customers retain ng^(t) in ARR
Each distinct customer of Type N acquired within the period imparts some variable sales and marketing cost — we call this p
Let m be the count of type M customers acquired within the period
Let λ be some number such that 0 < λ < 1; type M customers have ACV of $λ
Let j be some annual rate of contracted dollar retention such that 0 < j < g. Then we can say in some year t, these m customers have retained mλj^(t) in ARR
Each distinct customer of Type M acquired within the period imparts some variable sales and marketing cost — we call this k
Then further suppose there is some short-run fixed component of sales and marketing expenditure. This number is fixed in the sense that it remains constant as we vary the count of customers (n+m) acquired in the period. We can call this number z.
Let’s just assume gross margins are 100% for simplicity (Note that if we wanted to talk about LTVs in terms of gross profit, we could arbitrarily rescale the starting ARR, and the model would be functionally identical.)
Transforming our assumptions into useful statements
The starting ARR of the cohort: n + mλ
Average ARR per customer is (n+mλ) / (n+m)
At any number of years elapsed since acquisition t, the retained ARR of the cohort is ng^(t) + mλj^(t)
The total sales and marketing expenditure to acquire the cohort is z + np + mk
The blended CAC for the cohort is therefore just (z + np + mk)/(n+m)
Because j, g < 1, we can observe that summing g^(t) and λj^(t) across all t gives us convergent geometric series, and we can in turn gauge average LTV within each type of customer (note that Type M has unambiguously lesser LTV):
LTV for Type N: 1 / (1 - g)
LTV for Type M: λ / (1 - j)
Average LTV across types is merely a weighted average: n/(n+m) * 1/(1-g) + m/(n+m) * λ / (1 - j)
Just to revisit what’s important here — Saasco has the option to sell to two kinds of customers. One kind of customer is unambiguously worse on practically every measure that matters than the other.
What happens if we sell only to the good customers?
Suppose the heads of finance and sales have put their heads together and observed that Type M customers are unambiguously worse than Type N customers. They decide not to sell to Type M — as a matter of policy — for a quarter. If we take the liberty of assuming the business is otherwise identical, then the model simply sets m = 0, and we can get the following observations:
The starting ARR of the cohort: n
Average ARR per customer is $1
At any number of years elapsed since acquisition t, the retained ARR of the cohort is ng^(t)
The total sales and marketing expenditure to acquire the cohort is k + np
The blended CAC for the cohort is therefore just (z + np)/(n)
Because g < 1, we can observe that summing g^(t) and across all t gives us a convergent geometric series, and we can get the following average LTV: 1 / (1 - g)
Strangely, they realize that the economics of the business have gotten worse. Everyone expected the business to be smaller — but quite surprised to see that kicking out the worse-than-average customers has diminished their average LTV:CAC ratio.
How does this happen?
What could have gone wrong selling only to the good customers?
I want to start by setting up an inequality that identifies the scenarios under which LTV:CAC worsens. Simply put, I want to compare LTV:CAC when m > 0 to LTV:CAC when m = 0.
To be extra clear here:
LTV | m > 0: n/(n+m) * 1/(1-g) + m/(n+m) * λ / (1 - j)
CAC | m > 0: (z + np + mk)/(n+m)
LTV | m = 0: 1 / (1 - g)
CAC | m = 0: (z + np)/(n)
I don’t think it’s necessary to show the expanded form here. It will be sufficient here to identify one example combination of variables. Shortly, I’ll pick values for n, m, λ, etc. such that we get the desired outcome.
Solving for a more concrete example (part I)
To make my point extremely concrete, I want Type M to have LTV < CAC for m > 0, n = 0. It is very easy to imagine business analysts (wrongly) making the case against customers for whom LTV < CAC.
Recall that LTV for Type M is just λ / (1 - j)
And the blended CAC in a world where m > 0 is (z + np + mk)/(n+m). Setting n = 0 gives us a CAC of (z + mk)/m = z/m + k.
Then we only need to find λ, j, z, m, k such that λ / (1 - j) < z/m + k.
I’ll (somewhat arbitrarily) pick figures as follows:
Set λ = 0.8
Set j = 0.8
Then we get the following expression:
0.8 / (1 - 0.8) < z/m + k
0.8 / 0.2 < z/m + k
4 < z/m + k
We know in this case we want m > 0, so I’ll just say set m = 5, and set k = 1
Then we can say 4 < z/5 + 1, which in turn implies z > $15. So let’s just set z = $20.
In summary, I’ve picked a set of variables where building a business solely out of five Type M customers would lose money. These variables are as follows:
Set λ = 0.8
Set j = 0.8
Set m = 5
Set k = 1
Set z = $20
Solving for a more concrete example (part II)
Recall we set up the following expressions for LTV and CAC, varying whether Saasco allows itself to sell to Type M customers.
LTV | n, m > 0: n/(n+m) * 1/(1-g) + m/(n+m) * λ / (1 - j)
CAC | n, m > 0: (z + np + mk)/(n+m)
LTV | n > 0, m = 0: 1 / (1 - g)
CAC | n > 0, m = 0: (z + np)/(n)
We can plug in the values we just used in describing Type M customers to simplify these expressions:
LTV | n, m > 0: n/(n+5) * 1/(1-g) + 5/(n+5) * 4
CAC | n, m > 0: (20 + np + 5)/(n+5)
LTV | n > 0, m = 0: 1 / (1 - g)
CAC | n > 0, m = 0: (20 + np)/(n)
We can now set up the inequality in a slightly more legible form. Again, we want n, g, p such that [LTV | n, m > 0] / [CAC | n, m > 0] > [LTV | n > 0, m = 0] / [CAC | n > 0, m = 0].
This is identical to saying the following:
[n/(n+5) * 1/(1-g) + 20/(n+5)] / [(25 + np)/(n+5)] > [1 / (1 - g)] / [(20 + np)/(n)]
We know we need g > j, so we can just say g = 0.9.
This gives us further simplification in our inequality:
[n/(n+5) * 10 + 20/(n+5)] / [(25 + np)/(n+5)] > [10] / [(20 + np)/(n)]
As it turns out, we can simplify the above even further: n(50-20p) > 400.
If we suppose that p = k, i.e. that the average Type N customers has equal variable acquisition costs to Type M customers, we can say p = 1 , which gives us n(30) > 400, or n > 40 / 3. So let’s just pick n = 14.
Solving for a more concrete example (part III)
To recap again, Saasco has the option to sell to two distinct types of customers: Type M and Type N. Against nearly every measure, Type M is worse than Type N. Moreover, we’ve set up the model such that Type M customers alone are unambiguously lossmaking. These are terrible customers. Nonetheless, layering these customers onto the good customers improves unit economics across the business.
We previously made the following observations about Saasco:
LTV | n, m > 0: n/(n+m) * 1/(1-g) + m/(n+m) * λ / (1 - j)
CAC | n, m > 0: (z + np + mk)/(n+m)
LTV | n > 0, m = 0: 1 / (1 - g)
CAC | n > 0, m = 0: (z + np)/(n)
LTV | n = 0, m > 0: λ / (1 - j)
CAC | n = 0, m > 0: (z + mk)/m
We also uniquely identified each of our variables:
Set λ = 0.8
Set j = 0.8
Set g = 0.9
Set m = 5 if sold any m
Set n = 14 if sold any n
Set k = p = $1
Set z = $20
Plugging these in gives us the following scenarios:
If we sell only Type M:
LTV = λ / (1 - j) = 0.8 / 0.2 = $4
CAC = (z + mk)/m = ($20 + $5) / $5 = $5
LTV:CAC = 0.8
If we sell only Type N:
LTV = 1 / (1 - g) = $10
CAC = (z + np)/(n) = ($20 + $15) / 14 = $35 / 14 = $2.50
LTV:CAC = 4
If we sell both Type N and Type M:
LTV = n/(n+m) * 1/(1-g) + m/(n+m) * λ / (1 - j) = (14/19) * ($10) + (5/19) * ($4) = $7.50 + $1 = $8.50
CAC = (z + np + mk)/(n+m) = ($20 + $14 + $5) / (19) = $2.05
LTV:CAC = $8.50 / 2 = 4.1
So even on average adding the loss-making Type M customers improves economics. This is not an obvious result.
The things that matter
Companies frequently try to avoid selling to bad customers. The expectation is that joint optimization of deal sizes, margins, and retention metrics will nudge a company closer towards profitability. Certainly, one would expect, eliminating customer segments with miserable LTV:CAC would improve LTV:CAC across the whole business. But this isn’t always true.
Operationalizing this (mistaken) tactic frequently uses one of the following policies:
Not selling deals to non-executive buyer personas
Setting minimum deal size thresholds
Instituting a ban on month-to-month billing terms
Setting gross margin floors on deals
Implementing decelerators on deals sold to less-than-optimal customer segments
I hope to have argued that these policies can be self-defeating. SaaS companies generally have enormous GTM fixed costs (e.g. rents, salaries, vendor obligations, some share of event marketing, etc.). Even deals with relatively poor economics can pay down this fixed cost burden; this in turn improves economics across the entire business.
With that in mind, it becomes somewhat straightforward to realize that margins, retention rates, contract sizes, and all the other go-to-market SaaS metrics are merely useful tools. They are not outcomes unto themselves.
The only outcome that matters is to deliver cash returns in a reasonably expeditious fashion. The optimal path to cash returns may run through sales to bad customers.