You model the price rise at ten percent across the base, assume ten percent churn, and the arithmetic clears comfortably. Revenue up, a few accounts lost, net positive.
Then the three that leave are the three you show investors. You modelled how many would go and never modelled which ones, and only the second question had an answer.
Why these three models
The decision is whether to raise prices on customers you already have, and by how much on whom. The features that fire are a response driven by the customer’s private information rather than by chance, an installed base where value is concentrated in a few accounts, and an increase that spends something accumulated.
Three lenses. Adverse selection produces an equilibrium answer about who leaves. Concentration produces a random answer about why the loss is not proportional to the count. Reputation decay produces a cycle answer about what an increase costs beyond the accounts it loses. The first two together are the reason a model that clears on averages can be badly wrong.
1. The leavers are selected, not sampled
A price rise is not a lottery run across your base. Each customer decides using information you do not have: what alternatives they hold, what switching would cost them, how much slack is in their budget, whether someone internal has been arguing to replace you anyway.
That is a selection process, and it runs in a direction you should expect. The accounts that leave are the ones with the best outside options, which typically means the larger and more sophisticated buyers, the ones with procurement functions, the ones a competitor has already approached. The accounts that stay are the ones with the fewest alternatives, which is not the same population.1
Notice the awkward consequence. If the ones with options leave, your remaining base is more locked in than before, which reads on the dashboard as improved retention. You will record a price rise that increased both revenue and retention, and the mechanism producing the second number is that your most mobile customers already left.
Building that alternatives column is a conversation, not a research project. Ask whoever sits in each account two questions. Who else did they evaluate before choosing you, and who has approached them since. Sales usually knows both and has never been asked to write them down. Where nobody knows, the absence is the answer for that account: a relationship in which you cannot name the competing options is one you have not been close enough to see.
A second column belongs next to it and is not the same thing. Some accounts have no substitute for you and still cannot absorb an increase, because the budget is set by somebody who has never heard your name. A distributor working to a fixed margin, a school on a fee schedule published a year in advance, an NGO on a signed grant line: none of them have alternatives and all of them will cut scope rather than pay more. Captive and able to pay are different properties, and collapsing them is how a carefully differentiated increase still loses revenue.
The instruction is to model the composition rather than the count. Before setting a number, sort the base by how many real alternatives each account has, which you already partly know from your own sales conversations. Then ask what the base looks like after the accounts in the top group leave. If the answer is a smaller, less prestigious, more captive set, you are not raising prices, you are trading the quality of your customer list for margin, and that may still be correct but it should be a decision rather than a discovery.
2. The loss is concentrated, so the count is the wrong denominator
The second lens is arithmetic, and it compounds the first.
Revenue in most customer bases is heavily concentrated: a small number of accounts contribute a large share of the total, so the mean account is far larger than the median one and the average describes almost nobody.2 Losing ten percent of your accounts therefore tells you nothing about how much revenue you lost. It could be two percent or it could be forty.
Combine that with selection and the exposure is specific. If value is concentrated in your largest accounts, and your largest accounts are also the ones with the most alternatives, then the selection process is aimed directly at the concentration. That is the case worth modelling and it is exactly the case an average-based model hides.
Run the increase against your top five accounts individually, because the base does not have a typical customer.
Do the arithmetic in public, because it usually reverses the decision. Take a base of two hundred accounts where the top five carry fifty five percent of revenue. A ten percent increase across everyone, with ten percent of accounts leaving, reads as a nine percent gain if you assume the leavers are average. They are not average. Lose one of the top five and you have given up eleven percent of revenue against a ten percent gain on the rest, and the exercise is negative before you count what replacing that account costs.
So write it on one page: the increase, and beside it your single largest account, with the question of whether the gain survives losing that one relationship. If it does not, the price rise is a bet on one customer and should be discussed as one, by the person who holds that relationship rather than by the person who built the model. Founders who run this generally find the decision was never about pricing policy at all.
Do not reach for a distribution here. The operational content is simply that you should compute the loss account by account for the largest ones rather than applying a churn percentage to a total, and that you should do it for the specific accounts rather than for a representative one. Five rows of arithmetic replace a model that cannot be right.
3. What the increase spends
The first two lenses cost you accounts. The third costs you something that does not appear in the churn number at all.
Standing with a customer behaves like a flow rather than a balance. It is topped up by delivered outcomes and drawn down by asks, and it decays on its own when neither happens. A price rise is a draw, and its size depends on what has been paid in recently rather than on what was paid in over the whole relationship.3
Which explains a pattern founders find puzzling: the same increase lands quietly with one customer and badly with another of similar size and tenure. The difference is usually not the number. It is whether anything was delivered in the preceding quarter that the customer noticed, and whether the last three interactions were requests or contributions.
The size of the increase also matters less than what it is measured against. The same absolute change reads as routine indexation next to one reference point and as a betrayal next to another, and you have some control over which reference is in the customer’s head when the email arrives.4 Announce what changed alongside the price, or the price is the only thing that changed.
The operational form is a sequencing rule rather than a communication tactic. Do not raise the price in a quarter where the last thing you delivered was an apology. Schedule the increase to follow something the customer can point to: a capability they asked for, a resolved problem, a documented result. This is not manipulation and it is not a discount in disguise. It is the recognition that you are drawing on an account and that the account has a balance you can observe.
The same logic sets the notice period. A long notice period is itself a top-up, because it hands the customer time and control, which is worth something to them and costs you only patience.
There is a version of the sequencing rule for founders with nothing to point to, and it is not to invent something. If the last quarter delivered nothing the customer noticed, say so internally and delay the increase by a quarter while you fix that. A price rise announced into an empty account is the most expensive available way to discover the account is empty.
Then plan the response before you send. Roughly a third of accounts will ask for something rather than accept or leave, and what they ask for is usually a longer term, a payment schedule or a reduction in scope. Decide in advance which of those you will grant, because deciding it account by account under pressure produces a set of inconsistent terms that your customers will eventually compare with each other. The concession worth pre-approving is almost always time: a longer notice period, or the increase staged over two steps six months apart, costs you a few months of the gain and keeps the relationship intact. The one to refuse is a permanent exception, which converts a price rise into a discount you then administer forever and cannot withdraw without running this whole decision again.
What the three say together
- Sort the base by how many real alternatives each account has, using what your sales team already knows.
- Run the increase account by account for your top five, not as a percentage against a total.
- Ask what the base looks like after the mobile accounts leave, and decide whether you want that base.
- Sequence the increase to follow a delivered outcome, and give notice that is longer than convenient.
Where they disagree
The selection argument and the concentration argument point to opposite pricing structures.
Selection says raise prices where alternatives are weakest, which means differentiating: a larger increase on the captive accounts and none on the mobile ones. Concentration says your revenue lives in a handful of large accounts, and those are precisely the mobile ones, so a policy that spares them is a policy that raises prices only where there is little money.
Both are correct and together they describe the actual trap: the accounts you can raise prices on without losing them are the ones where the increase does not matter, and the accounts where it would matter are the ones you would lose. Founders discover this halfway through and resolve it by applying a uniform increase, which is the one option neither model supports.
The way out is usually not a price change at all. If your large accounts have alternatives and your small ones do not, the binding problem is that you are not differentiated for the buyers who matter, and a price rise is a test of that rather than a fix for it. Run the increase on a small segment first and treat the result as a measurement of your position, not as a revenue exercise.
What none of them contain
None of the three sees the customers who never arrive. A higher price changes who considers you at all, and that effect shows up over the following year in a pipeline that looks slightly different for reasons nobody attributes to the pricing decision made twelve months earlier.
None of them handles the customer who stays and reduces scope. Churn analysis is binary and much of the real response is neither leaving nor accepting: it is dropping two seats, declining the upgrade, or slowing payment. That response is invisible to a model that counts logos and it is often larger than the churn.
And one property the ensemble will not produce: your customers talk to each other. All three lenses treat each account as deciding privately on its own information. In a small market they compare notes, and a price rise handled badly with one account arrives at the next one before your email does.
The one action that survives the ignorance: before you set the number, write your top five accounts in a row with two columns, what they pay and how many credible alternatives they have. If your revenue and your customers’ options are concentrated in the same rows, the increase is a test of your differentiation and should be run on a small segment first rather than announced to the base.
Who has to move
This is the founder’s call and the input has to come from whoever sits in the accounts, because the alternatives column is knowledge that lives in sales conversations and nowhere else in the company. The instinct is to model the increase in the spreadsheet, where every customer is a row of the same height and a churn assumption applies to all of them equally. The cheapest first test is the five-row grid, which takes one conversation to assemble and usually shows that the churn percentage in the model was answering a question nobody had.
Sources and notes
- Jeffrey Carpenter and Andrea Robbett, Game Theory and Behavior, MIT Press. Adverse selection, in which a change in terms alters the composition of the population that accepts them because the response depends on private information the other side holds, is developed in the chapters on asymmetric information. Used in section 1. The classic form concerns insurance and lending; the application to a price rise on an installed base is the same mechanism, with the customer’s outside options in the role of the private information.
- Albert-László Barabási, Network Science, Cambridge University Press, on heavy-tailed distributions and the divergence between mean and median where a small number of cases account for a large share of the total. Used in section 2 for the general property only. No functional form is claimed for any customer base, and none is needed: the operational content is to compute the largest accounts individually rather than to apply a rate to a total. On why the stronger label should be avoided, see Anna D. Broido and Aaron Clauset, Scale-free networks are rare, Nature Communications 10, 2019, article 1017, https://www.nature.com/articles/s41467-019-08746-5, which reports that only a small minority of nearly one thousand network datasets show the strongest evidence of scale-free structure.
- Martin W. Cripps, George J. Mailath and Larry Samuelson, Imperfect Monitoring and Impermanent Reputations, Econometrica 72(2), 2004, pages 407 to 432, developed at book length in George J. Mailath and Larry Samuelson, Repeated Games and Reputations: Long-Run Relationships, Oxford University Press, 2006. The result is that under imperfect monitoring a reputation is temporary rather than permanent, so standing must be replenished rather than accumulated once. Used in section 3 for treating goodwill as a flow with a recent balance rather than as a lifetime total.
- Sanjit S. Dhami, The Foundations of Behavioral Economic Analysis, Oxford University Press. Reference dependence, under which a change is evaluated against a reference point rather than in absolute terms, is why the size of an increase matters less than what it is measured against and what accompanies it. Used to support the sequencing and notice-period recommendations in section 3. Direction only; no magnitude is claimed, and the loss-aversion coefficients quoted in business writing are contested.
A note on a number this article does not give. There is no safe percentage. The tolerable increase depends on how differentiated you are for the specific accounts that carry your revenue, which is exactly what the five-row grid measures and exactly what no benchmark can supply. Get the number from a segment test, not from a survey of what other companies did.
Joshua Agonya Pi’Rwot, Founder.