Four customers signed in six weeks after nine months of nothing. Nobody ran a pilot. Two of them named the same reference account without being prompted.
That is not the market discovering you. That is four people copying one decision, and the one decision may have been wrong.
Why these three models
The decision is what to conclude from a sudden run of adoption, and what to spend on the strength of it. The features that fire are buyers who can see each other’s choices, a weak private signal about whether you are any good, and a short sequence of outcomes that looks like a trend.
Three lenses. Cascades produce a complex answer about why the run started and why it can stop just as abruptly. Statistical discrimination produces an equilibrium answer about what buyers are actually pricing when they cannot evaluate you directly. Luck versus skill produces a random answer about how much of the run means anything at all. The first explains the shape. The second explains why you were cheap to copy. The third stops you from spending against it.
1. After a short run, buyers stop using their own information
An information cascade forms when people decide in sequence, each can see what the earlier ones chose but not why, and each has only a weak private signal of their own. Past a small number of visible choices in one direction, the rational move for everyone afterwards is to follow, and to ignore what they themselves know.1
Two properties of this matter commercially and both are counterintuitive.
The first is that cascades start on almost nothing. It does not take a majority. A short early sequence in one direction is enough to tip everyone who comes later, which means the run you are watching may rest on two decisions made for reasons neither person would defend.
The second is that cascades are informationally thin. Once people are copying rather than evaluating, their choices stop adding evidence. Two hundred adopters in a cascade contain roughly as much information about whether the product is good as the two who started it, because the other one hundred and ninety eight were not looking at the product. This is why cascades reverse so violently. A single credible public fact can flip the whole chain, since there was never much underneath it.
The founder-facing test is a question you can ask directly. When a new customer signs, ask what they looked at before deciding. If the honest answer is who else uses it, you are inside a cascade and your adoption numbers are not evidence about your product. If the answer is a trial, a spreadsheet, or a specific comparison they ran themselves, the number means what you think it means.
Finding that you are inside a cascade tells you what to do next, and it is not to celebrate. A chain you did not build reverses on one credible public fact, so the work is to give each copier a reason of their own before that fact arrives. Have every one of them run something small they chose: their own numbers before and after, a pilot in one branch, anything that leaves them holding evidence rather than a recommendation. It costs your team time in accounts that have already signed, which is why almost nobody does it, and it is the only thing still standing when the chain breaks.
Find out where the copying happens, because in most East African sectors it is a specific room rather than a market. An association WhatsApp group. A monthly meeting of procurement officers. Two or three people whose view carries the sector. That concentration is why the run started on almost nothing, and it is also why one bad implementation reaches every prospect you have inside a week. If you can name the room, you can name the account whose failure would cost you the rest of them, and that account gets your best engineer rather than your newest one.
2. What they are actually pricing is your category
Cascades explain the sequence. They do not explain why buyers had such a weak private signal in the first place, and the answer to that is usually structural rather than accidental.
When evaluating an individual is expensive and unreliable, decision-makers substitute the average of an observable group. That is statistical discrimination, and it is not the same thing as prejudice; it is a rational response to a screening cost, which is precisely what makes it durable.2 The buyer who has never bought from a two-year-old company in your city is not evaluating you. They are pricing the category you visibly belong to.
The consequence founders miss is that this cuts both ways in the same market. If your category’s average is poor, your individually excellent product is discounted to the average and you cannot argue your way out, because the buyer’s whole point is that they do not trust individual claims. If your category is currently fashionable, you are being upgraded to an average you did not earn, and that is what a lot of easy early traction actually is.
You are not being evaluated, you are being averaged, and the fix is a signal the average cannot fake.
The way out is not better positioning. It is a costly, verifiable act that a bad member of your category could not afford to imitate. A guarantee that pays out. A reference the buyer picks rather than one you supply. An audited number. Anything that is expensive precisely in proportion to how untrue it would be. Cheap claims cannot separate you from the average, because the average makes those claims too.
The costly signal has a failure mode that is expensive to learn late. A guarantee that pays out attracts the buyers most likely to claim it, and written loosely it draws claims from the accounts that were never going to implement anything. Scope it to something you control: a refund tied to a delivery milestone that is yours, not to a business outcome that depends on the customer doing their half. It still costs a weak imitator more than it costs you, which is the entire requirement, and it stops you funding the buyers who learned nothing.
If an audited number is out of reach this year, the cheap partial version is one customer who will take an unscreened call. Agree in writing that any prospect may ring them without telling you first, then give prospects the number instead of arranging an introduction. You have given up control of the reference, which is exactly why it carries information. It costs nothing, and a competitor with a weaker product cannot copy it, because they would not survive the first call.
3. Most of the run is noise
The third model is the one nobody wants applied to their own results.
Any short sequence of outcomes mixes skill and luck, and the weight is not intuitive. Where the outcome depends heavily on chance, sequences that look like clear trends occur routinely with no underlying change at all. The practical consequence is regression: extreme runs are followed by ordinary ones, and the ordinary period gets attributed to whatever changed in between, which is usually something you did.3
Early-stage selling sits high on the chance side of that continuum and founders rarely accept it. The number of things outside your control in any given month is large: a competitor’s outage, a budget cycle opening, one champion changing jobs. None of those are your product and all of them arrive as sales outcomes with your name on them. A run is a measurement of the market’s noise before it is a measurement of you.
Note the direction of the error. Six good weeks after nine bad months feels like a discontinuity, and the temptation is to explain it: the new messaging, the conference, the pricing change. Sometimes it is. In a market with cascade dynamics, a run of four is exactly what the cascade model predicts from a single arbitrary starting choice, so the run is explained twice over before any of your changes are needed as an explanation.
The discipline is to size the sample before interpreting it, and the pull in the other direction is strong enough to be treated as a default rather than as an occasional lapse: people draw confident general conclusions from very few cases, especially when the cases are recent and vivid.4 Four is not a sample. It becomes one when the mechanism is identified, which means at least one of the four bought after evaluating you directly rather than after seeing the others. One independent buyer is worth more evidentially than fifteen copiers, and this is the single most useful reframing in the article.
Turn the split into a spending rule, so it survives an optimistic week. Count last quarter’s signings in two columns and write the ratio down. Four copiers and one evaluator means your evidence base is one, and hiring two more sales people against a base of one is a bet on the room rather than on the product. Three evaluators out of five is different: something is working, and the money goes into finding more buyers who resemble those three, not into amplifying the reference that started the run.
Then decide now what you will say when the run stops, because it will stop and that meeting will otherwise happen unprepared. The default explanation will be that something you changed broke it: the price, the messaging, the new rep. Usually nothing broke and the chain simply ran out of people in the room. Before you reverse any decision, split the stall the same way. If evaluator conversion is flat and only the copying stopped, leave the change alone and go and find the next room.
What the three say together
- Ask each new buyer what they examined before deciding. Copiers and evaluators go in different columns.
- Count only the evaluators when you assess whether the product is working.
- Identify what average you are being priced against, and what costly signal would separate you from it.
- Before spending against the run, ask what the run would look like if nothing about you had changed. If it would look the same, you have no evidence yet.
Where they disagree
Cascade logic and statistical discrimination give opposite advice about early customers.
The cascade view says the identity of your first visible customers is nearly everything, because the chain forms from a short opening sequence, so you should spend heavily to land specific visible names even at a loss. The averaging view says visible names change only the group you are sorted into, and if the underlying category signal stays weak the discount returns as soon as attention moves. One says buy the logo. The other says the logo is a rental.
The reconciliation is about durability. A visible customer bought at a loss is worth it only if it converts into something a competitor cannot copy: a public outcome with a number attached, a reference who takes calls, a contractual commitment. A logo on a page is the cheap version and it decays. The two models agree that early visible customers matter and disagree completely about what you should extract from them, and the disagreement is resolved by what you actually get in writing.
What none of them contain
None of the three distinguishes a cascade from a genuine network effect, and they look identical from inside the company for at least two quarters. In a cascade, each adopter makes the next adopter more likely and nothing else changes. In a network effect, each adopter makes the product more valuable. The observable difference is not in the adoption curve; it is in whether existing customers use the product more as new ones arrive. That is a usage question, not a sales question, and no adoption metric will answer it.
None of them contains the seller’s own effect on the sequence. All three treat you as an object being evaluated. In practice you choose which customers are visible and in what order, which means you are partly writing the sequence that the cascade model treats as given.
And none of them handles the market where buyers talk directly to each other rather than merely observing choices. Communication changes the mechanism entirely: a cascade built on observed actions is fragile, while one built on shared and discussed experience is not, and the second is far more valuable and much slower to build.
The one action that survives the ignorance: for every customer signed in your best recent stretch, write down in one line what they examined before deciding. Then recount your traction using only the ones who examined the product. If that number is small, nothing has gone wrong yet, but you now know which number to grow and it is not the one on the dashboard.
Who has to move
This is a founder and head of sales problem together, because the instinct on both sides of that pairing is the same and it is wrong: a run of wins triggers spending to extend it. The cheapest first test is a question added to the existing sales notes, asked of every new customer, costing one line each. Within a quarter you will know whether you are being chosen or copied, and the two situations call for entirely different budgets.
Sources and notes
- Jeffrey Carpenter and Andrea Robbett, Game Theory and Behavior, MIT Press. Information cascades and the sequential-decision experiments that test them are treated in the chapters on learning and social information. Used here for the two properties in section 1: that cascades can form on a short opening sequence, and that once formed they carry little additional information and are therefore fragile to a single credible public fact. The underlying result is due to Bikhchandani, Hirshleifer and Welch, and to Banerjee, both 1992.
- Jeffrey Carpenter and Andrea Robbett, Game Theory and Behavior, MIT Press. Statistical discrimination, in which decision-makers use group averages as a substitute for costly individual assessment, is developed alongside signalling and screening. Cited in section 2 for the mechanism and for the point that it is a response to screening cost rather than to preference, which is why better claims do not dislodge it and costly verifiable signals do.
- Michael J. Mauboussin, The Success Equation: Untangling Skill and Luck in Business, Sports, and Investing, Harvard Business Review Press. The treatment of where an activity sits on the luck-skill continuum, and of regression to the mean as the consequence, underpins section 3. The specific claim used here is directional: in activities with a large chance component, short sequences routinely resemble trends, and the sample size required before a sequence is informative is larger than intuition suggests.
- Sanjit S. Dhami, The Foundations of Behavioral Economic Analysis, Oxford University Press. The tendency to draw strong conclusions from small samples, and to underweight the base rate in favour of a vivid specific case, is covered in the part on bounded rationality. Used in section 3 for the direction of the error only, with no magnitude claimed.
A note on a number this article does not give. There is no sample size at which a run becomes trustworthy. The threshold depends on how noisy your market is and how visible buyers are to each other, both of which differ by category. What transfers is the split between evaluators and copiers, and that split you can count from next week without any statistics at all.
Joshua Agonya Pi’Rwot, Founder.