The avoided conversation has a specific shape. The founder already suspects the unit does not work yet, and already suspects that the growth line on the slide is being paid for by something nobody has priced. Volume is the reflex. Volume is also a multiplier, and it carries the sign your contribution margin already has.
Three questions settle it. What does one unit actually earn after every cost that would vanish if that unit had not happened. Who sets the input prices that make the answer look survivable. And which of those two numbers have you arranged not to compute.
Why these three models
This piece runs the Wire Model: score the features of the decision, route to a small ensemble of formal models, then force the ensemble to produce dated actions. The scores that mattered:
- Single-lever equilibrium shift (0.9). The decision is a one-variable move. Push volume, or fix the unit. Trace where each one lands.
- Regime-break risk (0.8). Fuel, foreign exchange, mobile-money tariffs, the price of capital. Each one is set outside your company and changes on somebody else’s calendar.
- Cognitive and affective distortion (0.8). The number is avoided rather than missing, and avoidance has a documented shape.
- Historical-analog density (0.7). Scaling a below-cost unit while a third party covers the gap is one of the best-recorded failure structures in venture.
- Strategic actors with asymmetric information (0.4). Real, but small. The founder is mostly withholding from the founder.
That routes to comparative statics (the arithmetic), regime-switching (the structure), and behavioral (the attention). They span three outcome types, equilibrium, cycle-regime and complex, so their errors point in different directions and partly cancel.
Two models got folded rather than shipped. Adverse selection sits inside the first card, because the marginal customer is a margin question and comparative statics is already the calculus of margins. The governance layer, the luck-to-skill read on a single good month, sits in the blind spot. Neither earned its own card.
The framework: the unit carries a sign, and volume multiplies it
1. The arithmetic: volume multiplies the sign
Profit is quantity times contribution per unit, minus fixed cost. Volume touches one of those three terms. If contribution per unit sits below zero, every additional unit moves you further from the answer, faster, with more operational load and more working capital tied up.
The rebuttal is always scale. Fixed-cost absorption is real and bounded, a one-time gain with a ceiling you can compute today. Variable-cost reduction is real and conditional: it needs a named mechanism at a named volume. A freight rate that steps down at 40 tonnes a month. A route that cuts kilometres per drop once it carries 18 stops. Name the mechanism and the trigger volume, or you have a wish.
Then the harder correction. The marginal unit is worse than the average unit, and volume is bought at the margin.
eBay ran large-scale field experiments turning paid search on and off. Standard observational methods produced returns on investment above 4,100% with no controls, and above 1,400% with time and geographic controls. The experimental estimate was negative 63%, with a 95% confidence interval running from negative 124% to negative 3%. Brand keyword ads showed no measurable short-term benefit. New and infrequent users responded to ads, while frequent users, who were buying anyway, absorbed most of the spend.1
Read that as a unit-economics finding. The acquisition cost on your dashboard averages customers you bought with customers who arrived anyway. The incremental cost of the next customer is a different, worse number, and it is the only one volume touches.
Selection also degrades the asset that produces your organic demand. Groupon deals matched against Yelp reviews showed reviews mentioning daily deals carrying star ratings roughly 10% lower than other reviews, and merchant ratings dropping by an average of 0.12 after an offer.2 Small, until you price it. For independent restaurants, a one-star increase in Yelp rating is worth 5% to 9% more revenue.3
Volume buys more of a negative unit, faster, from customers who like you less.
Comparative statics, the arithmetic lens
Assumes: input prices, mix and demand hold still while you move one variable.
Fits because: single-lever equilibrium shift scored 0.9. The choice is volume or margin, and the sign decides it.
Breaks when: the ceteris paribus fails. A genuine step change in cost, a new supply contract, or a technology that resets variable cost makes the static trace wrong in magnitude and sometimes in direction.
Counteracts: the belief that growth is evidence of a working business.
May reinforce: premature harvesting, and killing a market you were three months of density away from serving profitably.
2. The regime: someone else prices the input that makes your unit work
Comparative statics holds everything else still. The regime model asks what happens when everything else moves, because the terms holding your unit up are set by parties who never asked your opinion.
Kenya makes the mechanic visible. On 16 March 2020 the Central Bank of Kenya waived charges on transactions between mobile money wallets and bank accounts. Between March 2020 and October 2022, monthly volume between payment service providers and banks rose from 18 million transactions worth about Ksh.157 billion to over 113 million worth Ksh.800 billion, increases of 527% and 410%. Charges returned, at reduced levels, on 1 January 2023.4 Every business whose take rate lived inside that free window had a core line item sitting on a regulator’s calendar.
Nigeria makes it faster. On 29 May 2023 the incoming president said the fuel subsidy was gone. By that Wednesday, petrol at NNPC stations in Abuja moved from 195 naira a litre to 537 naira, close to three times the price.5 Every delivery, logistics and agent-network unit costed at 195 was repriced overnight by one sentence in a speech.
Currency is the same class of variable, and it separates the two numbers founders keep conflating. Jumia’s full year 2024: orders reached 22.7 million, up 6% year on year, while gross merchandise value fell 4% as reported to $720.6 million and rose 28% in constant currency. Revenue fell 10% as reported to $167.5 million and rose 17% in constant currency. Loss before income tax was $97.6 million against $98.6 million the year before.6
More units. More local-currency value. A smaller company in the currency the investor pays in, and a loss that barely moved. You operate in one regime and report in another, and only one of them reaches a term sheet.
Your break-even is conditional on a regime that someone else can end with a press release.
So compute the unit twice. Once at today’s input prices, once at the prices that applied before the current subsidy, waiver, peg or promotional rate. The gap is your exposure, and it has an owner: a regulator, a telco, a processor, a landlord, or the fund that has not yet said no to your bridge.
Regime-switching, the structure lens
Assumes: the world runs in states with different cost matrices, and transitions are exogenous to you.
Fits because: regime-break risk scored 0.8, and the 2020 to 2023 window supplied three clean transitions.
Breaks when: the switch is endogenous. If your own scale is what pulled the input price down, treating it as a regime hands away a lever you actually hold.
Counteracts: extrapolating a cost base you were lucky to have.
May reinforce: fatalism about pricing, and waiting for policy instead of renegotiating with a supplier.
3. The avoidance: the founder declines the number
Information enters the utility function directly, which creates an incentive to avoid it even when it is free, useful, and carries no strategic cost.7 That is a documented result, and it has been measured on real money. A panel of daily online account logins at Vanguard across 2007 and 2008 found logins falling by 9.5% after market declines, with attention dropping further when volatility rose.8 People look at their money less precisely when looking matters more.
The founder version is procedural, which is what makes it invisible. Cohort files that stop getting refreshed. Acquisition cost blended across paid and organic so the paid number never stands alone. Last-mile cost booked to operations rather than to the order. A denominator that drifts from delivered orders to gross merchandise value, because one of them grew. Each is a decision about what to leave uncomputed, made once, then inherited by every deck afterwards.
Investors read the avoidance faster than they read the number. A founder who needs two weeks to produce contribution margin per order has answered the question already, and the answer travels.
A metric you cannot produce in ten minutes is a metric you have decided not to know.
Behavioral, the attention lens
Assumes: people deviate from rational information use in stable, predictable directions.
Fits because: cognitive and affective distortion scored 0.8, and avoidance rises exactly when the news is bad.
Breaks when: the metric is computed automatically and reviewed on a fixed calendar by someone other than the founder. Structure removes most of the effect.
Counteracts: the assumption that a missing number is an accident of bandwidth.
May reinforce: treating every accounting choice as motivated, which burns trust inside a finance team that made a defensible call.
GEER: the levers, cheapest first
Five channels carry the exposure: price, variable cost per unit, mix and selection, regime dependence, and measurement. Pull the cheap, reversible levers first.
- Freeze the unit in one sentence. One delivered order. One active agent-month. One disbursed loan. Write it down and stop renegotiating it. Hits measurement. Costs an hour.
- Compute contribution on last month’s actuals. Revenue per unit minus every cost that would have disappeared if that unit had not happened, from the ledger and the bank. Costs a day.
- Reprice the regime inputs at pre-subsidy levels and put both numbers on one page. Hits regime dependence. Costs half a day.
- Split cohorts by acquisition price. Organic, paid, discounted. Compare repeat rate and refund rate across the three. Hits mix. Costs a day.
- Run one holdout. Switch off paid acquisition in one city for two weeks and count what still arrives. Hits measurement. Costs two weeks and some nerve.
- Raise price on one segment. The fastest positive move on contribution, and the only one that needs nobody else’s agreement. Hits price. Costs one segment’s goodwill.
- Renegotiate the largest variable cost with a written volume trigger. Hits variable cost. Costs weeks.
- Retire the worst geography or SKU. Hits mix, hard. Costs revenue and morale.
No-lever flag: if contribution stays negative and no lever above moves it inside two quarters, you are running a research project that happens to have customers. That is a legitimate thing to run. It needs to be financed as one and said out loud to the people funding it.
RADAR: the portfolio, dated
DO NOW, by T+3 days. Reversible, and dominant across every scenario.
- Write the unit definition in one sentence. Have your finance lead sign it.
- Compute contribution margin per unit for the last complete month from ledger and bank data.
- Build the subsidy register. Every input in that unit whose price is set by someone else, with the entity that sets it and the earliest date it can change.
- Recompute the unit at pre-subsidy input prices. One page, both numbers side by side, dated.
HEDGE, by T+14. Cheap insurance against the arithmetic being worse than you think.
- Run the paid-acquisition holdout in one channel or one city.
- Test a price rise on the segment with the highest willingness to pay and the lowest cost to serve. Test it, do not announce it.
- Get your largest variable-cost supplier to put a volume-triggered rate in writing, even unsigned.
- Produce the cohort split by acquisition price and read the discounted cohort’s repeat rate against the organic cohort’s.
DEFER AND TRIGGER. Irreversible, so wait, and pre-commit the observable trigger now.
- Defer: a new city, a new category, a step-up in paid acquisition, a warehouse lease, any headcount that scales with volume.
- Trigger to scale: contribution margin per unit positive for two consecutive months on the frozen definition, and the discounted cohort repeating at no worse than 80% of the organic cohort. Then spend, in that order.
- Counter-trigger: at T+28, contribution still negative and no lever has closed more than a fifth of the gap. Hold volume flat, stop buying growth, and re-plan the quarter as a margin project.
If you are the investor. DO NOW: ask for the unit definition in writing before you accept a single number. HEDGE: request one month of order-level raw data and recompute it yourself. DEFER: valuation, until you have seen the unit priced at pre-subsidy input costs and at a capital cost the company will actually face.
CHAIN: what usually happens next
Match on structure, not sector. The reference class is companies that scaled a unit priced below its delivered cost while a third party covered the gap. Daily deals, meal kits, ride-hailing, informal-retail distribution, agent networks and last-mile e-commerce all sit in it.
The base rate is current. Of 431 venture-backed companies that shut down since 2023, 385 had identifiable causes: ran out of capital 70%, poor product-market fit 43%, bad timing or macro conditions 29%, unsustainable unit economics 19%. The median company had raised $11 million.9
Read the top and fourth lines together. Running out of capital is the death certificate. Unsustainable unit economics is the cause of death. They are one event observed at two distances, which is why the first number is large and comfortable and the second is small and specific.
Second order, with a local case. Copia Global spent twelve years building order density across Kenyan informal retail, raised more than $123 million, entered administration in May 2024 after failing to close new funding, and was wound down.10 Volume was never the input that was missing.
Third order: survivors reprice, the benchmark contribution margin resets upward, and the next founder still carrying an unpriced subsidy finds the round harder than the last cohort did. Present-state modifiers push the same way. Capital is dearer, and the 2023 Nigerian fuel and currency moves plus the Kenyan tariff normalisation hit cost bases inside the same eighteen months.
Subtract the counterfactual before you over-diagnose. Some of what presents as a unit-economics failure is a demand failure wearing a cost mask. If nobody buys at the price that clears, cost work will not save it, and the honest answer is a different product.
Matrix-break flag. Cheap capital was itself a regime, and all three models assume input prices move slowly. When capital reprices, the tolerated subsidy shrinks and every company that was one scale story away gets reclassified in the same quarter. The other break runs the opposite way: if automation genuinely collapses a variable cost line, the base rate under-predicts you. Test that claim the way you test the rest. Name the cost line, the percentage, and the month it appeared in the ledger. Absent all three, the base rate stands.
What this ensemble cannot see
Four things, and they are large.
Magnitude. Comparative statics gives direction reliably and size badly. It will tell you a price rise helps. It will not tell you how much volume walks.
Your elasticity. Willingness to pay is local, segment-specific, and unobservable from a spreadsheet. Every number in this piece came from someone else’s market.
Luck in a good month. One month of positive contribution is one draw. Early-stage results sit close to the luck end of the luck-to-skill continuum, so two consecutive months is the minimum evidence, and even that regresses.
The subsidies still unnamed. Unpaid founder salary. A co-founder’s savings. A rider absorbing waiting time nobody bills for. Sixty-day supplier credit extended free because the relationship is old. A grace period a regulator has not yet chosen to end. These are the ones that never reach a register, because nobody sends an invoice for them.
Do the thing that survives all four. By T+3, write the unit in one sentence and compute it twice, at today’s input prices and at pre-subsidy prices. Put both on one page and date it. If the second number is negative, you have just chosen your quarter.
Sources and notes
- Blake, T., Nosko, C., and Tadelis, S. “Consumer Heterogeneity and Paid Search Effectiveness: A Large Scale Field Experiment.” Econometrica 83(1), 2015, 155-174. Working paper with the full ROI tables: NBER w20171. The OLS estimates above 4,100% and 1,400%, and the experimental estimate of negative 63% with a 95% interval of negative 124% to negative 3%, are from the paper’s return-on-investment section and Table 4. Author copy: Berkeley Haas.
- Byers, J. W., Mitzenmacher, M., and Zervas, G. “Daily Deals: Prediction, Social Diffusion, and Reputational Ramifications.” 2012. Full text. The 10% lower ratings among reviews mentioning daily deals, and the average 0.12 star drop after an offer, are in the reputational-ramifications section.
- Luca, M. “Reviews, Reputation, and Revenue: The Case of Yelp.com.” Harvard Business School Working Paper 12-016. Full text. A one-star increase in Yelp rating leads to a 5% to 9% increase in revenue, driven by independent restaurants.
- Central Bank of Kenya, “Reintroduction of Charges for Mobile Money Wallet and Bank Account Transactions,” press release, 6 December 2022. Full text. Waiver dated 16 March 2020; PSP to bank monthly volumes and values, and the 527% and 410% increases, are as stated in the release; charges effective 1 January 2023.
- Princewill, N. “Gas prices nearly triple in Nigeria as new leader triggers panic-buying by halting subsidies.” CNN, 1 June 2023. Report. The Abuja NNPC retail move from 195 naira to 537 naira a litre is quoted directly from the piece.
- Jumia Technologies AG, “Jumia Reports Fourth Quarter 2024 Results,” 20 February 2025. Jumia’s investor site and the SEC-hosted copy of the release both refuse a standard browser request; the full release text, including the full-year table, is mirrored at Value The Markets. Context on the South Africa and Tunisia exits: TechCabal.
- Golman, R., Hagmann, D., and Loewenstein, G. “Information Avoidance.” Journal of Economic Literature 55(1), 2017, 96-135. Full text.
- Sicherman, N., Loewenstein, G., Seppi, D. J., and Utkus, S. P. “Financial Attention.” Review of Financial Studies 29(4), 2016, 863-897. Panel of daily online account logins at Vanguard across 2007 and 2008. Full text. The 9.5% fall in logins after market declines is in the abstract and Section 3.
- CB Insights, “The Top Reasons Startups Fail,” analysis published 5 March 2026. Report. 431 venture-backed shutdowns since 2023, 385 with identifiable causes; percentages sum above 100 because companies carry multiple causes.
- “Copia enters administration after failing to secure funding.” TechCabal, 24 May 2024. Report. Founded 2012, more than $123 million raised across seven rounds, administration May 2024.