A profitable dealership sets out to optimize. Nothing reckless — each quarter, one defensible change. Steer ad spend toward the cheapest cost-per-lead. Pay the floor on units moved. Reward the reps who close fastest and the campaigns that fill the lot for the least money. CAC falls. Volume climbs. Every dashboard agrees it's working.
By Q4, conversion is up and the average deal is quietly shrinking — and the explanation is always available: inventory mix, seasonality, the market. By Q7 the lot is full, CAC has never been lower, and the business has never made less money. The same quarter the metrics peak, the dealership is at its nadir. Goodhart's Law has run to completion: optimize a proxy hard enough and it stops standing in for the thing you wanted.
Here is the part that matters: none of it is about the customer's information problem. Akerlof wrote the lemons market from the buyer's side. The operator sits in the opposite chair — it could see quality if it looked, but it optimized toward the two signals that were easiest to measure and cheapest to move, across both the advertising and the sales floor. Those were never the signals that predicted revenue. The peaches were always in the data. The system was tuned to look away from them.
The repair is not less volume. It is connecting the signals that run the length of the experience — the ad that brought them to the lot, the path they walked, the rep who met them, the terms they took — to the revenue each one finally produced. Find the signals that mark the peaches. Then point the volume at them. That is Performance Revenue Optimization.
Power-law performance is not an accident of any one industry. It is the equilibrium state of any market where customer types vary substantially, information about those types is asymmetric, and acquisition systems cannot distinguish types at the point of sale. These papers were written from the buyer's side of the market. Read them from the operator's side — the side doing the acquiring — and they stop being a theory of why buyers get cheated and become a theory of why acquisition systems, left to optimize freely, select against their own best customers.
In a used-car market, sellers know if their car is a lemon. Buyers cannot tell, so they price every car at the average expected quality. That price is too low to attract sellers of peaches, who exit.
The remaining pool is more lemon-heavy, prices fall further, more peach sellers leave. The market drifts toward lemons by structure — not because anyone behaved badly, but because it cannot distinguish types at the point of sale.
Spence answered adverse selection: high-quality types can credibly separate themselves by sending signals that are differentially costly — easier for them to send than for low-quality types.
Education was the canonical example: the credential is valuable not because it transfers skill, but because completing it is differentially affordable for high-productivity workers. The cost asymmetry is the signal. Funnels can be engineered the same way.
Why does sustained advertising work even when the ads convey no information? Because only firms with strong repeat-purchase economics can afford it. The spending itself is the signal.
This extends signaling to the investment layer: the amount, pattern, and consistency of spend is read by the market as evidence of durability. The same logic runs in reverse — which spend produces revenue is itself a signal worth instrumenting.
The power law is not a claim to take on faith. It is a property you can measure in your own portfolio — and if it is there, it invalidates the forecasting math most acquisition systems silently assume. The distribution does not just describe the revenue. It decides which statistics are even allowed.
Lifetime value across a recurring-revenue portfolio commonly follows a power law: a small fraction of customers — often the top 5–15% — drive the majority of multi-year revenue. It is the empirical shape of most portfolios where customer types vary substantially and acquisition cost is comparable across types. The first step is to measure your own tail index, not to assume it.
When a system cannot distinguish types, it values every transaction at the population average, E[V]. Transactions worth more than the average can't be priced for, so they're under-served and exit. Acquisition optimized on a blended average converges on the lower tail — the same dynamic, now produced by the operator's own targeting rather than a buyer's pricing.
A signal s separates high from low performers if and only if it is strictly more costly for the low type to produce. The cost asymmetry must be strict; signals equally costly to both convey nothing. This is the empirical bar applied to every candidate signal: does it correlate with above-baseline performance, and does the differential survive the test?
Mean-based LTV, CPA targets, and σ-based confidence intervals all assume a stable mean and finite variance. A power-law portfolio in the α<2 range has neither — variance is infinite, so the intervals are invalid. And the sample mean is median-biased low: in most finite samples it sits below the true long-run mean, so LTV calibrated on observed data understates the customers you haven't booked yet — and the thin tail it's missing is exactly where the revenue is. Optimization built on that mean then spends away from the tail it can't measure. The remedy is distributional — estimate the tail, forecast in quantiles, and steer toward the signals that mark a tail landing.
The Galton board is one of the few classical statistics visualizations with the right properties for this argument. We adapt it across five layers — each adding one mechanism that explains why the performance distribution stays power-law, and what signals can do to identify the right tail before acquisition cost is committed.
Every recurring-revenue portfolio where customer LTV varies substantially across tiers and acquisition cost is broadly comparable exhibits the same power-law shape. The dealership is teaching material. The real applications are everywhere — and the steeper the power law, the more the signal architecture is worth.
Merchant acquisition costs land in a comparable range across small, mid-market, and enterprise merchants — the sales motion differs but cost-per-acquired-merchant is not dramatically different. Residual revenue diverges by 100–1000×.
A single mid-market merchant processing $50M in annual volume produces more monthly residual than 500 small merchants combined. A processor optimizing for cost-per-merchant-acquired is mechanically guaranteed to underperform one optimizing for portfolio-weighted residual — the optimization is pointed at the metric that doesn't predict revenue.
The power-law shape is even sharper than the dealership case. The signaling architecture matters more, not less.
Initial ACV looks comparable across customer types. Year-one revenue does too. Renewal rates, expansion, and multi-year contract size diverge by an order of magnitude.
Acquisition dashboards report health while cohort-revenue analysis three years later reveals the portfolio rotation the optimization quietly produced. The damage shows up in net revenue retention — too coarse and too lagged to drive in-flight decisions.
Acquisition cost per engagement looks reasonable. Initial engagement size varies but not hugely. Multi-year relationship value, referral revenue, and renewals separate the portfolio dramatically.
Lead-gen optimized for cost-per-opportunity systematically attracts one-off project buyers over ongoing-relationship buyers. The acquisition looks similar; the revenue tail does not.
Premium pricing across small-business segments looks comparable per policy. Loss ratios, retention, and multi-line cross-sell diverge by tier.
Optimization targeting cost-per-bound-policy systematically attracts higher-risk, lower-retention segments. Underwriting cleans up some of the damage. Marketing produces all of it.
Account-opening cost per business customer is broadly comparable. Deposit balances, product attachment, and treasury revenue follow an extreme power law.
Campaigns optimized for new-account count silently rotate the portfolio toward customers who generate fees but never become franchise relationships. Track franchise-relationship probability at the source, not account count.
Initial lease economics look comparable across segments. Renewal rates, equipment expansion, and service revenue diverge sharply by customer maturity and scale.
The CAC-reduction trap is severe here because the front-end contract obscures the variance in the back-end stream that actually produces the portfolio's profitability.
The signaling work is the answer. Instrument signals throughout the full funnel for those correlated to performance above baseline. Build them in as early as conversion data permits. Defend them as competitors imitate them. Measure them continuously, source by source, not in aggregate — then amplify the volume that lands in the tail.
ABPRO is the reporting backbone behind Performance Revenue Optimization. It is built from first principles around two foundational layers — investment and revenue — with everything between treated as signal-preservation infrastructure rather than a reporting endpoint.
The whole point of tying investment reporting to revenue reporting is to test which acquisition signals actually predict revenue — and to do it before a CAC-lowering optimization has rotated the portfolio toward the wrong tail.
It is the difference between chasing a cheaper front-end ratio and protecting the back-end one that pays the bills.
The intellectual lineage. None of these were written about marketing. All describe the dynamic this method operationalizes.