Analysts who took apart Casper's 2020 IPO filing found the same customer valued two ways: about $843 in lifetime revenue, and about $427 in lifetime gross margin once its 50.7% gross margin was applied (venturetwins' S-1 analysis, derived from the filing). True contribution, after fulfillment and returns, sits lower still.
Against a blended acquisition cost of roughly $302, the first number implies a comfortable business and the second implies an LTV:CAC of about 1.4x. Casper IPO'd on the second reality. That gap is the subject of this post.
Lifetime value exists for one job, setting a ceiling on what you can pay to acquire a customer, and it only does that job when it is built on contribution margin instead of revenue. Every number below names its source.
The short version: LTV is not a trophy number. Its one job is to cap acquisition spend, which means it has to be built on contribution margin after returns, discounts, and variable cost, not on revenue. Revenue-based LTV overstates the real figure by 30 to 65% in practitioner audits, and a single averaged LTV hides that a fifth of your customers carry most of the value.
Customer lifetime value is the total contribution margin a customer generates across their whole relationship with your store, discounted to today. It exists to answer one question: what is the most you can afford to pay to acquire one more customer like this one. Everything else people do with LTV, from board slides to valuation, is secondary to that ceiling.
The mistake that breaks it is computing LTV on revenue. Revenue is not money you keep. A practitioner audit of the common inflation mechanics put revenue-instead-of-contribution overstatement at 30 to 65%, before returns and discounts are even counted (Eightx, practitioner estimate).
Casper is the clean illustration: about $843 of lifetime revenue per customer became about $427 once its 50.7% gross margin was applied, and true contribution, after fulfillment, returns, and the roughly 20% of gross revenue Casper refunded or discounted, would sit lower still (venturetwins, derived from the S-1). Bill Gurley made the same argument a decade earlier: an LTV formula should discount marginal contribution, all future variable costs included, never gross revenue (Above the Crowd, 2012).
Here is the arithmetic on an illustrative store. Say a customer places 4 orders over two years at a $60 average order value. Revenue LTV reads $240. Now apply the costs revenue ignores: a 55% gross margin leaves $132, fulfillment and returns take it to roughly $95, and the variable marketing you spend to earn those repeat orders pulls contribution to around $75.
The revenue number said $240. The number you can bank against acquisition is about $75. If your blended CAC is $70, revenue LTV tells you you're printing money at 3.4x while margin LTV tells you you're barely above water. Only one of those is true. (Figures illustrative.)
The honest answer: the famous LTV:CAC target, 3:1, was never an ecommerce finding. It comes from SaaS venture guidance. David Skok, whose essay popularized it, describes 3:1 as an “early guess” he later validated against SaaS businesses, not ecommerce ones (forEntrepreneurs, “SaaS Metrics 2.0”).
The “3:1 average for ecommerce” that circulates online traces to First Page Sage, whose disclosed sample is 74% B2B firms and mostly organic-channel data, which describes neither DTC nor paid acquisition. No peer-reviewed or named-dataset study establishes 3:1 for a Shopify store.
The metric sophisticated operators watch instead is contribution-margin payback: how many months of a customer's margin it takes to earn back their acquisition cost. The DTC practitioner target is 3 to 6 months, with anything past 12 treated as risky (Eightx).
Payback is harder to game than a ratio because it forces you to state the two things LTV models fudge, real contribution margin and real CAC, and it tells you when the cash comes back, which is what constrains a store that isn't venture-funded. Inside one “healthy” blended LTV:CAC, channel economics can run from roughly 8:1 on SEO to 1.5:1 on paid social, so a single ratio can hide a channel losing money on every order.
Because customers are not distributed anywhere near evenly, and the average describes none of them. Across 339 public non-CPG companies, the top 20% of customers delivered on average 67% of sales, product firms at 0.67 and subscription firms at 0.59, with profit concentration running higher still (McCarthy & Winer, Marketing Letters, 2019).
The real rule is closer to 70/20 than 80/20, and it means one blended LTV number averages whales against one-timers into a figure that predicts neither.
Blue Apron is the cautionary case. About 70% of its acquired customers churned within six months, and the blended averages in its S-1 hid that decay behind healthy-looking aggregates; the stock fell about 70% after the IPO (Knowledge@Wharton on the CBCV method; the cohort-masking reading is Daniel McCarthy's own analysis of the sparse S-1 disclosures, not Wharton's).
The failure was not the churn alone. It was reporting an averaged, immature cohort as if it were the steady state, then funding acquisition against it.
Warby Parker shows the subtler version, in a company with good economics. Warby was profitable on the first order, yet independent analysis found its disclosed CAC understated true CAC by 20 to 40%, because the company divided marketing that also drives repeat orders across all active customers rather than only newly acquired ones (Theta CBCV analysis).
Its own S-1 showed CAC climbing from $27 to $40 in a single year (venturetwins teardown). A “good” filing can still get the denominator wrong.
If LTV sets the ceiling, retention raises it faster than anything else you can touch. The canonical customer-valuation paper computed the sensitivity directly: a 1% improvement in retention raised customer value by 2.45 to 6.75%, a 1% improvement in margin raised it about 1%, and a 1% cut in acquisition cost raised it only 0.02 to 0.32% (Gupta, Lehmann & Stuart, Journal of Marketing Research, 2004, at a 12% discount rate).
In round terms, a point of retention is worth roughly 5x a point of margin and up to about 100x a point of CAC.
That ordering is why mature operators obsess over cohort retention rather than shaving CPMs. The cheapest-looking win, negotiating ad costs down a few percent, sits at the bottom of the list. The expensive-looking work, earning the second and third order, moves the number that decides your acquisition ceiling.
It also explains why revenue-LTV optimism is dangerous: the same math that rewards real retention gains rewards imaginary ones, so an LTV built on an over-optimistic retention curve inflates exactly where it does the most damage.
You don't need a stochastic model. In a peer-reviewed test, a plain hiatus rule, where a customer is treated as gone after a fixed silence, classified active-versus-dead customers correctly 83% of the time for an apparel retailer, beating the Pareto/NBD model's 75% (Wübben & Wangenheim, Journal of Marketing, 2008).
The exotic models won only for aggregate purchase-volume forecasting, not for the customer-level call an operator makes. ASOS's own data-science team later abandoned the BTYD framework for practical reasons, and even their production machine-learning model reached only a 0.56 rank correlation on individual lifetime value (Chamberlain et al., KDD 2017).
The takeaway is not that models are useless. It is that rank-ordering customers by recency captures most of the signal, and that a fixed silence window, tuned to where your own repurchase curve flattens, gives you a defensible “lapsed” definition with no model at all.
The claims this file cut, and why:
“3:1 LTV:CAC is the proven benchmark for ecommerce.” Its lineage is SaaS venture guidance its own author calls an early guess (Skok, forEntrepreneurs). No ecommerce dataset validates it. The widely-shared “3:1 average” from First Page Sage rests on a sample that is 74% B2B, so it is SEO content, not a DTC benchmark.
“CAC is up 222% since 2013.” The figure is SimplicityDX's loss-per-new-customer, not CAC, from a press release with no disclosed methodology (Business Wire, 2022). Cite CPM trends from a named dataset instead, or don't cite it.
“iOS 14.5 cut ad performance about 40%.” The measured −38% was reported ROAS on Meta's dashboard, confounded with attribution loss from day-of-conversion recording (Common Thread Collective). True incrementality change is unknown, so the number measures reporting, not performance.
“Value-based lookalikes lift ROAS by X%.” Meta documents the feature and implies better results, but no public, methodologically described lift exists (Meta developer docs). Every percentage in circulation is an unaudited vendor anecdote.
“Discount-acquired customers always have lower LTV.” The best field-experimental evidence found the opposite for first-time buyers: deeper introductory discounts raised their long-run purchasing, while the negative effect fell on established customers (Anderson & Simester, Marketing Science, 2004). The folk claim confuses the channel a discount runs on with the discount itself.
“Predictive ML lifetime value beats a fixed-window rule.” Not at the individual level. BTYD models lost to a simple recency rule on customer-level tasks (Wübben & Wangenheim), and even ASOS's production model reached only 0.56 rank correlation. Rank-ordering is feasible; precise individual LTV prediction is not solved.
ASK THE AI. Paste into ChatGPT or Claude with numbers you already know.
· “My average order value is [X], my gross margin is about [Y]%, and a typical customer places about [Z] orders before going quiet. Walk me through revenue LTV versus contribution-margin LTV, and tell me the most I should pay to acquire a customer if I want payback inside 6 months.”
· “Here are repeat-purchase counts for my last few monthly cohorts: [paste]. Show me how a single blended LTV would mislead me versus reading these cohorts separately, and tell me which cohort I should worry about.”
· “My blended CAC looks fine, but one channel might be losing money. Given these per-channel spend and new-customer numbers: [paste], compute channel-level CAC and tell me which channel a healthy blended number could be hiding.”
On contribution margin, always. Revenue lifetime value overstates the real figure by 30 to 65% in practitioner audits before returns and discounts are even counted (Eightx). LTV exists to cap what you can pay to acquire a customer, and you can only spend margin, not revenue. A revenue-based ceiling authorizes acquisition spend your store cannot actually fund.
There is no validated ecommerce benchmark; the famous 3:1 is imported SaaS venture guidance its own author calls an early guess (Skok, forEntrepreneurs). Watch contribution-margin payback instead. The DTC target is 3 to 6 months to earn back acquisition cost, with beyond 12 months treated as risky. Payback is harder to inflate than a ratio.
Probably not to start. A plain recency rule matched or beat the sophisticated Pareto/NBD model at telling live customers from dead ones, 83% versus 75% for an apparel retailer (Wübben & Wangenheim, 2008). Models help for aggregate forecasting, but for the customer-level decisions you make daily, rank-ordering by recency captures most of the signal.
Because a fifth of your customers carry most of the value. Across 339 public companies, the top 20% delivered about 67% of sales (McCarthy & Winer, 2019). An average blends whales and one-timers into a figure that describes neither, and it is how averaged, immature cohorts hid Blue Apron's roughly 70% six-month churn until after the stock fell.
MetricsNavigator builds these from your store's real order history, margin, cohorts, payback, and customer concentration, when you connect your store. The honest LTV, without the spreadsheet.
