ArticleIssue #193 | Marked To The Wrong Market
As difficult as it is to believe, the end of 2026 is in sight (and on the extended forecast). Black Friday is 19 days from now. The holiday shipping cutoff for most brands begins in about 6 weeks.
Between now and then – over the next 4-5 weeks, most brands will run some of the deepest discounts of their year, stretch return windows into January, switch on every payment-plan option they can support and buy more first orders than in any other month. This isn’t just an eCommerce/retail phenomenon, either – the same thing happens in B2B, home services (where I’m already getting offers for “no interest + no payments until 2030”), even Senior Living (lock in 2026 rates).
In every case, brands offering those discounts will fold the customers/clients acquired from them into the same trailing LTV, which governs how next year’s CAC targets are set.
The only problem is that the LTV for this incoming cohort is likely to be materially worse – but almost nobody will catch it for months. Why? Because every number that measures the first order or initial purchase will – truthfully – be up. The recap will say (quite truthfully) “record month” or “strongest close to Q4”. But the cohort acquired during this period is likely to produce something different ~9 months later, which will only be seen by a much smaller audience.
You might think I’m about to bash discounting (which, in fairness, I’ve done several times). I’m not, though I do encourage you to read the BFCM Offers That’ll Make Your CFO Smile.
I want to focus this issue on something else that is remarkably under-discussed in our industry: the fact that most brands price their customers off a signal whose meaning depends on how hard it was to send – and then spend 4 weeks making it as easy to send as possible.
A Purchase Is Evidence
Let’s talk about what a first purchase actually is, at its most fundamental level: evidence. A prospect voluntarily parts with a valuable resource (money), in exchange for a good/service you provide. That the exchange happened tells you something about whether they have a real need your product/service meets – presumably, they do, else why spend the money?
If we then call the prospects who do have the real need “fits”, it follows that Lifetime Value (LTV) is mostly what fits are worth, because they’re the ones who buy again. This even applies to purchases historically considered one-off, such as mattresses – yes, you only buy a mattress for yourself once every 5-10 years, but the overwhelming majority of homes in the US have multiple bedrooms/beds, most people do have friends, and virtually all mattress brands sell more than just mattresses.
If we continue this analogy, it stands to reason that how much a purchase tells you depends on what it costs to make. When buying is expensive – full price, shipping costs, a return process that takes effort, an intake process that requires real time/commitment, etc. – a non-fit either doesn’t bother or self-selects out once they realize what’s involved. But when buying is cheap, that abstention or self-selection doesn’t take place at nearly as high of a rate.
But don’t take my word for it – you can run this with your own business/your own customers. I’ll use hypothetical data below, but substitute in your own assumptions.
Imagine you have 100 prospective customers. You know ~20 are fits. Assume each fit is worth $120 in contribution over the 12 months after the first order and each non-fit is worth $0 (people do make mistakes multiple times). Now, vary the friction:
At high friction (full price, shipping costs, etc), 14 people buy; 10 of them are fits. If you reduce friction to BFCM levels, 46 buy (3.3x the first orders) – but only 14 of them are fits. Reducing friction bought 4 more “fits” + 28 “non-fits” – which means repeat contribution per buyer fell 57%. Yet, when you pull the report, every first-order metric on the dashboard improved.
The statistician Thomas Bayes actually developed a formula to address this very thing (I bet most of you thought you got out of statistics by going into marketing). Using Bayes’ theorem, we can define the share of buyers who are fits it.
p·f / [p·f + (1 − p)·n]
Where:
p is the share of prospects who are fits
f is the rate at which fits buy
n is the rate at which non-fits buy
Friction mostly impacts n. And because (1 − p) is the big term (most prospects aren’t fits) small changes in n dominate large changes in f.
You might read this and think the $0 is what’s producing the result you see above – but it isn’t. Even if you assume every non-fit generates $20 of repeat contribution, the BFCM-level cohort still lands 45% below the high-friction one. So, the size of the drop depends on the assumption. The direction doesn’t. There’s no plausible set of numbers under which 3x-ing the non-fit purchase rate leaves cohort quality on-par with higher-friction periods.
You might also read this and think it’s just the old line about coupons attracting deal-seekers – and you’d be (at least) half right. The discount is the only friction lever most people notice, but it’s hardly the only one impacting the decision. Free returns, BNPL, free-shipping thresholds, one-click checkout, artificial deadlines & scarcity + ads designed for the viewer to never evaluate anything all play a role, too. Each one lowers the cost of buying for fits and non-fits alike. Non-fits are far more sensitive to it, because a fit is predisposed to buy anyway. Every one of those levers raises first-order conversion rate AND dilutes what a first order tells you.
Put another way: what you really have are 2 markets AND 2 prices. The first-order market prices a transaction (what does it cost to get someone to buy this today). The customer market prices a relationship (what is this customer worth both today AND over a future period).
Under high friction, the two markets move together closely enough that the visible one (the transaction market) functions as a proxy for the other. But that correlation was never a property of the customer – it was a property of the friction. When you modify the friction, you disintermediate the market.
The Algorithm Buys What You Asked For
The platform makes this worse, for the same reason it’s good at its job.
In The Ratio Trap, I argued the fix for ROAS was to pass real contribution back to the platform instead of revenue, then constrain on a level. That’s still right. But first-order contribution is still a price from the first-order market. Tell a value-based bidder to maximize it and it will go find the people most likely to place a profitable first order – which, in a low-friction environment, is a population enriched for exactly the buyers who never place a second order. Nothing is broken. The machine is doing exactly what you told it to do.
The obvious next move is to pass predicted LTV instead. The platforms will happily accept it, and I’d love to tell you it solves the problem. It mostly doesn’t – and you can mathematically determine why.
The data invoice from the ROAS issue applies here, and it’s actually worse: a value signal carries (1 + CV²) times the data requirement of a bare count, where CV is the coefficient of variation of the values. For order values, most eComm brands land somewhere between 2x and 4x. Customer values are a different animal. If 70% of a cohort never orders again and the other 30% each spend exactly the same amount, the CV is already 1.53x – 3.30x the data, before anything interesting has happened. Real repeat customers don’t spend the same amount (a handful order 20x and become the crown jewels of the P&L), so customer-value CVs of 2 to 3 aren’t hard to reach. That’s 5x to 10x the conversions to learn as much from a predicted-LTV signal as from a count.
And that’s before the prediction is wrong, which early on it usually is. You’re asking the platform to bid on a model’s guess about a number you won’t observe for 12 months, trained on cohorts acquired under different friction than the auction it’s bidding into.
Very few accounts can afford to do that. So, for most brands, the bidding system can’t do the selecting. Something upstream of the auction has to.
November’s Average Customer Isn’t November’s Last Customer
Volume within a single cohort compounds the problem.
The platform delivered first to the people most likely to respond, and every additional impression after that went to someone less likely to care. A BFCM budget push does the same thing to customer quality. The first dollars find the people most ready to buy, and fits are overrepresented among them. Each additional dollar reaches further down the distribution.
Cohort LTV is an average over everyone you acquired in a given period (say, November) – and the customers bought with the last $50,000 you deployed sit below that average. Same structure as average versus marginal ROAS. A CAC ceiling set off average cohort LTV is too generous for exactly the spend it governs.
You might read this and think – (1) Sam has a point but (2) I need a new data pipeline to check it. You don’t. Split last year’s November cohort by acquisition week, or (for BFCM specifically) hourly spend level, then compare 90-day repeat rates. If the late-cohort repeat rate holds up, you’ve learned something good about your business. If it doesn’t, you’ve learned how much you overpaid for your late-month customers… which is worth knowing before you do it again this month.
Payback Is Duration
How much all of this costs depends on where in the customer relationship the value sits. Consider these two brands (Brand A + Brand B):
They have the same LTV, same CAC, same margin over CAC. Every dashboard either brand looks at reports them as the same business. Now, reduce expected repeat contribution by 25%. Brand A’s margin over CAC goes from $15 to $12.50 – down 17%. Annoying, but not the end of the world. Brand B’s goes to $5 – down 67%. Identical shock, 4x the P&L hit. That one hurts.
In finance, a bond’s duration measures how sensitive its price is to a change in rates. It’s driven by how far out the cash flows sit. CAC payback period is the same measure for a customer. Brand A recovers its CAC on the first order – a short bond. Brand B recovers 29% of its CAC on the first order and finances the rest on the expectation that the customer comes back – a 30-year bond, and the first thing hit by any shock to repeat behavior.
Brand B is also the business most likely to have built its LTV curve on early cohorts – customers acquired when friction was higher, who mostly found the brand because they were fits. Applying that curve to marginal customers acquired at BFCM friction is marking a long bond at a rate that no longer exists.
You might read this and think it’s a hypothetical. It already happened, in 2022, as a result of Apple: their App Tracking Transparency (ATT) update raised acquisition costs and degraded signal in the same window that rising rates repriced every future cash flow. The models that broke first were the long-duration ones: subscription, meal kits, anything built on losing money on the first order, but making it back at some future point. Plenty of those businesses weren’t badly run in any operational sense; they had simply marked their customers to the wrong market.
The Friction Is Coming Back
The decoupling ran on cheap evaluation as much as cheap buying – and evaluation is getting expensive again, on the buyer’s side.
In What The LLM Does Not Know, I talked about how the corpus is being overwhelmed with synthetic reviews, generated creative + 400 “blog articles” restating the same dozen claims. The end result? Every additional evaluation costs a buyer more time and returns less marginal information. When search gets expensive, buyers stop browsing and consolidate onto brands they already trust. That’s simply the math from the top of this issue, run in reverse: when evaluating is costly, a buyer who picked you out of the noise is sending a much stronger signal.
Brands that spent the last 5 years buying transactions at the lowest friction they could engineer now hold a customer file with weak evidence behind it. Brands that bought relationships hold the one asset whose price is going up.
If that’s not evidence for building a brand + a community, I don’t know what is.
OK, So Now What?
1. Measure the coupling. Take 3 years of cohorts and plot first-order characteristics against 12-month contribution: discount, AOV, channel, payment method, promo versus non-promo. Then monitor the correlation over time. If it weakened as your CAC fell, your LTV model is priced on a relationship that no longer holds.
2. Never blend promotional cohorts. Report BFCM (and other high promo) customers as their own cohort, permanently. They don’t belong in a trailing LTV that sets future targets or caps. This is the same principle as why you should never average brand search ROAS into your non-brand target.
3. Treat friction as a dial, not a cost. Every friction-reducing lever should be evaluated on 2 things: (1) first-order conversion rate, which it will always improve AND (2) on cohort repeat contribution. Some levers survive that test. Some turn out to be buying customers you’d have been better off never meeting. You can go bankrupt from success faster than you can from failure.
4. Design Offers That Select. A 40% sitewide code is worth the same to a fit and a non-fit. A full-price sampler, a refill bundle or subscribe-first pricing is worth far more to someone who expects to use the product for a year. Those offers lower the cost of buying for fits more than for non-fits – the opposite of what a percentage discount does. Creative changes your audience more than targeting does, and offers work the same way.
5. Compute your duration before you set the ceiling. Divide first-order contribution by LTV. Then cut repeat contribution 25% and recompute margin over CAC (just like I did above). If it goes negative, your November CAC ceiling is a bet that this month’s customers will behave like the ones you acquired in a fundamentally different time. That might be a bet worth making, but it probably isn’t.
A first order tells you someone was willing to buy at the price you set. It tells you very little about whether they’ll buy again. The cheaper you make that purchase, the less it tells you.
