A loss leader is not "a cheap product". It is a bet that a specific, nameable path carries buyers from the cheap thing to the profitable thing.
A supermarket sells milk below what the milk costs it. Nobody at the supermarket believes milk is a good business. Milk is a door. It pulls a person into a building where the rest of the basket carries the margin, and the milk is placed at the back so the person has to walk past the margin to reach it.1
That is the whole idea, and everything else in this course is a variation on it. Note what the supermarket has that a bad loss-leader proposal does not: it knows exactly where the money comes back, and it has physically engineered the walk.
Every real loss leader has three parts. If you can only name two, you do not have a loss leader, you have a discount.
| Part | The question it forces |
|---|---|
| The lead item | What are we giving away below its own cost, and what is that cost per user? |
| The recapture path | By what named mechanism does a person who took the lead item end up paying us? |
| The attach assumption | What fraction of takers travel that path, and how would we know we were wrong? |
The second one is where proposals die. "We give away X, and later they buy Y" is not a path. "They buy Y because X left them holding something they cannot get out of Y's competitor" is a path. The strongest version of a path is necessity: the lead item makes the paid item required rather than merely appealing. Gillette does not hope you will want blades. The razor is useless without them.2
The classic failure is the cherry-picker: the customer who takes the milk and leaves. In retail this is bounded, because walking into the shop is a cost and the basket is right there. In software it is unbounded and free, so an unusually large share of your takers can be pure cost. This is why a software loss leader lives or dies on the attach assumption rather than on the appeal of the free thing. A free product that everybody loves and nobody converts on is not a loss leader. It is a hobby with a server bill.
For twenty years, software had no meaningful marginal cost, so "below cost" quietly stopped meaning below cost. It meant "below acquisition cost": we spend to get you, we earn it back over your lifetime. The lead item itself was nearly free to serve.
AI products undo that. Every generated image, every transcribed hour, every long-context call is a real payment to a model provider. An AI loss leader has genuine variable cost per use, which puts it back in the world the supermarket lives in, where giving away more of the lead item actively hurts. Two consequences follow, and they matter for any AI product decision:
France bans selling below cost - revente à perte, banned since the law of 2 July 1963 and now sitting in the Code de commerce, with fines up to 75,000 EUR.3 It bites on resale of goods in an unchanged state: you buy a thing, you resell that same thing for less than you paid. Software and services you build yourself are not resale. So the rule is real, and it is not your problem. Keep the reflex, though - the habit of asking "is the obvious move actually legal here" is what Lesson 7 is about, and the document under review gets exactly this kind of question wrong.
Do these from memory, before scrolling up.
Read the DGCCRF page on revente à perte (five minutes, and it is the regulator's own words), then skim Loss Leader Pricing for the complement-margin framing. Both are listed with trust notes in RESOURCES.md.
Anything unclear, or want a worked example from a different industry? Ask your teaching agent. It wrote this lesson and can go deeper on any line of it.