Marketing — the Atlas · ch.26 · funnels & loops
📣 Chapter 26 · Part V · Growth & measurement

Where does the next customer come from?

Funnels leak; loops compound. This chapter is about the missing arrow at the bottom of the oldest diagram in marketing — and what happens to a growth model when you draw it in.

Here's the whole chapter in one line: a funnel spends users; a loop reinvests them — and that's the difference between renting growth and owning it. Everything below is the mechanics of the reinvestment.

1The oldest diagram in marketing

Picture a crowd of strangers at the top of a wide opening, and one customer dropping out of the narrow end at the bottom. That's the funnel, and it is older than radio. In 1898, an American ad man named St. Elmo Lewis wrote down the stages a salesman walks a prospect through: get their attention, hold their interest, build desire, close the action. AIDA. Every funnel you have ever seen in a pitch deck — awareness/consideration/conversion, see/think/do, TOFU/MOFU/BOFU — is Lewis's four stages wearing this quarter's labels.

The diagram earned its 130-year run because it encodes two true things. First, buying is a sequence: nobody purchases what they've never heard of, so something has to move a person from stranger to buyer, stage by stage. Second, each stage loses people — more hear of you than consider you, more consider than buy. The funnel's narrowing isn't pessimism; it's arithmetic.

Programmer's version: a funnel is a pipeline of filters. Ten thousand records in, each stage a predicate, a handful of rows out the bottom. And like any filter chain, the stages multiply: 30% × 40% × 10% end-to-end is 1.2%, and no single stage looks like the culprit.

But look closely at the shape and notice what it assumes. Value flows one way. Strangers in at the top, revenue out at the bottom, and the bottom connects to nothing. The customer, once produced, exits the diagram. Hold that thought — the entire modern critique of the funnel lives in that missing arrow.

2Pirate metrics — the funnel gets instrumented

For its first century the funnel was a picture you argued in front of. In 2007 a startup investor named Dave McClure turned it into a dashboard. His five stages spell AARRR — Acquisition, Activation, Retention, Referral, Revenue — "metrics for pirates," and the joke smuggled in a serious upgrade: every stage got a number, and every number got an owner.

  • Acquisition — they showed up. Visits, installs, signups. The top of the funnel, and the easiest number to buy.
  • Activation — they had the experience that makes the product make sense. Not "created an account" — reached the moment the value landed. The file synced. The first message got a reply.
  • Retention — they came back. The number McClure put in the middle on purpose, because everything downstream is a multiple of it.
  • Referral — they brought someone. The stage every earlier funnel forgot to draw.
  • Revenue — they paid. Last, notice — not first.

What AARRR got right wasn't the letters; it was the discipline. Stop debating taste, measure each stage as a conversion rate, and go find the leak — because in a chain of multiplied percentages, the worst stage dominates the product. It's the same instinct Hopkins had with coupons a century ago (Chapter 15): argue with receipts, not adjectives. AARRR is the coupon, generalized to the whole customer journey.

And one of the five letters is quietly not like the others. Acquisition, activation, retention, revenue all point down the funnel. Referral points up — its output is more acquisition. McClure drew a funnel with a loop hiding inside it, and it took the industry another decade to notice that the loop was the interesting part.

💡
Read AARRR as a type signature and the oddity jumps out: four stages are User → User — the same person, one step further along. Referral is User → NewUser — it mints the input type. A pipeline where one stage produces the pipeline's own input isn't a pipeline anymore. It's a recursion waiting to be noticed.
The salesman's staircase, drawn once, relabeled forever.
St. Elmo Lewis · Attention, Interest, Desire, Action · 1898
the moveWrite down the stages a prospect actually passes through — buying is a sequence, so selling should be too.
the insightEach stage loses people, so the diagram narrows. The funnel's shape is arithmetic, not decoration.
the blind spotThe bottom connects to nothing. The customer exits the picture the moment they're won.
the legacyAwareness → consideration → conversion, see/think/do, TOFU / MOFU / BOFU — a century of relabeling one drawing.
Same funnel — now with numbers, owners, and a joke.
Dave McClure · "Startup Metrics for Pirates" · 2007
the moveFive stages, five conversion rates, five owners: AARRR. The funnel stops being a picture and becomes instrumentation.
the upgradeActivation and retention get named as stages at all — the funnel finally extends past the sale.
the sleeperReferral: the one stage whose output is the funnel's own input. A loop, filed as a funnel stage.
the lessonMeasure every stage, find the leak, fix the worst one first. Hopkins's coupon logic, applied to the whole journey.

3The critique — funnels leak, and top-up gets pricier

Here's the uncomfortable property of every funnel ever drawn: it's a consumable. Users go in, some fraction converts, all of them — eventually — leave. Churn guarantees it. So a company that grows by funnel alone is running a bucket brigade for a leaky bucket: buy users, watch them drain, buy more. Growth continues exactly as long as the buying does, which is why funnel-only growth is best understood as renting your customer base month to month.

And the rent goes up. Part IV showed why: nearly every acquisition channel is an auction now — the feed (Chapter 21), the search page (Chapter 20), the shelf (Chapter 25). Auctions have a ratchet built in: your competitors' budgets are the price-setting mechanism, and each channel's cheap early audience gets bought first. So sustaining the same top-of-funnel volume costs more every year, while the funnel underneath quietly leaks at the same rate. CAC rises, saturation arrives, and the growth curve that looked exponential in year one is revealed as a straight line you were repainting monthly.

The diagnosis, in one sentence: the funnel has no memory. Nothing a customer does at the bottom makes the top any wider. Every cohort starts from zero, purchased at this year's auction prices. The fix isn't a better funnel. It's a different shape — one where the diagram's output feeds its input, so that this month's customers make next month's cheaper.

Before we draw it, run the funnel honestly once. The widget below is AARRR with sliders — notice how the stages multiply, and notice which stage is worth fixing first. Then notice the small arrow we've drawn that Lewis never did.

Interactive · the leaky funnel 10,000 visitors a month · slide each stage · find the leak worth fixing
25%
40%
10%
10%
presets
Slide the stages. The end-to-end rate is the product of every stage — and the worst leak dominates the product.

4Growth loops — the output feeds the input

Now redraw the machine the way the growth community (Balfour, Winters, and the Reforge school) started drawing it in the late 2010s: not a funnel but a circle. A new user arrives, gets value, and — as a side effect of getting that value — produces something that brings the next user. Output feeds input. Run it again. That's a growth loop, and once you have the lens, you find loops powering most of the compounding growth stories of the last twenty years:

  • The referral loop. User loves product → invites someone → who loves it and invites someone. Dropbox is the canon: give storage, get storage, both sides paid in the product itself. The loop's fuel is delight plus an incentive denominated in something the company mints for free.
  • The content loop. User asks or answers something → the artifact gets indexed → search delivers the next user to that artifact. Stack Overflow, TripAdvisor, Zillow: every question answered, review written, or listing viewed is an acquisition asset that works nights and weekends. (Chapter 20's search machinery is the distribution half of this loop.)
  • The UGC / marketplace loop. Buyers attract sellers; sellers' inventory attracts buyers. Two funnels, each feeding the other's top. Every marketplace you can name is this loop with a logistics department.
  • The paid loop. Revenue → ads → users → revenue. Yes, paid acquisition can be a loop — if the money actually cycles: payback fast enough to reinvest, LTV genuinely above CAC (Chapter 29 will interrogate that "genuinely"). If not, it's the leaky funnel again, wearing a flywheel costume in the board deck.

Programmer's version, and it's exact: the funnel is a fold; the loop is an unfold. A fold consumes a list of strangers and reduces them to a number at the bottom. An unfold takes a seed and generates the next element from the last one — a recursive function whose base case is churn. The reframe matters because it changes the engineering question from "how do we pour more in?" to "what does each user produce that could summon the next one — and where is that production breaking?"

5Loop math — k, the clock, and honest numbers

One number describes a loop's strength. For every user who arrives, how many new users does the loop generate? Invites sent times invite conversion, reviews written times readers converted — however your loop works, the product of its stages is its k-factor. And k has a threshold with a famous mystique: at k > 1, every cohort brings a bigger cohort, and growth is self-sustaining. An epidemic, in the flattering sense.

Time for the honest part. Sustained k > 1 is vanishingly rare. It has happened — early Facebook platform apps, a handful of viral games and video products, each for a window measured in months before the channel saturated or the platform closed the loophole. Planning your growth model around sustained k > 1 is planning to win a lottery that has paid out perhaps a dozen times in software history.

But here's what the mystique hides: sub-viral loops still compound. A loop with k = 0.4 turns every 100 acquired users into 67 more — 40 in the first cycle, then 16, then 6, a geometric series summing to k/(1−k). That's a 1.67× multiplier on every user you acquire by any other means: every paid user, every search visitor, every referral's referral. Your effective CAC just fell by 40%. Nobody would call k = 0.4 "viral," and it might be the highest-leverage number in the company.

And k isn't even the whole story — the loop has a clock. A loop that turns over weekly compounds ~13 times a quarter; a loop that turns over quarterly compounds once. Modest k on a fast clock beats impressive k on a slow one over any horizon that matters, for the same reason compounding frequency matters in finance. When you instrument a loop, you're measuring two things: the multiplier, and the cycle time. Most teams only measure the first.

📈
Evidence check. Published growth benchmarks put healthy, sustained referral k-factors around 0.15–0.4, with anything above ~0.7 exceptional and short-lived; the k>1 stories (early Facebook apps, viral-era games) lasted months, not years. The folklore numbers hold up as folklore: Dropbox's referral program was credited with roughly a third of daily signups at its 2009–2010 peak (Houston's own telling — treat the precision accordingly), and PayPal literally paid referral bounties because ~$20 a user undercut the era's ad prices. The durable finding isn't any single k — it's the arithmetic: at k = 0.4, the loop hands you 1/(1−k) ≈ 1.67× on every acquired user, compounding silently under whatever else you do.
Interactive · the loop compounder seed 1,000 users · set k and the clock · watch the series sum
0.40
4 wk
presets
Seed 1,000 users and let the loop run. Below k = 1 the series converges — the question is to what, and how fast.
⚠️
The k-factor trap. k is a product of stages, and the tempting stage to inflate is invites sent — so products spam address books, gate features behind "invite 3 friends," and juice k while torching the thing that made the loop work: a recommendation that felt like one. Chapter 13 priced this exactly — fabricated urgency and forced advocacy are loans against trust. A loop powered by resentment has a high k and a short life, and the churn it causes (remember: churn is the loop's base case) quietly eats the multiplier.

6Funnels for diagnosis, loops for strategy

So which picture is right? Both — they answer different questions. The funnel is a diagnostic instrument: when this month's number is bad, AARRR tells you where the leak is, stage by stage, owner by owner. Nothing beats it for finding the broken step. The loop is a strategy document: it answers where growth structurally comes from, and what compounding asset each new customer adds to. Diagnose with the fold; plan with the unfold.

In practice a real company runs a portfolio: a content loop earning search traffic, a referral loop multiplying everything, and paid top-up filling the gap while the loops spin up — each loop with its own k, its own clock, its own owner. The mistake isn't running paid; it's running only paid and calling the resulting straight line a growth curve. The other mistake is loop romanticism: waiting for compounding that never comes because the loop was never actually instrumented, just drawn.

And underneath every loop, one substrate: retention. A loop's cycles run through living users — churned users write no reviews, send no invites, generate no revenue to reinvest. Run a beautiful loop through a leaky product and you've built a slower funnel with better branding. That's why McClure put retention in the middle of the pirate's word, and it's why the next chapter hands the growth problem to the product itself.

The widget below races the two models honestly — same monthly spend, one with a loop bolted on. Watch where the loop line starts (behind — the loop's tax is paid up front) and where it ends. Then drag churn up and watch the loop stop mattering. That second lesson is the one growth decks omit.

Interactive · top-up vs loop — the 36-month race same budget · A buys 500 users/mo · B buys 400 and builds a loop
6%
0.40
Two teams, one budget. A rents users; B rents fewer and builds the machine that mints them. Slide churn and k to see who's ahead at month 36 — and when.
Give space, get space — the loop that named an industry.
Dropbox referral program · 2008 →
the setupAdWords CAC came in at multiples of the product's price. The funnel's rent was unpayable — so they stopped renting.
the loopInvite a friend, both sides get free storage — an incentive denominated in the product, minted at near-zero cost.
the resultReferrals credited with roughly a third of daily signups at peak; signups quadrupled in the folklore telling. Permanently cited, rarely replicated.
the fine printThe loop worked because the product did — storage you'd actually recommend. The incentive amplified advocacy; it didn't manufacture it.
Nobody runs one loop anymore.
The loop portfolio · 2026
the moveContent loop for search and AI answers, referral loop as a multiplier, paid top-up sized to fill the gap — run as a system, each loop with its own k and clock.
the instrumentationPer-loop dashboards: multiplier, cycle time, and the stage that's leaking. AARRR survives — inside each loop.
the disciplinePaid is a loop only if the money cycles: payback windows watched like uptime, reinvestment automatic.
the lessonThe funnel diagnoses, the loops compound, retention underwrites both — and the portfolio beats any single hero loop.

7The operator's picture

Close the chapter with the exercise that makes it real. Take your product and answer one question in writing: what does a new user produce, as a side effect of getting value, that could bring the next user? An artifact search can find? An invitation with a reason? Inventory another side wants? Revenue that genuinely recycles? If the honest answer is "nothing," you don't have a loop — you have a funnel and a hope, and your growth plan is a bid strategy.

If you do have a loop, instrument it like McClure instrumented the funnel — every arc of the circle is a conversion rate with an owner, the multiplier and the clock on one dashboard. Then run the checklist:

  • Name the loop. One sentence, output to input: "users who ____ produce ____, which brings users who ____."
  • Measure k honestly. The product of the arcs, not the vanity version. Sub-viral is fine; unmeasured is not.
  • Measure the clock. Cycle time is half the compounding. Shortening it is usually cheaper than raising k.
  • Check the substrate. Churn is the base case that ends the recursion. No loop outruns a leaky product.
  • Keep the funnel for diagnosis. When the loop slows, AARRR the arcs and fix the worst leak first.

One loop beats every funnel on a long enough horizon — and the strongest loops don't live in the marketing department at all. They live inside the product: the free tier that sells itself, the workspace that recruits its own members, the artifact that markets by existing. That handoff — from marketing-led growth to product-led growth — is the next chapter.

Check yourself

3 questions · instant feedback 0 / 3