Growth loops: the definitive guide

The mechanics: reinvestment, compounding, and the reverse gear

A loop has three parts — input, process, output — and one defining property: some fraction of the output routes back to the input without new external spend. That reinvestment step is the entire difference between a loop and a funnel. Miss it and you have a pump: every cycle financed from runway, forever.

Balfour’s Universal Growth Loop generalizes it to the company level — growth attracts resources, resources solve new problems, solved problems produce more growth; “self-reinforcing systems create compounding returns” — and supplies the symmetry the pitch decks omit: “The loop starts to reverse,” producing “compounding destruction.” The same math that compounds up compounds down. Reforge’s momentum framing adds that “the effects of losing momentum compound rather than increase linearly” — a stalled loop isn’t paused, it’s unwinding.

The quantitative core is older than the vocabulary. David Skok’s viral-growth math (formulated with Stan Reiss): two parameters govern everything — the viral coefficient K (invites sent × conversion rate per user) and the viral cycle time. And the counterintuitive result everyone forgets: “The most important factor to increasing growth is not the Viral Coefficient, but the Viral Cycle Time (ct) which should be made as short as possible” — because K is raised to the power of t/ct. Halving cycle time does what doubling K cannot. (Skok’s own sobriety note: “only a very small number of companies actually achieve true viral growth.”)

And the model loops replaced deserves its due: Dave McClure’s AARRR (2007) was never as naive as its critics needed it to be — the original deck carries the caveats “‘Precision’ is Illusion” and “we don’t know jack about your business.” The loops essay’s real charge against funnels is organizational: they “create strategic silos” — “Marketing brings in low quality users/leads at top of funnel to hit their goal, but that tanks retention.” The funnel’s sin isn’t the diagram; it’s the org chart it produces.

The taxonomy, with its attributions checked

The loop families below are what survive source verification — including one correction to the most-repeated “fact” in the genre.

The Racecar frame (Hockenmaier & Rachitsky) is the honest architecture: one self-sustaining growth engine “drives nearly all your growth long-term” (named engines: performance marketing, virality, content, sales), surrounded by things that are not loops and don’t pretend to be — Turbo Boosts (one-offs), Lubricants (conversion, brand, retention), and Kickstarts (“unscalable tactics for acquiring your first 1,000 users”). Note what the framework quietly concedes: the loop orthodoxy needed a category system that legitimizes non-compounding work. And a correction for the record: the ubiquitous four-word taxonomy “viral, content, paid, sales” as acquisition loops is universally attributed to Reforge — no Reforge-authored page using that literal phrasing could be verified. The closest verified formulation is Racecar’s four engines.

Network-effect loops are the strongest and worst-understood family. NFX’s Bible — “every new user makes the product/service/experience more valuable to every other user,” sixteen kinds, “the best form of defensibility” — self-limits more than its citers do: most products need critical mass, value comes from usage not size, and negative and asymptotic network effects exist. (Its “70% of all the value created in technology since 1994” is a first-party claim with no methodology on the page — quote it as NFX’s estimate, not a fact.) The sharpest critique of the category comes from inside it — Chen’s Cold Start Problem: network effects are “widely invoked in startup pitches but are poorly understood,” and “there are no universally accepted metrics proving whether network effects are occurring.”

Social-capital loops (Kevin Kwok on Superhuman): “Social capital — not personal utility — is what drives Superhuman’s acquisition loop,” and explicitly, “It does not have network effects.” Kwok also states the family’s limit: social-capital loops “traditionally do not scale,” and the remedy is to “sequence to a better loop.” Loops are a sequence, not a choice — the loop that starts you is rarely the loop that scales you.

Engagement loops (Eyal’s Hooked model): trigger → action → variable reward → investment, repetition shifting the trigger internal. Carry its ethics file with it, because it’s unusually well documented: Eyal’s own “we in the consumer web industry are in the manipulation business,” Ezra Klein’s “if we don’t talk about addiction, we are letting these companies off the hook,” and Eyal’s rebuttal — “We are not puppets on a string” — in the same conversation.

PLG loops: the term has a verified origin — Blake Bartlett at OpenView, defined as “a go-to-market strategy that relies on product features and usage as the primary drivers of customer acquisition, retention and expansion” — and a verified deflation: Wes Bush’s own second edition concedes PLG “is no longer a competitive advantage in 2026; it is ‘table stakes.’” Elena Verna names the breakdown conditions precisely — AOV above $10K, 1,000+ employee targets, high complexity, users without permission to trial: “But PLG is not enough in B2B” — and Kyle Poyar documents both the PQL fallacy (users “are generally not the same people as those who buy your product”) and the cannibalization risk of bolting on sales: “it ultimately leads to a disinvestment in PLG and subsequently a smaller and smaller pool of new prospects entering the sales pipeline.”

And the flywheel: Jim Collins’ page — “No single defining action… No one killer innovation… No miracle moment,” with the “doom loop” as the named failure state — never mentions virality, SEO, referrals, or software. Software growth borrowed the metaphor, not the research. Say “flywheel” for motivation; do the math with Skok.

The reality check: what the numbers actually say

Viral factors are small, and the exponential was a period artifact. Chen’s practitioner synthesis (directional — no disclosed dataset): most products run K of 0.2–0.3 or below, need >0.5 to be distinguishable, and “it’s basically impossible to create loops where the viral factor is >1 in the first session”. The classic viral loops died because “consumers became accustomed to the techniques, spam filters activated, platforms restricted the behavior.” The deeper math: invite conversion is discounted by the installed share of the network — at 50% penetration a 10% conversion behaves like 5% — and 50% retention converts the whole curve into what Chen labels the “shark fin”: the “EPIC FAIL” shape. The live demonstration: Clubhouse, still invite-only, fell from 9.6M monthly downloads in February to 900K by April — “Users soon realised it is always the same topics by the same people.”

Referral loops have real, modest, well-measured economics — nothing like the folklore. The canonical academic result (Schmitt, Skiera & Van den Bulte, Journal of Marketing 2011, ~10,000 bank customers over ~3 years): referred customers carried +25% daily margin, ~18% lower churn hazard, and ~€20 lower acquisition cost — statistically solid, and from German retail banking, not SaaS; port the direction, not the numbers (LTV referred vs non-referred is the metric to watch in your own data). The reward-size follow-up (field experiments, 400K+ customers): larger rewards acquire more customers but considerably reduce referred-customer profitability. Operationally, from ReferralCandy’s merchant network (vendor data, self-selected Shopify population): top-quartile share-action rate 4.64% of eligible prompt views, first referral typically at 14 days, and 69.2% of sharing flowing through chat apps and email — the loop runs through channels you cannot see. Meanwhile the numbers you’ve actually heard — Dropbox’s “3,900% growth,” “referred customers have 16% higher LTV,” “92% trust referrals” — none traces to a primary source. The Dropbox figure has no Dropbox or Houston artifact behind it; it lives exclusively on marketing blogs.

Word of mouth is the loop that shows up in buyer research. 73% of B2B SaaS marketing executives rank word of mouth first in consideration; 58% start shortlisting by asking peers. And the referral surface is migrating into machines: G2’s 2026 buyer data has review sites at 38% versus AI chatbots at 37% as shortlist sources, with 82% of buyers having sourced a recommendation from an AI chatbot in the past 24 months. The word-of-mouth loop increasingly closes inside a model’s answer — which is a distribution surface you influence with evidence, not ads.

And the composition of real GTM should humble the orthodoxy. Across ICONIQ’s 205-company survey: 57% of B2B SaaS runs sales-led as the primary motion, 38% hybrid — and 5% product-led. Self-serve typically produces ~10% of revenue (~20% at high-growth companies); partner and channel loops carry ~20% of revenue on average, rising to ~29–30% above $250M ARR. The loop content you read is written by and about the 5%.

How loops fail: the mechanisms

Fifteen failure modes are documented in the research base; these eight carry the clearest mechanisms.

1. A referral program on top of a product nobody loves. The program multiplies enthusiasm; it cannot create it. The operator summary: “even infinity multiplied by zero equals zero”; the ledger version: 8 months, ~$2,400, 3 referrals — two of them the same person. Gustaf Alströmer (YC): “The biggest misunderstanding about growth is that is somehow magically happen once you have a good product. It’s not the case” — it cuts both ways. (Program design does not substitute for the precondition.)
2. Incentives import negatively-selected users. From the person who ran Uber’s $300M+/year program: incentivized users “are usually MUCH WORSE than organic ones,” with LTV and engagement “half as good or lower, which is often enough to defeat the mathematics that justified the program in the first place” — plus pull-forward cannibalization of users you’d have gotten free.
3. Cash-paying loops get industrialized by fraud. Any loop paying for a verifiable-looking action creates arbitrage — Uber’s drivers and riders “colluded to scam Uber out of billions in incentives” — and conversion optimization strips exactly the friction that would catch it.
4. Calling a funnel a loop. No reinvestment step means every cycle is financed externally: “There is no concept of how to reinvest… This is unsustainable.” The paid version in the wild: $40K on ads for 12 customers at $450/month — a one-shot campaign wearing a loop costume.
5. A dozen weak loops instead of one strong one. “You’ll be tempted to draw a ton of loops for your product, but what that typically means is that you just have a ton of low-powered loops that aren’t sustainable… The fastest growing products are typically powered by 1-2 major loops.” Diagrams are cheap; compounding thresholds are not.
6. Assuming loop rates are stationary. Users habituate, competitors copy, each cycle reaches less-qualified people — the Law of Shitty Clickthroughs applies to invites too, and Reforge’s version: undifferentiated tactics’ “effectiveness decreases and always trends to zero.” A loop tuned just above K=1 slides below it silently.
7. The loop can’t compound because the bucket leaks. A loop reinvests retained users. Duolingo’s growth revival started exactly here: “Our first attempt at reigniting growth was focused on improving retention, i.e. fixing our ‘leaky bucket’ problem” — the fix compounded into 4.5× DAU over four years. Teams misdiagnose this as acquisition because the acquisition dashboard looks fixable.
8. The loop’s conversion step is owned by someone else. Viddy lost a third of its staff when Facebook’s algorithm changed; Zynga’s own IPO filing named the dependency; Gmail’s 2013 tab change produced “double-digit” declines in Groupon email opens — and Groupon’s email loop later declined from ~100% of the business to ~20%, a $135M annual gross-profit headwind (SEC-filed). The 2020s version is the content loop’s click step: ~60% of one programmatic-SEO build deindexed within four months, and clicks fall from 15% to 8% of searches when an AI summary appears. Platform-mediated loops carry platform risk in every cycle.

The counterexamples file

For every piece of loop orthodoxy, a documented case of the opposite. The pattern to watch: unscalable work earning the loop.

Questions operators actually ask