The Lean Startup

Before IMVU, five years of work met a public shrug

There.com was supposed to be a richly rendered online world. Its team spent five years and about $40 million building the technology before an October 2003 launch. A Stanford teaching case reports $20,000 in first-month revenue and flat sales over the next six months; the board then redirected the company from consumer socializing toward military simulation. Eric Ries, then a software engineer, had experienced a familiar technical reflex: when a product struggled, work harder on architecture, features, discipline, and quality.¹

Will Harvey and Ries carried that failure into their next company, IMVU. Steve Blank, an investor and director, required its leaders to attend his customer development course. Blank's distinction was that established companies execute known business models while startups search for them. Ries joined that search logic to agile software practice and ideas he drew from lean production. The result eventually became The Lean Startup, published in 2011 after several years of blogging, advising, and speaking about IMVU's methods.²

This origin establishes the book's question: when neither product nor customer is known, how can a team learn before its labor and capital harden an assumption into a company? It also establishes its evidentiary boundary. The Lean Startup is a participant's theory built substantially from a consumer-software venture, then illustrated with other cases. IMVU's later commercial success made the story influential; it did not by itself show which practice caused the outcome or where the practices would travel.³

Six months of good engineering answered the wrong question

IMVU's founders hypothesized that people would add three-dimensional avatars to the instant-messaging networks they already used. Network effects would follow because each customer would bring existing friends. The engineering team spent about six months making IMVU interoperate with multiple messaging services. ⁴

When users encountered the product, the premise broke. They did not want to invite their existing friends into an awkward avatar system. They were more interested in meeting new people. Ries later described much of the integration code as work customers refused to use. In a contemporary interview, he asked the question that became the method's center: if learning that the strategy was wrong was the useful result, why had it required six months of software? ⁵

The answer was not merely “code faster.” IMVU could have tested whether anyone wanted the proposed add-on with a description and a download button. Once an early product existed, the company bought roughly one hundred visits a day with five dollars of Google AdWords traffic. It followed each daily cohort through registration, download, use, return, and purchase. Product changes that looked like progress but left those behaviors flat no longer counted as evidence of progress.⁶

That experiment made several kinds of ignorance visible at once. The team did not know whether the value proposition attracted anyone, whether a visitor could become an active user, or whether repeated use could support revenue. A small, fresh cohort made changes easier to interpret than cumulative user totals. But the experiment could only answer questions expressed in its funnel. It could not discover a valuable outcome the team had not imagined, prove why behavior changed, or decide whether increasing that behavior was good.⁷

Build–Measure–Learn turns work into a claim

Ries names demonstrable progress through uncertainty validated learning. Instead of treating a feature, plan, or launch as an accomplishment in itself, a team identifies a consequential assumption and asks what observation could change its decision. Activity moves through Build–Measure–Learn, but design moves backward:⁸

decide what must be learned → specify evidence and a decision rule → build the smallest valid test → observe a comparable result → revise the strategy

A minimum viable product is therefore not the cheapest product a company can sell. It is the smallest intervention that can test an assumption without invalidating the result. IMVU's original integration was not minimal because most of its work was unnecessary to learn whether customers wanted the premise. A landing page could test attraction; a functioning beta could test use and payment; a split test could compare a particular change. Different questions require different artifacts.⁹

Innovation accounting makes that learning reviewable. Establish a baseline, change something tied to a hypothesis, compare cohort behavior, and decide whether to persevere or pivot. A pivot changes part of the strategy while retaining knowledge and a larger vision. At IMVU, abandoning the add-on for a standalone network was not a random restart: the team kept the avatar experience while changing the assumed social setting.¹⁰

The framework does not make the final decision automatic. Weak metrics can confuse correlation with cause. A team can optimize a funnel whose economics are poor, whose users differ from the population it hopes to serve, or whose outcome is harmful. Nor does disappointing evidence dictate whether persistence is courage or denial. The method disciplines judgment; it does not replace it. ¹¹

The experiment had participants, not just data points

IMVU reduced the founders' financial exposure by shifting part of the learning into public use. The company's beta was not merely unfinished. In Ries's own account it was buggy enough to crash some customers' computers, and IMVU charged for virtual goods while changing the product around them. The Stanford case, which is based largely on company records and participant interviews and uses disclosed composites elsewhere, reports that roughly half of early subscribers were eighteen or younger. Early adult users converted to payment more often, while some younger users sent cash through the mail; enabling payment through mobile-phone bills was followed by a rise in revenue.¹²

Those details complicate the word minimum. For the six-person company, a buggy release and five-dollar acquisition budget kept cash commitments small. For users, the same test involved crashes, attention, behavioral observation, personal data, and money. For young users and their families, payment access and online interaction also raised questions that a conversion funnel could not answer. The case discusses temporary user frustration as a risk the founders accepted, but it does not describe consent, child-safety review, or independent oversight of the experiments.¹³

Review should be proportional to consequence, not identical for every product comparison. Later research on the ethics of online controlled experiments distinguishes low-risk interface tests from experiments that may require consent, protection, compensation, or independent review. It proposes autonomy, fairness, non-maleficence, and beneficence as criteria alongside business impact. These are later ethical standards, not requirements supplied by Ries's book.¹⁴

In medicine, credit, employment, education, public infrastructure, safety systems, and ecologically consequential work, a defective test may be difficult to reverse or may concentrate risk on people unable to refuse. The minimum responsible experiment must be bounded by their exposure, not only by the startup's cost. A commercially successful result can still require stopping. ¹⁵

Later evidence supports parts of the method under conditions

Ries explicitly builds on Blank's customer-development process. His use of small batches and waste also draws from the Toyota Production System, but the translation is partial. Toyota was improving a production system with observable customers and products; a startup is trying to learn whether a product and business should exist. Unreleased code can resemble inventory, but a person in an A/B test is not inventory.¹⁶

Independent research has since tested related elements. A randomized trial of 116 Italian startups found that entrepreneurs trained to state theories and test predictions made more precise predictions and pivoted more often than a control group that received other market-feedback training. Another field study of early-stage teams supported key assumptions of hypothesis-based probing while finding team members' business education to be an important boundary condition.¹⁷

Those studies are stronger evidence for formal hypotheses and probing than IMVU's success story, but they do not validate every element of the book or every domain of application. Felin, Gambardella, Stern, and Zenger's critique argues that the method gives little guidance for generating a good hypothesis, that customers may not know what a genuinely novel venture could make possible, and that readily observable feedback can favor incremental changes. Bocken and Snihur's published response agrees that Lean Startup is not an ideation method but argues that, once a venture has a vision and initial model, experimentation can reduce uncertainty, engage stakeholders, and support collective learning.¹⁸

The durable contribution of The Lean Startup is thus not the injunction to move fast. It is the demand that work expose a consequential belief to evidence before commitment compounds. Read the IMVU story with two ledgers open: what the team learned, and what its learning asked other people to bear. A valid loop needs both.

Structured reading paths and evidence limits

The Eric Ries path locates the book within the author's participant experience and later synthesis. The Four Steps to the Epiphany is the explicit customer- development influence, while the Toyota Production System is a partial operating lineage for small batches and waste reduction. Those are also the two formal reading dependencies; the production metaphor does not make users experimental inventory or establish that startup search and mature manufacturing share one method.

Innovation, entrepreneurship, and renewal situates hypothesis testing among other ways ventures form and change. Learning, quality, and reliability asks when a quick test produces dependable knowledge rather than merely rapid feedback. Measurement, accounting, and control exposes who chooses the funnel, guardrails, and stopping rule. Organizational intelligence connects experimentation to institutional revision; benefit for all life requires a successful metric to remain answerable to people and living systems bearing the test's effects.

No structured impacts or typed relations are asserted for this work. The eight related paths and two dependencies establish lineage and editorial comparison, not universal efficacy. The evidence package combines Ries's participant account, a company-linked teaching case, contemporary profiles and interviews, independent field and randomized studies of selected practices, a peer-reviewed conceptual dispute, and later experimentation ethics. It contains no causal evaluation of IMVU's whole method and no direct testimony from its early users or their families. It also lacks worker, community, high-stakes-domain, and ecological outcome evidence sufficient to show that a commercially successful learning loop is responsible.¹³