Launch planning examples: five launches, what each tested and what teams misread

What to take away

Launch planning examples are easier to argue about than to run. The five real launches below, Google+, Mailbox, Coolest Cooler, MoviePass and Airbnb, each left a public record of the decision it forced.

Beside each real launch sits a small invented case, labeled as invented and carried through to the same kind of decision.

None of the invented cases is a success story. The interesting part of a launch is the argument that happens afterwards, and a plan either equips you for that argument or does not.

Counts appear where they carry the reasoning. Those counts are made up and small on purpose, because small is what a first launch produces.

Example one: the launch that beat its prediction and was still a failure

Google+ launched on June 28, 2011 by invitation only. Google reported about 10 million users within two weeks, a bar most first launches never reach.

Much of that growth came from placement inside Google's own products, including the bar across Gmail and Search. The consumer service closed on April 2, 2019.

The invented case below beat its numbers on a channel the team did not own.

Invented case. A scheduling service for dental practices. The plan predicted forty visits and two paid signups in week one. It got a hundred and ten visits and five signups.

The debrief found four of the five came from one forum post by an office manager. The plan listed six channels, and five produced nothing.

The launch was a success and the channel plan was wrong. The honest next step was not to celebrate but to find out whether that forum was reachable again.

It was not. The post was organic and could not be repeated on demand.

Decision. Treat the result as one channel with unknown repeatability, not as a validated launch. Spend the next two weeks trying to reproduce the reach, not building features.

Example two: the launch with nothing to observe

Mailbox, an email app built by Orchestra, launched free on iOS on February 7, 2013 with a reservation queue. That queue was reported at more than a million people.

Nothing was for sale, so the queue measured interest in a promise, not willingness to pay. Dropbox bought the company in March 2013 and shut Mailbox down in February 2016.

Invented case. A tool for freelance bookkeepers. Announced to a mailing list of three hundred, with a signup form and no purchase path.

Twenty-two people signed up. Nobody was ever asked for money, because payment was going to be added later. Six weeks on, the team could not say whether any of the twenty-two would have paid, and the list had cooled.

Decision. The launch produced a list, not a reading. Re-run with a price attached, accepting that the second ask to the same people is weaker than the first would have been.

The cost of the missing instrument was the list itself, which is a common way to lose the only asset you had. Getting a refusal early and cheaply is the point of pricing validation, and this launch bought the opposite.

Example three: the launch that could not ship

Coolest Cooler raised about $13.3 million on Kickstarter in the summer of 2014 from more than 62,000 backers. The campaign promised delivery in February 2015.

Many backers were still waiting in 2016. Kickstarter's terms require a creator who cannot deliver a reward to refund the pledge instead.

Invented case. A physical product, pre-sold from a landing page with a stated delivery window of six weeks.

The page did well. Thirty-one orders. The supplier's lead time turned out to be eleven weeks, which had never been confirmed in writing.

Decision. Notify every buyer, offer refunds, hold the ones who wanted to wait. That is not a choice: the order and shipment rule requires it of a seller who cannot ship within the stated window.

The launch also produced a genuine finding, which is that demand at that price exists, and the finding survives the mess. What does not survive is the assumption that a supplier quote given on a call is a commitment.

Example four: the launch that worked on the wrong people

MoviePass cut its monthly price to $9.95 in August 2017. Subscriptions passed 3 million by mid-2018, and the service shut down on September 14, 2019.

The price attracted the customers who cost the most, the people who see several movies a month. A fast-growing subscriber count hid that until the bills arrived.

Invented case. A course for new managers, promoted through the founder's own network.

Fourteen sales in the first week, well past the predicted five. Then the exit survey came back and eleven of the fourteen had bought because they knew the founder.

Decision. Discount the result to the three cold buyers and treat those three as the actual sample.

The plan's error was not in the channel but in the prediction, which had not separated warm from cold. A prediction with one number cannot detect this; a prediction with two can. Keeping those columns apart is the same discipline that makes early counts readable in market validation.

Example five: the launch nobody saw, and the reach work that follows

Airbnb is the case where the counts were small and the response was distribution work. In 2009 staff went door to door in New York to photograph listings, and the site added a way to cross-post a listing to Craigslist, where renters were already looking.

Both moves were distribution and supply work, not a rebuild of the product.

Invented case. A niche data service. Shipped on time, priced, instrumented, with a public announcement.

Nine visits in a week. Two of them the founder.

Decision. This is the result people find hardest, because it looks like rejection and is not. Nine visits is not evidence about demand, it is evidence about reach.

Reach is a separate body of work with its own plan, and the promotion and sales guidance for small businesses treats it that way: audience, channel, and message chosen deliberately rather than announced once.

The next move is a reach experiment, not a product change. Rebuilding the page is the temptation to answer a question nobody asked.

What you can honestly do with counts this small is set out in the discovery counts: a denominator of nine supports no conclusion at all.

Reading across the five

The pattern is one thing.

Four of the five launches produced a real reading about something, and in four of the five the team's first instinct was to read it as an answer to a different question. A written prediction makes that mistake visible, because the prediction names which question you were asking.

The sequencing that produces two separate readings from one launch is worked through in launch planning.

Running one of these yourself

Write five sentences from your own plan.

Name what will be tested, what would count as a negative result, what you will be tempted to conclude instead, who the audience actually is, and what you can see in week one.

If the third sentence is hard to write, spend time on it; everyone can predict the good outcome.

A plan's useful part is the sentence naming the plausible wrong conclusion you are most likely to reach, written with no stake in it. Apply that move to your own capacity, not the plan: that is what readiness is for.

Common questions

Are these real companies?

The five worked cases are invented, and none of their counts came from anywhere. The named launches are real: Google+, Mailbox, Coolest Cooler, MoviePass and Airbnb, with dates and figures taken from public reporting.

Real launch numbers are rarely published and, when they are, come from a different offer, price, and audience than yours.

Which of the five is most common?

Five, by a wide margin. Most first launches are seen by almost nobody, and almost all of them are interpreted as rejection.

How small a sample is too small to act on?

Any sample is enough to act on if the action is cheap and reversible. The threshold that matters is not a count but the cost of being wrong. Nine visits can justify trying another channel; it cannot justify rebuilding the product.

What if the prediction was right but for the wrong reason?

Write both down in the debrief. A right prediction from a wrong model will hold once and then stop, and the debrief is the only place that ever gets caught.