Platforms and the chicken-and-egg problem: which side first?

Suppose the owner of an electrical shop in Can Tho wants an app that links air-conditioner repairers with householders. Fifteen repairers sign up and nobody calls them. Or the reverse: two hundred households install the app and no repairer takes the work. Which side should come first? This is the problem of the platform business model, and I rely on one research chapter about how platform start-ups validate and launch.

How does a platform differ from an ordinary seller?

A platform must create value for at least two groups, and both have to want it. Göcke and Meier call this multi-sidedness: Airbnb has to be attractive to guests (demand) and to hosts (supply) alike (Göcke and Meier, 2021).

The other three characteristics hang together. Network effects raise a platform's value to each user as more users join, in two forms: same-side (more users on a social network) and cross-side (more sellers on eBay benefit buyers). Network effects create the chicken-and-egg problem, because a minimum number of users is needed before the virtuous circle can start. They also tend to produce winner-takes-all or winner-takes-most outcomes, where one or a few platforms dominate the market (Göcke and Meier, 2021).

Why is the chicken-and-egg problem so hard for platforms?

In my reading, each side has a reason to come only when the other side is already there, so at launch neither has a reason to arrive first. Göcke and Meier write that every start-up with a platform model built on same-side or cross-side network effects faces this at launch, and that it must reach a critical mass attractive to both sides (Göcke and Meier, 2021).

They also note that the platform sponsor can act as a supplier itself, which raises the benefit to customers sharply early on, when supply is thin (Göcke and Meier, 2021). In my view that is one way out of the chicken-and-egg problem, but it costs money and people. In the repairer example, it means the app owner's own shop taking some of the jobs. Two of the three launch methods below (immersion and anticipation) go in this direction; the third does not.

How small did Airbnb and Gojek start, according to public sources?

Both started very small, but the company pages give milestones, not the method for solving the problem.

The Airbnb newsroom says two hosts welcomed three guests into their San Francisco home in 2007; that in March 2008 Airbed & Breakfast launched at SXSW and received just two bookings; that in August 2008 its website launched in time for the Democratic National Convention and received 80 bookings; that its own payments system also launched in August 2008; and that in March 2009 it changed its name to Airbnb and expanded from rooms to apartments and whole homes (Airbnb, n.d.). Göcke and Meier cite Paul Graham's 2013 essay, which uses Airbnb as an example of doing manual work in the early days (Göcke and Meier, 2021). I have not read that essay.

The Gojek page says that in 2010 it started with 20 motorcycle taxis and one call centre, with a mission to remove friction from daily life; that in 2015 it launched the app, expanded beyond Jakarta, saw food delivery take off and raised Series A funding; that in 2016 it became Indonesia's first unicorn; and that by 2019 over 2 million driver-partners were registered (Gojek, n.d.). The page does not say how drivers or customers were recruited. From "one call centre" I infer that matching was done by hand at first, but that is my inference, not a statement in the source.

Do not read these two timelines as recipes. They show that the first transaction was small and could be done by hand. They do not show how enough people arrived on both sides.

Which three ways to start does the study describe, and what does each cost?

All three start-ups began by identifying one core transaction between the parties, and only then added features. For kawaloo (storage space), the core transaction is a host renting out spare space for a fee (Göcke and Meier, 2021). From there the study draws three methods:

The study also mentions Zappos (via Göcke and Weninger 2020 and Ries 2011, not one of the three cases): it began with a website whose supply was shoes picked by hand, with no link to any manufacturer, so it tested demand without testing whether suppliers wanted to join. The authors use it to show that a platform can start from one validated side (Göcke and Meier, 2021).

A hypothetical platform: an app linking air-conditioner repairers and householders

This example is entirely invented: the app name, the district and every number are mine, made up to illustrate.

Suppose "ThoGia" serves one district of Can Tho. Its core transaction: a repair request is placed, a repairer turns up, the householder pays. Warranties, reviews and scheduled servicing wait.

Following the study, test each side before matching them. Demand: run a sign-up page that takes requests, and interview householders. Supply: interview repairers about the minimum workload that would make them take jobs through the app. To probe network effects, the authors suggest showing users different signals about how many people are already on the platform (Göcke and Meier, 2021); here that means two pages saying "5 repairers already here" and "15 repairers already here".

Then measure both sides at once, because the study defines platform market fit as continuously attracting and satisfying demand from both sides (Göcke and Meier, 2021):

Suppose that in the first weeks there are 12 requests a week and only 5 are taken within 24 hours, while 15 repairers have registered and several say they are free. The two numbers say different things. Five out of 12 points to poor matching: repairers are present but in the wrong place or at the wrong time, or dislike the price. Twelve jobs for 15 repairers means supply already exceeds demand. Recruiting more repairers would not fix either. The shop owner should let the shop's two technicians take the jobs nobody has accepted, serving customers while learning why outside repairers refuse, and put effort into finding more requests. That combines anticipation with immersion. Every number and the reasoning in this paragraph are my illustration, not from any source.

When does it make sense to build a platform, and when not?

The fit and no-fit criteria below are my synthesis from the chapter, not a test the authors give. It fits if you can serve one side by hand in a small area, and the core transaction is simple enough to run manually before you automate. Anyyogi and LogCorp did most of their early work by hand, and kawaloo focuses on one neighbourhood (Göcke and Meier, 2021).

It does not fit if your market is one where users rely on a single platform and you cannot fund long losses. The study says winner-takes-all or winner-takes-most conditions can be recognised by single-homing, and that knowing whether users are single- or multi-homing affects capital needs, market entry and competitive positioning (Göcke and Meier, 2021); the part about funding long losses is my inference. It also does not fit if you cannot validate both sides: the authors warn that checking only one side because time is short comes at a high cost, and that building a minimum viable platform on guesses about both sides is a great risk (Göcke and Meier, 2021).

Going deeper: what this chapter is, and what it has not shown

The library's platform framework comes from "Development and validation of platform businesses in digital entrepreneurship" by Göcke and Meier (2021), in the book edited by Soltanifar, Hughes and Göcke (2021) (Göcke and Meier, 2021; Soltanifar et al., 2021). The chapter is open access and I read its text and two tables in full, but have not viewed the figures. I have not read the rest of the book, including chapter 4 on the venture pyramid on which this chapter builds.

The first limit is method: all three cases are recent German start-ups, described by the authors from interviews; they do not compare the three launch methods for effectiveness and there is no control group (Göcke and Meier, 2021).

The second limit is the company sources. The Airbnb and Gojek pages are the companies' own accounts, they do not explain the launch mechanism, and the Gojek page gives no customer numbers at its 2010 and 2015 starting milestones. The repairer example is invented to show how to read the measures, not Vietnamese market data.