पाठशाला Pathshala · विचार Vichār, The idea · Lesson 13 · Build

Network effects you can actually measure

Every deck claims a network effect and few can show one. Test it: if the product gets better as the local network gets denser, retention will rise with density. If it does not, stop asserting it.

Pathshala, The Founder Library · 11 October 2026 · 6 min read

Mumbai’s residential towers and a lit arterial road seen from high above at night. — Photograph: Rajkumarrr comics · Pexels

A network effect is the most valuable property a technology company can have and the most frequently claimed without evidence. It is also one of the easiest to test. If the product really gets better as more people use it, the people in the denser parts of your network will stay longer. If they do not, there is no network effect, whatever the deck says.

This lesson defines the effect precisely, explains why density rather than size is what to measure, and gives the test as a figure. It follows the [previous lesson on moats](/library/moats-that-matter-in-year-one), where network economies appear as a power that is seeded in year one and established only in takeoff. This is how you know whether the seed is growing.

What a network effect is, and what it is not

The definition a16z’s partners use is plain: the product or network becomes more valuable to users as more people use it. The same piece separates network effects from three things often mistaken for them: scale effects, where costs fall as the company grows but no user benefits from another; brand preference, where users choose the familiar name; and increasing returns in general. A company can grow fast, get cheaper to run and become famous without any user being better off because another joined.

Data network effects get the same scrutiny. NFX, whose network effects manual catalogues sixteen kinds, offers a one-line test: does increased usage lead to more useful data? If not, it is merely a scale effect. The manual also notes that value per user is asymptotic: the fifth review of a restaurant adds a lot more than the thirtieth. That flattening matters for the test below. A real network effect rises and then levels off.

Why density, not size

Most network effects are local. A rider in Hyderabad gains nothing from a driver in Kolkata. NFX describes ride-hailing exactly so: up to a point, more drivers benefit riders because of reduced wait times. The benefit lives inside the local network, so the total user count says almost nothing about whether the effect exists.

An aerial view of dense housing and a few high-rises among trees in Kochi. — A network effect lives inside a neighbourhood rather than a country. A rider in Kochi gains nothing from a driver in Kolkata. Photograph: Jomin Thach · Pexels

Andrew Chen’s word for the local unit is the atomic network, which an excerpt from The Cold Start Problem defines as the smallest network that can stand on its own. For a workplace chat product it may be a team of fewer than ten in a single company. For Uber’s early growth it was a single time and place, a station at five in the evening, before it was a city. Chen’s advice is to focus on density and set aside concerns about total market size. That is also the measurement advice. Count users per atomic network, not users in total.

The test: retention against density

Define the atomic network for your product: a city, a pin code, a campus, a company, a WhatsApp group, a category of seller. Assign every user to one. For each network compute a density measure that reflects what users actually benefit from: active suppliers per pin code, active colleagues per team, listings per category per city. Then take the cohorts that joined in the same month and compute their retention at a fixed point, ninety days or whatever your product’s natural period is, separately for each network.

Railway tracks run in parallel and cross one another seen from directly above in Bangkok. — More lines make a junction more useful only up to a point. Plot the value against density and look for the rise and then the flattening. Photograph: Markus Winkler · Pexels

Plot retention against density, with density on a doubling scale. If the product has a network effect, the denser networks retain better, the line rises, and at some density it flattens because the next user adds little. If the line is flat, users in dense networks are no better off than users in sparse ones, and there is no network effect to claim, however many users there are. The figure lets you see both shapes and how noisy real data makes them.

Two practical rules make the chart trustworthy. Use at least six or eight networks, spread across a wide range of density, because three points will fit any story. And make sure each network holds enough users in the cohort for its retention figure to mean something; a pin code with nine households will swing by ten points on one family’s decision. If you do not yet have enough networks of enough size, say so, and report the test as pending rather than passed. That is an honest position for a company in its first year, and a better one than a claim that collapses on the first question.

A network effect you cannot see in retention by density is a network effect you are asserting, not one you have.

The other signals worth tracking

Li Jin and D’Arcy Coolican of a16z listed sixteen ways to measure network effects in December 2018. Four are practical for an early company. Organic share: as the network grows, the share of new users who arrive without paid acquisition should rise. Cohort retention over time: newer cohorts should retain better than older ones at the same age, though the authors note that this often fails because early adopters are unusually motivated. Market depth and match rate: in a marketplace, whether a user who looks for something finds it, and how quickly. Pricing power: whether participants will pay more, through fees or take rate, as the network becomes more valuable.

They also suggest asking how many of your users use competing services at the same time. Heavy multi-tenanting, a seller listing on three platforms or a rider with two apps, means the network effect is weaker than it looks, because users are not locked in by the value of your network alone. Wherever sellers or drivers can list on more than one platform at no cost, this number deserves a place on the dashboard.

A worked example: home services across pin codes

A Hyderabad startup connects households with electricians and plumbers. Its deck claims a network effect: more providers mean faster bookings, which means more households, which attracts more providers. It has 40,000 households and 900 providers across sixty pin codes.

The team groups pin codes by active providers per pin code and looks at households who joined between January and March: what share booked again within ninety days. In pin codes with fewer than four providers, about 18 per cent rebooked. With four to eight, about 24 per cent. With eight to sixteen, about 29 per cent. Above sixteen, about 31 per cent, and no higher above thirty. That is roughly five points per doubling, flattening after three doublings: a real local network effect with a ceiling.

Then the team checks confounders. The denser pin codes are also the wealthier ones, and wealthier households might rebook more anyway. Comparing similar-income pin codes at different densities, the lift shrinks to about three points per doubling but does not vanish. The honest slide says so: a local network effect worth three points of ninety-day retention per doubling of provider density, saturating at about thirty providers per pin code. The strategy follows from the number. Concentrate supply to reach the saturation density in one pin code at a time rather than spreading providers thinly across the city.

Confounders that fake a network effect

Wealth and urbanity. Denser networks are often in richer, more urban places whose users would retain better anyway. Compare networks of similar income or city tier. Age of the network. Your oldest networks are densest and also have had longest to improve the product locally. Compare cohorts that joined in the same month. Where you spent. If marketing and sales effort was concentrated in some networks, density and retention may both be results of the spend. Selection. Users who choose to join a dense network may differ from those who join a sparse one. None of these makes the test useless. Each is a reason to compare like with like before you claim the line.

A monthly ritual: the density chart

On the first working day of each month, refresh one chart: retention of the latest complete cohort against density, by atomic network, with last month’s line faintly behind it. Beside it, three numbers: organic share of new users, match rate or its equivalent, and the share of your users active on a competitor. If the line is rising and steepening, the network effect is forming and you should concentrate on density. If it is flat, stop claiming the effect and look for the power you actually have. Either way you will know, which is more than most decks can say.

The worked example is illustrative. Sources were checked in October 2026. Nothing here is investment advice.