"We have network effects" is the most over-claimed and least-examined line in startup strategy. In practice it functions as a permission slip to stop working: the founder believes the graph is now defending itself, so the energy that built the position gets redirected somewhere else. That belief is usually wrong on the facts, and the redirection is what actually kills the company. A real network effect is narrower, weaker, and more perishable than the word implies, and treating it as a wall you own rather than a state you maintain is a category error with a body count.
Start with the definition, because almost every misuse begins by blurring it. A network effect exists when the product becomes more valuable to a given user as more other users join the same network and that user can interact with them. The telephone is the clean case: a phone is worthless with one subscriber and more useful to me specifically as the people I want to reach get lines. The value accrues to the same user from the same graph. Hold that definition rigidly, because most of what gets filed under "network effects" fails it.
Four things founders miscall network effects
Scale economies. Your unit costs fall as volume rises, so you can price below a smaller competitor. Real advantage, wrong label. This is a supply-side cost curve; it says nothing about whether the product gets better for me because you have more users. AWS is cheaper at scale. That is not a network effect, and it defends differently: a better-capitalized entrant can buy down the same curve.
Data feedback loops. More usage produces more training signal, the model improves, the product gets better for the next user. Genuinely powerful, and genuinely not a network effect in the interaction sense. My value does not depend on other users being reachable inside a shared graph; it depends on the vendor having ingested enough aggregate behavior. That is a scale-and-learning economy. It has its own decay function: signal saturates, and a competitor with a smarter architecture can match your quality on a fraction of your data. Conflating the two hides exactly the boundary condition that matters.
Brand. Trust and recall are durable, but they are a property of one relationship, you and the customer, not of users interacting with each other. Brand is real and, of the four, the most defensible, precisely because it is expensive to build and expensive to fake. That is why your best marketing looks like waste: a durable reputation is a costly signal in Zahavi's sense, a handicap only a serious operator can carry. It is defended by continued expenditure, not by a graph.
Switching costs. Data lock-in, integrations, retraining friction. This keeps users who would otherwise leave. It is a real moat and the one most worth engineering. It is also a wall around each user individually, not a network effect, and it decays every time an entrant ships a one-click importer.
The reason the distinction is not pedantic: these four decay along completely different curves and are attacked with completely different weapons. Call them all "network effects" and you will defend the wrong flank.
Even the real ones are weaker than the pitch deck
Grant a genuine, definition-passing network effect. It still has three properties that make "wall" the wrong mental model.
It saturates. Metcalfe's law, value proportional to n², is the number every deck implies and almost none obey. The n that matters to me is not the global user count; it is the count of people I specifically need to reach who are on the network. Once my relevant contacts are on WhatsApp, the ten-millionth stranger in another country adds zero to my value. The marginal user contributes roughly nothing past a threshold that is reached early and locally. So the wall does not keep getting taller as you scale. It plateaus at your relevant n, which is small, while your reported n keeps climbing and flattering the story you tell yourself.
It is local and clusterable. Because value is local, the graph is not one wall; it is thousands of dense clusters loosely stitched together. A competitor does not need to beat your global graph. It needs to win one dense subgraph where it delivers more value, achieve local liquidity there, and expand cluster by cluster. This is the classic disruption vector, and it is why global scale is nearly irrelevant at the point of attack. Facebook did not beat MySpace's larger network head-on; it saturated one campus at a time, where the local graph was denser and the relevant-n was 100% of the people you cared about. Slack did not out-scale email; it won individual teams. Your ten million users are no defense inside the challenger's beachhead of two thousand, because the two thousand only care about each other.
It is multi-tenant and low-loyalty. The telephone metaphor smuggles in an assumption that rarely holds now: that a user belongs to one network. Real users multi-home. A driver runs Uber and Lyft on the same phone and switches by surge. A creator posts to Instagram, TikTok, and YouTube the same afternoon. A merchant lists on three marketplaces. When multi-homing is cheap, the network effect stops producing lock-in even where it produces value, because the user harvests the value of your graph without leaving anyone else's. Loyalty is not a property of the network; it is a property of switching cost, which is a different moat that you have to build separately and that you probably assumed the network effect was giving you for free.
Put the three together and the picture inverts. The thing sold as a smooth, ever-rising wall is actually a plateau (saturation), full of independently winnable pockets (clusterability), around users who never agreed to be exclusive (multi-homing).
To be precise about the boundary, since a sharp reader will supply the counterexamples: some network effects genuinely are strong, and they are strong exactly where these three properties fail. They resist saturation when the relevant-n is effectively the entire graph rather than a local cluster, a global payment or messaging standard where the person you need to reach could be anyone. They resist clusterability when liquidity is expensive to replicate locally, so a challenger cannot bootstrap a beachhead cheaply: a stock exchange, a dominant developer platform, a two-sided market where thin local supply is worthless to demand. They resist multi-homing when the switching cost is high and the value of a second network is low, so users consolidate on one. Where all three hold at once, you have something close to a durable structural moat. The point is not that such moats never exist. It is that they are rare, they are specific, and the founder claiming one owes you an argument for why saturation, clustering, and multi-homing all fail in their case, an argument I almost never hear, because the word is doing the work the analysis should be doing.
Advantage is a metabolism, not a monument
Here is the frame that has held up for me across three very different domains. In biology, a living system does not maintain its ordered state by building a wall against entropy. It maintains it homeostatically, by continuously spending energy to hold a gradient that the second law is constantly trying to erase. Stop spending, and the organism does not sit frozen at its last good state; it decays toward equilibrium at a rate set by its environment. Body temperature, blood pH, membrane potential: none are assets you possess. They are processes you run, every second, or lose.
Competitive advantage is the same kind of object. It is not a monument you build once and stand behind. It is a metabolic state you hold against a market that is continuously trying to arbitrage it away. The reason this matters is not poetic. It changes the diagnostic question you ask about any claimed moat, from "how big is it?" to "how fast does it decay if we stop feeding it?"
That decay-rate question is the whole test. Take any advantage you believe you have and ask: if we stopped trying, froze the roadmap, stopped fighting fraud, stopped recruiting the supply side, stopped shipping, how long until it is gone? If the answer is a quarter, you do not have a moat; you have a treadmill you had mistaken for a wall. That is not a reason to despair. Every real advantage is a treadmill. The point is to know you are on one, so you keep running and keep the energy pointed at the thing that actually decays slowest.
I watch this directly in Kommerce, the commerce operating system I build for cash-on-delivery markets. In trust-scarce economies the defensible position is not a graph of buyers and sellers; it is the operational homeostasis of trust itself: fraud rates held down, delivery reliability held up, the confidence a merchant has that money arrives. None of that is owned. Every point of it is a gradient held against constant pressure. Fraudsters adapt weekly, couriers churn, a single bad delivery cohort re-teaches a merchant to distrust the whole rail. Stop spending energy on any of it and the "moat" is gone in weeks, not years. The advantage is real, and it is entirely a process. There is no version of it that survives neglect.
The same discipline runs through the causal-inference work in my research life, and it sharpens the strategic point. In a system with feedback, correlation between "we have a big network" and "we are winning" tells you almost nothing about whether the network is causing the win. The network may be a downstream marker of the real driver, relentless execution, the way an epigenetic clock reads out biological age without being the mechanism that ages you. Kill the execution and watch what happens to the marker. Founders who mistake the readout for the mechanism stop doing the thing that was actually producing the advantage, then are shocked when the readout follows.
What the wall metaphor makes you do
The category error is not academic; it changes behavior in two predictable and expensive ways.
First, it makes you complacent on defense. Believing the graph defends itself, you under-invest in the clusters, exactly where the disruptor is achieving local liquidity. By the time the aggregate numbers move, the challenger owns a dozen dense subgraphs and has the reference customers to take the next dozen. You defended the average while losing the margins, because the average was the number that felt like a wall.
Second, and this is the failure I have watched more often, it licenses the wrong expansion. Convinced the core is safe behind its network effect, the team pours its finite energy into new markets, new products, new user segments, none of which have crossed the local liquidity threshold, all of which consume the metabolic budget the core actually needed. This is how most startups die of indigestion rather than starvation: not from too little, but from swallowing more surface area than they can metabolically defend, on the theory that the original position no longer needs feeding. It always needed feeding. The belief that it did not is the whole mistake.
So retire the sentence. "We have network effects" is, at best, a claim that one of four different advantages is present, each with its own decay curve and its own attacker. At worst it is an anesthetic. Replace it with the only question a moat is really asking: how much energy does it take to hold this gradient, and are we still willing to spend it? A moat you can stop maintaining was never a moat. It was a wall you painted onto a treadmill, right before you stepped off.