The Data Flywheel: When Usage Makes the Product Measurably Better

A data flywheel is a loop in which product usage creates data, that data improves the product on a measurable metric (accuracy, relevance, match rate, prediction error), and the better product attracts more usage. It is a moat only when the improvement is real, compounds with usage, and cannot be matched by a rival who buys or synthesizes similar data.

Most pitch decks claim one. Few can name the number that improves. That gap is what this page is about.

Data flywheel vs. data network effect vs. Amazon's flywheel

Three ideas share the word "flywheel" and get mixed up constantly.

Amazon's flywheel is a business model loop. Jeff Bezos sketched it on a napkin in 2001, borrowing the flywheel metaphor from Jim Collins, who describes progress as turn upon turn of a heavy wheel, each turn building on work done earlier. Amazon summarizes the loop simply: "Lower prices lead to more customer visits, and more customers increase sales volume and the ability to invest in better customer experience." No data is required for that loop to work.

A data flywheel is narrower. The loop runs through a model or an algorithm: more usage, more data, a better model, more usage. It can exist even when users never interact with each other.

A data network effect is the strongest form. NFX defines it this way: "When a product's value increases with more data, and when additional usage of that product yields data, then you have a Data Network Effect." The same NFX manual warns that data network effects tend to be weaker than many people, especially venture capitalists, want to believe, because more data does not automatically mean more value. For the broader family, see network effects.

The diminishing returns problem

The best-known critique is a16z's essay The Empty Promise of Data Moats. Its core points:

• Most "data network effects" are really scale effects. The authors write that "there generally isn't an inherent network effect that comes from merely having more data."
• The cost of getting new, useful data goes up while its marginal value goes down. After a point, new data only covers the long tail of rare cases.
• The lead can shrink over time. In the essay's words, the moat "erodes as the data corpus grows and the competition races to catch up."

NFX makes the same point about reviews: on Yelp, "the 5th review adds a lot more value than the 30th."

So a real data flywheel needs two things a fake one lacks: a metric that keeps improving with more data, and data that keeps changing (fraud patterns, traffic, taste, new content) so the dataset never goes stale.

Why it matters for startups vs. big companies

For a big company, a data flywheel defends a lead it already has. Its scale means the model has seen more edge cases than anyone else.

For a startup, the math runs the other way. The incumbent already has more data. A startup's data flywheel only works if it collects data the incumbent does not have: a narrow vertical, a new workflow, an outcome that happens inside the startup's product and nowhere else. That is why this test depends on the proprietary data test. A flywheel built on public data is a flywheel anyone can spin.

It also matters in AI. If a competitor can plug in the same foundation model, the only lasting difference is what your own usage teaches the system. See beyond the AI model.

Types of data flywheels

1. Recommendation flywheels. Clicks, plays and purchases improve ranking and personalization (Netflix, Spotify, TikTok, Amazon).
2. Outcome-labeled flywheels. The product sees how things turned out (a chargeback, a correct answer, an actual arrival time) and learns from the label (Stripe, Duolingo, Uber).
3. Real-time crowd flywheels. Every user is also a sensor, and the data expires quickly (Waze). This is the case closest to a true data network effect.
4. Edge-case collection flywheels. A deployed fleet finds rare situations a lab never would (Tesla).
5. Feedback flywheels. Users accept or reject suggestions, which tells the system what good output looks like (Grammarly, search ranking).

Data flywheel examples

• Google Search. Queries and clicks show which results satisfy users. In the 2025 US antitrust remedies ruling, a federal judge ordered Google to share certain search index and user interaction data, including user click data and search queries, with qualified rivals: a sign of how much that data is worth.
• Netflix. Netflix's own researchers reported that the recommender system influences choice for about 80% of hours streamed, and that personalization and recommendations save the company more than $1B per year. NFX still calls its data network effect marginal next to content inventory.
• Spotify Discover Weekly. Launched in July 2015, it looks at the songs you play, finds users who added those tracks to their playlists, and pulls more songs from those playlists. Other people's playlists improve your recommendations.
• TikTok For You. TikTok says the feed ranks videos using signals like videos you like or share, accounts you follow and comments you post, and is built to "continuously improve, correct, and learn from your own engagement."
• Amazon recommendations. Item-to-item collaborative filtering treats items as related when buyers of one are more likely than the average customer to buy the other. More purchases make those relationships sharper.
• Stripe Radar. Radar's fraud models are trained on more than 70 trillion data points across the Stripe network, and Stripe reports a 32% average reduction in fraud. Stripe also says 92% of charges come from a card it has seen before: data a new processor does not have.
• Duolingo Birdbrain. Birdbrain predicts whether a learner will get a given exercise right, estimating both learner ability and exercise difficulty, then uses that prediction to build personalized lessons. Every answer is a labeled data point.
• Waze. Waze uses real-time GPS data and crowdsourced hazard reports from its driver community to route around jams. NFX calls it a strong case because nearly everyone who consumes the data also contributes it, and it must be continuously updated.
• Tesla. Tesla describes its autonomy approach as a pipeline of fleet data collection, labeling, simulation, neural net training and vehicle processing. Cars on the road surface rare scenarios for training.
• Uber. Uber's DeepETA model predicts the gap between the routing engine's estimate and the real arrival time, learning from completed trips. Uber reports it improved production metrics over the previous model.
• Grammarly. Grammarly's privacy policy states that its algorithms analyze your text, writing behavior and usage data, and that using the product gives it permission to use your writing to improve its products.

Notice the pattern. The strongest cases (Stripe, Waze, Duolingo, Uber) learn from outcomes that happen inside the product. The weaker ones learn from signals a well-funded rival could also gather once it has users.

How a startup builds a data flywheel from zero

1. Pick the metric first. Write down the number that should improve with usage: fraud caught, match acceptance rate, prediction error, answer accuracy. If you cannot name it, stop here.
2. Capture the outcome, not just the input. Logging what users typed is easy to copy. Logging whether the match worked, the loan was repaid or the suggestion was accepted is not.
3. Start narrow. A small, self-contained market gives you dense data on one problem faster than an incumbent's broad but thin data. Win one vertical, then expand.
4. Close the loop quickly. Retrain or re-rank often enough that users can feel the improvement. A loop that updates once a year is a report, not a flywheel.
5. Measure the curve. Plot your metric against data volume. If it has flattened, more users will not widen your lead, and you need a different moat.
6. Prefer perishable data. Data that goes stale (prices, traffic, fraud patterns, inventory) forces rivals to match your current usage, not your historical archive.

Common mistakes and fake data flywheels

• "We'll have the most data." Volume is not a moat if the metric plateaus. That is the a16z critique in one line.
• No named metric. "The AI gets smarter over time" is not evidence. A chart of accuracy against usage is.
• Data that is public or purchasable. If a rival can scrape it, license it or generate it with a model, the flywheel is shared.
• Confusing a business loop with a data loop. More customers lowering your costs is a scale advantage, closer to Amazon's original flywheel. It is not a data flywheel.
• Counting the foundation model's improvement as yours. If your product gets better when the model vendor ships an update, so does every competitor's.
• Flywheels "at scale." If the improvement only appears with millions of users you do not have, it is a plan, not a moat.

How Moat Check scores it

Moat Check asks one question for this test: "Does that data measurably make the product better as usage grows?"

• Pass (10 points): A named metric (accuracy, match rate, relevance) improves with more usage, and that is true today or built into the design. Stripe Radar's fraud reduction and Duolingo's exercise predictions are the model.
• Partial (5 points): The loop is plausible but not proven. The improvement only appears "at scale," the metric is vague, gains have flattened, or the data that would drive it is not yet being captured.
• Fail (0 points): More data does not visibly improve the product. The value comes from features, content or a third-party model, not from what usage teaches the system.

This test sits next to proprietary data: one asks whether the data is yours, the other asks whether it makes the product better. A strong score usually needs both. Moat Check applies a fixed framework using AI and web research; treat the result as a starting point, not investment advice. For the full framework, see all moat tests, the methodology and what a moat is.

Check your startup's moat

Questions

Keep reading

• Switching Costs: Why Customers Stay, and How to Earn It
• Distribution Moat: How Startups Own Channels Rivals Can't Buy
• Platform Risk: What Happens When the Platform Ships Your Feature
• All 10 types of moats
• What is a moat in business?