The Dot Plot.

A latticework reading of Y Combinator's guide to individual user visualization — which mental models the tool amplifies, which it quietly overturns, and what new ones to add.

▶ Play video

Video: Y Combinator / Dave Fontenot

What this video is really about.

Every founder tracking a dashboard is, in some sense, reading a map. The map may be accurate — the DAU line really is going up — and still be useless as a guide to where the product needs to go next. Dave Fontenot's fourteen-minute tutorial on dot plots is not really about a chart type. It is about the gap between the map and the territory, and the specific form that gap takes when the territory is individual human behaviour.

The video's central claim is precise: aggregate metrics don't just obscure detail — they actively deceive. A DAU graph that holds flat at two or three users per day is compatible with a product everyone loves and returns to daily, or with a product that keeps cycling through churned users who each try it once. You cannot tell which world you are in without looking at individuals. The dot plot is the mechanism that makes individual inspection cheap enough to do routinely.

Three kinds of edits to the latticework follow. Some familiar models come out sharper. Others don't survive the argument intact. And a handful of new structural observations earn a place in the toolkit — ones the canon doesn't yet have a name for.

Old models, sharper edges.

The clearest amplification is of the map is not the territory. A DAU graph is a map. It compresses every individual's session into a single point in a single series. Fontenot's move is to zoom in — not to abandon aggregation, but to hold the individual and the aggregate in the same view simultaneously. The dot plot is a map that is explicitly shaped like the territory: one row per person, one cell per day, exactly mirroring the structure of the thing it represents.

Signal vs. noise gets one of its clearest product-specific illustrations. The noise here isn't random — it's structural. A metric like DAU has a real signal buried inside it, but the aggregation introduces a specific kind of noise: the mixing of users who behave differently. Separating the weekday-listener cohort from the weekend-listener cohort is not a refinement of the DAU signal; it is the removal of noise that was preventing the signal from being readable at all.

Redundancy also appears, in an interesting form. Fontenot recommends layering multiple symbols into the same grid — a ring for the first-use day, different marks for different events. Each additional symbol is a redundant encoding of different information into the same spatial structure. The grid doesn't grow; the density of information per cell increases. The result is that a single glance at a row tells you more than a page of log output would.

Models that do not survive intact.

The most quietly subversive moment in the video is the PayPal story. Fontenot attributes it to Max Levchin: faced with a fraud problem they couldn't algorithmically characterize, PayPal built a visualization and had humans stare at it. The humans couldn't articulate the pattern — they didn't know what fraud looked like — but they could point at the screen and say that's different. This bends the common explicit knowledge bias: the assumption that insight must be articulable to be actionable. The fraud detectors worked precisely because the pattern was visible before it was nameable.

The availability heuristic gets undermined in the other direction too. Fontenot's target audience — founders — is good at constructing stories about typical users. The stories feel available because they are; the founder built the product for someone. But the dot plot reveals that the distribution of actual user behaviours is often bimodal or stranger — weekday-only listeners and weekend-only listeners who would each need completely different product decisions to retain. The "typical user" story is a creature of the aggregate. Zoom in and it evaporates.

Finally, going up and to the right as a proxy for health takes a serious hit. Any DAU graph that draws new users through the top of the funnel and bleeds them out the bottom can show a flat-but-growing line for months. The champions who bought the product are gone; new users cycle through. The line says nothing about which world you are in. Fontenot's B2B churn story is the illustration: $80,000 signed, three of ten seats activated, the champion who drove the purchase left the company — and the only place all of that was visible was the dot plot.

New entries for the latticework.

The video's most portable new model is what might be called champion risk — the specific B2B failure mode where adoption depends on a single internal advocate, and usage never spreads wide enough to survive that person's departure. Champion risk is distinct from churn for dissatisfaction: the product may be working fine, the champion may have loved it, but the organizational surface was never enlarged. The dot plot reveals it as a pattern: ten seats contracted, three ever activated, usage sporadic. No metric captures this as directly.

The PayPal story contributes pattern-first detection: the discipline of exposing data visually before you have a hypothesis about what you are looking for. Classical scientific method runs the other direction — hypothesis first, then look for confirming data. PayPal's fraud team inverted this, and it worked. The dot plot applies the same inversion to product analytics: show every individual's behaviour, without filtering, and let anomalies surface before you have named them.

Third is structured sampling as a management practice. For products at scale, Fontenot describes printing out dot plots for random samples of user segments and handing sheets to team members: "iOS users in France." "Web users in the US earning over $80k." The discipline isn't the visualization; it's the deliberate, repeated, human act of looking at randomly selected individuals across defined segments. This is closer to an ethnographic method than a product-analytics method. It belongs in the latticework alongside confirmation bias as its corrective: you cannot cherry-pick individuals to inspect if you are randomly sampling.

And finally, event fidelity — the observation that the quality of everything downstream depends entirely on choosing the right event to plot. Log "opened the app" and you measure habit; log "processed an invoice" and you measure value delivered. These are different things, and the dot plot will faithfully reveal whichever world you are tracking. The mistake of charting the wrong event is not a technical mistake; it is a conceptual one, the same error as picking the wrong outcome variable in any causal analysis.

When to reach for which.

The latticework, after the dots.

The dot plot is not a novel technology; it is a novel commitment. The commitment is to resist the temptation of the aggregate — not to abandon it, but to hold it alongside the individual view. Most product teams have aggregate data by default and individual data by exception. The dot plot inverts the default.

Most founders just like ignore this. But I think it's the most important signal to figure out if you've built something that people want. — Dave Fontenot, Y Combinator

What the latticework gains from this episode is a sharper version of an old lesson: the map is always a lossy compression, and the specific form of the loss matters. Aggregation loses individuals. The way you recover them is not to disaggregate everything — that's intractable — but to build visualizations that are structurally isomorphic to the thing you care about. One row per person. One cell per day. The dots tell you what the line cannot.