ExpertsFMHow To Better Understand Your Users
Y Combinator Startup Podcast · July 9, 2026
5 Selected Lessons:
framework · 1 heuristic · 1 reasoning_pattern · 1 case · 1 priority · 1
Two essential tools for understanding users: cohort retention curves + dot plots
Source
Y Combinator Startup Podcast — How To Better Understand Your Users (July 9, 2026) · at 8:49 ▸
When this applies
when deciding which measurement tools deserve limited founder/team attention to truly understand whether a product is working
The takeaway
Combine a cheap aggregate health-check tool with a cheap disaggregated explanatory tool rather than relying on either alone.
Why it works
Cohort retention curves answer the aggregate 'do people stick around' question cheaply, while dot plots answer the disaggregated 'how and why' question; together they form a complete diagnostic loop from detection to explanation to action.
One level up
The most valuable measurement systems pair a macro signal with a micro explanation so you know both that something is wrong and why.
Where it stops holding
This stops working when the product is too early/pre-retention to have meaningful cohorts, in which case dot plots alone may suffice.
Detail
hierarchy:
- Cohort retention curves — aggregate signal of whether groups of acquired users stick with the product over time
- Dot plots — disaggregated signal of how individual users actually use the product, which features drive value, and where to fix things
rationale: Retention curves tell you IF there's a retention problem; dot plots tell you WHY and give you the specific questions to ask users and features to build/fix.
In their words
So cohort retention curves and dot plots are, in my experience, two of the most important tools that you've got to understand your users.
Dot plot: individual-user visualization grid
Source
Y Combinator Startup Podcast — How To Better Understand Your Users (July 9, 2026) · at 0:00 ▸ · at 4:00 ▸ · at 6:41 ▸ · at 8:49 ▸
When this applies
when you need to understand how individual units of a system behave over time, not just the aggregate trend
The takeaway
When an aggregate metric could mask heterogeneous underlying behavior, build a visualization that preserves individual-unit granularity so pattern recognition can happen visually.
Why it works
Aggregating data across users averages away the very behavioral variance that signals product-market fit or failure; plotting each user as a row against time as columns preserves that variance so the human eye can spot patterns.
One level up
Trust human pattern recognition over summary statistics when the signal you need is in the variance, not the mean.
Where it stops holding
This stops working when you have too many users to visually sample or when the event you choose to track doesn't actually represent value in the product.
Detail
procedure:
- Make a 2D grid: one row per user, one column per time period (usually days)
- Pick one event that represents value in the product (e.g., listening to a song, processing an invoice)
- Place a dot in each cell where that user performed the event on that day
- Mark the user's first-use day with a distinct symbol (e.g., a ring)
- Optionally encode additional dimensions (features used, device type, geography, income) with different symbols or cell shading
- Sort/filter rows by attributes of interest to surface patterns
- Scan the grid visually to let pattern recognition surface insights an aggregate chart would hide
benchmarks: {"team_size_fit":"works with 10-20 users up to billions via sampling"}
rationale: Aggregate charts like DAU/MAU lump all users together and tend to trend up-and-right even when individual users are unhappy or not getting value; dot plots expose the underlying heterogeneity.
In their words
What you basically do is just make a two dimensional grid, like a spreadsheet, where there are a bunch of columns and a bunch of rows.
Segmenting users by behavioral timing to find distinct use-case cohorts
Source
Y Combinator Startup Podcast — How To Better Understand Your Users (July 9, 2026) · at 4:00 ▸
When this applies
when a visualization reveals that your user base splits into behaviorally distinct subgroups you hadn't previously defined
The takeaway
Let the data surface natural segments before imposing your own hypothesized personas; then use the discovered segments to drive product and targeting decisions.
Why it works
Visualizing raw per-user timing data lets a human notice temporal clustering (e.g. weekday vs weekend) that no predefined segment or survey question would have captured.
One level up
Discovery should precede categorization — build tools that let patterns emerge rather than forcing data into assumed buckets.
Where it stops holding
This stops working when the sample of users is too small or too homogeneous for any real temporal pattern to exist.
Detail
approach: Scan individual-user rows for timing clusters (e.g., weekday-only vs weekend-only usage)
signals:
- consistent usage on specific days of week
- single-day trial with no return
- bursts of activity following a specific feature use
rationale: These clusters suggest different underlying use cases or user types, which can inform redesign, targeting, or prioritization decisions.
In their words
there's a couple of users who seem to kind of only use it on the weekends. That's an interesting observation that might help me redesign my product in a different way or target different users,
Aggregate growth metrics can mask unhappy users
Source
Y Combinator Startup Podcast — How To Better Understand Your Users (July 9, 2026) · at 0:00 ▸ · at 6:41 ▸
When this applies
when any top-line metric is trending positively while you have no visibility into whether the underlying units (customers, users, accounts) are actually satisfied
The takeaway
A rising aggregate count is not evidence of satisfaction; it can simply reflect acquisition outpacing dissatisfaction.
Why it works
Summing or counting across a growing user base produces an upward trend almost mechanically as new users are added, independent of whether existing users are deriving value or churning quietly.
One level up
Never let growth in volume substitute for evidence of value delivered per unit.
Where it stops holding
This stops working as a warning when your product has so few users that you can already observe them individually without any aggregation distortion.
Detail
rule: Don't trust an up-and-to-the-right aggregate chart as proof of product health.
threshold: Any amount of growth can produce this illusion, per the speaker.
rationale: Aggregate metrics lump all users together, obscuring whether individuals are actually enjoying or benefiting from the product.
In their words
And if you have any amount of growth, those graphs tend to be going up into the right, even if users aren't actually enjoying using your product.
PayPal fraud detection via human pattern-spotting on raw visualizations
Source
Y Combinator Startup Podcast — How To Better Understand Your Users (July 9, 2026) · at 6:41 ▸
When this applies
when you face an anomaly-detection problem but don't yet know what the anomalies look like
The takeaway
When you don't know what pattern you're looking for, give humans a raw visual representation of the full data and let them flag anomalies, then investigate those flags.
Why it works
Humans are good at perceiving 'that's different' in a visual field even without an articulated rule, so raw visualization plus human eyes can find signal before you have a formal model.
One level up
Before building automated detection, use human perception as a cheap first-pass anomaly detector to discover what the patterns even are.
Where it stops holding
This stops working when the data volume is too large for humans to visually scan, or when the fraud/anomaly signal is not visually distinguishable at all.
Detail
situation: PayPal had a major fraud problem at launch but didn't know what fraud patterns to search for.
decision: They built a visualization/graph of all transactions and had humans stare at the screens to spot anomalies visually.
outcome: Humans could flag 'that thing there is different, probably fraud' without knowing the exact rule, enabling the team to dig in and identify real fraud cases.
lesson: Raw visual pattern recognition by humans can precede and inform formal detection systems when the patterns are unknown.
In their words
while the humans didn't know what exactly was going on, they were able to look at the screen and say, That thing happening there, that's different and probably fraud.