The Marginal User Method
Find the user on the cusp of converting, then go watch the worst-case version of them.
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 95%
Instead of optimizing for the average user, Adriel Frederick targets the marginal user — the person right on the cusp of taking the action you want. Data locates them; direct observation explains them. The technique pairs a quantitative search (where is traffic high but conversion terrible?) with a deliberate trip to the extreme worst-case user, whose experience exposes the complete list of product defects at once.
Origin
Developed by Adriel Frederick on Facebook's growth/user-acquisition team while designing global registration, where the extreme case was a user on a feature phone over EDGE, far from a Facebook data center.
Core principles
- 01The marginal user is the person just on the cusp of taking the action you want.
- 02Data gives you a clue where they are and hints at what's wrong — it never gives the answer.
- 03The worst-case user shows you everything that's wrong; the marginal user tells you what to fix first.
- 04Funnel analysis is blind to problems orthogonal to the funnel.
- 05Quantitative and qualitative are complements, not substitutes.
How to run it
- 1
Use data to locate the pocket of marginal users
Look for a segment with high demand but poor conversion — e.g. a country with lots of registration traffic but terrible conversion rates. High traffic proves they want in; low conversion proves you're failing them.
Pro tip Segment by geography, device class, and network condition, not just by funnel step.
- 2
Go to the extreme worst case inside that pocket
Identify the most poorly-served user in the segment — the feature phone on a slow network, far from your infrastructure, in a language you may not support. Enumerate everything wrong with their experience.
Pro tip The worst case is a diagnostic instrument: it surfaces the whole defect list (language detection, phone number formatting, latency) in one pass.
Watch out Don't try to fix the worst case wholesale — some of its problems are structurally unattackable.
- 3
Go watch and talk to real users, not just the funnel
Observe the person signing up. Ask what they're doing and why. This is where you find the problems the funnel cannot show you, because they sit orthogonal to the steps you instrumented.
Pro tip Ask the deceptively dumb follow-up: 'does anyone in the real world call you that?' — the answer often changes the product.
Watch out Data-rich companies breed the illusion that you're 'swimming in answers' and just need to tease them out. You aren't.
- 4
Subtract the unfixable barriers to define what's actually marginal
From the worst case, mentally remove the barriers you can't attack (device, network). Ask: if this person had the best phone and a great connection in that country, what would still be broken? Whatever remains is your prioritized, attackable list.
Pro tip This subtraction converts an overwhelming defect list into a shortlist of the barriers closest to being resolved.
- 5
Fix the attackable barriers and make that user your North Star
Ship the fixes that unblock the marginal user — correct language detection, country detection for phone formatting, latency reduction. Making the marginal user successful makes a long tail of similar users successful.
In the wild
Frederick watched someone in India sign up for Facebook for the first time. Asked what name they'd enter, they said their full legal name. He asked whether anyone in the real world calls them that — no. That meant outbound friend requests wouldn't be accepted (nobody recognized the name) and inbound searches would fail. A funnel analyst would have chased the friend-request accept rate and built a mini-funnel around it, never seeing that the root cause sat far upstream at name entry.
→ A problem completely invisible in the funnel data was surfaced by watching one person, revealing a root cause several steps back from where the metric was bleeding.
Hunting a country with high traffic and terrible registration conversion, Frederick pushed to the extreme user: a feature phone, on an EDGE connection, far from a Facebook data center. Inspecting that single experience yielded the full defect list — wrong language, country not detected so phone numbers formatted wrong, punishing latency.
→ One worst-case experience produced the complete prioritized fix list for the whole market's registration flow.
Common mistakes
Using data alone to identify and diagnose the marginal user
Data tells you where they are and roughly how bad it is. It cannot tell you why. Problems orthogonal to the funnel are invisible to funnel analysis; you have to go watch.
Optimizing for the average or the loudest user
Averages hide the cusp. The people already converting don't tell you what to build; the people who nearly convert and fail do.
Trying to fix everything the worst case reveals
Some barriers (device class, network generation) are unattackable. Without the subtraction step you drown in a list you can't act on.
Is it for you?
Best for
Growth PMs and founders with meaningful traffic but weak conversion in a specific segment or geography, especially international expansion.
Not ideal for
Pre-traffic products with no conversion data to segment, or teams that cannot get access to real users to observe.
From the transcript
“for me it's a person who is just on the cusp of taking the action you want to take”
“something I caution people against though is don't use the data alone to figure out who the marginal user is”
“there's a problem that's orthogonal to that funnel that you can't see from looking at the data and you have to go look at the…”
“you fall into the Trap of thinking that you're swimming in answers because you have all this data”
From the episode
Humanizing product development
Adriel Frederick (Reddit, Lyft, Facebook)