Proxy Metric Selection
Goal on simple short-term inputs you can prove drive the long-term output you actually want
- Difficulty
- Moderate
- Time to result
- ~weeks to results
- Steps
- 5
- Confidence
- 95%
Lachs's method for choosing what a team is goaled on: never goal on the long-term outcome itself (retention), because it can't be moved or measured meaningfully in a sprint. Instead identify the short-term inputs that plausibly drive it, prove the link by experiment, and goal on those. Keep each metric simple and intuitive even at the cost of accuracy — an understood metric beats a precise composite nobody can move.
Origin
Jessica Lachs, learned across a decade of metric-setting at DoorDash — explicitly 'mostly from bad metrics', including a failed Merchant Health composite score.
Core principles
- 01You learn more from picking the wrong metric than the right one
- 02Retention is a terrible thing to goal on — find its inputs instead
- 03A proxy metric must be experimentally shown to drive the long-term output, not assumed to
- 04Simple and imperfect beats perfect and uninterpretable
- 05Three understandable metrics beat one composite score
- 06Sequence the inputs by impact; hitting the top three is usually 95% of the value
How to run it
- 1
Name the long-term output you actually care about
State the true business outcome (retention, engagement, lifetime value) plainly — then accept you cannot goal a team on it because it cannot be moved or measured on an experiment's timescale.
Watch out Teams goaled directly on retention thrash: the feedback loop is longer than the planning cycle.
- 2
List the candidate short-term inputs
Ask what actually drives the output. For merchants: active hours on the platform, photo coverage, an accurate and robust menu. Each must be something a team can move within weeks.
Pro tip Inputs a team can act on directly are worth more than inputs that are merely correlated.
- 3
Test the link by experiment
Run experiments to confirm the short-term input actually moves the long-term output. A proxy that doesn't drive the outcome is worse than no proxy — it manufactures confident, wrong work.
Pro tip Quantify the elasticity: what does a 20% improvement in this input buy you?
Watch out Skipping this step is how vanity metrics enter the goal stack.
- 4
Refuse the composite
Do not weight inputs into a single coefficient-laden score. If nobody can say whether a 0.1 increase is good, the metric is meaningless. Ship the two or three raw inputs as separate goals instead.
Pro tip The test: can someone across the company talk about this metric and have an intuition for it? If not, simplify.
Watch out Data scientists gravitate to composites because they are more 'correct'. Correctness that nobody can move is worthless.
- 5
Rank the inputs and attack them in order
Quantify which input matters most, goal a team on that one, and materially move it before moving to the next. Getting one through three right is usually 95% of the value; the remaining inputs can wait.
Pro tip Keep a team on a metric long enough to build real expertise in its levers rather than rotating them between metrics.
Watch out Chasing all six inputs at once produces movement on none.
In the wild
DoorDash built a Merchant Health score combining active hours, image coverage, menu accuracy, and other inputs into one weighted number. When it read 35, nobody could say what 35 meant or how to move it. The team replaced it with a simple leading question — how many new merchants get an order within their first seven days — plus direct goals on individual inputs like photo coverage and accurate open hours.
→ Three understandable metrics replaced one composite; teams knew exactly what they were trying to move, and the score's residual precision was judged not worth the loss of intuition.
Rather than goaling teams on retention, DoorDash identifies the inputs that drive retention and experiments against those, letting teams iterate quickly on things they can actually move inside a quarter.
→ Fast iteration cycles on measurable inputs, with retention treated as the output to be validated rather than the dial to be turned.
Common mistakes
Goaling on the long-term outcome
Retention and LTV move too slowly to support experimentation. A team goaled on them cannot learn within its planning cycle and will attribute noise to its own work.
Building the composite score
Weighted composites optimise for statistical completeness and destroy interpretability. If a metric can't be discussed intuitively across the company, it won't drive real outcomes.
Assuming the proxy without testing it
A short-term input is only a valid proxy if experiments show it drives the long-term output. Untested proxies encode a hypothesis as a goal.
Is it for you?
Best for
Data, product, and growth leaders setting quarterly team goals in a business where the outcomes they care about (retention, LTV) are too slow to goal on directly
Not ideal for
Situations where the true outcome is fast and directly measurable (e.g. checkout conversion on a single-step flow), where a proxy adds indirection for nothing
From the transcript
“ultimately you want to find a shortterm metric you can measure that drives a long-term output”
“retention is a terrible thing to goal on because it's like it it it's almost impossible to to drive in a meaningful way in a…”
“we find proxy metrics for long-term outcomes”
“so I always encourage folks just pick something simple even if it's not perfect”
“it's simpler to have a composite metric but it was so hard to understand what it was and how to move it that it it…”
“doing the things that matter first and most quickly like is a competitive advantage in my opinion”
From the episode
Building a world-class data org
Jessica Lachs (VP of Analytics and Data Science at DoorDash)