A proxy is a stand-in: you use one thing to represent another when the thing you care about is hard to observe or measure directly. Revenue is a proxy for value created; NPS is a proxy for loyalty; clicks are a proxy for engagement. Proxies are everywhere in business and analysis. The question is never "do we use proxies?" — we must. The question is whether the proxy is valid: does it track what we care about, and under what conditions does it break? A good proxy is correlated with the target and is cheaper or faster to measure. A bad proxy is noisy, gamed, or correlated for the wrong reasons.
The risk is proxy failure: optimising the proxy while the target moves the wrong way. Goodhart's Law — when a measure becomes a target, it ceases to be a good measure — is the classic warning. The strategic discipline is to name the target, choose a proxy that tracks it, and monitor for divergence. When the proxy and the target can diverge (e.g. short-term revenue vs long-term value), the proxy is dangerous. When they're aligned or the proxy is cheap to check, the proxy is useful. The same metric can be a good proxy in one context and a bad one in another.
Section 2
How to See It
Proxy reasoning shows up whenever someone uses a measurable quantity to stand for something they can't or don't measure directly. The diagnostic: what do we really care about, and what are we actually measuring? Look for "we use X as a proxy for Y" or for metrics that are clearly stand-ins — engagement rate for value, pipeline for revenue, citations for impact. When incentives are tied to a metric, ask whether that metric is a valid proxy for the outcome the organisation wants. The pattern: target (hard to measure) → proxy (easier) → decision or incentive. Watch for when the proxy drifts from the target.
Business
You're seeing Proxy when a company uses "monthly active users" as a proxy for engagement. The target might be "users getting value." MAU is easy to measure; value is hard. The proxy works until growth tactics inflate MAU with low-value users. Then MAU and value diverge — the proxy fails. The same logic applies to "qualified pipeline" as a proxy for future revenue: it can be gamed or become noisy when definitions shift.
Technology
You're seeing Proxy when engineering uses "lines of code" or "story points" as a proxy for productivity. The target is output that matters — features, reliability, impact. The proxy is cheap to measure. When incentives attach to it, developers may optimise lines or points rather than impact. The proxy ceases to track the target. Better proxies might be deployment frequency, incident rate, or user outcomes — each with its own failure modes.
Investing
You're seeing Proxy when an investor uses "growth rate" as a proxy for quality. The target is sustainable value creation. Growth can be a valid proxy when unit economics are sound; it becomes a bad proxy when growth is bought with unsustainable burn or one-off cohorts. The discipline is to check whether the proxy still tracks the target at the scale and stage in question. Proxies that work at seed often break at scale.
Markets
You're seeing Proxy when a regulator uses "capital ratio" as a proxy for bank safety. The target is systemic stability. The ratio is measurable and standardised. When banks optimise the ratio by shifting risk off balance sheet or into complex instruments, the proxy and the target diverge. The proxy failed in 2008. The lesson: when the proxy is gamed, strengthen the proxy or add other measures that are harder to game.
Section 3
How to Use It
Decision filter
"Whenever you use a metric or indicator to stand for something you care about, name the target and the proxy. Ask: does the proxy track the target in this context? Can it be gamed or can it diverge? If incentives attach to the proxy, monitor for divergence and be ready to change the proxy or the incentives."
As a founder
Name the outcomes you care about — retention, value delivered, sustainable growth — and the proxies you use — NPS, usage, revenue. Check alignment: when we optimise the proxy, does the target improve? When they can diverge (e.g. revenue up, retention down), either improve the proxy or add a constraint. The mistake is incentivising a proxy that can be gamed. Sales comp on pipeline without conversion or retention creates bad behaviour. The second mistake is never validating: do we ever measure the target directly? Spot-checks and cohort analysis can confirm whether the proxy still holds.
As an investor
Portfolio companies and markets are full of proxies: growth for quality, multiples for value, engagement for retention. Ask: what is the target, and is the proxy still tracking it? Proxies break when the world changes — new competition, saturation, regulation. When you see a company optimising a proxy (e.g. user growth) while the target (e.g. monetisable engagement) is unclear or weakening, that's a red flag. Demand evidence that the proxy and target are aligned.
As a decision-maker
When you set targets or KPIs, you're choosing proxies. Make the target explicit: "We care about X. We measure Y as a proxy." Then design so the proxy is hard to game and periodically validate with a direct measure of X where possible. When someone proposes a new metric, ask what it's a proxy for and under what conditions it would fail. The discipline reduces Goodhart effects and keeps incentives aligned with intent.
Common misapplication: Treating the proxy as the target. Once you forget that the metric is a stand-in, you optimise the number and lose the outcome. Keep the target in view. Review whether the proxy still tracks it.
Second misapplication: Using one proxy forever. Relationships between proxy and target change — markets, products, and behaviour shift. Proxies that worked at launch may fail at scale. Revalidate periodically and be willing to change the proxy or add new ones.
Bezos insisted on metrics that tracked customer value: "We want to be customer-obsessed, not competitor-obsessed." Amazon uses proxies like repeat purchase rate, delivery speed, and defect rate — and explicitly avoids optimising proxies that can diverge from long-term value (e.g. short-term profit). The "working backwards" press release is a way to keep the target (customer outcome) in view so proxy choices stay aligned. Bezos also warned against "proxy abuse" in internal metrics: when a metric becomes the goal, it often stops measuring what it was supposed to measure.
Jim SimonsFounder, Renaissance Technologies, 1982–present
Renaissance uses vast arrays of data as proxies for future price moves. The target is predictive signal; the proxies are historical patterns, order flow, and other quantifiable series. The entire enterprise depends on proxies remaining valid out of sample. Simons' discipline: rigorous backtesting and constant monitoring for when proxies break. When the relationship between proxy and target shifts (regime change), the model is updated or retired. Proxy validity is the core risk in quantitative investing.
Section 6
Visual Explanation
Proxy: Use a measurable stand-in (proxy) for a hard-to-measure target. Validate that the proxy tracks the target; monitor for gaming and divergence.
Section 7
Connected Models
Proxy sits at the intersection of measurement, incentives, and inference. The models below either explain when proxies fail (Goodhart, vanity metrics), how to interpret them (signal vs noise, correlation vs causation), or what can go wrong (confounding, map vs territory).
Reinforces
Signal vs Noise
A proxy is useful when it carries signal about the target and isn't mostly noise. Signal vs noise is the same idea applied to data: we want measures that reflect the underlying quantity. Noisy proxies are bad proxies. Improving measurement and filtering noise improves proxy validity.
Reinforces
Correlation vs Causation
A proxy is correlated with the target; that correlation may or may not be causal. If the correlation is coincidental (e.g. both driven by a third factor), the proxy can break when the third factor changes. Correlation vs causation reminds us to ask why the proxy tracks the target — and whether that link is stable.
Reinforces
Confounding Factor
Confounding is when a third variable drives both proxy and target. The proxy appears valid until the confound changes. Controlling for confounds or choosing proxies with a causal link to the target reduces this risk. Proxy choice is a form of confound management.
Leads-to
Goodhart's Law
Goodhart's Law: when a measure becomes a target, it ceases to be a good measure. That's proxy failure under incentives. The proxy was valid when it was just observed; when we optimise it, behaviour adapts and the proxy can diverge from the target. Proxy discipline includes anticipating Goodhart and designing metrics that are hard to game.
Section 8
One Key Quote
"When a measure becomes a target, it ceases to be a good measure."
— Charles Goodhart, 1975 (Goodhart's Law)
The quote is the canonical warning about proxy failure. The measure was a proxy for something we cared about. Once we make it the target of incentives or strategy, behaviour shifts to optimise the measure, and the link between measure and original goal can break. The practitioner's job: choose proxies that are hard to game, or accept that they will need to be updated when they become targets. Keep the real target in view.
Section 9
Analyst's Take
Faster Than Normal — Editorial View
Name the target. Before you adopt a metric, write down what you actually care about. "We care about X. We're measuring Y as a proxy." If you can't state X, you're at risk of optimising a number that doesn't map to an outcome. The target might be retention, value delivered, or long-term margin — make it explicit.
Check alignment under incentives. The moment you tie compensation or promotion to a proxy, you change behaviour. Ask: can someone move the proxy without moving the target? If yes, you have a Goodhart risk. Either change the proxy, add constraints (e.g. revenue and retention), or accept that you'll need to revise the proxy when gaming appears.
Revalidate over time. Proxies that worked at one stage can fail at another. Growth as a proxy for quality may hold at seed and break at scale. Engagement as a proxy for value may hold for one product and break when the product changes. Schedule periodic checks: does the proxy still track the target? If not, change the proxy or the target.
Use multiple proxies when stakes are high. Single proxies are easier to game. When the decision or incentive is important, use several proxies that would require different (and conflicting) behaviours to game. No single metric is perfect; a dashboard of aligned proxies is more robust.
Section 10
Test Yourself
Is this mental model at work here?
Scenario 1
A company ties sales comp to 'pipeline created.' Reps start inflating pipeline with low-quality leads. Conversion and revenue drop.
Scenario 2
A team measures 'lines of code' to track productivity. They later switch to 'deployments per week' and 'incident rate.'
Scenario 3
An investor says 'we care about long-term value creation' and uses revenue growth as the main metric for portfolio companies.
Section 11
Summary & Further Reading
Summary: A proxy is a measurable stand-in for a hard-to-measure target. Use proxies deliberately: name the target, choose a proxy that tracks it, and monitor for divergence and gaming. When a proxy becomes a target, it can cease to be a good measure (Goodhart's Law). Revalidate over time and prefer proxies that are hard to game when incentives attach. Connected ideas include signal vs noise, correlation vs causation, Goodhart's Law, and vanity metrics.
Practical guide to choosing metrics that matter. Emphasises the difference between vanity metrics and actionable metrics — in effect, bad vs good proxies — and how to pick proxies that track business outcomes.
Muller on the misuse of metrics as targets. Covers Goodhart's Law, Campbell's Law, and the damage done when proxies are optimised without regard to the underlying goal. Historical and institutional examples.
Kahneman on substitution: we often answer an easier question (the proxy) when we're asked a hard one (the target). The same mechanism appears in proxy choice. The book provides the cognitive basis for why we use proxies and when we misuse them.
Doerr on OKRs. OKRs are a framework for choosing and aligning proxies: objectives are targets, key results are measurable proxies. The book stresses alignment and periodic review — the same discipline that keeps proxies valid.
Reinforces
Map vs Territory
The map is not the territory; the proxy is not the target. Map vs territory warns against confusing the representation with the thing. Proxy is the same caution applied to metrics: don't treat the stand-in as the real thing. Update the map when the territory changes.
Tension
[Vanity Metrics](/mental-models/vanity-metrics)
Vanity metrics are proxies that look good but don't track outcomes that matter. They're proxies that have failed — they're easy to optimise and don't move the target. The tension: vanity metrics are often the easiest to measure and report. The discipline is to prefer actionable proxies that track the target over impressive ones that don't.
