A key result becomes unarguable when it names one metric, a starting baseline, a numeric target, a direction of travel, a deadline, and a named source of truth — the "from X to Y by when" format. Drop any of those six parts, and the KR reopens for debate the moment someone doesn't like how the quarter ended.

Quick answer: Write every key result as metric + baseline → target, by deadline, measured in [source of truth]. If two reasonable people in the same review could disagree on whether it was hit, one of those parts is missing.

Vague key results don't fail loudly. They fail quietly, three months later, when someone says "well, I think we basically hit it" and nobody in the room has the standing to disagree. That failure is entirely preventable — but only if you catch it at write-time, not at review-time.

The Six-Part Anatomy of a Bulletproof Key Result

Every measurable key result contains six explicit parts: a single metric, a baseline (where you start), a target (where you're going), a direction (up or down), a deadline, and a source of truth (who measures it, and where). Leave one out, and the KR becomes an opinion instead of a commitment.

This isn't a stylistic preference. In Measure What Matters, John Doerr traces the format back to Andy Grove's Intel in the 1970s, where a key result had to be a number you could point to on a dashboard, not a feeling. Goal-setting research from psychologists Edwin Locke and Gary Latham backs this up directionally: specific, quantified goals consistently outperform vague "do your best" goals across decades of workplace studies. Vagueness isn't neutral — it actively depresses performance.

PartQuestion it answersExample
MetricWhat exactly are we counting?D30 retention rate
BaselineWhere do we start?41%
TargetWhere do we need to land?52%
DirectionUp or down?Increase
DeadlineBy when?End of Q3 (Sep 30)
Source of truthWho measures it, and where?Amplitude cohort report, owned by growth analytics

Metric: one number, not a bundle

The metric has to resolve to a single number, not a cluster of related things you're hoping all move together. "Improve activation and engagement" is two metrics wearing one KR's clothes. If you can't say the word "and" while stating your metric, you're closer to done. This is also where the difference between an outcome and an output matters most — see the full breakdown in outcome vs. output OKRs if your team keeps writing KRs that are really just shipped features in disguise.

Baseline: measured, not assumed

Most disputed KRs fail here first. A baseline that's "about 40%, I think" isn't a baseline — it's a guess dressed up as data. Pull the actual number from the source of truth before the cycle starts, write it into the KR text itself, and timestamp it. If nobody can produce that baseline today, that's a signal you don't have this metric yet — see the proxy-metric trap below.

Source of truth: the tie-breaker nobody argued about

This is the part teams skip most often, and it's the one that causes the loudest arguments in review. When engineering's dashboard says 49% and the data team's warehouse says 52%, "we hit the target" becomes a negotiation. Name the exact dashboard, query, or report in the key result itself, so there's only one number in the room.

Direction and deadline: the two parts everyone assumes are obvious

Direction isn't always self-evident. "Improve friction score" could mean higher is better (a satisfaction-style score) or lower is better (an effort score) depending on how the underlying survey was built — write "reduce" or "increase" explicitly, every time, even when it feels redundant. The deadline deserves the same discipline: "by end of quarter" is fine, but "eventually" or "ongoing" isn't a deadline, it's a way of never having to be wrong.

For a deeper walkthrough of how these six parts fit into a full OKR cycle — objectives, cascading, scoring — see the complete guide to advanced OKRs.

Three Weak Key Results, Hardened Into Ones You Can't Wiggle Out Of

Weak key results share a pattern: a verb ("improve," "increase," "grow") attached to a noun with no baseline, target, deadline, or owner. Hardening them means running each one through the six-part anatomy above until the sentence can only be read one way.

Weak KRWhat's missingHardened KR
"Improve retention"Metric, baseline, target, deadline, sourceIncrease D30 retention for new signups from 41% to 52% by Sep 30, measured in the Amplitude cohort dashboard
"Increase engagement"Metric (which action?), baseline, deadlineGrow weekly active users completing the core action from 2,400 to 3,200 by Nov 15, per the WAU-core-action report in product analytics
"Improve onboarding"Metric, direction, source of truthReduce time-to-first-value (signup to first completed project) from a median of 6.2 days to 3.0 days by Oct 31, measured in the activation funnel report

"Improve retention" — the classic

"Retention" isn't one metric; it's a family of them (D1, D7, D30, cohort curves, revenue retention). Picking the wrong window is its own failure mode: a D30 improvement can mask a D1 collapse, and a healthy logo-retention number can hide a shrinking revenue-retention number underneath it. Name the window, the baseline, the target, and the deadline, and the argument disappears.

"Increase engagement" — the vaguest word in product

"Engagement" tempts teams into vanity metrics: logins, page views, time-in-app. None of those necessarily mean the user got value. Anchor the metric to the action a customer takes when they're actually doing the job they hired your product for — which is exactly the diagnostic question the Jobs to Be Done framework is built to answer.

Amplitude's North Star Framework makes a similar case: pick the one action that structurally ties customer value back to growth, rather than a basket of activity metrics that all move for different reasons. If you can't name that one action, "engagement" will keep meaning something different to every reader of the doc.

"Improve onboarding" — fix the moment, not the vibe

Onboarding KRs often measure satisfaction ("users feel onboarding is smooth") instead of behavior. Behavior is measurable; feelings, without a validated survey instrument, usually aren't. Map the metric to a specific point on the customer journey — time-to-first-value, drop-off at step three, the moment the user first says "oh, that's what this does" — and you get a number instead of a mood.

The Proxy-Metric Trap: When "Measurable" Isn't Actually Measurable

Teams reach for proxy metrics when the real outcome is hard to instrument, and the proxy quietly becomes the actual target — even though nobody agreed it should be. The fix is picking metrics that are already instrumented (or cheap to instrument) and resistant to gaming, not whatever number happens to already exist in a spreadsheet.

This trap shows up constantly with soft goals:

  • "Improve code quality" becomes "reduce story points per bug," which teams game by writing smaller, less useful tickets.
  • "Improve customer trust" becomes "raise NPS by 5 points," even though the survey has an 8% response rate and swings 15 points month to month on sample noise alone.
  • "Improve morale" becomes "engagement survey score," measured once a quarter by a tool nobody remembers opting into.

None of these proxies are automatically wrong — the failure is treating them as settled when nobody actually decided they were an acceptable stand-in for the real thing. A proxy metric that moves without the underlying outcome moving is a task wearing an outcome's clothes.

Author and OKR coach Christina Wodtke's core argument in Radical Focus: a key result has to be an outcome you can defend at review time, not a task you can quietly complete without moving anything real.

Not every proxy is a trap. A leading indicator — a number that reliably predicts the outcome before the outcome itself is measurable, like early feature adoption predicting D30 retention weeks later — earns its place if the correlation has actually been checked. The trap is a proxy nobody has validated, adopted purely because it was already sitting in a spreadsheet somewhere.

Three questions separate a real metric from a proxy that will get argued about in the next review:

  1. Does it already exist, or would we have to build tracking for it? If instrumentation is the actual project, say so — don't let the KR quietly become "ship the analytics event" wearing a metric's name.
  2. Can it move without the real outcome moving? If yes, someone will find that path under deadline pressure, and the number will stop meaning anything.
  3. Would a skeptical stakeholder accept this as proof? If your VP of Sales would raise an eyebrow at "story points" as evidence of quality, they're right to.

Ambiguity about which number counts isn't just a product-team problem, either — it's an organizational pattern. Gallup's long-running employee engagement research has repeatedly found that fewer than half of employees can say clearly what's expected of them at work. A key result with no agreed source of truth is that same failure, just scoped to one team for one quarter.

How to Pressure-Test a Key Result Before the Cycle Starts

The fastest way to catch a weak KR is to interrogate it before the quarter starts, not during the review when the incentive to reinterpret it is highest. Run every draft KR through a short set of questions with the whole team in the room, not just the KR's author.

Ask, out loud, before anyone commits:

  1. Can we pull this number today? If the answer is "not yet," instrumentation is a prerequisite task, not an assumption.
  2. If it lands exactly on target, does everyone in this room agree we succeeded? Silence or hedging means a part of the anatomy is still missing.
  3. Whose dashboard wins if two teams report different numbers? Name that source of truth now, not in the review.
  4. Could this number move for the wrong reasons? — layoffs inflating retention, a pricing change inflating revenue-per-user, a support policy change inflating NPS.
  5. Does hitting this KR trace up to something the company actually cares about? A team-level threshold that's disconnected from the objective above it is how cascading OKRs without waterfall breaks down into busywork nobody can defend.

This pressure-test matters most at the moment of scoring. Google's publicly documented OKR guidance treats an average score of roughly 0.6–0.7 across a set of key results as a good, appropriately ambitious quarter, not a shortfall. That convention only works if every KR in the set was written precisely enough that a 0.65 means something specific, rather than a number three people negotiate into existence after the fact.

Loose KRs are exactly how a cycle degrades into what's sometimes called goal theater — a review meeting full of confident scorecards that don't actually reflect what happened. A hard threshold, agreed before the quarter started, is what keeps the scoring conversation about the work instead of about the wording.

Where Prodinja Fits: Making the Threshold Non-Negotiable From the Start

Most of this discipline is a writing habit — but it's also a modeling problem, and tools shape habits. In Prodinja, an Outcome is modeled as an entity that requires a baseline, a target, and a deadline before it can be saved. A key result gets committed as "from X to Y by when" rather than left as a fuzzy aspiration someone fills in later.

It's designed to force the six-part anatomy at the moment the KR is written, not retrofit it during a tense review three months later.

That's a structural nudge, not a guarantee — a team can still pick a bad metric or a gameable proxy inside a well-formed Outcome. But it does close off the most common failure mode this article opened with: a KR that shipped with no number attached at all.

Key Takeaways

  • A measurable key result needs all six parts: metric, baseline, target, direction, deadline, and source of truth — missing any one reopens it to interpretation.
  • Write the baseline as a measured number pulled from the source of truth before the cycle starts, never as a remembered estimate.
  • Name the exact dashboard or report that will settle disputes inside the KR text itself, not as a side conversation during the review.
  • "Improve X" is a task pretending to be a key result; it needs a specific metric, a starting point, and a target before it counts.
  • Proxy metrics aren't inherently wrong, but they need to survive three tests: does it already exist, can it move without the real outcome moving, and would a skeptic accept it as proof.
  • Pressure-test every KR before the quarter starts by asking whether everyone in the room would agree, right now, on what hitting the target would look like.
  • Anchor engagement and onboarding metrics to a specific customer action or journey moment, not a general feeling of "better."

Frequently Asked Questions

What's the difference between a key result and a metric?

A metric is just a number you can track; a key result is that metric attached to a baseline, a target, and a deadline. "Weekly active users" is a metric. "Grow weekly active users from 2,400 to 3,200 by Nov 15" is a key result. Every KR needs a metric, but not every metric someone tracks deserves to become a KR.

How many key results should one objective have?

Most OKR practitioners converge on two to four key results per objective, which is roughly the range John Doerr describes as workable in Measure What Matters. Fewer than two rarely captures the full picture of what "done" means; more than four tends to dilute focus and makes the objective harder to score honestly at review time.

What if we genuinely don't have a baseline yet?

Then measuring the baseline is this cycle's actual task, not a footnote. Write a short instrumentation KR ("Ship tracking for time-to-first-value and establish a baseline by [date]") instead of guessing at a number for the metric you actually care about. A guessed baseline produces a KR nobody can defend once real data shows up.

Can a key result be binary — done or not done?

Occasionally, for genuinely binary milestones (a certification, a launch, a compliance deadline), but treat this as the exception. Binary KRs are easy to game by shipping a hollow version of the thing, and they tend to measure output ("we launched it") rather than outcome ("it worked"). Where possible, pair a milestone with a measurable result that shows the milestone mattered.

How precise does the target number need to be?

Precise enough that two people looking at the same dashboard reach the same verdict, not necessarily precise to the decimal. "52%" is fine; "51.7% to 52.3%, allowing for normal sampling variance" is over-engineering. The goal is removing subjective interpretation, not chasing false precision the underlying data can't support.