Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Growth Marketing › Growth Fundamentals

Changing a North Star Metric

When a north star metric stops leading revenue or starts being gamed: telling it from a loose definition or weak input, and re-basing targets and tests.
Guideadvanced13 min readUpdated

What a north star is for, and what changing it costs

The north star and OKRs framework treats the north star as a count of value delivered that runs ahead of revenue, so every team works towards the same number and sees a problem before it reaches the accounts. It sets five tests for a candidate (it reflects customer value, leads revenue, can be measured now, can be moved by teams and is hard to game), breaks the measure into inputs and names its limits: teams optimise whatever is watched, guardrails have to sit beside it, and the gap in time between marketing's work and the north star can be long. One of its listed mistakes is changing the north star every cycle, since a measure swapped at each planning round cannot build a trend that teams believe.

That last point is the cost of any change, not only a frequent one. A north star is useful because of its history, read month against month and cohort against cohort, and changing it breaks that comparison. It also unsettles what is built on it: key results, the input tree, the growth model, the weekly meeting and the experiment backlog, which scores impact against it or the loop stage feeding it, as the experimentation cadence guide requires.

So the first question is narrower than "is there a better metric?". Has the current north star stopped doing its job, and would a smaller repair bring it back? A north star that falls while still running ahead of revenue is giving early bad news, which is its job. Neither that, nor a new leader's preference, nor the start of a planning year is a reason to change it.

Is it the north star, its definition or an input that is wrong?

Four ordinary explanations can make a sound north star look broken, and each has its own check.

  • The measurement changed. A fixed tag, a redesigned consent banner or a release that moved the counted event steps the number up or down on a date; the measurement change log is the first place to look.
  • The mix changed. The total can move because the blend of customers behind it shifted while every segment behaved as before; the guide to explaining why a metric moved separates mix from rate.
  • The counts are small. A segment or a week built on a few customers jumps about by chance; the guide to reading small numbers puts a range around such figures.
  • The effect has not arrived yet. The framework warns that work on the inputs can take a long time to show in the north star. Where the north star counts something a growth loop produces, the page explaining how a growth loop compounds shows why the customers a change sets off keep arriving over later cycles.

With those ruled out, three faults remain.

A loose definition. The idea still stands for value and runs ahead of revenue, but the counting rule lets in events that carry none: a route that fires the counted event without the core action, duplicated records, test accounts, an action counted when started rather than completed. The tell is that cohorts with more of a strict count, restricted to the cases that carry value, still go on to more revenue and retention, while the loose count no longer separates them. The repair is a tighter definition.

A weak or misaimed input. The north star and revenue still move together, but the north star has stalled, or a team has been pushing an input that does not feed it. The tell is in the tree: an input rose and the branch above it did not. The repair is to refocus the input, change who owns it or correct the arithmetic between levels.

The north star itself. Even counted strictly, the measure no longer stands for what customers pay for, or no longer moves before revenue does, as when the core action, the basis of pricing or the kind of customer has changed. The tell is that the strict count has parted from revenue and retention too, cohort after cohort, with the ordinary explanations ruled out. This, or a north star that now fails another of the framework's five tests, is the fault that calls for a replacement.

What you seeLikely causeThe check that separates itWhat changes
A step on a single dateMeasurement changeThe change log for that dateThe record, and history restated if needed
The total moves while every segment holdsMix shiftMix and rate split by segmentHow the movement is reported
Swings in a small segment or weekNoiseRanges and pooled periodsNothing yet
Inputs moved, north star has notLag, or a weak inputThe loop's cycle time and the tree's arithmeticPossibly the input and its owner
The count rises without customers getting more valueLoose definitionA strict count set against revenue and retentionA new version of the definition, not a new north star
A strict count has also parted from revenueThe north star no longer fitsThe criteria check below, by cohortThe north star and what hangs from it, with its history kept under its own name

Signs a north star has stopped doing its job

Read the framework's five tests in reverse: each criterion that chose the north star also names a way it can fail later.

It no longer leads revenue. The north star climbs while renewals, repeat orders or retained revenue stay flat or fall. A watched number rising while the outcome it was chosen for does not is the sign the Goodhart's law framework tells teams to look for, and a north star in that state is turning into one of the vanity metrics it was chosen to displace.

It can be moved without creating value. A feature, an incentive or a stretched definition pushes the count up while customers get nothing more. The framework's limits section lists this risk; the sign that it has happened is a rise that a sample of the records behind the count, read as the Goodhart's law framework advises, cannot explain by anything customers did differently.

It no longer reflects customer value. The product, the pricing or the customer base changed, and the counted action is no longer the moment customers get what they pay for.

Teams can no longer move it. After a reorganisation or a product change, teams cannot each name an input beneath it that they control, so it is watched but steers nobody.

It can no longer be counted cleanly. Tracking or account changes mean it cannot be read by cohort from existing data without a correction each time.

A loose definition can produce the first two signs and a tracking change the last, so check the definition and the change log first, and read every sign by cohort over enough history, pooled as the small numbers guide advises.

Checking a replacement against the choosing criteria

A replacement faces the same five-criterion table the framework uses for a first choice, but each criterion can now be tested against what went wrong.

Criterion from the frameworkThe question for a replacement
Reflects customer valueDoes it count the value the old north star missed, and leave out what the old count let in?
Leads revenueHas it moved before revenue in this business's own history, not only in theory?
Measurable nowCan it be read back over the history from existing data, by cohort, so the switch keeps a trend?
Movable by teamsCan each team name an input beneath it that it controls?
Hard to gameWould the route that inflated the old count inflate this one too?

The second row needs evidence rather than argument, and the method belongs to the leading indicator entry, whose section on choosing indicators that predict sets out the check. Run it on the business's own data and record the result; a candidate chosen because it looks right or moves quickly risks repeating the original problem. The fourth row carries the time limit the growth fundamentals hub sets for a north star: a candidate whose inputs no team can move this quarter is a good report and a poor north star.

Read every candidate by Cohort, using the table the cohort analysis tutorial builds, so that a young cohort's enthusiasm does not hide an older one's decline. Then set it beside the bridge from the growth accounting tutorial: a candidate that rises with new customers while the retained base shrinks may be following acquisition rather than value.

Switching over: key results, the input tree, the model and tests already running

Once the decision is made, pick a switch date and re-base what hangs from the north star, from the definition outwards. The academy exercise on re-basing a craft kit club's paid programme after its north star failed practises the diagnosis, the criteria check and this order for Google Ads and Meta, with bidding moved one platform at a time onto an earlier paid step that the subscription system sends back.

The definition. Write the new measure into the metrics layer as the metrics layer guide sets out in its section on versioning and change, which decides whether the change is a new version or a new metric with its own name, how history is restated and how the old one is retired. Add the switch to the measurement change log. The metric definition change checklist puts these re-basing steps beside the sign-off, dashboards and data tests that any redefined metric needs, as one list to tick before the switch date.

Key results. The framework's mid-cycle check applies directly: each key result tied to the old north star, or to an input that no longer feeds the new one, is kept, rewritten with the change noted, or dropped openly, and by the framework's own rule each rewritten one states where its data comes from and the reading it starts from. Google's re:Work guide to OKRs, as the framework reports it, wants objectives that are concrete as well as ambitious, key results that can be measured and verified, few of them per team and visible across the organisation, and grading on a simple scale from no achievement to full achievement. The old north star can stay on as a KPI that is watched rather than targeted, the distinction the comparison of OKRs and KPIs draws.

The input tree. The tree beneath the old north star does not carry over by default, and the new measure may need a branch nobody owns yet. Rework the arithmetic between levels with the KPI tree framework beside the KPI trees guide, reassign owners, and keep the guardrails beside the branches they protect.

The growth model's north star row. The spreadsheet growth model tutorial turns the active base into the north star through a conversion row read from the business's own usage data. Rebuild that row from the new measure's own history, run the back-test again and list the new relationship among the shaky assumptions in the hub's growth model document until it has been checked.

The weekly pre-read. The weekly growth meeting tutorial builds its pre-read from the north star and its inputs. From the switch date the new measure leads, the old one sits beneath it as a diagnostic, and the switch is marked so that nobody reads the step as performance.

Tests already running. Leave them on the primary metric, guardrail metrics and decision rule that the experiment hypothesis template fixes before a test is built, and call them against that rule. Record a finished test whose metric no longer feeds the north star as it stands, with a note, write the next tests against the new inputs and score the backlog again.

Telling people, and checking that the change held

Write the decision down before announcing it. A decision log entry on the marketing decision log template records the change in the template's own columns, with the fault diagnosis, the criteria check and the cohort readings as its evidence and a date for checking it. Record a Baseline for the new measure at the switch as well.

Then tell everyone who reads the north star, in one note, what changed, why, from which date, what happens to each team's key results and inputs, and where the old measure can still be found. Making the change visible across the organisation matches what the framework, citing re:Work, recommends for key results.

On the check date, ask three questions. Has the new measure gone on moving ahead of revenue in the cohorts acquired since the switch? Has anything started inflating it without creating value? Can every team still name the input it owns? If the north star counts what a loop produces, set the check date late enough for the loop's cycle time. A replacement that disappoints deserves the same diagnosis as the original before anyone proposes another; swapping again at the next planning round is the framework's listed mistake.

A family history app replaces its north star (illustrative)

Illustrative example — not a real result

Olvetrisk Family Archive moves from trees started to households where relatives contribute

Olvetrisk Family Archive, an invented family history app sold by monthly household subscription, has steered by family trees started each week. After an April feature that imports a tree from a file, trees started rise over 12 weeks from about 2,100 a week to 3,600, while the share of new households still paying at month six stays at 52% and households where a relative other than the subscriber adds to the tree stay at about 9,800 a month, out of about 41,000 paying households.

The growth lead rules out the ordinary explanations (no tracking change in the log, a rise held in every channel for 12 weeks, renewals read only in cohorts old enough to reach month six), then the smaller repairs. A strict count, trees with at least 10 people added by hand in the first week, had drifted from renewals before the import: in cohorts that reached month six before April, households starting such a tree were still paying at 53%, against 51% for the rest, so a tighter definition would not help; and the inputs beneath the north star had all moved as planned.

The candidate is contributing households: households where a relative, on their own account, added a person, photo or story to the tree in the month. Against the framework's five criteria, it counts the family filling in the tree together, which is what subscribers describe paying for; it is logged by household from launch; product and marketing can each name an input beneath it; and an imported file cannot create it. The precedence check the leading indicator entry describes, run on 24 monthly cohorts, finds households with a contributing relative in their second month still paying at month six at 74%, against 41% for those without one, with the gap present in 22 of the 24.

The analyst adds contributing households to the metrics layer as a new metric under its own name, restates two years of history and logs the switch for 1 August. At the mid-quarter check the key result on trees started is rewritten, with the change noted, as "contributing households rise from 9,800 to 10,600 by the end of September", with its baseline and data source; the others are kept or dropped openly. Two running tests finish on the primary metrics written before they started. The input tree becomes new households, relatives invited and invitations accepted; the model's north star row becomes contributing households as a share of paying households, 24% over the last year; and the weekly pre-read opens on the new measure with trees started beneath it and 1 August marked.

The decision log entry records both checks, the growth lead as owner, the expected effect, a guardrail on complaints about invitation emails and a check date of 15 February, when the first cohort acquired since the switch reaches month six and can show whether contributing households still lead renewals.

Common mistakes

  • Replacing a north star that only needs a tighter definition. If a strict count still follows revenue, version the definition instead.
  • Changing it because it brings bad news. A falling north star that still runs ahead of revenue is doing its job.
  • Skipping the ordinary explanations. Measurement changes, mix shifts, small counts and lag can each make a sound north star look broken.
  • Choosing a replacement because it moves. Speed is not evidence; run the leading indicator check on the business's own cohorts.
  • Grading old key results against the new measure. Keep, rewrite with the change noted, or drop each one openly.
  • Moving running tests onto the new north star. Finish them on the rules written before they started.
  • Leaving the input tree and the model untouched. Both were built for the old measure.
  • No decision record and no check date. The next disagreement then reopens the choice from memory.

The growth fundamentals hub shows where the north star sits in the growth model, beside the loops it summarises and the experiments aimed at its inputs.

Frequently asked questions

How do I know whether to replace a north star or just tighten its definition?

Count the north star strictly, keeping only the cases that carry the value it was chosen to represent, and set that strict count beside revenue and retention by cohort. If cohorts with more of the strict count still go on to more revenue and retention while the loose count no longer separates them, the definition is at fault and a new version of the same metric will do. If the strict count has parted from revenue as well, once a measurement change, a mix shift, noise and lag have been ruled out, the north star itself has stopped doing its job and a replacement is worth checking.

What happens to experiments that are already running when the north star changes?

They finish on the primary metric, guardrails and decision rule written before they started, and are called against that rule as planned. Moving a live test onto the new north star turns a planned reading into a search for a flattering one. Record each finished result with a note where its metric no longer feeds the north star, and write the next tests against the new inputs.

Is it a mistake to change a north star at all?

No. The mistake the north star framework names is changing it every cycle, because a measure that keeps being swapped cannot build a trend anyone can trust. A north star that no longer reflects customer value, or no longer runs ahead of revenue, misleads the teams steering by it and should be replaced, but only after the diagnosis, a check of the candidate against the original criteria and a recorded decision with a date for confirming that the new measure still leads revenue.

Terms used on this page

Frameworks

Where this page is used

Part of these learning paths

Sources

  1. Guide: Set goals with OKRs — Google re:Work

Related