
How to Set KPIs: Our Standard for Every Metric
On this page
Analytics & Data
Making a KPI SMART or giving it a stretch target does not make it actionable. To answer how to set KPIs: give each one an owner, a written formula and source, a baseline with its noise band, a target judged against that noise, an action threshold tied to a named action, and a counter-metric. A target can be hit by chance; the action rule tells someone what to do when the evidence warrants attention.
George T. Doran's original SMART paper put the action plan alongside the objective:
It is the combination of the objective and its action plan that is really important.
His original letters were Specific, Measurable, Assignable, Realistic and Time-related. Assignable meant specifying who would do the work; Realistic meant considering the resources available. Achievable, Relevant and Time-bound are later variants.
Doran also said not every objective needed all five criteria, and not every objective had to be quantified at every management level. Our reading is that SMART is useful for writing an objective, but it supplies no action threshold, counter-metric or noise check. Sources were read on 29 October 2026.
MUST: Give every KPI a definition and a decision
K1. Name one owner
Requirement. Put one person's name beside each KPI. The owner is responsible for checking the measure, deciding whether its threshold has been crossed and recording the action. Other people can do the work; one person owns the decision.
Why. Doran's Assignable criterion asks who will do it. A team name can identify contributors, but it leaves an unanswered question when a number crosses its threshold on a busy Monday: who opens the record and starts the review?
Test. Does every KPI row contain exactly one person's name, with no team name in the owner field? Count the rows that fail and assign each one before the next review.
K2. Write the formula and source of record
Requirement. Write the numerator, denominator, period and source for every KPI. Define when an item enters the count and what happens when a record changes later. Name one system as the source of record.
Why. Two plausible formulas can produce different answers while sharing the same label. For example, close rate = new customers / leads, counted by the month the lead arrived. That cohort definition keeps a customer who closes later connected to the lead period that produced them.
CPL = channel spend / leads, CAC = channel spend / new customers, and payback months = CAC / monthly gross margin per customer. Ad platforms can each claim credit for the same customer, so use the CRM as the count of record. The metrics to pick by business type gives these formulas context.
Test. Can two people independently compute last month's value from the written definition and get the same number? If not, record where their inputs or timing differ and resolve that difference before setting a target.
K3. Separate the target from the action threshold
Requirement. Our rule is to write the action threshold and the named action before looking at the next result. The target states where you want the measure to go over a defined period. The threshold states when the owner must do a particular thing.
Why. A quarterly target of 24 booked jobs per week on average does not tell the operations manager what to do after a poor week. A rule such as “if the weekly count is under 10, review missed calls and lead response times that week” does. The decision card for turning a measure into an action gives the threshold a place beside the decision it supports.
The threshold needs a direction, a period and an owner. Record the action even if the investigation finds no change is needed. A threshold crossing starts a review; it does not automatically prove the cause or prescribe a fix.
Test. For every KPI, can you state what the named owner does on Monday if the latest settled value crosses the line? The answer must name an action and the information the owner will inspect, not just say that the KPI turns red.
K4. Set the target against a baseline and its noise
Requirement. Establish a baseline before choosing a target. Calculate its ordinary variation, then state the target over a period long enough to distinguish progress from a lucky observation. An industry benchmark can prompt a question; it cannot tell you how much your own process varies.
Why. Say a hypothetical 25-person home-services company recorded these booked jobs over 12 weeks:
22, 18, 25, 20, 17, 23, 21, 19, 26, 20, 18, 23
The total is 252, so the mean is 21 jobs per week. The absolute differences between consecutive weeks are 4, 7, 5, 3, 6, 2, 2, 7, 6, 2, 5. They total 49 across 11 moving ranges, giving an average moving range of about 4.45.
The NIST individuals chart method uses 1.128 as the constant for a moving range of two observations. The following arithmetic gives the natural process limits for this hypothetical baseline:
average moving range = 49 / 11 = about 4.45 multiplier = 3 / 1.128 = about 2.66 limits = 21 plus or minus (3 x 4.45 / 1.128) limits = 21 plus or minus about 11.85 = about 9.2 to 32.8 jobs
These are process limits, not a promise that every future week will fall inside them. They describe the variation estimated from this baseline and are useful only while the process and measurement remain comparable.
A quicker count check gives a narrower sense of an ordinary swing: the square root of 21 is about 4.6, so roughly 16 to 26 jobs is an ordinary week. That approximation and its sample-size implications are explained in what analytics is for a small business. The two calculations serve different purposes: the square-root check is a rough count heuristic; the wider process limits use the observed moving ranges.

The target line at 25 sits inside the ordinary week-to-week band around 21. One week at the target does not establish that the process has improved.
A target of 25 jobs in one week was met or beaten twice in the baseline, at 25 and 26. Hitting it once therefore cannot establish a change. An average of 24 per week over the next quarter is a different, time-bound target; judge it across the quarter and against the baseline pattern, not from one good week.
For a signal of change, watch for a point beyond a process limit or eight consecutive weeks on one side of the center line. Those are two of the NIST Western Electric rules; using additional rules also increases false alarms. Investigate a signal before declaring a cause.
Test. Does each target have a baseline period, a calculated noise band and a target period? For a count target inside the ordinary single-period band, can you show how many baseline periods already met it and specify what sustained pattern would count as evidence of change?
K5. Protect the target with an independent counter-metric
Requirement. Pair each KPI with a counter-metric that can reveal an unwanted way of improving the main number. Our rule is that the counter-metric must come from a source the measured team cannot edit. Give it its own baseline and an action rule.
Why. Donald Campbell described the pressure placed on a measure when it becomes consequential:
The more any quantitative social indicator is used for social decisionmaking, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor.
The later popular paraphrase of Goodhart's law makes a related point: turning a measure into a target can damage its value as a measure. A counter-metric checks whether the desired result is being achieved at the expense of something else. It is a warning instrument, not proof of intent.
The documented Wells Fargo case shows why the source matters. In September 2016, the CFPB announced a $100 million fine and reported that employees were spurred by sales targets and compensation incentives. Roughly 1.5 million deposit accounts may not have been authorized, and roughly 565,000 credit card accounts were applied for that may not have been authorized.
In February 2020, the Department of Justice announced the $3 billion resolution, describing pressure to meet unrealistic sales goals from 2002 to 2016. Its account also said staff altered customer contact information, preventing customers from learning about unauthorized accounts and preventing employees from reaching them for satisfaction surveys.
Our reading of the Wells Fargo facts is that a satisfaction-survey guardrail can itself be defeated if the measured team controls the contact information. Use a source that team cannot edit, and check whether the source captures the cases you need.
For booked jobs, a useful counter-metric could be the share needing a return visit within 30 days, taken from a job log the sales team cannot edit. If that share rises above its own baseline upper limit, pause the push for more bookings and investigate. A rise alongside bookings calls for investigation; it does not by itself prove the target was gamed.

Compare the target with the counter-metric at the same review. If both bookings and return visits rise, investigate whether the push for bookings is affecting job quality.
Steven Kerr's 1975 paper on rewarding A while hoping for B warns about goal displacement. The 1995 Academy of Management Executive reprint states: “Many managers seek to establish simple, quantifiable standards against which to measure and reward performance. Such efforts may be successful in highly predictable areas within an organization, but are likely to cause goal displacement when applied anywhere else.” The practical question is whether the behavior rewarded by the KPI is actually the behavior the company wants.
Test. Does every KPI have a counter-metric, its own source and an action rule? For each source, can you answer yes or no to whether the measured team can edit the records used to calculate it?
SHOULD: Make the KPI useful at the review
K6. Pair an outcome with an earlier indicator
Requirement. Pair each lagging KPI with a leading measure. In our definitions, a lagging metric records an outcome after the fact, such as revenue, churn or CAC. A leading metric is an earlier number that should predict that outcome, such as qualified leads or quotes sent.
Why. An earlier measure gives you a chance to investigate before the outcome arrives. The word “leading” is a hypothesis, though, not a property conferred by putting a metric on a dashboard. Check past periods to see whether changes in the proposed leading measure preceded changes in the outcome often enough to be useful.
For the hypothetical home-services company, booked jobs can be the leading measure and weekly gross margin from accounting the lagging outcome. The pair still needs a historical check: more bookings may not predict more gross margin if job mix or delivery costs change.
Test. For each lagging KPI, is a proposed leading measure named, and have you compared the two across past periods? Record the periods checked and whether the earlier movement actually preceded the outcome.
K7. Match the review cadence to the data
Requirement. Review each KPI on a schedule that matches how quickly the process can move and how quickly the data settles. Record the period being reviewed and the cutoff for late data.
Why. A daily inspection of a weekly count mostly reads noise, in our judgment. It also invites an owner to react before the full period exists. Weekly booked jobs deserve a weekly review; a quarterly target deserves a quarterly assessment of the average.
Freshness matters even for daily data. Google Analytics data freshness guidance says daily data for standard properties typically takes about 12 hours, while processing can take 24 to 48 hours and report data may change during that window. Mark a recent value as provisional if you review it before that window has passed.
Test. Does one review period contain enough settled data for a move outside the stated noise band to be possible? Check the source's processing delay and mark each review as settled or provisional.
K8. Limit each owner to three to five KPIs
Requirement. Our rule is three to five KPIs per owner. Count measures requiring that person to make decisions, not every number available on a report. Put useful background measures on a monitor list.
Why. More measures create more claims on the same review time. Kerr described how visible behaviors can attract attention while less visible contributions receive less credit. A list dominated by easily counted activity can leave quality or cooperation without an owner.
The OKR authors' page on John Doerr's site describes three to five key results per objective. That is a separate convention, not evidence that three to five KPIs per owner is a universal optimum. Our limit is a practical rule for keeping decisions manageable.
Test. Count the KPIs assigned to each person. If an owner has more than five, identify which decisions overlap and move at least one measure to the monitor list or assign it to another appropriate owner.
K9. Label committed and stretch targets separately
Requirement. Give a committed target and a stretch target separate labels. Our reading is that a stretch target belongs in planning, while an action threshold belongs in operations. Both can be useful, but they answer different questions.
Why. In their 2002 research review, Edwin Locke and Gary Latham wrote:
We found that specific, difficult goals consistently led to higher performance than urging people to do their best.
Their reported meta-analysis effect sizes ranged from .42 to .80 for specific, difficult goals compared with asking people to do their best, and from .52 to .82 for goal difficulty. The paper also describes limits: people need feedback on progress and commitment to the goal, and performance can level off or decrease when ability limits are reached or commitment lapses. A difficult goal alone does not specify an operational response.
The Google OKR playbook published on Doerr's site distinguishes committed OKRs, expected to score 1.0, from aspirational OKRs, expected to average 0.7 with high variance. OKRs are objectives and key results used for change; a KPI can track the health of an ongoing process. A company's KPI target need not inherit the OKR scoring system.
Ordóñez and colleagues' 2009 paper Goals Gone Wild identified potential side effects of goal setting, including narrow focus and unethical behavior. Locke and Latham published a reply, and the authors debated the evidence. Set ambitious goals with feedback and guardrails, and keep the routine action threshold explicit.
Test. Can a reviewer identify the committed target, any stretch target and the action threshold as three separate fields? Is the action required at the threshold stated even if the stretch target is missed?
AVOID: Keep measures that earn their place
K10. Retire a KPI that never changes a decision
Requirement. Our rule is to retire a KPI or demote it to a monitor list after two quarterly review cycles, or six monthly reviews, with no threshold crossing and no logged action. First check whether its threshold is set too wide. Record the last action triggered and review that date at each retirement check.
Why. A measure can remain useful background information without occupying a slot on an owner's decision list. Equally, a KPI that never crosses its threshold may have a badly chosen line. Review the baseline and the intended decision before tightening a threshold or retiring the measure.
Definitions and source systems also change. Use measurement checks for keeping definitions honest over time when a process change makes an old baseline questionable. Iron Goo's operations service provides an example of a monthly reporting cadence: it reports on chatbot deflection, escalation and agent operations, including what changed and what ships next, and retires skills that turned out to need a human.
Test. Does the KPI list have a “last action triggered” date column? Count the measures with no crossing and no logged action across the specified reviews, then check each threshold before deciding whether to retain, demote or retire it.
K11. Do not wire a KPI straight to pay before the safeguards exist
Requirement. Establish the counter-metric and the action threshold before connecting a KPI to compensation. Document what the payment rewards and what behavior the guardrail is intended to detect.
Why. Compensation can focus attention on a measure, and that same pressure can encourage people to improve the number in ways the company did not intend. The CFPB described sales targets and compensation incentives as part of the Wells Fargo case. The OKR authors' page on Doerr's site describes OKRs as divorced from compensation, which is one way to manage that tension.
This is a design judgment, not a claim that every performance incentive fails. An owner can test the proposed incentive against the formula, the source and the counter-metric before money makes the measure harder to change.
Test. For every KPI connected to pay, are the threshold, named action, counter-metric and independent source documented? Count any missing fields before the next compensation decision.
A completed KPI card for weekly booked jobs
This card uses the hypothetical 25-person home-services company and the 12-week baseline above. The owner name is fictional. The return-visit limit must be calculated from that metric's own baseline before the card is used.
| Field | Completed card |
|---|---|
| KPI | Booked jobs per week |
| Owner | Maya Lee, operations manager |
| Formula | Count jobs with a confirmed date created in the CRM during the week, Monday through Sunday. |
| Source of record | CRM |
| Pair | Leading: booked jobs per week. Lagging: gross margin per week from accounting. Check the relationship against past periods. |
| Baseline | 12 weeks; mean 21 jobs per week; natural process limits about 9.2 to 32.8. |
| Target | Average 24 booked jobs per week over the next quarter. |
| Action threshold and action | Any week under 10 jobs, or eight consecutive weeks under 21: Maya reviews missed calls and lead response times that week. |
| Counter-metric and source | Share of jobs needing a return visit within 30 days, from the job log the sales team cannot edit. If the share rises above its own baseline upper limit, pause the push on bookings and investigate. |
| Cadence | Weekly, after the CRM count for the completed week is settled. |
| Last action triggered | None yet in this hypothetical card; enter the date and action when triggered. |
| Retirement check | Each quarter. After two quarterly cycles with no threshold crossing and no logged action, check whether the threshold is too wide, then retain, demote or retire the KPI under our rule. |
Put your next steps in order
Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.



