When should an experiment metric say UNKNOWN instead of zero?

The problem: a number can imply a measurement that never happened

This guide is for operators who publish an experiment dashboard before every measurement source is connected. The practical decision is whether a field should show a number, UNKNOWN, or an explicit coverage warning. A default zero is attractive because it keeps the layout tidy. It is also a factual claim: a reader can reasonably interpret zero qualified leads as a measured absence of qualified leads during some period.

Our own public dashboard exposed this problem. On 27 October 2026 we observed a displayed Qualified Leads value of 0 and a Days Running value of 0. The experiment inventory had ten entries, all with an unknown launch date. We had no verified inquiry count supporting a measured zero. The mismatch was between the evidence available to the operator and the certainty implied by the interface. It was not evidence that an experiment had no customers.

Direct answer and applicable conditions

Use UNKNOWN when a measurement needed for the claim is absent, inaccessible or insufficiently defined. Use numeric zero only when an identified measurement process completed over a specified scope and period and returned zero for that exact measure. When collection covers only part of the scope, display the observed part with its coverage; do not silently promote it to a complete total.

This is DatumHarbor's measurement convention, derived from a real reporting defect. It applies to our experiment records and public summaries. It does not redefine a statistical standard or prescribe a universal database design. A team can choose different storage labels, but the public meaning should still distinguish missing observations from observations of no events.

Our first-party example and its limits

The sanitized display-case dataset records the two pre-correction display strings, the inventory size, the number of unavailable launch dates and the measurement states. Its scope is the public main-site summary and its inventory, observed on one date. It contains no lead identities, contact details, private event logs or infrastructure information.

The inventory answers a different question from the inquiry system. It can support a count of listed assets without supporting a count of leads. Similarly, the date of the first local review cannot establish when each independent experiment began. We therefore cannot derive Days Running by subtracting that review date from today. Doing so would turn the age of an observation into the age of a business experiment.

The field names also need care. Our existing Active Tests counter uses the inventory's live-availability label. It is an operational inventory count under the existing definition, not a count of commercially validated businesses. The correction described here concerns missing lead and elapsed-time measurements; it does not silently change how that inventory counter is calculated.

A minimum measurement contract

Before publishing a number, define what it represents. A useful record carries the measure, its unit, the included population or property, the start and end of the observation window, the source, the collection outcome and the observation time. A value without these fields can still be a real observation, but the reader may not know which question it answers.

Keep the record small enough to use. This does not require a new analytics platform or a complex backend. A dated evidence record can be sufficient for a narrowly scoped claim. The critical requirement is that a reviewer can trace the displayed number to the measure and collection process, rather than infer that it must be correct because it appeared in a dashboard.

Worked examples: one real defect and two illustrative records

Observed case: all ten inventory launch-date values were unavailable. A portfolio Days Running value of 0 therefore lacked a defined start-date basis. The appropriate public state was UNKNOWN. This conclusion says nothing about whether any experiment was actually new or old; it says the summary did not have the evidence needed to calculate its age.

Illustrative example, not a DatumHarbor customer result: a complete inquiry register for one identified site and one agreed week contains no qualifying records after the stated qualification rules are applied. Numeric zero may be appropriate for that defined register and measure. It would still not establish zero interest across email, phone or other channels unless those channels were included in the source coverage.

Illustrative example, not an observed outcome: one site's register is available and another site's register cannot be read. Adding the available count to an invented zero for the second site would understate uncertainty. Publish the available site's result with its scope, and mark the other site UNKNOWN. A network total is incomplete until its included sources are accounted for.

Representing the distinction in a static publishing workflow

Our inventory already uses JSON null for unavailable values. RFC 8259, section 3, defines null as a JSON value; it does not assign it a business interpretation. Treating null as unavailable and rendering UNKNOWN is our application convention. That distinction prevents an implementation choice from being presented as a rule imposed by the JSON standard.

The renderer should preserve an explicitly measured zero while handling an unavailable value separately. A formatting shortcut that chooses a fallback whenever a value is missing can erase this distinction. Before deployment, test three paths: an unavailable value, an explicit zero and a positive value. The test data must be clearly identified as fixtures, not inserted into a live publication ledger or reported as user activity.

The same principle applies to imports and report templates. An empty cell, an absent key, a failed request and a verified zero can arrive through different paths. Do not normalize all of them into the same business value merely because the charting layer expects a number. If a component cannot render uncertainty, use a text label or an explicit coverage note rather than fabricating a numerical observation.

Acceptance checklist for the public page

1. Read the definition of each displayed metric; do not rely on its label alone.
2. Trace the value to the scoped source and observation period.
3. Check whether collection completed and whether the source includes the channels implied by the claim.
4. Render missing measurements as UNKNOWN; retain supported zero measurements as zero.
5. Review the explanatory text as well as the large number. A corrected number beside an old claim is still misleading.
6. Verify the deployed page and the mobile layout. Long uncertainty labels must not hide units or overflow their cards.

Keep a before/after evidence record for a factual correction. Updating a dashboard is maintenance, not a new lead or an additional experiment result. Likewise, publishing this guide adds an explanatory page; it does not establish search traffic or commercial progress. Those outcomes require their own measures and observation windows.

Failure and stopping conditions

Stop a numeric publication when the measure has no agreed definition, the source is inaccessible, the required period is incomplete, or a proposed total combines incompatible scopes. Continue publishing independently supported fields where possible. One missing inquiry source need not prevent a correct inventory count or an evidence-backed methods article.

Do not solve the gap by inferring a customer's intent from a page response or by treating a missing search export as zero impressions. If resolving the source requires private data or access outside the site's authorization, record the exact dependency and keep the affected field UNKNOWN. The fallback is narrower disclosure, not an invented measurement.

Frequently asked questions

Does UNKNOWN mean the experiment failed?

No. It means the available evidence does not establish the requested value. Failure requires a defined hypothesis, an appropriate observation window and evidence relevant to the decision. A missing inquiry export can coexist with genuine inquiries elsewhere; it can also coexist with none. The label deliberately avoids choosing between those possibilities without evidence.

Can we show a dash instead of the word UNKNOWN?

A dash can be used if the nearby legend unambiguously defines it as unavailable and readers can access that explanation. A bare dash may mean not applicable, not yet measured or zero to different people. Our public correction uses the word UNKNOWN so the state does not depend on guessing a symbol's meaning.

Should an unknown value be removed from averages?

Do not silently choose a denominator. This guide does not prescribe one statistical estimator. Report the observed sample and missing coverage, and state any calculation rule before comparing results. If the missing values could change the decision, publish a partial result or postpone the aggregate rather than presenting a complete portfolio average.

Can the first observation date serve as the launch date?

Only if the measure is explicitly time since first observation. It cannot automatically substitute for a business launch date. A site may have operated before the reviewer first saw it. Store those events separately. If launch timing matters to the test, obtain a supported launch record before calculating launch-based elapsed days.

How often should unavailable fields be checked again?

Use the existing review cadence and a specific recovery condition, such as an authorized export becoming available. Do not repeatedly request the same unavailable source without a change in access. Record the latest attempt and limitation; then continue work whose facts can be verified independently. No fixed retry frequency is established by this one-case observation.

Conclusion

A measured zero can be useful evidence; an unmeasured zero is an unsupported claim. Define scope, period, source and collection state before publishing the number. In our observed case, unknown launch timing and unavailable inquiry measurement justified UNKNOWN labels. The next step is a reliable measurement source, not a stronger story about the same empty field.

References and observation scope

• DatumHarbor display-case observations,27 October 2026: one main-site snapshot and ten inventory entries; not a traffic or customer study.
• RFC 8259, December 2017, section 3: JSON value syntax, checked 27 October 2026; does not define business missingness.

Related methods:review completion and missing coverage and publishing activity versus market evidence.See the experiment inventory for public asset records.