Missing data in AI output: preserve unknowns explicitly
Missing data in AI research should remain missing until the original material supports a value. Use a separate status field to distinguish not disclosed, not applicable, not found within scope and extraction failure. A blank cell, a model’s guess and a reported zero are different observations. Combining them can silently change a comparison or calculation.
Define a status vocabulary before extraction
The W3C Metadata Vocabulary for Tabular Data describes metadata for interpreting tables, including null representations. It does not prescribe the research statuses below. These statuses are an editorial convention that preserves why a value is absent. Keep the numeric value separate from that explanation so a spreadsheet does not accidentally turn a label into zero.
| Status | What it means | Next check |
|---|---|---|
| Reported | A value was located in the cited source | Verify unit, scope and period |
| Not disclosed | Checked source explicitly omits or does not provide the item | Record exactly which source was checked |
| Not found in scope | Search did not locate it in the defined material | Expand or revise search if important |
| Extraction failed | Relevant material exists but could not be read reliably | Inspect the original table or image |
| Not applicable | The field does not fit this entity or reporting basis | Record why the field is inappropriate |
Avoid treating “not found” as proof that a company never discloses the information. Your claim is bounded by the documents, dates and search method inspected. A model may have access only to an excerpt, while the value exists in an unprovided note. Preserve that distinction rather than converting a limited search result into a statement about the issuer.
A hypothetical customer retention table
Assume you are comparing three invented subscription businesses. Company A reports retention of 92 percent. Company B’s uploaded scan contains a retention table that extraction failed to read. Company C reports a different renewal measure with an incompatible population. Only A supplies a comparable observed value in this task. Recording B and C as zero would make their apparent operating performance much worse without evidence.
Keep A’s assumed 92 with its definition. Give B a null value and extraction-failed status, and ask a reviewer to inspect the scan. For C, keep the reported renewal measure in a separate field; mark the requested retention definition unavailable in the checked scope. Do not average the three entries or rank the companies on a metric you only observed for one of them.
Make calculations respect the status
Before calculating, define eligibility: which statuses and definitions can enter this result? For a like-for-like average, require reported values with matching populations and periods. Report the eligible count beside the result. If a missing item is material, present a range or leave the calculation unresolved instead of filling the gap with a plausible industry value.
Use a record with value, unit, period, status, reason, source position and reviewer. In a CSV, test what your analysis tool does with an empty numeric field. Some import or charting settings may display an absence as a gap, omit a row or apply a default. Verify the actual behavior with a small deliberate missing example before relying on the chart.
Repair the evidence without rewriting history
When B’s scan is manually checked, add a new extraction record with the source position and who checked it. The old failure remains useful evidence about your pipeline. If the original scan is genuinely unreadable, keep the unresolved status rather than using the model’s earlier suggested figure. A later disclosure is a new observation, not proof that the old report contained the value.
This method cannot establish what should have been disclosed under a particular jurisdiction’s rules. It makes a narrower promise: readers can tell a real zero from a missing value, understand the limits of the search and see which inputs were excluded. That is enough to prevent many misleading comparisons without pretending incomplete data is complete.
FAQ
Can I use zero when a company does not report an expense?
Not merely because a search returned no value. A reported zero needs a source and definition; absence of a result is not that source.
Is not applicable the same as not available?
No. An inapplicable metric does not fit the task or entity. An unavailable metric might fit, but sufficient evidence was not obtained.