Consider a reporting handover where the client asks where “cost per qualified inquiry” came from. The analyst points to a spreadsheet, the spreadsheet points to an export and the export contains no qualification field. Somewhere along the path, advertising cost was joined to sales records. That missing explanation is the data lineage problem.
For ChatGPT campaign reporting, lineage describes how a source observation becomes the number a decision-maker sees. It includes retrieval, transformations, joins, filters, definitions and ownership. You do not need an expensive catalog to begin. A short, accurate map of the actual pipeline is more useful than a sophisticated diagram that omits the manual spreadsheet step.
Follow one important number backward
Choose a metric that can change the campaign budget. Start with its published label and locate the cell, query or calculation that produces it. Then identify every input until you reach the original advertising report and any business-system records.
Record each step in sequence. For example: platform spend response, accepted report table, campaign-to-client mapping, joined qualification outcomes, calculated efficiency measure, rounded client display. This is an illustrative pipeline, not a statement that AthillyAds performs those joins or imports sales outcomes.
At each step ask what can change. The report can be refreshed, the mapping can be edited, a lead can be reclassified and a formula can be revised. A lineage map that records only file locations misses the operations that explain why the final number moved.
Give transformations names and owners
A transformation should have a clear purpose and a reproducible rule. “Clean data” is too vague. “Remove duplicate report rows using campaign, date and country as the identity” tells the reviewer what happened and where an error might arise.
Assign an owner for the rule and a location for its implementation. If a person manually excludes a row, record the reason and the evidence. Manual work is not inherently invalid, but an undocumented manual adjustment cannot be reproduced reliably.
Keep business decisions separate from technical repairs. Excluding an invalid duplicated row is different from deciding that a lead is outside the sales definition. The first changes data integrity; the second changes the outcome population. Both need documentation, but they should not hide under one “cleanup” label.
Mark the boundaries between systems
Joins are often where a reporting story becomes uncertain. Specify the fields used to connect a ChatGPT campaign to an internal record and what happens when there is no match. Preserve unmatched rows in an exception view instead of silently dropping them from the result.
Record the grain on each side of the join. Campaign-day spend joined to individual leads can repeat the spend across every lead unless the aggregation is designed carefully. A correct-looking query can therefore inflate a financial numerator while leaving the output table plausible.
Use stable campaign identities for names and account scope. Include checks that enrichment does not unexpectedly change source totals. The lineage map should point to those checks, not merely state that the data was validated.
Trace one number backward
- Published metric
Identify the formula behind cost per qualified inquiry.
- Join
Check grain, key and unmatched records.
- Origin
Recover the source value and rule version used.
Preserve the version of the rule
Suppose qualification changes from “sales reviewed” to “sales accepted”. A later report can legitimately produce a different denominator from the same leads. Record the definition version and effective date, plus whether historical reports were restated.
Code changes need a similar record. A formula repair can affect several past periods at once. Link the revised output to the transformation version used and preserve a note explaining the correction. This allows the team to distinguish a real change in campaign performance from a change in how performance was calculated.
The metric dictionary defines what the measure means. Lineage explains how the implementation produced it. Neither replaces the other: a correct definition can be implemented incorrectly, and a perfectly reproducible pipeline can calculate the wrong business concept.
Test the map with a second person
Ask someone who did not build the report to reproduce one published value from the documented inputs. They should be able to locate the source, apply the stated rules and explain any difference due to rounding or later source updates.
If they need an undocumented password, hidden local file or verbal instruction, the map has revealed an operational dependency. Fix that dependency through the team’s approved access and storage arrangements. Do not solve it by copying secrets into the lineage document.
Also test a failed case: an unmatched campaign, a missing outcome value or a duplicated input row. The map should show where the issue becomes visible and who resolves it. Happy-path documentation alone does not protect a client report from realistic errors.
Keep the published claim traceable
Store the relevant lineage version with the reviewed report snapshot. The client does not need every implementation detail on a slide, but the team should be able to substantiate the slide when asked.
Use OpenAI reporting as the source for platform field definitions. Your transformations and business joins remain separately owned analytical work. A trustworthy ChatGPT campaign metric has a path that can be followed from the displayed number back to its evidence, with no unexplained leap between systems.
Sources and scope
Document transformations and ownership between source observations and published campaign metrics so errors can be located and reproduced.
Working methods and examples are editorial suggestions. Check current platform requirements and available features before implementation.