JuicyAgentsMetric Mismatch Finder
Find out why two reports show different numbers
Metric Mismatch Finder is an AI agent you set up in the ChatGPT or Claude you already use. Give it the two numbers and what you know about each report, and it ranks the likely causes of the gap with a quick check for each; with row-level exports of both reports, it also builds a bridge from one total to the other.
Pick your AI for the steps and the prompt
About each AI:ChatGPT WorkClaude app
Checked 29 Oct 2026. Not run by JuicyAgents yet. Written from each AI’s official docs; the example is illustrative.
Why 1,240 and 3,980 differ, 4 causes ranked, 1 unknown
- CausesPaid order versus any login; September versus 90 days; customer IDs versus account IDs
- D1’s address joinMay duplicate customers, so run a distinct count
- D2’s test-account ruleUnknown, listed after the ranked causes
- GapNo split of the gap without rows
- Dictionary entriesTwo draft dictionary entries wait for Noor and Sam; no official definition is chosen
How it works
What it does for you, and what it leaves to you.
Two reports show different numbers for what looks like the same thing, often under the same label. Different populations, date rules, filters or joins can hide behind that label, and teams lose time arguing about which number is right instead of finding the cause.
Start from the two numbers that disagree. The assistant walks the same cause checklist every time: what is counted, dates and time zones, filters, deduplication, money adjustments, refresh timing, joins that duplicate rows and sampled analytics data. It ranks the causes against what you supplied and gives each a check you can run in a few minutes. With row-level exports of both reports it builds a bridge from one total to the other and leaves any unexplained remainder visible. It never picks the official number: different definitions can be valid for different questions, and the report owners decide the labels.
You stay in charge. Report owners confirm how each number is calculated and approve any new label or dictionary entry. An analyst runs the suggested checks on the real data before a report is changed or a number is restated.
You get
- Ranked causes, each with a check you can run in a few minutes
- A reconciliation bridge from one total to the other, when both exports are supplied
- Rows found only in one report or only in the other
- Label suggestions and draft data-dictionary entries for the owners to approve
- Questions for the report owners
It never
- Say which number is right or declare an official definition
- Estimate or split the gap without rows to count
- Force an unexplained remainder to zero or spread it across causes
- Query databases, change SQL or edit dashboards
- Edit or overwrite the exports you upload
The causes it ranks
The example above, in words.
Ranked causes: paid order versus any login; September versus 90 days; customer IDs versus account IDs; D1’s address join may duplicate customers, so run a distinct count. Unknown: D2’s test-account rule. No split of the gap without rows. Two label suggestions and two draft dictionary entries wait for Noor and Sam; no official definition is chosen.
Settings
The setup prompt starts with the defaults, marked here.
What you have
Report settings and definitions give ranked causes with checks and no arithmetic. Row-level exports of both reports let your AI build a bridge from one total to the other.
- Report settings and definitions (default)
- Row-level exports
Report type
Marketing and ad platforms adds causes such as attribution, conversion date and modeled conversions.
- Business reports (default)
- Marketing and ad platforms
Your working rules
Your reporting purpose, known owners and terminology. Do not include personal row-level data.
Optional. Empty unless you fill it in.
Check the first result
Each setup page has a fictional exercise to paste first. A good answer to it meets all of these.
- Paid order versus any login, September versus 90 days, and customer versus account each appear as a supported cause with a quick check.
- D1’s address join is a possible duplication with a distinct-count check, not a proven error.
- D2’s missing test-account rule stays unknown.
- No bridge, split or estimate of the 2,740 difference is made without rows.
- Both dictionary entries are marked Draft and name Noor or Sam as approver; no official definition is chosen.
- Removing an essential input causes a question or a blocked item, not an invented fact.
Questions people ask
- Is Metric Mismatch Finder free?
- The setup guide and prompt are free. In Claude it works on any plan, including Free, which allows up to five projects. In ChatGPT it needs Work on the web, which comes with Plus, Pro, Business, Enterprise or Edu.
- Do I need to export the data?
- No. With report settings and definitions alone, it ranks the causes and says which export or count would settle each one. For a bridge between the two totals, attach row-level exports of both reports as CSV or XLSX, with an ID column they share and personal fields removed.
- Does it tell me which number is right?
- No. Different definitions can be valid for different questions. It suggests clearer labels and drafts data-dictionary entries, marked Draft until the report owner approves them.
- Does it work for ad platform and CRM numbers?
- Yes. Choose Marketing and ad platforms to add causes such as the attribution model, click date versus conversion date, modeled conversions, consent choices and CRM deduplication.
Have us set it up
Rather not set it up yourself?
We set up Metric Mismatch Finder in your own accounts and the AI you already pay for, test it on your real work, and hand it over with a short walkthrough.
From $299
- Runs in your accounts; you own everything
- Tested before handover
- Money back if we can’t make it work