ChatGPT basics ChatGPT basicsFoundations

ChatGPT file uploads and data analysis: check the spreadsheet

Give ChatGPT the source file, metric definitions, desired output and validation checks. Use Work for a finished analysis or workbook, then inspect the calculations and source reconciliation. This guide includes a fictional campaign dataset with known totals so you can test the result.

Overview

I would judge an AI-generated campaign report by the denominator before judging the chart. A beautifully labeled conversion rate can still divide the wrong two columns. That is the sort of mistake that survives a meeting because the slide looks reassuring.

The habit I want is the reconciliation line: a visible connection between the source rows and the reported totals. If that line is missing, the analysis is still a draft.

Give the file a business definition

Give the file a business definition

A column named leads does not explain whether it contains people, form submissions or campaign memberships. A date may refer to creation, conversion or export time. These definitions change the answer, so they belong with the file rather than in the analyst's head.

For the fictional training dataset in this guide, leads are unique campaign-attributed leads and opportunities are qualified opportunities attributed to each campaign. Pipeline is unweighted attributed pipeline value, not booked revenue. All spend is USD for one illustrative period. Those constraints make the arithmetic interpretable.

I would ask ChatGPT to restate the definitions before making recommendations. If the real export lacks them, resolve that gap or keep the conclusion narrow. A correct formula applied to the wrong business definition is still the wrong analysis.

TipName the unit of each row: a lead, an opportunity, a campaign or a daily observation.

Choose a reviewable output

Choose a reviewable output

ChatGPT Work can produce files such as spreadsheets, documents and presentations from supplied sources. Request the file type and structure you actually need. For a campaign review, a workbook with source data, calculated metrics and a short summary is easier to inspect than a paragraph of unexplained numbers.

Keep the raw inputs visible and unchanged in their own section or sheet. Put derived fields beside clear formulas. Put recommendations somewhere else. That separation makes it possible to find out whether a disagreement comes from the data, the arithmetic or the judgment.

In desktop, supported previews and annotations can help with focused revisions. On the web, review the generated file and download it when necessary. The interface may differ, but the standard remains the same: the reviewer can trace a claim to a calculation and then to its source.

Reconcile the calculations with the source before using the result to make a decision. 01 / Source: Original rows and definitions; 02 / Calculation: Formulas and exclusions; 03 / Reconcile: Counts and totals match; 04 / Decision: Evidence plus limitations
Reconcile the calculations with the source before using the result to make a decision. Open diagram

TipSpecify the sheet names and checks in the original request.

Use the fictional campaign dataset

Use the fictional campaign dataset

Download campaigns.json and attach it to the task. It contains four fictional campaigns: Search, LinkedIn, Webinar and Partner. It is training material, with no real customer performance implied. The definitions are included in the file so the task starts with explicit context.

The known totals are $4,000 spend, 150 leads, 18 opportunities and $245,000 pipeline. These figures give you an independent check. If the report returns different totals, investigate before evaluating the narrative. A disagreement here is useful: it exposes a problem early, on material whose answer you know.

For a real export, create similar reconciliation checks before asking for interpretation. Count rows, identify duplicates and record totals for the relevant fields. The small training example teaches the habit; the habit is what transfers to a messy production file.

TipSave the original file so you can compare it with the generated workbook.

Calculate the weighted aggregate correctly

Calculate the weighted aggregate correctly

Overall cost per lead is total spend divided by total leads: $4,000 / 150, approximately $26.67. Overall cost per opportunity is $4,000 / 18, approximately $222.22. The lead-to-opportunity rate is 18 / 150, or 12 percent. These are arithmetic checks on the fictional data, not performance targets.

Do not average campaign-level cost per lead to get the overall figure. Search is $20, LinkedIn $60, Webinar $15 and Partner $20. Their simple average is $28.75, which answers a different question because it weights each campaign equally rather than each lead. This is exactly the kind of plausible number a reviewer should challenge.

The pipeline-to-spend ratio is $245,000 / $4,000, or 61.25. Call it that. It is not revenue return, profit or causal proof that the campaigns generated incremental business. A ratio becomes misleading when the label claims more than the data supports.

TipFor aggregate rates, sum the numerator and denominator before dividing.

Make missing values visible

Make missing values visible

Real exports introduce nulls, duplicate rows, mixed currencies and partial periods. Ask the agent to report how each was handled. A blank opportunity count should not silently become zero unless that is the agreed data rule. The difference changes both the total and the interpretation.

For division by zero, require an explicit unavailable result rather than a misleading zero cost or conversion rate. Preserve the reason in a note. For duplicates, define the key that makes two records the same and retain an exclusion log. “Cleaned the data” is too broad to review.

I would run a second test with one missing field and one duplicate before using the workflow on a real campaign export. You are testing how the analysis fails, which is often more informative than watching it succeed on four tidy rows.

TipRequire an exclusions sheet or list whenever rows are removed or transformed.

Ask for a recommendation within the evidence

Ask for a recommendation within the evidence

The fictional data can support descriptive comparisons. It cannot establish campaign quality over time, incremental lift or sales capacity. A useful recommendation says what the observed ratios suggest and what additional evidence would change the decision.

For example, LinkedIn has a higher cost per lead in this sample and a different opportunity mix. That can justify investigating audience, lead quality and deal progression. It does not automatically justify cutting the channel. The sample is small and contains no experiment or time series.

Ask for the recommendation to separate observations, hypotheses and next checks. This keeps the output useful without forcing certainty. The best analysis may end with a specific investigation rather than a budget change. A spreadsheet is allowed to reveal a question (finance will survive).

TipRequire one piece of evidence that would reverse the recommendation.

Inspect the file as well as the numbers

Inspect the file as well as the numbers

Open every sheet and check the formulas, formats and labels. Confirm that currency appears consistently and percentages are not displayed as raw decimals without explanation. Inspect whether long labels are cut off and whether the summary points to the correct calculation range.

If the workbook is intended for another application, open it there. A preview is useful, but it is not a substitute for checking the artifact in its destination. For a document or presentation, verify that the same numbers survive the transfer from workbook to narrative.

Ask ChatGPT to report what it actually checked. A successful file save establishes that a file exists. Formula verification, reconciliation and visual inspection are separate checks. Keep them separate in the completion note so the reader can understand the remaining uncertainty.

TipA completion message should name the output file and the checks performed.

Where file analysis goes wrong

Where file analysis goes wrong

The most common failures are silent assumptions: averaging ratios, merging periods, treating blanks as zeros and calling pipeline revenue. They often survive because the report is otherwise coherent. Put these checks in the request before the first run.

Another mistake is letting the agent overwrite the source. Keep raw material separate and write derived output to a new file. For repeated analysis, preserve the definitions and checks in a Project or reusable skill, while supplying a fresh dated export each time.

Once the fictional example reconciles, try a small approved real export and manually check a sample. Expand only when the same logic holds. Which number in your weekly report can someone trace all the way back to the source rows?

How to set it up

How to set it up

Download and attach the example

Use the fictional campaigns.json linked above. Read its definitions before starting.

Request the workbook

“Create a campaign analysis workbook with Raw Data, Metrics and Summary sheets. Preserve the source. Use formulas for CPL, cost per opportunity and lead-to-opportunity rate. Reconcile totals and distinguish pipeline from revenue.”

Check the known answers

Verify $4,000 spend, 150 leads, 18 opportunities, $245,000 pipeline, $26.67 aggregate CPL and 12% conversion.

Test a messy input

Add a missing value or duplicate in a copy. Require explicit handling and an exclusion log, then inspect the resulting file.

FAQ

Frequently asked questions

What should I upload?

The smallest approved source file that contains the relevant records, plus metric definitions and the reporting period.

Can ChatGPT create a workbook?

Work supports producing spreadsheets and other files. Available tools and previews depend on the surface and account.

Why not average campaign CPL?

A simple average weights campaigns equally. Overall CPL requires total spend divided by total leads.

Is pipeline divided by spend ROI?

It is a pipeline-to-spend ratio. It does not establish realized revenue, profit or causal return.

What should happen to blank cells?

Apply an explicit agreed rule and disclose it. Do not silently convert unknown values into zero.

Should I trust the chart?

Check its source range, labels and calculations before accepting the visual conclusion.

Is the example real company data?

No. It is fictional training data with known arithmetic checks.

How do I repeat this monthly?

Preserve definitions and validation checks, use a fresh dated export and review changes to the source structure.

Sources

Sources & further reading

ChatGPT and Codex change quickly. This page was last reviewed September 22, 2026; verify time-sensitive details against the official docs above before relying on them.

Put it to work

Related GTM workflows

Use these existing playbooks to explore the business workflow. Adapt their tool-specific steps to your chosen environment and check the result.

Get the AI-for-GTM playbook in your inbox

New ChatGPT and Codex guides, use cases, and prompts every couple of weeks.

Subscribe →