ChatGPT Data Analysis Prompts: 20+ That Actually Work (2026)

By Paul Park · September 4, 2026 · 9 min read

You uploaded your spreadsheet, typed "analyze this data," and ChatGPT answered with a summary so vague it could describe anyone's sales figures. That is not a model problem; it is a prompt problem. Most ChatGPT data analysis prompts fail because they hand the model a pile of rows and no instructions about what the columns mean, what question matters, or what a useful answer looks like. This guide collects 20+ data analysis prompts that actually work: a four-part formula you can apply to any spreadsheet, copy-paste templates for cleaning data, writing Excel formulas, finding trends, and building reports, plus the verification steps that keep the numbers honest.

Why "Analyze This" Prompts Fail

Ask ChatGPT to "analyze this" and it will guess what you want, because it does not know what your columns mean, what business question you are trying to answer, or how much precision you need. The result is the average of every vague data request in its training data: three generic observations, one sentence about growth, and zero numbers you can act on. Real users describe the same frustration on Reddit's ChatGPT communities: the model is good at guided analysis but gets confused the deeper you go, and it starts inventing structure when the dataset is messy. Every failure traces back to the same root cause, which our guide to stopping generic AI responses covers in depth: garbage context in, generic output out.

The Data Brief: A 4-Part Formula for Every Analysis Prompt

Instead of a one-line request, give ChatGPT a Data Brief with four layers: role, data, task, and output format. Role tells the model how to frame the analysis; data describes your file, columns, and any caveats; task states the exact question; output format controls how the answer comes back so you can actually use it. This is the same idea as the context sandwich in our guide to structuring AI prompts, adapted for spreadsheets.

Prompt 1 — The master Data Brief template. Keep this saved and fill it in for every file:

Role: You are a senior data analyst who explains findings to
non-technical people. You never invent numbers that are not in
the data.

Data: I am uploading [FILE NAME] with columns [LIST COLUMNS]. It
covers [TIME PERIOD / SCOPE]. Known caveats: [MISSING ROWS, DUPLICATES,
CHANGED DEFINITIONS].

Task: [ONE SPECIFIC QUESTION, e.g. "Which product category drove
the revenue growth last quarter, and was it volume or price?"]

Output format: Answer in three parts: (1) a 3-sentence plain-English
summary, (2) the supporting numbers with the exact rows or totals you
used, (3) a list of anything in the data that makes the answer less
certain. If you cannot answer from the data, say so instead of guessing.

Compare what happens with and without the brief:

How to Prepare Your Spreadsheet Before You Paste

ChatGPT analyzes whatever you give it, including your formatting mistakes. A little preparation prevents most wrong answers. Clean files produce dramatically better analysis, which is why experienced users recommend cleaning the data before you upload it. Five minutes of prep saves an hour of correcting nonsense:

• One header row. Row 1 must contain unique column names like date, region, revenue. Delete title rows, merged cells, and notes above the headers.
• Consistent types. Make dates actual dates, numbers actual numbers, and blanks truly empty. Mixed formats, like "1,234" in one cell and "1234" in another, make models misread values.
• Export CSV for big files. A clean CSV parses more reliably than a formatting-heavy workbook, and it is the format the official data-analysis flow expects for uploads.
• Anonymize before uploading. Replace customer names, emails, and other personal data with IDs. You should never paste sensitive data into a chat you do not fully control. Comma-separated values have been a standard interchange format for decades for a reason: the CSV format is plain text that any tool can read, which also makes it the safest format to share with an AI.
• Sample enormous files. If a workbook has hundreds of thousands of rows, upload a random sample or a pre-aggregated summary, and say clearly that it is a sample.

ChatGPT Data Analysis Prompts That Actually Work

The templates below are grouped by job. Fill in the brackets, upload your file, and adapt the output rules to your audience. Prompt 1 is the skeleton; Prompts 2 through 22 are ready-to-use specializations.

Dataset audit prompts

Prompt 2 — Understand what you just uploaded.

Here is my file [FILE]. Before any analysis, audit it and report:
(1) column names with their data types, (2) row count and time range,
(3) missing values per column, (4) obvious duplicates, and (5) any
values that look wrong, like negative quantities or future dates.
Do not summarize trends yet; just describe the data's condition.

Prompt 3 — Define the metrics before you measure them.

Here is my file [FILE]. I plan to analyze [GOAL, e.g. customer
retention]. Propose 3-5 concrete metrics I can calculate from these
columns, define each one in one line, and flag any column that is
missing or too messy to support it.

Data cleaning prompts

Prompt 4 — Diagnose dirty data.

Audit my file [FILE] for data-quality problems: inconsistent
spelling, duplicate rows, mixed date formats, blank cells, trailing
spaces, and numbers stored as text. For each problem, show the column,
the number of affected rows, and an example. Do not change anything
yet; give me a fix list first.

Prompt 5 — Write a cleanup plan I can approve.

Based on your audit of [FILE], propose an exact cleanup plan: what
to delete, what to standardize, and what to leave alone because it
might be meaningful. For ambiguous cases, ask me a question instead of
deciding. I will approve each step before you apply it.

Prompt 6 — Detect and deduplicate rows safely.

My file [FILE] may contain duplicates. Find rows that are exact
duplicates and rows that are near-duplicates (same customer or same
transaction with small differences). Show me counts and examples, then
recommend a deduplication rule that keeps the most complete record.
Do not delete anything yet.

Prompt 7 — Standardize messy categories.

List every unique value in the [CATEGORY] column of my file and flag
groups that mean the same thing but are spelled differently, like
"NY", "N.Y.", and "New York". Propose one canonical label for each
group and show the mapping table.

Exploration and insight prompts

Prompt 8 — Find the trends that matter.

Using my file [FILE], identify the five most important trends in
the data. For each trend: state it in one line, show the supporting
numbers, name the time period where it started, and say how confident
you are. Rank them by business impact, not statistical noise.

Prompt 9 — Explain an outlier honestly.

Here is my data [FILE] and the metric I care about is [METRIC].
Find the biggest outliers. For each one, show the value, the date
range, and up to three plausible explanations that are consistent
with the other columns. Tell me which explanation you can actually
support from the data and which is speculation.

Prompt 10 — Segment without overfitting.

Segment my customers or rows in [FILE] into 3-5 groups using the
available columns. Name each segment, give its size and the values of
its defining columns, and state one assumption behind the grouping.
If a segment is too small to trust, say so.

Prompt 11 — Compare periods correctly.

Compare [PERIOD A, e.g. last quarter] with [PERIOD B, e.g. the
quarter before] in my file [FILE]. Use the same number of days for
both periods, adjust for known differences I list here [NOTES], and
report the change as both absolute and percentage. Flag any period
with missing data that makes the comparison unfair.

Excel and Google Sheets formula prompts

Prompt 12 — Generate a formula from a plain-English request.

In Excel, I want to [TASK, e.g. "sum revenue for the West region in
March"]. My columns are [COLUMNS]. Write the formula, explain what
each part does in one line, and give a version that works with both
Excel and Google Sheets if the syntax differs.

Prompt 13 — Fix a formula that returns an error.

This formula returns [ERROR OR WRONG RESULT]: [PASTE FORMULA].
My data looks like [SAMPLE ROWS]. Diagnose why it fails, then give me
the corrected formula and a test case I can use to confirm it works.

Prompt 14 — Build a formula from a description, not a column list.

Pretend you cannot see my sheet. I will describe the layout in
words: [DESCRIBE COLUMNS AND ROWS]. Write an Excel formula that
[GOAL]. Use cell references only from my description, and if my
description is missing something, ask before guessing.

Prompt 15 — Turn a manual process into a formula.

I currently do this by hand in Excel: [DESCRIBE THE STEPS, e.g.
"look up the price from the price tab, multiply by quantity, and
subtract the discount if the code starts with VIP"]. Write a single
formula or a small set of formulas that automates it, and show me
where to put each one. Refer to the function reference if needed:
[Google Sheets function list] is the official catalog.

Chart and report prompts

Prompt 16 — Choose the right chart.

For my data [FILE] and the message [MESSAGE, e.g. "market share
shifted from product A to product B over the year"], recommend the
best chart type and explain why. Then describe exactly what data
series go on each axis so I can build it myself.

Prompt 17 — Write a chart title that says something.

Here are the numbers I plan to chart: [DATA OR SUMMARY]. Write 5
chart titles that state the finding instead of describing the axes,
like "Category B overtook Category A in June" rather than "Revenue by
category". Keep each title under 60 characters.

Prompt 18 — Summarize the data for a busy reader.

Here is my analysis of [FILE]: [PASTE FINDINGS]. Write a 150-word
executive summary for a manager who has not seen the data. Lead with
the single most important number, then the reason it happened, then
one recommended action. No jargon, no hedging words like "interestingly".

Business decision prompts

Prompt 19 — Stress-test a conclusion.

My tentative conclusion from [FILE] is: [CONCLUSION]. Act as a
skeptical reviewer. List the three strongest reasons it could be
wrong, what additional data would confirm or refute it, and one
alternative explanation the current data cannot rule out.

Prompt 20 — Turn findings into decisions.

Based on these findings from [FILE]: [PASTE FINDINGS]. Recommend
three concrete actions ranked by expected impact and effort. For each
action, state what you would measure to know it worked. Keep
recommendations to what the data actually supports.

Prompt 21 — Forecast with visible assumptions.

Using the time series in [FILE], project [METRIC] for the next
[PERIODS]. State every assumption you make about seasonality, growth
rate, and one-off events. Show a low, middle, and high scenario, and
label the projection as an estimate, not a prediction.

Prompt 22 — Explain the analysis to a non-technical audience.

Explain my analysis of [FILE] to [AUDIENCE, e.g. "a store manager
who has never used Excel"]. Use an analogy from retail or everyday
life, define every term, and end with the one thing they should do
differently tomorrow. No formulas, no statistics jargon.

How to Check ChatGPT's Numbers Before You Trust Them

ChatGPT predicts text; it does not run a certified audit on your workbook. Even with a perfect Data Brief, it can transpose columns, round silently, or produce a number that looks right and is not. That risk is why hallucination prevention matters more in data work than anywhere else, and our guide to reducing AI hallucinations walks through the general defenses. For spreadsheets specifically, four checks catch almost every error:

• Ask for the method. Every numeric claim should come with the formula, the row range, or the filter used. If ChatGPT cannot show its work, treat the number as a guess.
• Cross-check one total yourself. Pick the metric that matters most and verify it in Excel or Sheets with a simple formula before you build anything on it.
• Beware plausible answers. The dangerous outputs are not obviously wrong; they are neatly formatted and slightly off. Compare against a second calculation whenever the decision is important.
• Set the honesty rule in the prompt. Add "If you cannot answer from the data, say so" to every template. Models comply with an explicit permission to be uncertain.

You can also make the honesty rule persistent with ChatGPT custom instructions, so every analysis session starts with the same verification defaults instead of relying on each prompt to remember.

When Your Prompt Is the Problem, Not the Model

If your analysis requests keep coming back generic, look at what you actually typed. A rambling "look at this and tell me what you see" request will produce a rambling answer no matter which model you use. The fix is to restructure the request before you send it: name the file, list the columns, state the question, and specify the output. That restructuring is exactly what a prompt enhancer automates. Prompt Helper Gemini is a free Chrome extension that adds an Improve button inside ChatGPT, Gemini, Claude, Grok, and Perplexity; one click turns a messy data question into a structured prompt with role, context, and output rules. It does not touch your spreadsheet and it does not analyze data for you, but when you are tired and your prompt has gone lazy, it rebuilds the Data Brief structure in seconds. Free tier: 5 prompt enhancements and 5 Ask questions per week.

FAQ: ChatGPT Data Analysis Prompts

Final Thoughts: Better ChatGPT Data Analysis Prompts, Better Decisions

Every ChatGPT data analysis prompt in this guide shares one lesson: the model is only as precise as the brief you give it. Describe the file, state the question, demand the method, and verify the numbers, and ChatGPT becomes a genuinely useful analyst instead of a confident generalist. Start with the master Data Brief in Prompt 1, run the audit in Prompt 2 on your messiest spreadsheet, and add the honesty rule to everything you send. If you want the full method for improving prompts themselves, our guide to improving ChatGPT prompts is the natural next read.