Data analysis prompts: 20 ready-to-use templates for AI

Monday, 9:10 a.m. The marketing lead exports the Meta Ads report, pastes it into ChatGPT and types: "analyze this for me". What comes back is four paragraphs explaining that "CTR is an important engagement indicator" and suggesting she "test new creatives". Nothing wrong, nothing useful. She closes the tab and builds the deck by hand, like every Monday.

The problem is rarely the AI. It is the request. A data analysis prompt works like a brief for a newly hired analyst: if you don’t say what the business is, what changed, what a good result looks like and what format you want, you get the obvious back, in excellent grammar.

And more people are making that request every year. According to the TIC Empresas 2025 survey by Cetic.br, released in June 2026, 17% of Brazilian companies used some form of AI in 2025. Among those, use of text generation jumped from 20% to 30% in a single year. Adoption is growing faster than the quality of the questions.

A good data analysis prompt has five parts: business context, a description of the data (source, period and what each column means), the exact question, the answer format, and a request to check the calculations. Without context and a specific question, AI describes the spreadsheet instead of analyzing it. With all five parts, it delivers a diagnosis you can act on.

Why doesn’t “analyze this data” work?

Because the AI doesn’t know what normal looks like for you. A $48 CPA is excellent if you sell a $1,997 course and a disaster if you sell $79 T-shirts. Without your target and margin in the prompt, the model has no way to judge, so it does the only thing left: it describes.

The second reason is ambiguous columns. The “Results” column in Meta Ads can mean leads, purchases or landing page views, depending on the campaign objective. “Revenue” in GA4 may or may not include shipping. If you don’t define it, the AI picks an interpretation and runs with it, confidently.

The third is arithmetic. Language models are bad at mental math, especially long sums and weighted averages. When the tool runs code (ChatGPT’s data analysis mode, for example), the math becomes reliable, but only if you explicitly ask it to calculate with code and show the formula. That one detail removes a large share of made-up numbers.

The 5-block structure of a data analysis prompt

Every prompt that works, from the simplest to the most elaborate, carries these five blocks. You don’t need a paragraph for each. Sometimes one line is enough.

Block	What to include	Example
Context	Business, target, timing	“Fashion e-commerce, $180 average order value, CPA target under $35, pre-Black Friday period.”
Data	Source, period, column definitions	“Meta Ads export from Sept 1 to 30. The Results column means purchases.”
Question	One decision to make	“Which three ad sets should I pause this week?”
Format	Structure, length, audience	“A table of ad sets and three recommendations in under 150 words, for the director.”
Check	How to validate the result	“Calculate with code, show the CPA formula and confirm total spend ties out to $7,800.”

In practice, the question block is where most people get stuck. “How did September go?” is not a question. It is an invitation to summarize. “Which campaigns should I pause?” is a question, because it requires criteria and ends in an action. If you’re struggling to phrase it, our guide to data analysis questions by business area has 30 examples.

A useful trick for the context block: end the prompt with “before answering, tell me what else you’d need to know about the business to make this analysis reliable”. Half the time the AI points to exactly the piece you forgot.

Base template to copy: “You are a senior data analyst. Context: [business, target, timing]. Data: [source, period, what each column means]. Question: [decision I need to make]. Format: [table, length, audience]. Calculate using code, show the formulas and check that totals match [known value]. If information is missing, tell me what is missing instead of assuming.”

20 ready-to-use prompts for data analysis with AI

The prompts below assume you’ve already pasted the base template with context and data. Replace anything in brackets with your own case. They follow the order an analysis usually takes: understand first, then compare, then hunt for what looks off, and finally communicate.

Diagnosis: understand what you have

1. “Before any analysis, list data quality issues in this dataset: missing dates, duplicate rows, extreme values and columns with inconsistent formats. Don’t fix anything yet, just flag it.”
2. “Calculate [revenue, orders and average order value] by [channel]. Show the table and tell me whether any channel accounts for more than 50% of the result.”
3. “Break down the revenue change between [August] and [September] into three effects: order volume, average order value and product mix. Show how much each one explains in dollars.”
4. “Using the columns [sessions, carts, checkouts, orders], tell me which funnel stage has the largest proportional drop-off and compare it with the previous month.”
5. “If I could only track one number from this dataset every week, which would it be and why? Justify it with this data, not with generic best practices.”

The first prompt looks bureaucratic, but it saves the most time. A duplicated row in an orders table inflates revenue silently, and no later question will catch it. We cover these errors in more depth in data quality: the silent errors in your analyses.

Period comparison: separate signal from calendar

6. “Compare [last week] with the same week last year and with the average of the four previous weeks. Highlight cases where the two comparisons point in opposite directions.”
7. “Normalize results per business day before comparing [October] with [September]. How much of the difference disappears after normalizing?”
8. “List the five [campaigns] that grew the most and the five that dropped the most in absolute value, not percentage. Ignore items under [$500] in the base period.”
9. “CPA went from [$32] to [$41]. Separate the effect of CPM, CTR and conversion rate. Which of the three explains most of the increase?”
10. “Build a month-by-month table of [metric] with percentage change. Flag months outside the range of the mean plus or minus two standard deviations.”

Prompt 8 exists for a practical reason: ranking by percentage always puts the campaign that went from $40 to $160 at the top. It grew 300% and changes nothing about your month. Asking for absolute values with a minimum floor cuts that noise.

Anomalies: find what’s off before your client does

11. “Identify days when [spend] was more than 30% above or below its 7-day moving average. For each one, tell me whether [conversions] followed.”
12. “Is any [campaign] spending for more than [5] consecutive days without a conversion? List them with accumulated spend over that period.”
13. “Find pairs of metrics that usually move together in this dataset and show the days they diverged, such as clicks rising while sessions fall.”
14. “Compare total [Meta Ads conversions] with total [CRM orders] over the same period. Calculate the percentage gap and list likely causes, ordered by probability.”
15. “Which customers bought more often than average over the last [90] days and stopped in the last [30]? Give me the list and the value they represent.”

Prompt 14 tends to be the most revealing for anyone running paid media. Some gap between the ad platform and the CRM is normal and explainable (attribution windows, view-through conversions, canceled orders). When the platform reports twice what the CRM records, something in tracking is broken. If the drop is in sales, the full playbook is in sales dropped? a data diagnosis in 6 steps.

Reporting and decisions: turn analysis into action

16. “Write the executive summary of this analysis in five sentences for [the sales director]: what happened, why, what to do and the risk of not doing it.”
17. “Turn these numbers into three recommendations with estimated impact in dollars. For each one, tell me which assumption, if wrong, would invalidate it.”
18. “Play a skeptical executive and point out the three biggest weaknesses in this conclusion: [paste conclusion].”
19. “Suggest the best chart to show [this comparison] to a non-technical audience and explain in one sentence why it works better than a table.”
20. “Based on what this analysis couldn’t answer, list the questions I should investigate next week.”

How do you stop AI from making up numbers?

No prompt eliminates the risk, but a simple loop reduces it a lot. The rule: the AI calculates, you check two numbers, and only then does the conclusion count.

A concrete example. You have four campaigns with spend of $3,200, $2,150, $1,480 and $970. The total is $7,800, and you know that because it’s on the invoice. Put that figure in the prompt and ask the AI to confirm the sum before any analysis. If it says $7,950, you’ve found a duplicated row or a misread, and avoided an entire conclusion built on a false number.

Three other habits make a difference:

• Ask it to say “I don’t know”. The line “if information is missing, tell me what is missing instead of assuming” cuts down invented answers considerably.
• Upload the file instead of pasting. Pasting thousands of rows into the chat raises the risk of silent truncation. For large datasets, upload the file or aggregate first (by day, by campaign).
• One question at a time. Prompts with five questions get five shallow answers. Chaining works better than stacking.

Where prompts stop being enough

Good prompts solve the question side. They don’t solve the data side. If your information lives across Meta Ads, Google Ads, GA4, a CRM and a finance spreadsheet, you become the company’s manual ETL: export, clean, merge and paste every week. That’s where the routine dies, usually by week three.

This isn’t only a small-business problem. In Brazilian industry, 41.9% of companies with 100 or more employees were already using AI in 2024, according to IBGE’s PINTEC Semestral survey, released in September 2025. In the same survey, lack of qualified staff was cited as a difficulty by 54.2% of companies adopting advanced digital technologies. The tools arrived before the skills and, almost always, before the data was integrated. The cost of that isolation is covered in data silos: why your company decides in the dark.

Another limit: context is lost with every new conversation. You explain the target, the margin and what each column means, and next week you start from scratch. Saving the base template in a document helps, but it doesn’t replace having business context stored alongside the data.

Frequently asked questions

What is the best ChatGPT prompt for data analysis?

There’s no single best prompt, but there is a best structure. The ones that work include business context, a description of the data, a question that ends in a decision, the desired format and a request to check the math. The base template in this article covers all five blocks and works as a starting point for any case.

Can ChatGPT analyze an Excel spreadsheet?

Yes. When you upload the file, the data analysis feature runs Python code to read the spreadsheet, calculate and create charts. Always ask it to show the code or formulas used so you can check the logic. Without code execution, calculations are less reliable.

Can I paste customer data into ChatGPT?

Carefully. Under privacy laws such as Brazil’s LGPD or Europe’s GDPR, personal data like names, ID numbers, emails and phone numbers shouldn’t go into tools without a legal basis and without knowing where they’re stored. The safe route is to anonymize first, replacing identifiers with codes, or to use business plans with controlled data retention.

What is prompt engineering for data analysis?

It’s the practice of structuring instructions so AI delivers correct, actionable analysis instead of generic summaries. It involves giving context, defining the data, narrowing the question, specifying the format and building in checks. It doesn’t require programming, just clarity about the decision you want to make.

The question is yours. The grunt work doesn’t have to be

Knowing how to ask is the skill that will separate people who use AI to decide from people who use it to produce text. The 20 prompts above are good practice. But if every Monday you still export five reports to paste into a chat, the bottleneck is no longer the prompt. Sherlok connects Meta Ads, Google Ads, GA4, CRMs, spreadsheets and SQL databases and keeps your business context, so you can ask these same questions in plain language, directly on up-to-date data, and get the analysis ready.