File uploads and data analysis

0:00 Data analysis you can trust

Uploading a spreadsheet to ChatGPT and asking what does this tell me feels like magic. It is also one of the easiest ways to present a confident, wrong number to your boss. In this lecture you will learn a six step workflow that turns ChatGPT into a genuinely reliable analyst, and the questions that expose weak conclusions before anyone else does.

0:27 Why data needs a method

Why do we need a method for data? Because a chart always looks authoritative, even when the data behind it is a mess. Duplicate rows, text stored as numbers, a metric defined three different ways. Those problems do not announce themselves. They quietly change the answer. Think of it like cooking. The dish can look beautiful and still be made with the wrong ingredient. The six step method makes sure you check the ingredients before you serve the result.

1:01 Why code matters

ChatGPT handles two kinds of files differently. Documents it reads and reasons over. Data files, like spreadsheets and CSVs, it can analyse by writing and running Python code in a secure sandbox. That is the key. Numbers computed with code are far more reliable than numbers estimated in a paragraph. You can also pull files from connected apps like Google Drive or SharePoint, where your plan and admin allow.

1:31 Step 1: describe first

Step one, describe the data before analysing it. Ask for the number of rows, each column with its type and an example, missing values, duplicate rows and anything that looks like an outlier or data entry error. And say, do not analyse yet. This single step catches most of the problems that ruin an analysis.

1:55 Steps 2 and 3

Step two, clean deliberately. You decide the rules. Remove exact duplicates. Treat blank regions as unknown. Exclude test orders. And ask how many rows each rule affected. Step three, define your metrics. Engagement rate equals likes, comments, shares and saves, divided by impressions. Revenue is gross sales minus refunds, excluding VAT. Ambiguous metrics produce confident, wrong comparisons.

2:20 Steps 4 and 5

Step four, analyse with a specific question. Which content format had the highest median engagement rate in the third quarter, by platform? Show the table and the code. Median is often more honest than the average when a few viral posts skew everything. Step five, visualise simply. Sorted bars, labelled axes, no three D effects, and ask for an exportable file when you need to share it.

2:49 Step 6: verify

Step six, verify. Spot check two numbers yourself in Excel or Sheets. Ask to see the rows behind the top result. And always ask how many items are in each group. A chart that says Stories performed best looks very different once you learn it is based on three posts.

3:11 Read results critically

Read every result critically. Small samples are anecdotes, not trends. Tuesday posts may perform better because of what you posted on Tuesdays, not the day itself. Analysing only completed campaigns hides the ones cancelled early. And mixed currency columns, dirhams, riyals and pounds, must be converted with a stated rate and date. A great habit is to ask, what are the three biggest reasons this conclusion might be wrong?

3:41 Simple example

A simple example. Upload a small sales file with fifty rows and ask, before any analysis, to describe the data. ChatGPT reports two duplicate rows and one blank region. You tell it to remove the duplicates and label the blank as unknown. Then ask for total sales by month, with the code. Pick one month and add it up yourself in a spreadsheet. It matches. You now trust the result, and you know exactly why. That is the whole method in miniature.

4:17 Worked example: a UAE brand’s social review

A social media manager in the UAE exports six months of Instagram and TikTok analytics. ChatGPT finds fourteen duplicate rows and a column where some impressions were stored as text. She sets cleaning rules, defines engagement rate and asks for the median by format and platform. Carousels beat reels on Instagram for her account. She checks the counts, twenty two carousels and forty one reels, spot checks two values in Sheets, and presents it as a pattern worth testing, with a four week experiment plan, not as a law. Four weeks later, she reruns exactly the same cleaned workflow on the new export. Because the cleaning rules and the metric definition were saved as a recipe, the comparison is like for like. The carousel lead holds on Instagram but shrinks, and on TikTok nothing changes. Her report says so plainly, with the post counts beside each figure, and her manager approves a small shift in the content mix rather than a dramatic one.

5:28 Privacy first

Before uploading, remove names, emails, phone numbers and order numbers you do not need. Use your organisation's approved workspace for company data, not a personal account. And delete working files from chats when you are finished, where appropriate. If a colleague later asks for the source data, you will be glad you kept only what the analysis needed.

5:53 Try this now

Try this now. Export a real dataset that contains no personal data, your own social analytics, website traffic by page, or a public dataset. Upload it and work through all six steps, describe, clean, define metrics, analyse against one clear question, visualise simply, and verify. Spot check two numbers yourself in a spreadsheet, and ask how many rows sit behind the top result. Save the prompts as a data analysis recipe in your prompt library or a Project, so next month's analysis takes half the time.

6:30 Common mistakes

Let's name the common mistakes, because you will see them in other people's reports. Asking for conclusions before checking data quality. Using the average when one viral post or one huge order distorts everything, so ask for the median too. Mixing currencies without a stated conversion rate and date. And presenting a result based on a handful of rows as if it were a trend. If you avoid those four, your analysis will already be better than most.

7:04 Watch me do it, part 1

Let me run the UAE social review. I upload the six month export and paste the describe first prompt. Rows, columns with types and examples, missing values, duplicates and outliers, and do not analyse yet. ChatGPT runs code and reports fourteen duplicate rows and an impressions column where some values are stored as text with commas. I set the cleaning rules. Remove exact duplicates, convert impressions to numbers, and exclude two test posts published from our staging account. I ask how many rows each rule affected, and it reports fourteen, thirty one and two. Now I trust the table I am about to analyse.

7:49 Watch me do it, part 2

Next, the metric. Engagement rate equals likes, comments, shares and saves, divided by impressions. Then the question. Median engagement rate by format and platform, with the number of posts in each group, and show the code. On Instagram, carousels lead, from twenty two posts, against reels from forty one. I ask for a simple sorted bar chart. Then my check. I open the cleaned file in Google Sheets, filter Instagram carousels, and calculate the median myself. It matches. I check one reel figure too. Both match, so the finding goes into my report, framed as a pattern to test over four weeks.

8:33 Recap and next step

Recap. Describe the data first, clean deliberately, define metrics, analyse against a clear question, visualise simply and verify. Question every conclusion and protect personal data. Your next step: run the six steps on a real, non sensitive dataset, spot check two numbers, and save the prompts as your reusable data analysis recipe.