ChatGPT Analytics: Data Analysis, Workspace Usage and Your Traffic

ChatGPT analytics means one of two things. Most often it is using ChatGPT to analyze your own data: you upload a spreadsheet and it writes and runs Python to clean it, summarize it and chart it. In ChatGPT Enterprise and Edu it also means the workspace analytics dashboard, where admins track how many people use ChatGPT and for what.

There is a third meaning that none of the top results cover: measuring the visitors ChatGPT sends to your website and what they buy. This guide covers all three, with a worked example, the failure modes to watch for, and where the popular guides disagree.

Which ChatGPT analytics do you need?

Start from the question you are trying to answer. Each one leads to a different tool and a different section below.

How ChatGPT data analysis works

When you attach a data file and ask a question, ChatGPT does not guess the numbers from the text. For most analysis tasks it writes Python and runs it in a sandboxed, stateful Jupyter notebook, using libraries such as pandas for tables and matplotlib for charts. The results come back as text, interactive tables or charts, and you can open the code behind any answer. OpenAI’s data analysis help article describes the current behavior.

The feature has been renamed twice. It launched as Code Interpreter, became Advanced Data Analysis, and is now simply part of ChatGPT: you no longer switch it on, it runs when a question needs it. Older tutorials that show a separate toggle describe the 2023 interface.

Files it can read

Limits on the number and size of files vary by plan, model and workspace settings, so a file can upload and still be too large or messy to analyze completely. Split big exports by month or by sheet if the answers look incomplete.

What it is good at

• Describing a dataset. Row counts, column types, missing values, ranges, outliers and a plain-English summary of what each column holds.
• Cleaning and reshaping. Removing empty rows, fixing date formats, splitting columns, pivoting wide tables into long ones and giving you the result as a downloadable CSV.
• Aggregations. Counts, sums, averages and proportions by group, plus crosstabs of two categorical columns.
• Statistics. Correlations, t-tests, chi-squared tests, regressions and factor analysis, when you name the method you want.
• Text analysis. Labeling survey answers or reviews as positive, negative or neutral, or grouping open-text feedback into themes. Expect some misclassified rows; it is better at the overall distribution than at each individual label.
• Charts. Bar, line, pie and scatter charts can be switched to an interactive view; other chart types, such as heatmaps, may come back as static images. You can ask for colors, labels and sort order in plain language.

How to analyze data with ChatGPT in 5 steps

This is the method that produces answers you can defend in a meeting, not just plausible ones.

Step 1: Write down the decision

Open with what you are deciding and which metric settles it: “I’m deciding which landing pages to update first, based on conversions per visit over the last 90 days.” That one sentence tells ChatGPT what a finished answer looks like, which columns matter and which ones to ignore.

Step 2: Prepare and upload a clean file

Put descriptive headers in the first row, one record per row, and nothing else on the sheet. Delete totals rows and notes. Remove or hash names, emails and phone numbers before you upload. Then attach the file with the paperclip or plus button. A prebuilt GPT such as OpenAI’s Data Analyst works the same way; it is a convenience, not a requirement.

Step 3: Ask for a plan before an answer

Ask ChatGPT to describe the data, list its assumptions and propose the calculations before running them. You catch misread columns at this stage instead of after a chart is built on them. Add your ground rules: name the method if it matters, ask it not to treat correlation as cause, and ask it to flag missing data or odd spikes.

Step 4: Iterate on the output

Follow-up prompts are where most of the value is. Ask for a different grouping, exclude a segment, switch a pie chart to a bar chart, rename a column or add a second chart. ChatGPT keeps the notebook state between messages, so each step builds on the last one.

Step 5: Verify and export

Open the code behind the answer, then recompute two or three numbers yourself in a spreadsheet. If they match, download the cleaned table, the chart (interactive charts download as HTML; use the image export for slides) or the notebook code so you can rerun it next month.

Other ways to bring data in

• Inside Excel or Google Sheets. ChatGPT is available as a sidebar in both, if your workspace allows it. It is better suited to workbook mechanics, such as fixing formulas or building a KPI tab, than to statistics.
• Connected apps and the Data plugin. Where your plan and workspace support it, ChatGPT can query connected warehouses, BI tools and drives, then turn the findings into charts, dashboards or report drafts. Results are only as good as the access and metric definitions you give it.
• Pasting a table. Fine for a few dozen rows. For anything bigger, upload a file so the numbers go through code rather than being read from text.

How accurate is ChatGPT for data analysis?

When the task is pure computation, accurate. A peer-reviewed study archived on PubMed Central compared exploratory factor analysis run by ChatGPT with the same analysis in R on simulated datasets. The computed statistics matched R, and repeating the analysis a week later with the same prompt gave the same results. The author’s caution was about judgment: deciding how many factors a multidimensional structure has, or whether an item belongs in a scale, still needs a researcher.

That matches day-to-day experience. The arithmetic is reliable because Python does it. The weak points are upstream and downstream of the code: which rows went in, which method was chosen, and how the result is explained. The failure modes section below lists the specific ones.

Workspace analytics for ChatGPT Enterprise and Edu

Workspace analytics is the usage dashboard for organizations on ChatGPT Enterprise or Edu. It shows adoption and engagement at the workspace level so admins can see who is active, which GPTs and projects get used, and how usage changes over time. The details below come from OpenAI’s workspace analytics documentation.

Where to find it and who can see it

• Open Workspace settings > Workspace analytics, or go to chatgpt.com/admin/usage.
• Workspace owners, workspace admins and members with the analytics viewer role can view it. The viewer role is how you give a team lead or analyst the dashboard without admin rights.

What each section shows

The power users list has a precise definition: the top 20% of message senders who send 75 or more messages a week and use 3 or more different tools a month. Task insights hides SCIM groups with fewer than 10 members, and it is off by default in workspaces with EU data residency.

Exports and data freshness

Under Export data you can generate CSVs for Users, GPTs, Projects and the Impact survey. Custom date ranges can cover up to 12 months. Task insights cannot be exported yet. The data is not real time: it refreshes every 1 to 24 hours, typically within 6 to 12, with a target of up to 48 hours, and the dashboard shows a last-updated timestamp in UTC.

The exports are where workspace analytics becomes useful for a business case. Join the Users CSV with your HR or identity data by email and you can compare adoption by department, role or office, which the dashboard only does if SCIM groups mirror your org chart.

Which plans include which ChatGPT analytics

According to OpenAI’s pricing page, data analysis is limited on Free and included on Go, Plus and Pro, with higher upload and message limits as you move up. The same page lists no analytics dashboard, admin console or Compliance API on any individual plan. Workspace analytics as documented is an Enterprise and Edu feature; when OpenAI announced the improved dashboard, comments from Team (now Business) customers asked for it to be extended to them, so check your own admin settings rather than assuming.

Prices differ by country and change often, so this guide does not quote them. For data analysis, the practical difference between plans is how many files and messages you get before hitting a limit, not a separate analytics add-on.

Can your employer see your ChatGPT activity?

On a work account, partly through analytics and fully through compliance tooling. Workspace analytics is aggregated: it does not show message text, file contents or item-level compliance records. But the Users export is per person. It includes each member’s name, email, message count, rank within the workspace and usage by model, GPT, tool and project. So your admin can see how much you use ChatGPT, not what you wrote, from analytics alone.

Raw conversation logs are a separate product. OpenAI points organizations to the Compliance API for legal, security and retention workflows, and that does give access to content. The safe assumption: anything typed into a company workspace can be reviewed by the company. Keep personal questions on a personal account.

What not to tell ChatGPT

• Customer or employee personal data: names, emails, phone numbers, addresses, ID numbers.
• Payment card details, bank details, passwords, API keys and access tokens.
• Health, legal or HR records, including hiring materials.
• Anything your contracts, data processing agreements or company AI policy exclude.

For analysis you rarely need the identifying columns at all. Drop them, or replace them with a hashed ID, before uploading.

Where the popular guides disagree, and what is true now

Several of the pages that rank for this topic were written for older versions of ChatGPT, and they contradict each other:

The rule of thumb: for anything about features or limits, trust OpenAI’s help center over a third-party tutorial, and check the date at the top of the page you are reading.

Failure modes and how to debug them

When a ChatGPT analysis is wrong, it is almost always one of these. Each has a quick test.

The third meaning: ChatGPT as a traffic source in your analytics

If you run a website, “ChatGPT analytics” can also mean measuring the people ChatGPT sends to you. When someone clicks a link in a ChatGPT answer, the visit usually arrives as a referral from chatgpt.com, and links cited by ChatGPT often carry utm_source=chatgpt.com. In Google Analytics 4 that traffic lands in the Referral channel by default unless you create a custom channel for AI assistants; our AI traffic analytics setup guide walks through the channel group and regex.

Three things make this traffic harder to measure than it looks:

• Missing referrers. Clicks from the mobile and desktop apps, or copied links, can arrive with no referrer and no tag. Those visits are counted as Direct, and no tool can honestly recover them.
• Bots are not visitors. OpenAI also fetches pages with automated agents when it answers questions or builds its search index. Those requests show up in server logs, not as sessions in a JavaScript analytics tool, and they are not people reading your page.
• Sessions are not outcomes. A few hundred ChatGPT visits are worth nothing or a lot depending on what they do next. The number to report is conversions and their value by landing page.

That last step is what our AI conversion tracking guide covers in depth. SEOConversion does it with one script: it attributes conversions and revenue to ChatGPT, Perplexity, Claude, Gemini and Copilot at the landing-page level, and leaves visits with no referrer as Direct rather than guessing.

Worked example: analyzing your ChatGPT traffic with ChatGPT

This example joins the two halves of the topic: you export your ChatGPT referral data from your analytics tool, then use ChatGPT data analysis to decide which pages to work on. All figures are illustrative.

You export 90 days of sessions and lead form submissions by landing page, filtered to source chatgpt.com. The file has four pages plus a “Grand total” row the export added at the bottom. You delete that row, upload the CSV and send this prompt:

I'm deciding which landing pages to improve first for visitors from ChatGPT, based on lead value per session.
Each lead is worth $150. For each landing page, compute conversion rate (leads / sessions), lead value (leads x 150)
and value per session (lead value / sessions). Compute the totals from sums, not averages. Show one table sorted by
value per session, then explain the formula for each column.

ChatGPT returns this table. Here is how each column is computed, so you can check it:

Now run the two-number check:

• Biggest number: total lead value. 20 leads × $150 = $3,000. Matches.
• Small number: the setup guide. 4 / 340 = 1.18%, and $600 / 340 = $1.76 per session. Matches.
• Trap avoided: the average of the four page rates is (7.50 + 7.50 + 1.67 + 1.18) / 4 = 4.46%, which is not the real overall rate. The real rate is 20 / 600 = 3.33%. If ChatGPT had reported 4.46%, you would ask it to recompute from sums.
• Trap avoided: had the Grand total row stayed in the file, ChatGPT would have counted 1,200 sessions and 40 leads, double the truth.

The $150 lead value is the input that makes the whole table meaningful. If you do not have one yet, our guide on how to calculate conversion value shows how to derive it from close rate and average deal size.

Frequently asked questions