A .sav file holds the survey answers and also what makes them readable: the wording of each question and the label of each value. To ask Claude for crosstabs from that file you need a piece that reads all of it and does the arithmetic, because a language model writes well but is not a calculator. This walkthrough uses the Survey Coder connector, which we make: Claude asks, a tabulation engine computes, and you check the bases.
What you need
- A Claude account that can use connectors.
- A Survey Coder account. Creating one is free and the connector costs nothing.
- Your .sav file. You do not need SPSS installed, and you write no syntax.
Step 1. Connect Survey Coder to Claude
In Claude, open Settings → Connectors, find Survey Coder in the directory and click Connect. You sign in with your Survey Coder account and authorize access; Claude never sees your password. Details are on the connector page.
Step 2. Upload the .sav with the link Claude gives you
Type something like “I want crosstabs from a .sav file”. Claude returns a link to upload the file in the portal. This is the step that trips most people the first time: a file attached to the chat does not reach Survey Coder, it has to go through that link.
When the upload is done, go back to the conversation and say so. Claude checks the file and shows a summary of what it will import; nothing is imported until you confirm. If the study has no open-ended questions to code, say “tables only”: on that path nothing is coded and no coding credits are spent.
Step 3. Start with the totals
Before crossing anything, look at the main questions in total. You can ask for several in one request, up to twelve:
- “Show me Q12, Q13 and Q17 in total, with their bases.”
If you do not remember the variable names, ask “which questions does the study have?” and Claude lists the variables with their labels.
Step 4. Ask for the crosstabs
Now the same questions by the cuts you care about. Each request takes up to four banners:
- “The same three questions by gender, age and city.”
The answer comes back as a table: one column per segment, the base of each column and, on category rows, letters that mark which differences are statistically significant. An example with made-up brands and figures:
| Brand bought most often | Total | Monterrey (A) | Guadalajara (B) | Mérida (C) |
|---|---|---|---|---|
| Brand A | 41% | 58% BC | 39% | 31% |
| Brand B | 26% | 24% | 36% AC | 17% |
| Brand C | 33% | 18% | 25% | 52% AB |
| Base | 700 | 180 | 260 | 260 |
Read it this way: in Monterrey, Brand A (58%) is significantly higher than in Guadalajara and Mérida. Between Guadalajara (39%) and Mérida (31%) the difference does not reach significance, which is why neither carries the other's letter. For a refresher on reading a table like this, see the guide to cross tabulation analysis.
Step 5. Cut by the answer that matters
Demographics are the starting point. The cuts that usually explain a study are different: people who would buy against those who would not, promoters against everyone else, people who chose a brand against those who did not. You ask for them the same way, in one sentence:
- “Compare promoters (9 and 10) with everyone else on Q17 and Q22.”
- “I want to see people who bought Brand A against those who did not, on the Q22 attributes.”
- “Build a group of people who bought the brand but would not recommend it, and compare it with the rest.”
Claude builds those groups for the request and does not save them to the project, so you can try several cuts and leave nothing to clean up. If one of the questions is open-ended and already coded, you can cross its codes by these groups too.
Step 6. Ask for the dashboard or the deck
With the numbers in the conversation, the rest is the assistant's job: “with these results build a dashboard with charts and the three main findings”, or “prepare five slides for the client”. Ask it to quote the base next to every percentage and to point out low bases; the connector already hands it that information.
What to check before you deliver
- The base of every percentage. A filtered question has a smaller base than the full sample. Confirm the base matches the people who were asked the question.
- Low bases. Columns with few cases come flagged. 60% of 25 people is not a finding.
- Top 2 Box and mean. Survey Coder only computes them when it can identify the direction of the scale, from its labels or because you declared it. When it cannot, it returns the full distribution instead of assuming the highest number is the best.
- Significance. The letters appear on category rows and on battery summaries. On a single question, the Top 2 Box and mean rows come without letters, and NPS carries no test.
- The conclusions. The numbers come from the engine; the interpretation is the assistant's and yours. Read the tables before you accept a headline.
What it does not do
- It does not read files attached to the chat: the .sav goes through the upload link.
- This tables path is for .sav files.
- Each request takes up to four banners and twelve questions. For more, make another request.
- Requests do not replace the deliverable: for the full table book in Excel, ask Claude to generate it or do it in the portal.
What about ChatGPT?
Survey Coder also has an app in ChatGPT. We verified this walkthrough in Claude; what you can do from ChatGPT today is described on the connector page.
Frequently asked questions
Can I analyze an SPSS file with Claude without owning SPSS?
Yes. Survey Coder reads the .sav with its variable and value labels; you need no SPSS license or installation, and you write no syntax.
Does Claude compute the percentages?
No. Survey Coder's tabulation engine computes them: percentages, bases and significance tests between columns. Claude receives those numbers and explains them. The charts, decks and conclusions are built by the assistant from them.
How much does it cost?
The connector is free and, during the open beta, tables cost nothing. Coding open-ended questions costs one credit per response, and new accounts start with 250 free credits.
What about data privacy?
The file stays in your Survey Coder project, under your organization's permissions. The results Claude receives in order to answer you are processed by Anthropic under your Claude account's settings. Details are on the privacy page.
Next steps
Connect Survey Coder from the connector page. If you only need a quick crosstab with no account, try the free crosstab tool. And to see what is worth leaving to AI and what is not, read AI survey analysis.