NotebookLM Data Tables: I Fed It a Messy Spreadsheet

NotebookLM's new Data Tables feature is sold on one clean line: turn your messy data into tidy tables. So I grabbed the messiest spreadsheet I'd been avoiding for months, the one with merged cells and three header rows and a column somebody named "misc," and tried to feed it in. NotebookLM wouldn't take it. Not because the data was too messy for the AI, but because it was an Excel file, and NotebookLM won't open the door for a raw .xlsx at all.

That's the thing no promo post tells you. Data Tables is real, it works, and the wall you hit first has nothing to do with how good Gemini 3 is. It has to do with what NotebookLM will and won't let through the front door. If you've got a genuine mess sitting in Excel and you're hoping to point this feature at it today, here's what actually happens and how to get around it in one sitting.

What Data Tables actually is (and where it came from)

Data Tables is NotebookLM's newest Studio output, announced in December 2025 alongside its switch to the Gemini 3 model. You ask, in plain language in the Studio panel on the right, for a table with the rows and columns you want. NotebookLM then scans the sources already sitting in your notebook, all of them at once, and fills the cells from your own material. When it's done you can export the whole thing to Google Sheets.

Read that again, because it's the part that trips people up: it builds the table from sources already in your notebook. Not from a data file you hand it in the moment. The feature is a synthesizer of stuff you've already loaded, and "stuff you've already loaded" is exactly where a messy Excel file gets stuck.

What I wanted to do (and why I picked the ugly file)

I wanted the one thing the marketing promised: point it at a real mess and get order back. My test file was a spreadsheet of small experiments I'd tracked over a year, tool names spelled three different ways, dates in two formats, notes crammed into a single "misc" column. Version one of that sheet was terrible, and it stayed terrible, because I'm the kind of person who tracks things and never cleans them up.

That's the honest use case. Nobody needs an AI to tabulate data that's already tidy. The whole pitch of "messy data into tables" only means something if you can throw the actual mess at it. So I did.

Where it went wrong: the file wouldn't go in

The wall isn't the AI, it's the upload. NotebookLM flat out does not accept a .xlsx file as a source. Not "it struggles with it." It won't take it. The formats it does read are Google Sheets through Drive, PDF, DOCX, plain text, Markdown, and a few others, but the native Excel format is not on that list.

So I did the obvious next thing and exported to CSV. Here's the weird part: the .csv extension itself can get rejected, even though the comma-separated content is exactly what NotebookLM should love. Other people have hit the same snag and written it up under titles like "why does NotebookLM hate CSV." The data is fine. The three letters after the dot are the problem.

The workaround that actually got my data in

The fix is to stop handing NotebookLM a spreadsheet file and start handing it something it recognizes as a source. Two routes worked, and both take minutes once you know them.

The clean route is Google Sheets. Import your Excel file into a blank Google Sheet, then add that Sheet to your notebook through Drive. NotebookLM treats Sheets as a first-class source, so this sidesteps the whole .xlsx wall.

The scrappy route is text. Export the spreadsheet to CSV, and if the .csv upload gets refused, rename the file to .txt or paste the rows straight into a text source. NotebookLM's parser reads the comma-separated content happily once the extension stops scaring it off.

Make a table with one row per experiment and columns for tool name, date, what I tried, and the outcome. Pull every value from my sources, standardize the tool names to one spelling each, and put dates in YYYY-MM-DD. Leave a cell blank if the source doesn't say.

That last instruction matters. Telling it to leave a cell blank when the source is silent is the difference between a table you trust and a table that quietly guesses. It's the same discipline that keeps Notion AI useful when you point it at your own scattered notes: the tool is only as honest as the constraints you give it.

So what about the table itself?

Once your data is a real source, asking for the table is the easy part Google actually delivers on. The natural-language request in the Studio panel produces a structured table, and the export-to-Sheets button is right there in the overflow menu next to the output. That much is documented and consistent across every account.

One caveat I won't paper over. At launch in December 2025, Data Tables was available only on Google AI Pro and Ultra, plus most paid Workspace editions, and the free tier was promised in the coming weeks. Whether free accounts have it right now, in July 2026, I can't confirm as a hard fact, so do the five-second check yourself: open a notebook, look at the Studio panel, and see if Data Tables is listed before you build a plan around it. Google also hasn't said which Gemini 3 variant is under the hood, and the sort-and-filter controls the tech press mentioned are described as coming, not shipped.

What I will not do is invent a cell-by-cell accuracy score for you. There's no trustworthy community data yet on how faithfully it reads genuinely messy input, and making up a number would be worse than saying so. The checkable, reproducible finding is the one I hit at the door: the mess it can't handle is your file format, not your data. This is the same reason I put Sora 2 through a real test instead of trusting the sizzle reel. The demo is always clean. Your Tuesday isn't.

Would I do it this way again?

Yes, but I'd convert to a Google Sheet before I even opened NotebookLM. Knowing the input rule up front turns a frustrating afternoon into a two-minute setup. Data Tables is genuinely useful once your material is a valid source, and building the table out of files you've already loaded, rather than a fresh upload, is a reasonable design once you understand it.

The gap is honesty in the marketing. "Turn your messy data into tables" reads like you can drop your worst spreadsheet in and walk away. The truth is you clean up the container first, then it does the clever part. That's still worth it. I just wish the promise matched the doorway, the way it does with every tool that quietly makes you tidy the input before the magic.