Machine learning, without leaving the spreadsheet.

Five of the 38 data tools learn from your rows rather than summarize them: a neural network, logistic regression, a classification tree, k-means clustering and association rules. They train in your file, and the same data gives the same answer twice.

Method guide for Google Sheets Five learning tools free and uncapped

Install free Browse all 379 templates

Machine learning in Google Sheets, five ways

The neural network is a small multi-layer perceptron: one hidden layer with a tanh activation, features normalized for you, and a train and test split so you get a score on rows the model never saw. It predicts a number or a label, and it reports the honest test figure next to the training one, which is how you catch it memorising instead of learning.

Logistic regression is for a yes or no outcome. It fits by IRLS and hands back coefficients, p-values and an odds-ratio column, so instead of a black box you get a sentence like “each extra support ticket raises the odds of churn by half again”.

The classification tree learns a stack of yes or no questions that sorts rows into categories, and then prints the questions. You choose how deep it may go and how many rows a branch must keep, a quarter of your rows are held back and scored separately, and the rule it finds is one you can read out loud and apply by hand.

K-means clustering groups rows that resemble each other, with k-means++ starting points, z-score normalization and five restarts so one unlucky start cannot decide your segments. Seeded, so the clusters you show the room today are the clusters you show them next week.

Association rules is market basket analysis by the Apriori method: one row per basket, one column per item, and it finds what genuinely turns up together. It ranks by lift rather than by confidence, because confidence rewards whatever is simply popular and in a cafe every rule would end in coffee.

Which one to reach for

Four questions, five tools. All five are in the Data Analysis panel, all five are free, and all five write their report into a tab of your own file.

Predict a number

Neural Network a curved relationship you cannot write as a formula. Reports test R² and RMSE on held-out rows.
Regression a straight-line relationship you want to explain rather than only predict. On the statistics page.

Predict a yes or a no

Logistic Regression churn, default, conversion. Odds ratios, p-values, a confusion matrix and an accuracy figure.
Classification Tree the same question when you need the rule itself, in if-then form, rather than a coefficient.
Neural Network the same question when the pattern is not a straight line. Reports held-out accuracy.

Find the groups

K-Means Clustering segments nobody labeled in advance, from the columns you already have.

Find what goes together

Association Rules items that turn up in the same basket, ranked by lift, with support and confidence beside it.

It repeats

Four of the five carry a seed, and the seed is on the panel: the neural network, logistic regression and the classification tree use it to fix which quarter of your rows is held back for scoring, and k-means uses it to fix where the clusters start. Same rows, same seed, same answer, on your machine and on the machine of whoever you send the file to. Association rules needs no seed, because Apriori draws nothing at random: the same baskets give the same rules every time. A result somebody else can reproduce is a result you can defend in a meeting.

The engines themselves are not AI, and we will not call them that. They are arithmetic you can check, and more than 1,500 automated tests check them on every commit against published textbook values and independent implementations. The validation page shows the method.

Every result explains itself

A trained model that only prints weights is a homework exercise. Every run ends with a What this means card: the score that matters, what it implies about your data, and the caveat. On the free plan that sentence is written by the add-on from its own figures, inside your file.

Pro adds an AI reading of the same result, written by Google’s Gemini API from the shape of the result alone, with a See what was sent panel showing the exact payload. No cell values, no column headers, no formulas.

Free or Pro

All five are free, and no plan buys a bigger one. There is no trial clock and no card, and training a network on your whole export costs nothing. They sit under the same one ceiling as the rest of the Data Analysis bench, on both plans alike: a run reads at most 500,000 cells, because the whole block is copied out of the sheet before anything starts. K-means fits at most 50 clusters.

Pro covers the five simulation engines and the AI readings. Pro is $199/year, and a $9 Day Pass covers a week if you have one decision to make.

Worked examples to start from

Small, real datasets you can read end to end before you point any of this at your own export.

The other 6 models in the library that use one of these:

Try it in your own sheet

Open Sortia in Google Sheets and choose Start from a template.
Pick one of the models above, and it loads with the inputs filled in.
Change the assumptions to fit your situation and press Run.

Never used Google Sheets? Start here goes the whole way, in seven steps, and assumes nothing.

Install free All 379 templates

Other methods: Monte Carlo Decision trees Schedule risk Critical chain Optimization under uncertainty Statistics Forecasting Optimization What-if analysis