Data Storytelling Teaches You to Present a Finding. Nobody Teaches You to Find One.

Every guide to data storytelling teaches you the same five things. Know your audience. Pick the right chart. Remove the clutter. Guide the eye. Wrap it in a narrative.

All of that is correct, and all of it assumes something that is almost never true: that you have already found the story.

You have not. You have a number. Finding the story inside it is a completely different skill, it is much harder, and almost nobody teaches it.

Here are the four questions that do the work, and a real trial where every single write-up missed the story sitting in the document.

The trial where everyone missed it

In late 2025, Google ran a properly rigorous randomised controlled trial of an AI tutor in Sierra Leonean classrooms. Preregistered. 1,763 students. A real control group. Outcomes measured by an independent contractor, blind to who got the treatment.

Maths scores rose. The intent-to-treat effect was +0.258 standard deviations, which is a good result, and I want to be clear that the study is better evidence than almost anything else in this field.

Every write-up led with that number.

Now read further into Google's own technical report. Disclosed in the results: students who entered with stronger maths skills benefited more. For each additional standard deviation of baseline ability, the treatment effect grew.

The tool worked. And it worked disproportionately for the students who were already ahead.

For a development intervention, that is not a footnote. An AI tutor that amplifies existing advantage is a fundamentally different policy object from one that closes gaps, and a ministry with a constrained budget needs to know which one it is being sold.

Google published it. The coverage dropped it.

The average was the press release. The distribution was the story.

Question 1: Who did it work worst for?

This is the most powerful question in analysis and it takes four seconds to ask.

An average is a compression. It throws away the thing you most need to know, which is who this actually helped and who it did not. Any intervention, any model, any feature can produce a healthy average while doing nothing at all for a third of the people it was built for.

The moment somebody quotes you an average effect, ask them for the heterogeneity. If they cannot tell you who it worked worst for, they have not measured it, and they are selling you a number rather than a finding.

Question 2: What is the confidence interval?

"It worked" and "it worked, but the lower bound is nearly zero" are not the same sentence, and only one of them is honest.

In that same Sierra Leone trial, the effect was +0.258 SD with a 95% confidence interval running from 0.027 to 0.488. The result is statistically significant and I am not disputing it. But the lower bound is close to nothing at all. The honest reading is "somewhere between barely detectable and quite large, most likely moderate."

That is a different story from the one the headline tells, and it is the one you would want if you were spending public money.

Question 3: What is the baseline?

Beating nothing is not an achievement. The only meaningful question is whether you beat the dumbest thing that could possibly have worked.

I learned this the hard way, in public.

I built the first machine learning model for crop yield prediction in Sierra Leone, my home country. Twenty five years of national crop statistics. XGBoost, gradient boosting, random forests, with a strict anti-leakage protocol and a proper walk-forward evaluation.

And it lost. No model I trained on the crop statistics could beat naive persistence, the trivial baseline of simply assuming next year's yield looks like last year's.

I could have tuned until the numbers flattered me. I could have quietly not published. Instead I reported it, because the failure was the finding: the data the country currently collects does not contain enough signal to forecast the harvest. That is not a modelling problem, and no amount of modelling fixes it. It is a data problem, and a policymaker needs to know which of the two they have.

Adding free satellite rainfall data then cut the error by a third. The failure is what pointed at the fix.

If a paper or a vendor does not tell you what baseline they beat, assume there is a reason.

Question 4: What is in the appendix that is not in the summary?

Read the appendix. Read the footnotes. Read the limitations section that everyone skips.

That is where the findings live that the authors were obliged to disclose but would prefer you did not dwell on. It is not usually dishonesty. It is gravity. The headline goes at the top and the complications sink.

In the Sierra Leone trial, the report also disclosed that the model changed in the middle of the study. Students used one version of Gemini for the first six weeks and a newer one for the final three. Google said so, to their credit. But it means the intervention was not a fixed object, and there is no clean way to attribute the result to either version.

That is a real limitation, it was disclosed, and it was almost entirely absent from the coverage.

This matters more than your chart choice

We spend enormous energy arguing about bar charts and pie charts. That argument is not unimportant, but it is downstream of something far more consequential.

A beautiful chart of the wrong number is worse than an ugly chart of the right one. It is more persuasive, and it is wrong, which is the most dangerous combination available to an analyst.

Finding the story is the part of this job that cannot be automated, cannot be outsourced, and does not appear in any tutorial. It is judgment. It is asking the question nobody in the room wants to ask, about the number nobody put on the slide.

Start with the four questions. They will make you unpopular occasionally, and correct much more often.

Key takeaways

1. The story is rarely the headline number. It is usually the one underneath it.
2. Who did it work worst for? The single highest-value question in analysis.
3. Ask for the confidence interval. A point estimate is a summary of a range.
4. Ask what baseline they beat. If nobody names one, there is a reason.
5. Read the appendix. The complications sink, and that is where they land.
6. Volunteer the number that weakens your case. It is the fastest way to be trusted with the rest.

Which of these four have you actually asked in a meeting? And what happened when you did? 👇

---

The four questions above are Part 1 of my free guide, Data Storytelling for People Who Might Be Wrong, which also covers how to present a result when the data disagrees with you. Download it free.

Sources: Google DeepMind and Fab AI, technical report on Guided Learning in Sierra Leone. My crop yield model and its null result: arXiv

.13959, code on GitHub.