What Google's Learn Your Way gets right that most "AI personalisation" features don't
Tom W Dixon · Senior Product Manager and Digital Platform Lead
Most products that claim to "personalise with AI" mean a recommendation carousel or a chatbot bolted onto a help centre. Google's Learn Your Way is a rarer thing: a generative AI feature built around a specific pedagogical theory, then tested against a control group before anyone called it a success. It turns a static textbook chapter or PDF into mind maps, narrated slides, audio lessons and adaptive quizzes, tailored to the learner's grade level and interests. What makes it worth a product manager's attention isn't the format-shifting. It's the discipline behind the claim.
The feature is not the interesting part
Reformatting content into different modalities is not new. Plenty of tools will turn a document into a slide deck or a podcast. What's different here is that Google Research ran an actual efficacy study before shipping the pitch. Students using Learn Your Way scored 11 percentage points higher on a long-term recall test than students using a standard digital reader, drawn from ten OpenStax source materials spanning history to physics.
That's a specific, falsifiable number. Not "students loved it" or "engagement went up." A measured outcome against a defined control, on the thing the product is actually meant to improve: retention.
Most roadmaps skip the control group
I've sat in enough roadmap reviews to know how "personalisation" usually gets justified. Someone shows a demo, the demo looks good, the feature ships, and success gets measured by usage of the feature itself rather than the outcome it was supposed to drive. Usage is not evidence. It tells you people clicked, not that anything got better.
Learn Your Way's team also tracked self-reported comfort: 100% of students felt more comfortable being assessed after using the tool, against 70% for the control group. And 93% wanted to keep using it, against 67%. Those are secondary signals, correctly labelled as secondary. The primary claim rests on the recall number, not on the sentiment.
If you're building anything that claims to adapt to a user (content, pricing, onboarding, anything), the question worth stealing from this is simple: what's the control group, and what's the one outcome metric that would prove the adaptation actually worked?
Grounding matters more than the model
Learn Your Way runs on LearnLM, a family of models Google built specifically for pedagogy and folded into Gemini 2.5 Pro, rather than a general-purpose model prompted to "act like a tutor." That's a deliberate constraint, not a limitation. A general model optimised for plausibility will happily generate a confident explanation that's pedagogically wrong. A model grounded in learning science principles, tested against a specific outcome, is a narrower tool built to do one job well.
The lesson transfers past education products. The more specific the outcome you're trying to drive, the less a generic "AI-powered" layer earns its place, and the more it's worth asking whether the underlying model or process was actually built for that outcome, or just pointed at it.
Coverage in Forbes framed this as Google reinventing the textbook. That's the headline version. The product management version is smaller and more useful: ship the feature, but ship the evidence with it. If your team can't say what the control group was, you don't have a personalisation feature. You have a guess with a UI.