Is generative personalization the same as a recommender system?

Not quite. A recommender selects items from a catalog, while generative personalization can produce the artifact itself. The two overlap, because a large language model can also act as a recommender, including generating item identifiers directly rather than scoring a candidate list.

Does generative personalization mean a model writes the final content each person sees?

Sometimes, but often not. In the most common pattern today the model generates candidate variants and a bandit or ranker still selects which one to show, so the personalization step is selection. In other patterns the model does produce the answer or the interface directly.

Why is generative personalization harder to measure?

Because every person can see a unique artifact, there is no fixed set of variants to compare. You measure the policy that selects or generates, using randomized tests and a holdout, rather than the individual artifact.

What is generative personalization? | The Personalization Guide

Generative personalization is personalization that produces the tailored artifact with a model, rather than selecting it from a set of options prepared in advance. Earlier systems chose among a fixed inventory: which product to show, which article to surface, which subject line to send. A generative system can instead write the product description, compose the email, or assemble the interface for one person at the moment of the request.

The shift is a real one, but it is easy to overstate. Much of what is sold as generative personalization is still selection underneath, with generation only widening the set of things to choose from. The useful move is to separate the patterns.

Three patterns, easy to confuse

Large language models enter personalization in three distinct ways, and conflating them hides what a system actually does.

Generation feeding selection

The model generates candidate variants, and a bandit or ranker then chooses which one to show. Generation widens the set of options, but the personalization step is still selection. This is the most common and the most measurable pattern today, because the thing being chosen among is concrete and the choice can be tested.

The model as a recommender

The model acts as the recommender itself. This includes generative recommendation, where a model produces item identifiers directly rather than scoring a list of candidates. Industrial systems of this kind exist, but the reported results come from the companies that built them, so read them as vendor claims rather than settled findings.

Generating the experience itself

The model produces the artifact a person sees: an assistant’s answer, a bespoke explanation, or an interface assembled on the fly. Here personalization and content production merge, and there is no pre-authored variant to point back to.

Why generation moves the bottleneck

When any person can be shown a unique artifact, the hard part is no longer writing the content. Writing becomes cheap and close to unlimited. The hard parts become selecting and measuring what was produced, keeping it on brand and safe, and paying for model inference at the latency a live page or session requires. A system that can generate a thousand variants still has to know which one helped, and that is a measurement question, answered with randomized tests and a persistent holdout rather than with the volume of output. How personalization is measured covers the methods, and why most reported lift is not trustworthy.

Generation also sharpens older concerns. A fixed inventory of variants is auditable: you can list what a person could have seen. A generative system has no such list, which makes it harder to know what was shown and why, and raises the question of whether the system is optimizing to persuade a particular individual. Those are the questions the critics page returns to.

Generative personalization and generative UI

Where the generated artifact is the interface itself, laid out per visit rather than filled into a fixed template, generative personalization overlaps with generative UI. This guide covers what is selected or generated for whom and why; the sibling Generative UI Guide covers how interfaces are composed at runtime. They are two cuts through the same shift: one is about the decision, the other about the surface it renders on.

The practical test for any generative claim is which pattern it names. A system that generates subject lines a bandit then picks among is doing something real and measurable. A system that claims to generate a bespoke experience for every person is making a larger claim, and the first thing to ask is where the record of what was shown lives, and how the effect was measured.