Skip to main content

AI personalization lab

Design
recommendations
people actually want.

Try out different ways of recommending things to people, and see what each one would do to your catalog — before you build any of it.

Start experimentingExplore the live sample
  • Make up the audiences
  • Same result every time
  • See why each pick was chosen
TasteLab — Prototype personalization before you build it. Featured on Product Hunt.

Strategy control

Seed 424242

Trades a little relevance for materially wider catalog exposure.

Good matches

71%+21

Beyond the hits

73%+61

Variety

91%+12

Catalog reached

72%+50

Discovery Explorer

Taste constellation for the Discovery Balanced strategy, audience Discovery Explorer. Relevance proxy 71%, novelty 73%, diversity 91%, catalog coverage 72%, popularity concentration 0.49. Exposure reached 12 categories. Top recommendations: 1. Estuary Bells by Solveig Rand; 2. Small Hours Percussion by Batid Collective; 3. Interference Pattern by Ilya Bregman.

Luma Audio

Discovery Balanced · Discovery Explorer

  • Estuary Bells#1
  • Small Hours Percussion#2
  • Interference Pattern#3
  • Uphill Both Ways#4
  • Cold Open#5
Open in TasteLab

These are real results from a made-up catalog we ship with the product, so you can try it without uploading anything. Nothing here is invented — the numbers come out of the same engine you would use.

One catalog, three realities

One catalog.
Three completely different realities.

Same catalog, same listeners, same everything. All we change is how the recommendations get chosen — and what people end up seeing becomes a completely different product.

Content catalog

54 items
  • Slow Harbour LightsAmbient
  • Kitchen Table TapeLo-fi
  • Quiet MachineryModular
  • Weekday DevotionNeo-Soul
  • Harbour at FourField Recording
  • CounterweightJazz Fusion
  • Neon ArithmeticSynth Pop
  • Vespers for a WeekdayChoral
  • Third BellOrchestral
  • The Long Way RoundSpoken Word
  • NorthlineDowntempo
  • TessellateSynth Pop
  • Brass for an Empty HallJazz Fusion
  • Practice Room, 6amOrchestral
  • Housing CooperativeLo-fi
  • Beginner MindSpoken Word
  • Low Tide ChoirChoral
  • Repeat CustomerSynth Pop

Strategy

Popularity First

The control arm for every comparison. It ignores individual preference almost entirely, so it shows what your catalog looks like when nothing is personalised. Expect strong apparent relevance and very poor catalog coverage.

What this strategy pays attention to

  • Their taste 5%
  • What is popular 85%
  • What is new 10%

Good matches

49%

Beyond the hits

12%

Variety

79%

Catalog reached

22%

Recommendation stream

Discovery Explorer
  1. 1Thirty Six HoursMarguerite Okoye · Neo-Soul · 4:01AFFNOV
  2. 2Meridian BlueAshen Vale · Downtempo · 5:05AFFNOV
  3. 3Everything In ThreesKestrel Ono · Synth Pop · 3:18AFFNOV
  4. 4Slow Harbour LightsVera Ostlund · Ambient · 6:52AFFNOV
  5. 5Neon ArithmeticKestrel Ono · Synth Pop · 3:31AFFNOV
  6. 6Paper RadioNils Aberdeen · Lo-fi · 2:56AFFNOV
  7. 7Weekday DevotionMarguerite Okoye · Neo-Soul · 4:28AFFNOV
  8. 8NorthlineAshen Vale · Downtempo · 5:22AFFNOV

Only 12 of your 54 items ever get shown to anyone. The rest may as well not exist.

These are real results from a made-up catalog that ships with TasteLab, so you can try everything without uploading anything. They show how a strategy behaves, not what your actual users will do.

Filter-bubble chamber

See the bubble before
your users feel it.

A filter bubble is when the same small handful of things gets shown to everyone, over and over. Drag the handle: on the left is what happens when you just rank by popularity, on the right when you deliberately spread things out.

Discovery Explorer

Taste constellation for the Discovery Balanced strategy, audience Discovery Explorer. Relevance proxy 71%, novelty 73%, diversity 91%, catalog coverage 72%, popularity concentration 0.49. Exposure reached 12 categories. Top recommendations: 1. Estuary Bells by Solveig Rand; 2. Small Hours Percussion by Batid Collective; 3. Interference Pattern by Ilya Bregman.

Concentrated

12 of 54 items ever shown

Everything piles onto 6 categories and 6 creators. Everyone ends up seeing more or less the same things.

Balanced

39 of 54 items reached

Spread across 12 categories and 12 creators, and the matches are just as good.

How much of your catalog is seen

22%72%

More is better

Seeing the same thing twice

60%0%

Less is better

Creators who get any attention

50%100%

More is better

Works for brand-new users

32%53%

More is better

These are real results from a made-up catalog that ships with TasteLab, so you can try everything without uploading anything. They show how a strategy behaves, not what your actual users will do.

Synthetic audience studio

Model the users you
do not have data for yet.

Describe the kinds of listener you care about — someone who wants new things, someone who sticks to favourites, someone who just signed up. These are made-up profiles you invent to test against. They are not real people and not market research.

Selected segment

Discovery Explorer

Actively seeks novelty, moves between genres in a single session and responds well to emerging creators. Punishes repetition faster than any other segment here.

Prefers ModularPrefers Field RecordingPrefers Jazz Fusion

What this person is like

We know what they have liked before
Open to trying new things
85
Likes the unfamiliar
88
Sticks to what they know
12
Wants variety
82
Listens deeply
72
Fine seeing repeats
18

These are real results from a made-up catalog that ships with TasteLab, so you can try everything without uploading anything. They show how a strategy behaves, not what your actual users will do.

The trade-off mixer

Tune relevance, discovery
and diversity deliberately.

Seven sliders, one live simulation. Move any of them and the whole thing runs again right here in your browser — the map and the numbers change because the recommendations genuinely got recalculated, not because something faded between two pictures.

The controls

Simulating
35

How much we care that it suits this person.

10

Reward for going past the obvious hits.

20

Avoid showing several similar things in a row.

20

How much we lean on what everyone likes.

10

How much we lean on what was just released.

10

Mix in some deliberate wildcards.

30

How hard we try not to show the same thing twice.

Only the balance between the sliders matters, not the absolute numbers — turning everything up is the same as leaving everything alone.

Discovery Explorer

Taste constellation for the Hybrid strategy, audience Discovery Explorer. Relevance proxy 65%, novelty 47%, diversity 92%, catalog coverage 74%, popularity concentration 0.51. Exposure reached 12 categories. Top recommendations: 1. Static Bloom by Ilya Bregman; 2. Bright Interval by The Halcyon Unit; 3. Estuary Bells by Solveig Rand.

Good matches

65%

Beyond the hits

47%

Variety

92%

Catalog reached

74%

Crowding

0.51

Happy surprises

35%

What they would see

  1. 1Static BloomModular
  2. 2Bright IntervalJazz Fusion
  3. 3Estuary BellsField Recording
  4. 4Small Hours PercussionPercussion
  5. 5Two RiversOrchestral
  6. 6Thirty Six HoursNeo-Soul

These are real results from a made-up catalog that ships with TasteLab, so you can try everything without uploading anything. They show how a strategy behaves, not what your actual users will do.

Strategy comparison

Compare trade-offs,
not just scores.

The strategy that shows people the most fitting things can quietly mean most of your catalog is never seen by anyone. Here is what each one gives up.

All three, side by side

Good matches

higher is better

49%
65%
71%

Beyond the hits

higher is better

12%
47%
73%

Variety

higher is better

79%
92%
91%

Catalog reached

higher is better

22%
74%
72%

Crowding

lower is better

0.85
0.51
0.49

Happy surprises

higher is better

4%
35%
63%

Recommendation overlap

  • Popularity First ∩ Hybrid25%
  • Hybrid ∩ Discovery Balanced25%
  • Popularity First ∩ Discovery5%

How much of the catalog gets seen

  • Popularity First12/54 items
  • Hybrid40/54 items
  • Discovery Balanced39/54 items

Unique discoveries · Discovery Balanced

Items no other strategy surfaces
  • Interference PatternIlya Bregman · popularity 14
  • Uphill Both WaysThe Halcyon Unit · popularity 15
  • Cold OpenPriya Halloran · popularity 13
  • Second Person PluralChoir of Small Hours · popularity 12
  • Repeat CustomerKestrel Ono · popularity 57
  • Held Note for a Bright RoomVera Ostlund · popularity 21

Primary risk · Discovery Balanced: it trades 5 points of match quality for a lot more of your catalog being seen.

Recommended for this catalog

Discovery Balanced

Discovery Balanced

People still get good matches — 71% of the time, against 65% for Hybrid — but 72% of your catalog gets seen instead of a small slice of it, and attention stops piling onto the same few items.

We work this out from the numbers above, not from an opinion: among the runs whose matches are nearly as good as the best one, this is the one that gets the most of your catalog seen. The AI can explain a result like this, but it can never change a number.

Open the full comparison

These are real results from a made-up catalog that ships with TasteLab, so you can try everything without uploading anything. They show how a strategy behaves, not what your actual users will do.

Experiment brief composer

Turn a personalization hypothesis
into an experiment your team can ship.

Every completed simulation composes into an experiment plan, a product brief and an engineering outline — each one written against the metrics that run actually produced.

Discovery ranker · experiment brief

Luma Audio · seed 424242 · generated from a completed simulation

AI narrative · cited metrics
Hypothesis
Switching from Hybrid to Discovery Balanced will get about -2% more of our catalog in front of people, without people getting noticeably worse matches.
Audience
Discovery Explorer segment, plus a cold-start holdout for guardrails.
Control strategy
Hybrid — Blends affinity, behavioural association and popularity.
Treatment strategy
Discovery Balanced — Trades a little relevance for materially wider catalog exposure.
Primary metric
How many different items each person sees in a week.
Guardrail metrics
Things that must not get worse: how good the matches feel, how often people see repeats, whether they finish sessions, and how brand-new users get on.
Event instrumentation
Impression events with slot position and above/below-fold flag; click, save, skip and complete with item id and session id; a strategy-variant field on every event.
Rollout approach
Shadow-score both rankers for one week, then a 5% holdout, then 50/50 once guardrails hold for five consecutive days.
Success threshold
Agree the threshold before launch from your own baseline variance. TasteLab cannot supply it: these are simulation outputs, not measured effects.
Risks
In the simulation, attention spreads out more evenly. But a real catalog has licensing, availability and editorial rules this model knows nothing about, and brand-new users should be checked separately.
Recommended next action
Run the same comparison against your own catalog before designing the online test. If coverage does not move on real data, the ranker is not the constraint — retrieval is.

The implementation outline is framed as a starting point requiring validation against your own data — never as an automatically correct production architecture. Exports to Markdown, JSON, CSV and a printer-friendly page.

Instrumentation checklist

  • Impression events emitted for every rendered slot, not only clicks
  • Strategy variant recorded on every event in the funnel
  • Item and creator ids stable across the catalog pipeline
  • Cold-start users identifiable at request time
  • Exposure counts queryable per item and per creator
  • Guardrail dashboard live before the first ramp
  • Rollback path that does not require a deploy

Supporting insight

We reach -2% more of the catalog, and people barely notice a difference in how well things match.

That is the decision, as a number rather than an argument. Whether it is worth making depends on how much of your catalog currently earns nothing — TasteLab can show you that, and a real test with real users confirms it.

Catalog reached
74%72%
Good matches
65%71%

These are real results from a made-up catalog that ships with TasteLab, so you can try everything without uploading anything. They show how a strategy behaves, not what your actual users will do.

Start here

Prototype the experience
before building the infrastructure.

Explore how different recommendation strategies shape discovery, relevance and catalog exposure using a reproducible simulation.

Open the sample labCreate a project

These results come from a simulation, not from real people. They show how a strategy behaves on this catalog, which is useful for comparing options — but they are not a prediction of what your actual users will do. Try anything promising with real users before you rely on it.