01
Everyday sneakers
A little everyday comfort
THE LOODOS JOURNALFIELD NOTES № 001 — INTELLIGENT DISCOVERY2026
A little more
personal.
+
The art & science
of understanding
what matters now.
00 / THE STARTING POINT
between having a choice
and finding your own.
We have become very good
at showing people more.
What about showing them
what matters?
between having a choice
and finding your own.
01Same shelves.
ONE SIZE FITS NOBODY
Same shelves.
Different stories.
Every day, millions of people open the same apps, browse the same stores and encounter the same interfaces. But they arrive with entirely different intentions. One needs an everyday essential. Another is getting ready to go somewhere.
Traditional segmentation gives us a useful starting point: people with something in common. But a group is not a moment. A commuter can become a traveler by Friday. A useful experience needs room for that change.
LITTLE EXPERIMENT 01Same catalog. Different possibilities.
ON THE SHELF FOR ALEX
One order for everyoneA little bit of everything.
02
Weekend bag
Room for a new adventure
03
Quiet headphones
Your own little world
04
All-weather jacket
Ready for a change of plans
Change the visitor. Notice anything? The shelf stays exactly the same. Now try the contextual experience.
↳Illustrative ranking using fictional intent-match scores, not real customer data or AI inference.
A segment describes a group.
Intent describes a moment.
A few visitors are easy to understand. What happens when the little store becomes a very big one?
02A thousand products.
THE PROBLEM IS MATHEMATICAL
A thousand products.
A million possibilities.
Add people. Add products. The number of possible matches grows much faster than the shelves. And every match sits inside a changing world: new inventory, fresh behavior, different needs.
This is why personalization becomes an engineering problem. Maintaining another rule is easy. Keeping thousands of rules useful, consistent and current is a different task entirely.
02 / THE SCALE EXPLORERA few more people.
A few more people.
A lot more possibilities.
100,000
possible person–product pairs
1,000 people × 100 products↳A simplified view of candidate-pair scale. Real systems also account for sequence, context, constraints and changing inventory. Marks are aggregated, not one dot per person.
We cannot inspect every possibility by hand. But we can pay closer attention to the signals that matter.
03One click is a clue.
READING THE MOMENT
One click is a clue.
A sequence tells a story.
A previous purchase says something about the past. A search, a comparison and a product added to a basket offer clues about what someone needs now. Behavioral modeling brings those signals together.
Representations such as embeddings help express patterns in a form a model can use. Session intent inference then estimates likely needs from the available evidence. It is a prediction under uncertainty, never a direct reading of someone’s mind.
03 / A SESSION IN MOTION
Leave a few little clues.
Try a sequence of actions. The pattern matters more than a single click.
Your next click starts the story.
ILLUSTRATIVE INTENT MIX
No signals yet. Each intent begins with equal weight.
THE NEXT POSSIBILITIES
Everyday sneakers
Quiet headphones
Weekend bag
↳Deterministic scores over the most recent 20 actions. A production model would infer from learned representations; updating a prediction is different from retraining a model.
THE ENGINEERING DETAIL
Fresh signals need a fast path to the decision. Event quality, useful representations and low-latency inference all matter. A more complicated model is not automatically a better experience.
04The world changes.
FROM RULES TO LEARNING
The world changes.
Can the decision keep up?
Rules are valuable when a condition and its consequence are clear. “Do not recommend an unavailable item” is an excellent rule. But a growing list of if–then branches struggles to express the combinations that make a product relevant.
A predictive model learns relationships from examples. It can weigh several signals together, while explicit rules still enforce eligibility and business constraints. The practical system often needs both.
04 / TWO WAYS TO DECIDE
New session signal
Travel segment?
Everyday sneakers
Rule: if the visitor is not in the travel segment, show everyday sneakers.
↳The “model” is a deterministic teaching approximation. Real predictive systems use trained parameters; fresh signals change inference, while learning requires a separate online or periodic update process.
A rule says what to do.
A model estimates what may matter.
05“Best” depends on
THE ART OF CHOOSING
“Best” depends on
what you’re optimizing.
A product can be relevant, easy to buy, profitable or pleasantly unexpected. Those are different qualities. Maximizing one does not guarantee the others will follow.
Multi-objective ranking makes the trade-offs explicit. A weighted score is one way to combine priorities. Constraints, calibration and diversity across the whole recommendation list make production systems more nuanced.
05 / THE RECOMMENDATION WORKSHOP
What matters most?
Move the priorities. Make a different kind of “best.”
40.0% of the normalized mix
30.0% of the normalized mix
15.0% of the normalized mix
15.0% of the normalized mix
01
Quiet headphones
75.302
Everyday sneakers
72.503
Weekend bag
71.704
All-weather jacket
67.705
Pocket power bank
65.9Relevance matters. But relevance alone doesn’t tell the whole story.
Peek inside the scoring model
Score = Σ (normalized weight × synthetic product score). Weights sum to 100%; if all are zero, each gets 25%. Scores are fictional and range from 0 to 100. “Surprise” is an item-level proxy here; true list diversity also considers relationships between selected products.
| Product | Relevance | Conversion | Margin | Surprise |
|---|---|---|---|---|
| Everyday sneakers | 95 | 85 | 35 | 25 |
| Weekend bag | 76 | 70 | 90 | 45 |
| Quiet headphones | 85 | 90 | 65 | 30 |
| All-weather jacket | 68 | 60 | 55 | 95 |
| Pocket power bank | 65 | 78 | 45 | 65 |
↳Educational simulation with normalized weights and fictional product scores. These are not model predictions or calibrated purchase probabilities.
Even a carefully balanced score needs one more thing: a sense of what is happening around the customer.
06The right thing.
CONTEXT IS EVERYTHING
The right thing.
At the right little moment.
The perfect travel bag is less helpful when the destination forecast turns wet. A promotion is less useful when the item is unavailable nearby. A plan upgrade makes little sense when the current plan already fits.
Contextual enrichment brings these circumstances into the decision. It can adjust ranking, remove unavailable options or reveal that the most useful recommendation is no purchase at all.
06 / CHANGE THE WORLD
A DIFFERENT PLACE
Around the corner.
RELEVANT RIGHT NOW
Everyday sneakers
Available nearby. The neighborhood store has a 10% local promotion.
↳Synthetic locations, weather and usage. Nothing here reads your actual location, account or device data.
Understand the person.
Don’t forget the world around them.
07A little magic.
THE INVISIBLE SYSTEM
A little magic.
A lot of engineering.
The visible result may be a single useful suggestion. Behind it is a connected decision system: behavioral understanding, multi-objective optimization and contextual intelligence working together.
Retrieval narrows the possibilities. Inference and ranking put them in order. The interface makes them useful. Measurement asks whether any of it helped. Select a station to follow the signal.
07 / FOLLOW A LITTLE SIGNAL
01
What happened?
Searches, clicks, comparisons and purchases enter an event stream. Event quality, consent and a clear schema matter before any modeling begins.
A conceptual architecture, not a diagram of Loodos’ proprietary deployment. In practice, some steps run in parallel, while feedback closes the loop over time.
The experience feels simple because the system handles the complexity.
08Small moments.
THE VALUE OF GETTING IT RIGHT
Small moments.
Meaningful possibilities.
Better discovery should help people find useful things and help a business create sustainable value. Recommendation clicks alone cannot tell the whole story. Conversion, order value and repeat use belong in the measurement plan.
A result at a personalized touchpoint is not automatically a result across the whole business. Reach matters. So do costs, the baseline and the quality of the experiment. Try a few assumptions below.
10–25%
average order value increase
Up to 50%
conversion improvement at personalized touchpoints
The supplied brief attributes these figures to the Loodos article. The original PDF was not provided for verification. These are reported claims, not validated results or expected outcomes.
08 / A LITTLE WHAT-IFYour assumptions. An explicit calculation.
MODELED MONTHLY REVENUE
Baseline$150,000
With your assumptions$161,925
Illustrative difference+$11,925
Only 30% of sessions receive the assumed improvements.
See the assumptions and formula
Baseline = sessions × conversion rate × average order value. Scenario = baseline revenue from unaffected sessions + affected sessions × adjusted conversion × adjusted order value. Both lifts apply only to the affected sessions. Conversion is capped at 100%. The calculation assumes no change to the number of sessions and identical baseline behavior across both groups.
↳An illustrative revenue scenario, not a forecast or a net ROI estimate. Costs, statistical uncertainty, cannibalization and long-term effects are not modeled.
MEASURE WHAT MATTERS
Use controlled experiments and guardrails for customer satisfaction. A short-term lift can hide a long-term cost. Revenue is one useful view, not the full definition of value.
Not knowing everything.
Understanding what
matters now.
You changed the visitor. You followed the signals. You balanced the priorities and changed the world around the decision.
The most valuable discovery experience is not the one that offers everything. It’s the one that makes the right possibilities easier to find.
A little less searching.
A little more understanding.
A little more personal.