How We Built a Smart Car Matching Algorithm

From Simple Weights to a System of Common Sense

When we first started designing our car quiz, the task seemed straightforward:

ask users a few questions → assign weights to categories → rank cars.

But very quickly, we realized something fundamental:

People don’t choose cars like spreadsheets.

They think in terms of real life:

“I want a small car — not a truck.”

“I don’t want to spend more than I’m comfortable with.”

“I drive kids and clients — a two-seater is not an option.”

“I care about emotion, status, or control.”

If a system ignores those realities, it loses trust instantly. That realization led us to build a multi-layered matching algorithm that combines data-driven scoring, behavioral signals, and hard common-sense rules.

1. Raw Data

Hard specs: HP, MPG, Dimensions, Features.

2. Preferences

User priorities, weights, and “nice-to-haves”.

3. Common Sense

Hard constraints, logic rules, and deal-breakers.

Step 1Categories as a Foundation — Not the Truth

At the core of our system are 10 stable categories:

Practicality & Everyday Usability
Comfort & Cabin Experience
Performance & Driving Dynamics
Efficiency & Running Costs
Luxury & Status Feel
Technology & Innovation
Adventure & Capability
City-Friendly & Urban Life
Road-Trip & Long-Distance Comfort
Reliability & Ownership Confidence

Each car receives a 0–100 score in every category, derived strictly from real data: dimensions, horsepower, torque, fuel efficiency, pricing, towing capability, wheelbase, warranty coverage, and drivetrain.

This gives us an objective baseline. But categories alone are not enough.

Step 2One Answer ≠ One Signal

We quickly learned that almost no user answer maps to a single dimension. For example:

“I want driving excitement” affects Performance, but also Comfort, Technology, and even City usability.
“I care about status” touches Luxury, Comfort, and Technology.
“I want reliability” blends Reliability with Practicality.

So each answer in our quiz emits a Primary category signal and a Secondary category signal (with reduced weight). This approach prevents extreme distortions, preserves nuance, and produces a smoother, more human preference profile.

Step 3Normalization and Noise Control

Users answer different numbers of questions. Some preferences repeat more often than others. To prevent bias, we normalize category weights so that:

The total number of questions does not affect results
Strong opinions don’t overpower everything else
Every user ends up with a comparable preference vector

The outcome is a stable preference profile, not a pile of clicks.

The Goal: Mathematical Fairness. Whether you answer 5 questions or 50, your core preferences hold the same weight in the final score.

Step 4 & 5Why Weights Alone Don’t Work & The “Common Sense” Filters

We ran into a classic recommendation problem: A user chooses “Small & Agile,” but still gets recommended trucks and large SUVs. Why? Because weights change ranking, but they don’t forbid bad matches.

That insight pushed us to the next layer: Hard “Common Sense” Filters. We identified a set of questions where being wrong is unacceptable.

Size

If a user chooses “Small”:

Vehicles longer than ~190 inches are heavily penalized
Trucks, vans, and full-size SUVs nearly disappear
Compact cars receive a meaningful boost

Fuel Efficiency

If efficiency is critical:

ICE vehicles below 30 MPG are almost eliminated
EVs and hybrids rise to the top

Price

We adjusted thresholds for 2023–2025:

“Low cost” no longer means $25k
Penalties start later and scale gradually

Step 6Non-Negotiables

Some constraints are binary. In these cases, there are no trade-offs. A car either fits the lifestyle — or it doesn’t.

If you regularly drive kids or clients → Minimum 4 seats
If you need cargo space → Tiny trunks won’t pass
If you care about bad-weather stability → AWD matters

Step 7Behavioral Modes

The biggest leap came when we stopped looking at answers individually and started looking for patterns. A person is not a single choice — they are a behavior profile. We introduced dynamic modes that activate only when multiple signals align.

Sport Mode

Triggered By

Excitement, sound, low driving position, control.

Effect

Strong boosts for performance cars; penalties for utility-first vehicles.

Utility Mode

Triggered By

Work equipment, tools, crew, autonomy.

Effect

Trucks and vans gain priority; towing and payload matter more.

Luxury Mode

Triggered By

Status, silence, premium interiors, quality.

Effect

Luxury brands rise; mass-market vehicles de-emphasized.

Step 8Conflict Resolution

Real people are contradictory. So the system allows multiple modes at once, softens penalties when modes conflict, and enforces a safety floor to prevent soft penalties from erasing otherwise valid cars.

This keeps recommendations flexible, realistic, and non-aggressive.

A Matching System That Thinks Like a Human

What we ended up with is not a filter and not a leaderboard. It’s a decision model. Enthusiasts see driver-focused cars. Families never see two-seat sports cars. Professionals get tools, not compromises.

Ready to see it in action? Start the car match quiz, explore the car database, and compare cars side by side.

We stopped trying to find the best car. We started matching the right life.