Operating System · Free

Drop your thumbnail into a synthetic YouTube front page and make it compete against 5,840 real competitor thumbnails before you publish it.

Free usually means thin. This is 6,002 files and 1,011 MB, including 5,840 real competitor thumbnails that carry both an image on disk and a real title — the pair that makes one usable as a tagged competitor instead of a pixel blob at the pool median. Your thumbnail goes into a simulated feed beside 11 of them per impression, and 100 persona-agents either click it or scroll past. It runs on your machine on one dependency, with no API key, no account and no quota — and it is not an oracle: the agents model attention, not an audience, and they never watch your video.

Download Free

real competitor thumbnails carrying BOTH an image on disk and a real title — the pair that makes one usable as tagged competition

price. Instant download, one dependency (sharp), no API key, no account, no quota

persona-agents scan every feed, each with its own preference vector, position and attention bias, and Gumbel noise

real competitor decoys sit beside your thumbnail in every single impression

0 MB

the shipped zip: 6,002 files, counted from the zip's own central directory

bin tools: youtube-arena, arena-calibrate, build-arsenal-decoys, arena-autotag, the thumbnail factory, trend intel

How it works

YouTube CTR Arena · console

Simulated feed compar…active

100 synthetic viewer …active

Pixel attributesactive

Attribute ablationactive

5,840

real competitor thumbnails carrying BOTH an image on disk and a real title — the pair that makes one usable as tagged competition

Drop in your thumbnail

It faces 11 real competitors per impression

100 persona-agents scan and click

Read CTR, win-rate & mean rank

Fix the attribute the ablation blames

How it actually works

Exactly what it does — and when.

1For a thumbnail you are about to commit to → it builds a synthetic front page where your candidate sits among eleven real high-performing competitor thumbnails at shuffled positions, runs hundreds of impressions, and returns click-share, front-page CTR, win-rate versus real competitors and mean rank.

2For the question "why did nobody click" → it ablates each attribute to the pool median, re-measures the whole run, and reports the delta as that attribute's marginal contribution, so the answer is computed from the simulation rather than asserted.

3For 100 modelled viewers → each is a sampled archetype with its own preference vector over the attributes, its own position and attention bias, and Gumbel noise, and each clicks the best visible tile only if it clears their personal threshold — so scrolling past without clicking is a real outcome and the CTR is not inflated.

4For raw pixels → it measures contrast, colour pop, brightness, focal dominance, face and emotion energy, warmth, and clutter as a penalty, straight off your image, before any tagging is applied.

5For semantic signals pixels cannot see → it tags text clarity, curiosity, money cue, correlation, brand recognisability and number hook across your candidates AND the decoys, because scoring pixels alone is a loudness meter rather than a CTR model.

6For weights you should not have to guess → it mines attribute multipliers from the shipped arsenal's real view-lift and applies them with --calibrated, so focal dominance and face energy carry the weight the data gives them.

7For a batch of AI-generated candidates → 16 factory and intel tools generate on-topic thumbnails, reference-matched builds, face equations, icon tiles and galleries, and grow the reference pool further free off the public image CDN, resumably.

8For every run and every file → it runs locally with one dependency and no API key, no account and no quota, and seeded runs are reproducible — 1.01 GB, 6,001 files, download and open it immediately.

In practice · Four candidates, one seed, and the one change that moved the number

A creator has four thumbnail variants for the same video and no way to choose. They drop all four into the arena against the shipped pool of real high-performing thumbnails, at 100 agents and 400 impressions with the seed fixed at 7. Two variants lose outright on win-rate; the remaining two finish within a point of each other, which the docs tell them to read as a tie rather than a winner. The attribution block says colour pop is the biggest negative on both, so they raise saturation on the hero element only and re-run at the same seed — one variant now clears the other by six points, and the report attributes the gain to exactly the attribute they changed. They calibrate against their own niche, confirm the ranking holds, and publish that one. Nothing was uploaded and nothing was metered.

Use it to…

Real ways people put YouTube CTR Arena to work.

Settle version A against version B before you publish

Runs are seeded, so two versions of the same thumbnail compare honestly — click-share, front-page CTR, win-rate vs the 11 decoys and mean rank, measured the same way both times. Publishing is the only thumbnail test most channels ever run, and it is paid for with the upload.

Decide what to change next instead of redesigning on a hunch

The attribution block names the attribute costing you the most click-share, because that attribute was ablated to the pool median and the run re-measured. Repair that one named thing and run it again rather than regenerating the whole plate and hoping.

Build a decoy pool that looks like your front page, not everyone's

build-arsenal-decoys builds the tagged pool and takes --niche and --minViews. Your thumbnail then competes against the kind of tile it will really appear next to, and win-rate vs decoys starts meaning something specific.

This product

Publishing it and hoping

When you learn

Before you publish — CTR, win-rate, mean rank

After you publish, and it costs the video

The verdict

Why it wasn't clicked, attribute by attribute

"Looks good to me"

Competition

5,840 real high-performers as the decoys

Your thumbnail, alone, on a white canvas

Cost

$0 — one dependency, no account, no quota

A/B tooling on a monthly plan, if any

The shift it creates

Without it

Duct-taped manual workflows
Nothing that compounds over time
Rebuilding it from zero each time

→

With YouTube CTR Arena

It refuses to give you a score out of ten
"Why it wasn't clicked" is computed, not narrated
No-click is a real outcome, so the CTR is not inflated

What's inside

🥊 Simulated feed comparisons — evaluate your thumbnail against a reference pool you supply, with shuffled positions and a no-click outcome.

🧠 100 synthetic viewer profiles — configurable visual preferences, position bias, and seeded noise model choices within the simulation.

📊 Pixel attributes — contrast, colour, brightness, focal dominance, face/emotion estimates, warmth, and clutter measured from supplied images.

🔍 Attribute ablation — change one attribute to the pool median and rerun the simulation to inspect its modeled contribution.

🖼️ Bring your reference pool — third-party competitor images are omitted. Use your own or appropriately licensed images and metadata; original sample assets are included.

🎯 Optional calibration — fit attribute weights from the reference data you supply. Results depend on that data and do not establish real-world performance.

🏷️ Semantic tagging — record candidate and reference attributes such as text clarity, curiosity cues, branding, and number hooks.

🏭 Thumbnail tools — candidate building, galleries, comparison sets, tagging, and reference-pool utilities. Only import media you are authorized to use.

📚 Craft documentation — CTR psychology, recorded observations, arena design notes, and thumbnail prompt templates.

🔁 Reproducible comparisons — fix the seed when comparing revisions so random variation stays controlled.

⚙️ Local scoring — install sharp, check dependencies, and run on your machine. Local scoring does not require an API key or upload images.

📖 Simulation limits — these profiles model visual attention rather than a real audience. They do not watch your video or establish future CTR or revenue.

Who it's for

Creators publishing on a schedule who currently pick thumbnails by taste or team vote

Taste is not reproducible and a vote is a handful of people guessing. A seeded run against 5,840 real competitors is something you can re-run and get the same answer from, as many times as you like, for $0.

Thumbnail designers and freelancers who have to defend a choice to a client

"I like this one" loses arguments. Click-share and win-rate against real competitor decoys, plus the named attribute the ablation says is holding the plate back, is a defensible reason — and the client can re-run it.

Channel managers and small agencies running several channels

Seeded runs and per-niche decoy pools built with --niche and --minViews mean the same standard applies across every channel you handle, and the numbers are reproducible when someone asks how the call was made.

Questions, answered

✓ Free · no email✓ Yours to own✓ Use on every project✓ Built to ship results