Why we never optimize on view count alone
Rebuilding our ranker signal around CTR-by-surface and retention, not raw views.
Views are the metric every creator watches. Views are also the worst single signal you can use to tune a thumbnail engine. This post is about why we threw view count out of the ranker's loss function in the first month, and what we replaced it with.
01What views actually measure
A view on YouTube is the joint outcome of three things you don't directly control:
- Impressions — how many feeds you appeared in.
- CTR — what fraction of those feeds clicked.
- Click-through-to-watched-second — what fraction of clicks actually counted as a view by YouTube's threshold.
Optimizing on views collapses these three. A thumbnail that lifts CTR by 20% but drops impressions by 30% (because it underperformed on a single surface that drives most of your impressions) shows up as lower views in your studio. The thumbnail was better; the algorithm just stopped showing it.
02Why CTR-by-surface is the right unit
CTR is the part of the chain you actually control with a thumbnail. But "CTR" averaged across surfaces is itself a lie — homefeed CTR, suggested CTR, search CTR, browse CTR, shorts CTR all behave differently and respond to different visual cues. A thumbnail that lifts CTR on suggested by +12% might drop CTR on search by -4%, and the net effect on your channel depends entirely on which surface dominates your traffic mix.
// What the ranker actually ingests, simplified.
{
"video_id": "vid_018f7a",
"thumb_id": "thb_2c9",
"surface_ctr": {
"homefeed": 0.084,
"suggested": 0.117,
"search": 0.061,
"browse": 0.043
},
"surface_impressions": {
"homefeed": 18432,
"suggested": 41892,
"search": 9217,
"browse": 2104
}
}The ranker computes lift per-surface, then projects it through your channel's historical surface mix. The number that lands in the user's dashboard is a weighted average that reflects your traffic, not the average creator's.
03The retention guard
Even surface-decomposed CTR is incomplete. A high-CTR thumbnail that loses on retention quickly becomes a low-CTR thumbnail (YouTube's algorithm de-promotes videos with poor session behavior). So the ranker's loss function carries a retention guard — any variant predicted to lift CTR by less than its predicted retention drop, in discounted-views terms, is excluded from the top-3.
This catches the most common failure mode of generative thumbnail tools: maximizing for click without modeling what happens after.
When a creator says "your tool dropped my views," they almost always mean "the variant lifted CTR for a week, and the algorithm pulled back impressions in week two." The retention guard exists to keep this from happening twice.
04Why this matters at small scale
You can ship a CTR-only ranker on a channel with 1M monthly impressions and the noise will mask retention damage for a while. On a channel with 50k monthly impressions, the same ranker collapses fast — every mis-fire moves a meaningful share of the channel's traffic. Building for the small-channel case forces the surface and retention modeling that the large-channel case eventually needs anyway.
— Hana