This isn’t really an anomaly in the way it feels to a creator watching the numbers; it’s the expected outcome of YouTube treating satisfaction signals and engagement/retention signals as separate, non-interchangeable inputs. A video can score strongly on likes, comments, and shares (what YouTube’s documentation groups under satisfaction) while performing poorly on watch time and average view duration (engagement/retention), and the system does not average these into one blended score. It weighs them as distinct evidence about different things: satisfaction signals suggest people who watched enjoyed or valued what they saw, while retention signals suggest whether the video actually held attention. Weak retention typically still caps how far the video gets recommended, even when the satisfaction signals look good.
Why the two signal types don’t cancel each other out
YouTube’s own Creator documentation and public statements (including from YouTube’s Chief Product Officer and the Creator Insider channel, which YouTube runs to explain recommendation-system behavior to creators) describe the recommendation system as pulling from multiple categories of signal: performance signals like clicks and watch time, and satisfaction signals like likes, shares, “not interested” feedback, and survey responses shown to a sample of viewers asking how satisfied they were. These categories exist because they answer different questions. Watch time and retention answer “did this content hold attention as promised,” which matters because YouTube’s core business incentive (and its stated recommendation goal) is maximizing genuine viewer satisfaction and long-term platform engagement, not just short-term clicks. Likes and comments answer “did the people who stuck around feel strongly enough to react,” which is a real but partial signal, since it only comes from the subset of viewers who didn’t drop off before reacting.
A high like ratio and strong comment count with low average view duration usually means: the audience that stayed to the point of reacting genuinely liked the content, but a large share of viewers who clicked in didn’t stay long enough to become part of that reacting audience at all. Those two facts aren’t contradictory, they’re describing different populations within the same view count. The recommendation system isn’t confused by this; it simply has two honest, separate readings, one from people who left early (reflected in the retention curve) and one from people who stayed and engaged (reflected in likes/comments). Because YouTube has never published an exact formula for how these categories are weighted against each other, no creator or third party can state precisely how much a comment count offsets a weak retention curve. What’s consistently observable in practice, and consistent with YouTube’s own framing of watch-time-driven recommendations, is that weak retention limits how many additional impressions and recommendation slots a video gets tested against, regardless of strong reaction metrics elsewhere. Comments and likes can still help (they’re real satisfaction signal, and comments in particular can extend a video’s relevance window by keeping a video “active” in discussion), but they are not documented as a substitute for retention.
It’s also worth being precise about what “high like ratio” actually measures: it’s a ratio of likes among viewers who bothered to react at all, which is a small, self-selected fraction of total viewers. A high ratio among a small reacting group is compatible with the broader audience finding the video’s opening unengaging enough to leave, particularly if the drop-off happens early. This is why the two metrics can diverge without being in tension: they’re sampling different slices of the same audience.
A hypothetical example
Consider a hypothetical example: a channel called Harlow Home Repair uploads a 20-minute video on fixing a leaky faucet. Suppose it racks up a 98% like ratio and dozens of enthusiastic comments, but average view duration sits at just 3 minutes, far below the channel’s typical retention. Pulling the retention curve, hypothetically, shows a steep drop at the 45-second mark, right after the video’s slow, five-minute setup showing Harlow gathering tools and explaining general plumbing safety before getting to the actual fix. In this hypothetical, the comments that do exist reference the specific repair steps near the end of the video, suggesting the people who stuck around loved it, but most viewers who clicked from Browse or Suggested (per the Reach tab) left before reaching the useful part. Under the mechanism described above, Harlow’s team would reasonably treat this as a retention problem, not an engagement problem, since the reactions are already strong: tightening the intro to get to the actual repair within the first 30 seconds, rather than chasing more likes or comments, is the lever that would plausibly fix the underlying issue in this hypothetical scenario.
What to actually check when you see this pattern
Pull the retention curve first, not the summary average. Average view duration is a single number that obscures where the drop-off actually happens. In YouTube Studio Analytics, the audience retention graph shows exactly where viewers leave: a steep drop in the first 15-30 seconds points to a hook/intro problem, while a gradual decline throughout suggests pacing or content-density issues rather than a bad opening.
Compare the retention curve to your own channel baseline, not to an assumed universal benchmark. YouTube has never disclosed a “good” retention percentage, and it varies enormously by video length and format. What matters is whether this specific video’s curve is meaingfully worse than your typical upload’s curve, which tells you whether the issue is video-specific or systemic to your content style.
Check where the likes and comments are timestamped relative to the drop-off, where visible. If comments reference something near the end of the video, that confirms a loyal subset watched through, which supports the “engaged minority, disengaged majority” read rather than suggesting the retention number itself is somehow wrong.
Look at traffic source breakdown in the Reach tab. If most of the underperforming retention is concentrated in suggested/browse traffic rather than subscriber or search traffic, that points toward a mismatch between what the thumbnail/title promised and what the video actually delivers early on, since browse-traffic viewers have the least context going in and are quickest to leave if the opening doesn’t match expectations.
Treat the fix as a retention problem, not an engagement problem. Because comments and likes are already strong, there’s no reason to chase more reactions; the leverage point is the first 30-60 seconds of the video and whether it delivers on the title/thumbnail promise fast enough. Re-cutting the opening, tightening pacing, or restructuring where the core payoff lands in future uploads addresses the actual bottleneck the algorithm is responding to, rather than the metric that already looks fine.
The practical takeaway is to stop reading likes and comments as proxies for overall algorithmic health. They’re a real, separate signal about audience sentiment among people who stayed, but YouTube’s system is documented as also weighing retention independently, and retention is the input most directly tied to how widely a video gets recommended afterward.
