Versely
    Guides

    Mine rewatch spikes for your next thumbnail

    A bump above the retention line is a moment the audience chose. Turn spikes into thumbnails, clips, and re-hooks, then test them against a control.

    Versely Team••8 min read

    A local bump above the surrounding retention line is not noise, and it is not a compliment. It is a timestamp where a slice of the audience chose to rewind. Something on screen was replayable: a transformation, a line, a visual that did not land on the first pass. That moment is already a thumbnail, already a Short, already the re-hook the rest of the video is missing. Most teams note it and keep scrolling. The work is to harvest it on a schedule, then check it beat the control.

    The four-shape diagnosis (cliff, bleed, step, mesa) lives in audience retention analysis. A rewatch spike is the rare shape that is not a problem. This post is the harvest: how to tell a spike from a recovery, how to pull a still and a clip from the timestamp, and how to know whether the packaging actually won.

    Recognise a spike before you harvest it

    Open the native retention graph. On YouTube that is the audience-retention chart in Studio, absolute and relative. Versely's social analytics API will not give you this line. It will give you views, likes, comments, shares, and saves on posts that went out through Versely. The spike lives in the platform product. Steal from there, produce in here.

    A rewatch spike is a local rise above the line on either side. Viewers who were already watching went back, or new viewers joined at that moment in numbers large enough to lift the percentage. The graph ticks up, then resumes the decline.

    Two shapes get cashed in as spikes and are not:

    • A valley recovering. The line dipped, then climbed part-way back toward the previous level but never overshot the shoulders. That is a late payoff, not a replayable moment. Move the reveal earlier. Do not thumbnail the recovery.
    • A relative-retention blip against a weak absolute line. Relative view can tick up because the peer set is worse at that timestamp. If your absolute line is still falling through the "spike," you do not have a moment. You have a less-bad stretch.

    Write three numbers down before you touch an editor:

    1. Timestamp of the peak, converted from the percentage. At 25 frames per second (Versely's timeline default, not 24) a 10-minute video has 150 frames per percentage point. "Around 42%" is a window, not a frame.
    2. Width of the bump, in seconds. Under two seconds is often a flash. Four to twelve seconds is a usable clip. Thirty-plus is a segment, and you are no longer making a thumbnail.
    3. What is actually on screen. If you cannot name the replayable thing in one clause ("the sauce hits the pan and the colour flips"), it is not a spike you can package.

    The retention-curve glossary entry is the vocabulary. Your last twenty graphs are the prior.

    Harvest: still, clip, re-hook

    You are making three assets from one timestamp, in this order, because each one feeds the next.

    Still. Scrub to the peak, then step frame by frame. Motion means the frame 80 milliseconds off your target has closed eyes or a smear. Pull three candidates, not one: the peak, one beat before (anticipation), one beat after (aftermath). The anticipation frame is often the better thumbnail, because the peak is mid-action and reads as chaos at postage-stamp size.

    You can extract from a Versely clip with extract frames, or the longer stills workflow in extracting frames for thumbnails. Prefer the frame that already exists over a cold text-to-image of "the same moment." The grade, the character, and the lighting already match the video you are about to package. A separately generated image will not.

    Clip. Cut a short piece that starts on the spike, not a piece that contains the spike at 0:18. The spike is the hook of the derivative, not a reward for people who waited. Pad with one beat of setup and one beat of payoff, then stop. Do not pull the whole chapter the spike sat inside. Repurpose by spike, not by heading. That is the rule that keeps Shorts from opening on a topic title card.

    Re-hook. Write one sentence that names the replayable thing without explaining it.

    Cold open, no logo: the pan sauce flips from brown to neon green on contact.
Hold the still for one beat, then play the motion.
On-screen text, under six words: "this is the colour change."
Do not say "wait for it."

    That sentence is the first frame of the next long-form, the first three seconds of the Short, and the alt text of the thumbnail. If you need a generated variant of the still (tighter crop, cleaner face, room for type), run it through the thumbnail generator with the extracted frame as the reference, not as a memory. Keep type out of the bottom 15% and the top 10%. Those are platform chrome.

    The original video also wants the spike earlier than it currently sits. If the bump is at 6:10, that moment is doing the job a re-hook at 1:40 was supposed to do. Lift a still from it onto the timeline as an insert, or open the next episode on it. The editor is the place; do not regenerate the whole cut to move one beat.

    Test it against a control

    A harvested spike feels right. That is not a result. The test is whether the new packaging beats the packaging you already have, on the metric the surface actually optimises.

    YouTube long-form, native Test & Compare. YouTube Help is the source: you may run up to three title and thumbnail variants. The winner is chosen by highest watch time, not by CTR. Tests can finish inside two weeks, faster at high impression volume. Shorts, Made for Kids, private, mature, age-restricted, and scheduled livestreams or Premieres (until converted) are not eligible. "No clear winner" leaves the first-uploaded variant in place.

    That watch-time rule is the feature. A spike-derived thumbnail can lose CTR and still win, because the people it attracted stayed. Do not compare the test winner to your CTR dashboard and declare the tool broken. Put the extracted still up against the current thumbnail and, if you have a third, against a type-led variant of the same moment. Pre-commit to the two-week window. Peeking is how the early-impression variant "wins" on Tuesday and loses on day eleven.

    Shorts, TikTok, Reels. There is no native three-variant tool. Isolate one variable: the first three seconds, or the cover frame, not both. Same slot, consecutive days; simultaneous posts cannibalise the same pool. Pick the metric before you look. For a hook test that is average watch time, or YouTube Shorts' viewed-versus-swiped-away, at a fixed 24-hour checkpoint. Run at least three pairs before you conclude. One pair is feed lottery.

    Ad surfaces are a different protocol. If the spike is going into a paid unit, use the ad thumbnail testing playbook and keep this organic test separate.

    A weekly mining pass

    Do this on the same day you already read analytics, or it will not happen.

    1. Sort the week's long-form by watch time, not views. Open retention on the top two and the bottom two.
    2. Mark every bump that clears the shoulders of its own line. Ignore the first 15 seconds; that regime is hook and misclick.
    3. For each mark, harvest three stills, one clip, one re-hook sentence.
    4. Pick one still as a YouTube variant, or one clip as a Short. One, not four, so you can tell which test ran.
    5. Log the result against the control at the pre-committed checkpoint. If it loses, keep the still. Spike stills are reference frames for the next generate even when they lose a packaging test.

    You are looking for a moment the current audience already voted for, then making it easier to recognise from the outside. The spike is the vote. The thumbnail, the clip, and the re-hook are how you stop burying it at 6:10.

    FAQ

    Does a rewatch spike mean the whole video is good?

    No. A spike can sit on a curve that is otherwise a slow bleed or a mid-video cliff. Harvest the moment. Do not skip the rest of the diagnosis. If the only replayable beat is at 8 minutes, the first seven minutes are still the reason most people never saw it.

    Should I thumbnail the spike on the same video, or cut it into a new one?

    Both, on different clocks. Same video: run it as a Test & Compare variant on long-form YouTube, watch-time winner, two-week window. New video: cut the 30 to 55 second piece that starts on the spike and publish it as a Short or a Reel. Do not replace the long-form with the clip. They have different jobs.

    Why not just generate a new thumbnail from a text prompt of the moment?

    Because the prompt will not match the grade, the face, or the lighting of the video the thumbnail is supposed to represent. Extract the frame, then generate only if you need a tighter crop or room for type. Continuity between thumbnail and first second is the reason relative retention in minute one holds. A beautiful mismatch is how you buy a click you immediately refund.

    Can I read rewatch spikes from Versely analytics?

    Not as a curve. The social-analytics routes return views, likes, comments, shares, and saves on posts published through Versely. Retention, including rewatch bumps, is a native-platform chart. Pull the timestamp there, then extract and test here.