VSL Page Build Tuned to the Retention Curve

On the video pages we operate, more than half the people who press play are gone inside ninety seconds. Where exactly they leave is the only diagnostic worth having, so this build starts by making that curve readable and ends by acting on it. Player and instrumentation: Video is served from Mux or Bunny Stream over HLS with the lowest rendition first, so playback starts on an Israeli mobile connection without a spinner. The poster frame is pulled from an actual frame containing a face, not a title card. Autoplay runs muted behind a large, visible sound-on control; iOS Low Power Mode refuses autoplay outright, so the tap-to-play state is designed rather than inherited. Player events, meaning play, ten seconds, each quartile, ninety-five percent, every seek and the sound-on tap, land in GA4 and in a raw event table we can query directly, because GA4 will not tell you what happened inside one specific ten-second window. Reading the curve: Three points matter. The first ten seconds tell you whether the ad and the video agree; a cliff there is a traffic problem and we will say so rather than rebuild a page that is working. The pitch moment tells you whether the promise survived the setup. The price moment tells you whether the offer is wrong, and no page change repairs that. The annotated curve ships with the build, and where the correct action is to recut the video, we name the timecode and stop, because editing is not work we do. The page around the video: Call to action visible from the first frame and pinned as the viewport moves. A full transcript below the player as selectable text, which serves the share of visitors who will read but never watch and gets the argument understood by search and answer engines. Proof sits adjacent to the claim it supports instead of being gathered into a logo strip. Page weight is held so the largest contentful paint is the poster image and lands inside 2.5 seconds on a mid-range Android over 4G. Hebrew and RTL: Subtitles are WebVTT rather than burned in, so a wording correction does not mean a re-encode and the English cut reuses the same master. Hebrew subtitle lines get explicit bidi marks wherever a line ends in a question mark or a Latin brand name, otherwise the punctuation renders on the wrong side and reads as a typo to every native speaker. Layout uses logical CSS properties, so Hebrew and English are one template with a direction attribute, not two codebases drifting apart. Refusals: No call to action hidden until minute twelve. No disabled scrub bar. No fabricated live-viewer counter, no countdown that restarts on reload, no exit interstitial that traps the back button. Autoplay never carries sound. These tactics move a two-week number and pay it back in refunds and in what people say about you afterwards. If they are on the brief, we are the wrong shop and there is no version of this engagement that includes them. Not included: Scripting, filming, editing, voice work and translation of the master. We take delivery of a finished video and a transcript. We do not run media and do not manage the campaign feeding the page. Not a fit: Products where nobody will appear on camera. A stock-footage video with a synthetic voice reads as exactly that to an Israeli audience and the first ten seconds show it. Offers small enough that a long video cannot earn its production cost. And any brief that opens with a competitor page to reproduce, which we decline on sight.