Journal
What we have measured, and who measured it.
Notes on making marketing video that gets watched. Every number here is attributed to a named study, and when the studies disagree the disagreement is the article.
- EngineeringHow to build a video editor from scratchA timeline is four hard problems wearing one widget: contiguity, a single clock, a preview that lies, and knowing what an edit destroys. · 12 min
- SearchBeing mentioned on YouTube correlates with AI visibility better than links doAhrefs put the coefficient at 0.737 across 75,000 brands. The same team added schema to 1,885 pages and measured nothing. Both findings are theirs. · 7 min
- SearchThe machine writing about your landing page cannot run your landing pageNo major AI crawler runs JavaScript, and a third of ChatGPT and Claude fetches returned 404. Whatever is in the first response is the whole document. · 7 min
- ProductionNobody benchmarks the one thing a product video is made ofPhyWorldBench scored twelve models across 12,600 videos on physics. The one benchmark for the text in the frame covers ten models and puts its leader at 0.37. · 8 min
- LawThe AI voice laws founders name do not reach a catalogue narrator, and one nobody names doesTennessee's ELVIS Act hooks on an identifiable individual and the FCC ruling is about telephone calls. New York General Business Law 396-b is the one to read. · 7 min
- ProductionA minute of generated picture costs $3 to $42, and the narration costs five centsPriced off the vendors' own pages only. The comparison tables are still quoting a Veo rate Google retired on 8 September 2025. · 7 min
- SearchWhy Google stopped showing your video, and what it takes to get it backGoogle narrowed video results to pages where the video is the main content. The requirements are short, public, and a marketing page with a video in it fails the first one. · 7 min
- LawThe registration form makes you disclaim the part you paid forThe Copyright Office has registered more than 7,000 claims with the AI content disclaimed. Registration is not a yes or no question, it is a field on a form. · 7 min
- ProductionThe seed is not why you get a different video twiceTwo identical requests differ because the endpoint was busier the second time. The seed is the one term in that product you actually control. · 7 min
- LawThe provenance label does not survive the upload, and the specification says soThe AI Act asks for machine readable marking and California will ban stripping it. A C2PA manifest is one box in an MP4, and the upload rebuilds the container. · 8 min
- ProductionA caption that cannot disagree with the scriptAI caption tools run recognition over finished audio and guess the words and the times. Forced alignment takes the text as an input and can only decide when. · 7 min
- DistributionThe feed watches on mute, and your video has to survive itFour studies, four incompatible numbers, and one conclusion none of them contradict: if the argument only exists in the audio, most people never hear it. · 7 min
- SearchThe hero video is your Largest Contentful Paint, and the default embed weighs 1,149 kBLCP is defined to take the poster image or the first video frame, whichever is earlier. So the video above the fold is not decoration. It is the measurement. · 8 min
- DistributionFive video statistics, chased to the document under themOne is 200 people surveyed by an agency that now disclaims it. One says 50, not 53. One points at a Cisco URL that stopped serving the file. · 7 min
- CraftThe safe zone is not a number, and the pixel guides know itTikTok says the safe zone shrinks as the caption gets longer. Meta's page returns HTTP 400 without a login. The exact pixel guides disagree by 350 pixels. · 8 min
- CraftYour autoplay is being refused, and the statistic about it is measuring something elseAll three engines have blocked audible autoplay since 2019. The Web Almanac figure everyone quotes about compliance counts attributes, not video elements. · 7 min
- CraftHow long should a marketing video be?Two platforms have published the answer from tens of millions of videos. The curve they describe is not the one the round numbers imply. · 8 min
- DistributionEvery per platform LUFS table you have seen is a rumourSpotify states its loudness target in writing. No YouTube page states one, and the tables printing both at -14 LUFS hide the asymmetry that decides the mix. · 8 min
- CraftA narrator and a music bed cannot share one loudness numberAES TD1008 cites listening panels that heard speech 2 to 3 dB louder than music at the same measured loudness, so its target moves with the speech fraction. · 7 min
While you are here