What actually separates a good video ad from a slideshow: 12 measured benchmarks

We picked three video ads we agreed were good and analysed them frame by frame: one holds a single continuous shot in a light palette, one is dark and cuts hard every few seconds, one is bright mint-green and built from dozens of short beats. They share almost nothing visually on purpose. If a measurement holds across all three, it describes technique rather than one look, one brand, or one editor's taste.

Every number below comes from that measurement, decoded from the actual files at a fixed frame rate into a resolution-independent grid. Where a number only held for two of the three ads, or one, we say so rather than rounding it up to a rule. Three ads is a small sample and these are silent 846x468 screen recordings, not an industry survey. Treat this as a working floor calibrated to three specific ads that were good, not a universal law of advertising.

1. Cut rate is not a quality signal

The strongest of the three ads never cuts at all. It holds one unbroken shot for 48.5 seconds, more than two-thirds of its 71.3-second length. The other two ads cut hard, holding their longest single shot for only 6.7 and 14.9 seconds, and they are also good.

We went in assuming cut frequency would predict quality, because that is the standard short-form advice. It does not. We built a maximum-shot-length check and then dropped it as a threshold, because it would have failed our best reference. What replaced it is a check on whether the frame keeps changing inside a shot, cut or no cut, which is finding 2.

Do this: stop counting cuts as a quality proxy. A single held shot can outperform five cuts if it never actually goes still.

2. Every ad is frozen 18 to 22 percent of the time, and that band is tight

Measuring frame-to-frame difference and counting anything below a near-zero threshold as "dead," the three ads land at 22.4%, 19.2%, and 18.2% of their running time (excluding the closing hold, see finding 12). Three unrelated visual styles converge on a narrow four-point band.

That is a real floor, not zero. An ad that never stops moving anywhere reads as relentless rather than premium. An ad that sits still for a third of its length or more starts reading as a slideshow with captions.

Do this: let the frame actually stop sometimes, deliberately, for well under a quarter of the runtime. Being still is not the failure mode. Being still for the wrong reason, or for too long in one place, is (see finding 5).

3. The gap between the busiest and calmest moments runs 18 to 35x

We measured the ratio of the loudest 10% of motion to the quietest 10%, per ad. It came out to 24.1x, 18x, and 35.2x. A video that drifts continuously at one middling speed, instead of alternating hits and holds, can post the same average motion and still score badly here, because it never has a quiet moment for a loud one to land against.

This is the number missing from most "just keep something moving" advice. Average motion is not enough on its own. It is the difference between a video's loudest instant and its quietest instant that reads as intentional pacing rather than nervous energy.

Do this: measure your busiest-to-calmest motion ratio, not just the average. If it sits near 1, the ad is humming at constant amplitude with nowhere to land emphasis.

4. The first three seconds move well above the floor

Measured mean frame-to-frame motion inside the opening three seconds came out to 14.6, 13.9, and 6.2 units. Even the calmest of the three openers cleared roughly double what we set as the minimum pass threshold for that window.

None of the three ads open on a static logo card or a slow fade. Something visible is already moving before the first second is out.

Do this: don't spend the opening three seconds establishing mood with stillness. Put motion, even small motion, on screen from frame one. The scroll does not wait for the second beat.

5. No freeze inside the body of any ad runs longer than 1.8 seconds

The longest single frozen stretch measured inside each ad's body (excluding the closing hold) was 1.8, 0.8, and 1.2 seconds. We set a working ceiling above that, with margin, at 2.7 seconds.

This is a different measurement from the overall dead-frame percentage in finding 2. An ad can sit at 18% frozen time in total while never freezing for more than a second at a stretch, spread across many short pauses, or it can hit the same total with one long stall. Only the second pattern reads as the video hanging.

Do this: if a beat needs to hold, cap any single freeze at around two seconds. Past that, split it, or move something small inside the frame while the main composition holds.

6. Frame density and contrast sit in a narrow band no matter how different the ads look

We measured how much of the frame is actually drawn on (edge density, via a standard edge-detection pass) and how much luminance range separates the darkest and brightest visible content (contrast, measured at the 99th minus 1st percentile of brightness, not the more common 95th minus 5th, because a headline over a dark field covers only a few percent of the frame and gets discarded at the 95th percentile). Edge density measured 28.6, 33.4, and 30.5. Contrast measured 146.8, 159.4, and 159.3.

Distinct colour palette count, by contrast, ranged far wider: 14, 18, and 27 distinct colours holding at least half a percent of the frame each. Palette size is a style choice. Density and contrast, across a light ad, a dark ad, and a mint-green ad, are not.

Do this: whatever your palette, keep the frame visually dense (avoid large flat empty regions) and keep real luminance contrast, not just colour contrast. A washed-out frame fails this even with a bold accent colour.

7. One hue carries almost the entire colour budget

In the dark, cut-driven ad, 96.04% of every saturated pixel across the whole video falls into one warm hue family (a red-orange core plus its immediate orange neighbour). In the mint-green ad, 95.4% of saturated pixels sit inside a single 20-degree hue band. Both ads allow exactly two additional hues to clear a 0.5% share of the frame, and each of those two accents is reserved for one specific job (one flags a single called-out word, the other flags only a cost or negative-money figure). Neither ad lets a third colour in.

The light, continuous ad follows the same discipline by eye (a consistent pain colour against a consistent solution colour), though we did not carry a saturated-pixel percentage for it forward from the files we measured.

Do this: reduce your extracted brand colours down to one dominant hue plus at most two accents, each doing exactly one job every time it appears. A rainbow of pulled brand colours needs to be narrowed, not honoured.

8. Text size is measured as a percentage of frame height, never in pixels

Standard headline text across the two ads we measured for this clustered tightly at 8.5% to 9.8% of frame height. One of the two ads gave a single hero word, used exactly once, 62.0% of frame height, by far the largest element in the entire video.

Both reference ads are 846x468. A generated ad is delivered at 1080x1920. A pixel size from one has no meaning on the other; a percentage of frame height does.

Do this: define every text size as a fraction of frame height, floor headlines around 8% or higher, and give at least one beat a hero word well above 30% of frame height so the ad has a moment of real scale, not just consistent captions.

9. Corner radius, and everything else soft, forms one consistent system

Corner radii across every card and the outer canvas measured 4.0% to 6.0% of frame height in one ad and 3.6% to 5.2% in the other. Neither ad has a single sharp corner (under roughly 5 pixels of radius) anywhere in it. The same discipline showed up in the two other soft-edged details we could measure directly: in the mint-green ad, the one pulsing glow effect we measured fell to half brightness at 19.2% of frame height out from its center; in the dark ad, the heading and sub-copy text inside the same card sit at roughly a 2.0x luminance step apart, a clearly graded two-tier hierarchy rather than two similar greys.

Do this: pick one corner-radius token somewhere in the 4% to 5% of frame-height range and use it everywhere: mockups, buttons, containers, the outer frame itself. Consistency reads as intentional design more than the specific radius chosen.

10. Text builds word by word, with a deliberate blank gap, never letter by letter

Across all three ads, on-screen text reveals one word at a time, staggered 100 to 250 milliseconds apart, each word fading in over 200 to 300 milliseconds with a small upward slide and an ease-out curve. None of the three ever overshoots or bounces. The one ad that holds a continuous shot goes further: it never crossfades one statement straight into the next. It holds, fades out over 200 milliseconds, leaves roughly a 200-millisecond blank gap, then cascades in the next line. That gap is doing work; it is what makes the next line land as a new thought rather than a correction to the last one.

Do this: build captions word by word on a consistent stagger and a consistent ease-out, and don't be afraid of a short blank beat between statements. Letter-by-letter reveals and instant statement-to-statement crossfades both showed up rarely, if ever, across the three ads.

11. A count-up number eases in and decays, it does not move linearly

In the one ad where we could measure this directly, a cost counter runs from $0 to $699,893 over 3.3 seconds. Sampling the value every 0.3 seconds, the amount added per step rises for the first third of the animation, peaks, and then decays toward the final number rather than adding a constant amount each step. That is the signature of an eased curve, not a linear tween.

We only have this measurement from one of the three ads, so we present it as a single data point rather than a three-way pattern. It is specific enough, and the underlying logic general enough, that we think it holds beyond this one instance: a linear counter reads like a spreadsheet updating, not a number being revealed.

Do this: ease any animated number in and out. Avoid a constant per-frame increment.

12. Every ad closes on a held, static end card

All three ads park on a completely static frame at the very end: 1.3 seconds in the dark ad, roughly 3.5 seconds in the continuous-shot ad, roughly 3.7 seconds in the mint-green ad, its longest hold of any beat in the whole video. We measure this separately from the in-body freeze checks in findings 2 and 5, specifically so a correctly built ad is not penalised for holding its own call to action.

Do this: let your last beat actually hold, for a second or more. Stillness at the very end reads as confidence, not laziness, as long as everything before it kept moving.

Where we checked this against our own output, not just the references

A quality bar is only useful if it changes a real decision. We ran our own renderer's output through the same measurement and it first scored 75 out of 100: deliverable, but below the bar, failing on exactly two of the checks above. Motion (finding 5's family of checks) measured 0.428 against a required 1.1381, and the dead-frame fraction from finding 2 measured 0.719, nearly four times the 18-22% band the references hold.

Two changes that looked obvious from watching the render, adding word-builds and a slow camera pan, only moved the dead-frame fraction from 0.719 to 0.514. It still failed. The measurement is what showed the dead time was hiding in the tail of each beat, after its entrance animation finished and before the next cut, not in the obvious places. Adding continuous motion under every beat, not just at its start, brought the dead-frame fraction to 0.048 and the motion score to 1.347, both inside the reference band. Re-scored later against a real captured brand kit rather than a test fixture, it held up stronger, not weaker, which is the result you want if the technique generalises rather than fitting one file.

The honest limits of this data

Three ads is a small sample, chosen because we judged them good, not sampled at random from a market. They are silent screen recordings at 846x468, not the ads' original master files, and one of the three is literally a recording of the ad playing inside a video-preview interface. The bar above is calibrated to these three specific ads and should be read as "at least as good as the weakest thing we called good," not as an industry average. A resolution-independent grid lets us compare an 846x468 landscape reference against a 1080x1920 vertical deliverable, which is a modelling choice that makes the comparison possible, not a proof that every property transfers perfectly between the two formats. Two of the twelve findings above rest on one or two ads rather than three, and we have said so at each one rather than rounding up.

If you want a video built against a measured bar rather than a guess about what "feels premium," start with your own site and let sizzledraft build the ad from it.