One edit, two soundtracks: building an honest A/B pair
A saree campaign delivered twice — with the voiceover and without it — so the sound-on and sound-off versions of a single edit can be judged against each other rather than against two different films.
Written from the artifact, not from an engagement record. Everything below describes the work itself and the craft inside it — nothing here claims what a particular client wanted, feared or achieved.
01 / 05
Context
The situation the work is built for.
Feed placements autoplay muted. That is not a claim about any particular advertiser, it is how the surfaces are built, and it means a spot whose argument lives in its voiceover is a spot that makes no argument at all to most of the people it reaches. The standard response is to caption the film and move on. The better response is to find out whether the voiceover was doing any work in the first place — which requires shipping both versions and looking.
02 / 05
Constraint
The limit the work had to survive.
An A/B test is only readable when exactly one thing changes. The moment the clean cut is re-edited to "work better without narration", the comparison is gone: two different films are now competing and the result says nothing about sound. So the constraint on this pair was strict — the picture edit is held identical across both files, and the only difference is the presence of the voice track.
03 / 05
Mechanism
How it actually functions, step by step.
Cut the spot once. The edit is the constant in the experiment and is never touched again after this point.
Export twice from that single edit — one master with the voiceover committed to the mix, one with the narration track pulled. The manifest records these two files as the same spot, which is exactly the property the test depends on.
Hold the clean cut to the same standard as the voiceover cut rather than treating it as a fallback. Every beat the narration would have explained now has to be carried by frame, pacing or on-screen text.
Ship both into the same campaign, in separate ad sets, so each variant reports on its own row and the platform cannot average them together.
Agree the kill rule before launch. Which variant gets cut, at what impression threshold, and on which metric — decided while nobody is invested in the answer.
04 / 05
Artifact
The file you can open and judge for yourself.
Look for
Saree Campaign — Voiceover CutConcept
The sound-on half of a matched pair: one apparel spot exported twice, so the same edit can be judged with narration and without it.
No caption file or transcript has been produced for this cut yet.Ask us for one
Saree Campaign — Clean CutConcept
The same apparel spot with the narration removed — the version that has to carry the offer on picture alone.
No caption file or transcript has been produced for this cut yet.Ask us for one
How the work is labelled
How this is served
Nothing on this page plays on load, and nothing here is streamed out of somebody else’s file store. These masters have not been transcoded onto our own origin yet, so the player says exactly that rather than hotlinking a share link to look faster. Ask and we will screen any of them for you directly, at full quality, today.
05 / 05
What we measure
The instrumentation — where a results number would sit.
Where the results section would be
We don’t publish outcome numbers we can’t show you the dashboard for.
This case study ends here on purpose. We hold no client-authorised outcome figure for this work — no denominator, no date, no permission to publish one — so there is no results section. Everything above is checkable against files you can open on this page. When a number is authorised, it will appear here with the denominator and the date attached, and not before.
So here is the mechanism and the artifact instead, and below them the instrumentation we would ship with this work — the metric, where the number comes from, and the decision it triggers when it moves. Ask us for the measurement plan and we will walk through it against your own account, before you commit a rupee of media spend.
The measurement plan
Agreed before launch, wired up before the first impression, and read per creative — not per campaign, which is where a losing variant hides inside an average.
01Thumb-stop rate — three-second views divided by impressions, per variant.
- Instrument
- Platform delivery report, one row per creative. Never read at ad-set level, because that is where the two variants disappear into an average.
- Decision it triggers
- If the clean cut stops thumbs at parity with the voiceover cut, the narration is not earning its cost on that placement and the budget moves to picture.
02Sound-on rate for the placement.
- Instrument
- The platform’s audio-on metric where the channel exposes it; otherwise the completion-rate gap between the two variants stands in for it.
- Decision it triggers
- Sets the mix strategy for the next flight. Below a low sound-on rate we stop paying for voice talent on that placement entirely.
03Retention at the offer beat.
- Instrument
- The per-creative retention curve, with the frame where the offer appears marked on it.
- Decision it triggers
- A drop at the offer beat is a copy problem, not an audience problem. The next variant changes the line, not the targeting.
04Cost per outcome, per variant.
- Instrument
- Ad account reporting with per-creative attribution, wired up before launch rather than reconstructed afterwards.
- Decision it triggers
- The pre-agreed kill rule cuts the loser at the impression threshold set on day one, instead of after a meeting about whether it deserves longer.
