Google video experiment results: 70% directional, 95% conclusive
Google labels video experiment results directional at 70% confidence and conclusive at 95%. When to act on each, and how to line up Sume ad variants first.
In Google Ads video experiments, results are called directional at 70% confidence and conclusive at a 95% confidence interval. Use a directional result to pick what to render next, and a conclusive one to retire the losing ad.
This is from Create a video experiment, read 2026-10-03.
What each label should trigger
The first two columns are Google's wording. The last is a planning suggestion for AI-generated variants.
| Label | Google's threshold | Suggested move |
|---|---|---|
| Directional | 70% confidence | Draft the next round of variants; do not retire the control yet |
| Conclusive | 95% confidence interval | Retire the loser; keep the winner's script and scene as the next control |
| Also offered | 80% confidence, still directional | Same as directional |
Set up before the test starts
Google says the campaigns must exist before you create the experiment, conversion tracking is required for cost per conversion or conversion rate, and Brand Lift requires a separate study first. Render and upload your variants before you click create so nothing changes after launch.
On Sume, submit every variant with mode: "async" and its own Idempotency-Key. Reusing a key only for an exact retry returns the original job instead of billing a second render, per Jobs and results.
A rule for the next round
Make the winner the new control and change one thing again: the hook line, the scene, or the call to action. With an avatar video that means a new video_inputs scene or script, rendered as a fresh job. Do not edit a finished video and call it the same ad.
Sources
Related posts
More in Use cases
- Google Ads video experiment: 2 arms or up to 10 AI variants?
Google's basic video A/B test has two arms and only covers Video Reach and Video View campaigns; Custom allows up to 10. How many Sume variants to render.
- GPT Image 2.5 icon set: one anchor icon, then reference it
Make a consistent icon set with GPT Image 2.5 on Sume: approve one anchor icon, pass it as input_references, change only the subject, and keep a style block.
- GPT Image 2.5 pixel art sprite sheet: prompt a grid, slice it
Ask GPT Image 2.5 for a 4x2 sprite sheet at 2048x1024 on Sume, then slice frames and snap to a true pixel grid with Pillow. Prompt and code included.
- GPT Image 2.5 menu board prompt: quote every line, then proofread
Make a restaurant menu board image with GPT Image 2.5 on Sume: a prompt that quotes each item and price, a size that fits, and a proofreading pass before print.
Written by Sume