Demand Gen: what to hold constant between two AI video ads

For a clean Demand Gen A/B experiment between two avatar videos, keep these identical: avatar_handle, aspect_ratio, quality, scene or background, caption style, and the script after the hook. Change one thing, such as the hook scene. Google's page says to choose campaigns that differ in only one variable.

The one-variable rule is from Create A/B experiments for Demand Gen campaigns, read 2026-10-03. Field names are from Sume's Generate avatar video.

Field-by-field checklist

quality is the easiest to forget: it defaults to plus when omitted, so leaving it off in one request and sending max in the other would be a second variable.

Match the lengths

Different hook lengths shift the total duration, which can change billing and pacing. Sume plans total duration in a 4-60 second window, so write the hook variants to about the same spoken length.

Preview before you pay twice

If you suspect the framing differs between the two renders, create an avatar video preview first. A preview shows first-frame stills, and generate-video can then override quality for the final render only, so the approved still is reused. See Avatar video previews.