Google Ads video experiment: 2 arms or up to 10 AI variants?

Google's basic video experiment compares two arms, a control and a treatment, and applies only to Video Reach campaigns (Efficient Reach) and Video View campaigns. The Custom option allows up to 10 arms. So render one extra video for a basic test, and as many as nine alternatives for a Custom test.

This comes from Create a video experiment, read 2026-10-03.

Arms, campaign types and metrics

The page lists the measurable outcomes too, so you can choose the clip difference to match one of them.

Metrics you can compare

Google's page names Absolute Brand Lift, CTR, conversion rate, cost per conversion, CPC, CPM, CPV and video view rate as success metrics. Conversion metrics need conversion tracking, and Brand Lift needs a separate study set up before the experiment.

Pick the arm count from the question, not from what you can afford to render: one hook against one hook is a basic test; four hooks is Custom.

Can your workspace queue ten renders?

Sume accepts jobs as queued while queue capacity remains, and moves them to processing under a per-workspace concurrency limit. Its static table lists accepted job capacity of 6 on Free, 24 on Pro, 48 on Startup and 120 on Scale, and says to prefer the effective generation_limits field over that table. Ten variants fit within Pro and above; on Free the seventh submission can fail with 429 queue_full. See Generation admission.