AI UGC test playbook: which of the 8 stages Sume covers

Sume covers four of the eight stages in a current AI UGC testing playbook: reference research on TikTok, an original avatar as the start frame, a first-frame preview plus the talking clip, and finishing with captions and product footage. It does not cover the brief, the test design, the upload to a scheduler, or the measurement. Those stay with you and your ad platform.

The playbook is a ChatbotX guide published on October 12, 2026. It walks through eight stages: product brief, hunting outlier videos, a local research board, reverse-engineering the winners, designing tests, a start frame and generation, animating and finishing, then publishing and measuring. This page maps each stage to what Sume ships, so you know where an API call helps and where it does not.

Which stages map to a Sume route?

The table below follows the playbook's own order. Only routes documented on docs.sume.com are listed.

What does Sume return for the research stage?

Trending video search returns ranked public TikTok metadata: the watch URL, an optional cover thumbnail, the author handle, metrics and a relevance score. It does not return the video file, so there is nothing to download for a face swap or captions. In production the query is required, the limit runs from 1 to 50, and the window can be yesterday, this week, this month, three months, six months or all time.

The playbook suggests a 30-day research window that you can widen to 90 days. The nearest Sume settings are this-month and last-3-months. Treat the metrics as a shortlist for your own review, not as a verdict on why a video worked.

How do the start frame and the clip work?

An avatar is created once, from a prompt, structured traits or a reference photo, and then reused by handle. A talking video takes a script or ordered scenes, and Sume accepts an estimated length of 4 to 60 seconds. The aspect ratio defaults to 9:16, and the resolution is 720p.

To check the start frame before you pay for the full render, create an avatar video preview first. It generates only the first-frame still. Preview stills do not depend on the quality tier, so you can approve one and then render at a different tier without a new preview. Changing the script, avatar, scene or aspect ratio needs a new preview.

• Use one avatar per final video; the current execution expects one resolved avatar and one shared scene.
• Store captions on the preview request if you want them applied when the final video is generated.
• Send every media field as a public HTTPS URL.

Where does an AI presenter stop being the right tool?

The VideoTok guide repeats the FTC point that a creator cannot describe a product experience they did not have, and that AI UGC should not imply false customer results. ChatbotX draws the same line: use AI UGC for testing, speed and localization, and use real creators when lived experience or testimonial proof is the point.

In practice, script the avatar as a presenter who demonstrates or explains the product. Do not script it as a customer reporting results. Finish by placing your real product still over the clip with timeline compose, which stacks or overlays one hosted still and one hosted video into a single MP4.

What do I run first?

Run one research search, create one avatar, and approve one preview. Then render two variants that differ in a single line of the script, which is the single-variable test the playbook asks for. Give each request its own Idempotency-Key so a retry cannot create a second clip, and poll the job until it completes.