TIKTOK SHOP · AUG 13, 2026 · 7 MIN
How to Extract Winning Angles From Competitor TikTok Shop Content With AI
The five-step AI workflow I presented on TikTok Shop UK's webinar: pull competitor videos through Cruva, score them with Gemini, and roll the recurring hooks and selling points into a Slack insights channel.
You can extract the winning angles from a competitor's TikTok Shop content with a five-step AI workflow: pick your models, select the brands and product IDs to monitor through Cruva, download the top 30 videos per product with yt-dlp, score each video with Gemini across hooks, selling points, lighting, and videography, then run a script that rolls the patterns appearing 66 percent of the time or more into a Slack insights message.
Key takeaways:
- Split the work across models. Claude runs the workflow, Gemini 2.5 Flash watches the videos, Gemini 2.5 Pro handles only the hard ones.
- Cruva supplies the Marketplace data, so you analyze the videos that actually sold rather than the ones that merely went viral.
- Cap retries at 3 per video so a broken file cannot quietly burn your token budget.
- The output is a Slack message, not a spreadsheet. The insights land where the creative team already works.
This morning I was invited to present on TikTok Shop UK's webinar. The topic was how to use AI to extract the winning angles from your competitors' TikTok Shop content. This is the workflow I walked through, written up for anyone who could not join live.
Why analyze competitor content instead of your own?
Your own creative library tells you what has worked for you. Your competitors' libraries tell you what is working in the category right now, including angles you have never tested. On TikTok Shop, the affiliate videos driving a competitor's GMV are public. The data on which of those videos actually converted is available through a tool like Cruva. Put those two facts together and you have a research pipeline most brands never build because it looks like manual work.
The manual version is watching hundreds of videos and taking notes. The AI version is below. It produces the same kind of insight, but it runs on a schedule and it does not get bored halfway through the set.
What is the workflow?
1. Select your AI model
Different steps need different models, and using one model for everything either wastes money or produces weak analysis.
- Use Claude Opus at medium reasoning effort as the orchestrator. It runs the scripts, talks to the MCPs, and writes the final message.
- Use Gemini 2.5 Flash for video analysis. It watches the video and returns the scored breakdown.
- Use Gemini 2.5 Pro only for tough-to-analyze videos, where Flash returns something thin or inconsistent.
- Use Cruva to pull in TikTok Shop Marketplace data, which is how you know which videos are worth downloading in the first place.
The reasoning effort setting matters more than people expect. Medium is enough for an orchestration task like this. Cranking it higher slows the run and does not improve the insight, because the analytical work is happening inside Gemini.
2. Select the brands you want to monitor
For the webinar we selected Medicube, Dr. Melaxin, and Beauty of Joseon, three K-beauty brands competing for the same shopper.
Before anything else, run a simple API pull on L30 day GMV for each brand to test the connection. If that call fails, nothing downstream will work, and it is better to find out in the first minute than after the video downloads.
Then select the PID, the product ID, for each competitor product you want to analyze. Analysis at the product level is what makes the output usable. A brand-level rollup blends the angles from a cleanser and a serum together, and the creative team cannot act on that.
Pull L30 day GMV for the Medicube, Dr. Melaxin, and Beauty of Joseon shops. Confirm the connection, then list the top PIDs by GMV for each so I can choose which to monitor.
3. Download the winning videos
Connect open source code like yt-dlp to Claude. Claude writes the download script, you approve it, and it runs against the list of video URLs Cruva returned for each PID.
Two limits keep this step under control:
- Download 30 videos per PID. Patterns repeat well before that point, and going deeper mostly adds videos that did not sell.
- Cap retries on analysis at 3 per video. A corrupted file or a video that Gemini cannot parse will otherwise retry forever, and every retry costs tokens.
4. Analyze and score the videos
Run the downloaded videos through Gemini. The prompt asks it to score each video across components such as hooks, selling points, lighting, and videography, and to describe what it saw for each component in plain language.
Watch this video and score it on hook, selling points, lighting, and videography. For each component, describe specifically what the creator did. Return the result as structured JSON.
The structured output is important. The next step is a script, not a person, and it needs consistent fields to compare across 30 videos per product.
5. Roll up and send the analysis
Run a script to identify the commonalities behind the winning creative. The rule I use: patterns that appear across hooks or selling points 66 percent of the time or more roll up into a key insights section at the top of the message. Everything else sits below it as supporting detail.
Create a Slack channel for these insights and push the message via the Slack MCP. The creative team gets a short summary of what the category's winning videos have in common, per competitor product, without opening a dashboard.
What does the output look like in practice?
For each monitored product, the Slack message reads roughly like this:
| Section | What it contains |
|---|---|
| Key insights | The hook and selling-point patterns that cleared the 66 percent threshold |
| Component scores | Average hook, selling point, lighting, and videography scores across the 30 videos |
| Notable outliers | Videos that scored high on one component while the rest of the set did not |
| Source list | The video URLs and the Cruva GMV data behind them |
The creative team reads the key insights and briefs creators against them. The rest exists so anyone who wants to check the reasoning can trace it back to the videos.
Where does this sit in the stack?
This workflow is the research front end for two things Clankers has written about before. The output feeds directly into a creative brief built with Claude Code, Cruva, and Gemini, because the winning angles are exactly what a brief needs to tell a creator. It also pairs with a competitive analysis run through the Cruva MCP, which covers the GMV and creator side of the same competitors.
If you want this running on your competitor set, the Clankers 90-Day Revamp installs it inside your stack alongside the rest of the creative pipeline.
Frequently asked questions
Which AI models do you use to analyze competitor TikTok Shop videos?
I use Claude Opus at medium reasoning effort to run the workflow, Gemini 2.5 Flash for video analysis, and Gemini 2.5 Pro only for videos that are tough to analyze. Cruva pulls the TikTok Shop Marketplace data that tells me which videos to download.
How many competitor videos do you need to analyze?
I download 30 videos per product ID. That is enough to see patterns repeat across hooks and selling points without burning tokens on a long tail of videos that did not perform.
What does the video scoring cover?
Each downloaded video runs through Gemini and gets scored across components such as hooks, selling points, lighting, and videography. The scores are what the rollup step uses to find commonalities behind the winning creative.
How do you decide which patterns count as key insights?
A script looks for patterns that appear across hooks and selling points 66 percent of the time or more. Those roll up into a key insights section of the Slack message. Anything below that threshold stays in the detail rather than the summary.
Can this run without a developer?
The workflow uses open source tooling like yt-dlp connected to Claude, plus the Cruva and Slack MCPs. The setup is mostly prompts and configuration, and the scripts Claude writes for you. It helps to have someone comfortable in a terminal, but it does not need an engineering team.
Written by Sohun Sanka, founder of Clankers, an operator practice that builds AI systems and automation for social-commerce brands and agencies inside the tools they already use. This post expands on a LinkedIn post Sohun published on July 22, 2026.
Turn competitor research into a weekly habit
If your creative team is still guessing at angles, this workflow replaces the guess with a scored, repeatable read on what the category is rewarding right now. The Clankers 90-Day Revamp diagnoses where your creative pipeline loses time, then builds and transfers the systems worth owning.