How to Sell on TikTok Shop With AI

How do the best brands actually win on TikTok Shop, and where does AI genuinely earn its keep in the process? ## Summary Matt Edmundson talks with Sohun Sanka, founder of Clankers and former head of growth at the TikTok Shop outreach tools Euka and Reacher. Having audited more than 2,000 affiliate programmes and seen the backend of over 150 agencies and 1,000 brands, Sohun has a clear read on what separates the brands that win on TikTok Shop from the ones quietly burning through samples. The conversation moves from the biggest mistake most of us make with AI (never showing it what good looks like) to the practical mechanics of creator commerce. Sohun explains why top affiliates post 15 to 21 times a week, why lighting matters more than copy in the first two seconds of a video, and how the best brands treat their affiliates like a sales team rather than a mailing list. On the AI side, he walks through Claude Code plan mode, turning repeat tasks into skills, and the single connection almost nobody has switched on that would save agencies 15 or more hours a week. - 03:14 Who Sohun is and what Clankers does - 07:36 The biggest mistake we are all making with AI - 08:33 Why you have to show AI what good looks like - 15:06 First steps on TikTok Shop before you touch AI - 17:42 Using AI to analyse video, lighting and hooks - 25:03 Frontier models versus the agents that connect them - 28:27 Claude Code plan mode and turning tasks into skills - 32:38 The hundred dollar versus two thousand dollar task - 44:12 Why consistency and volume win on TikTok Shop - 48:25 Treating your affiliates like a sales team - 50:43 The window fans affiliate making a million a year - 55:38 Where to begin with TikTok Shop and AI ## Show AI What Good Looks Like *08:33* The first mistake Sohun sees is that almost nobody has their processes written down clearly, and their data is rarely clean enough for AI to learn the business. Feed a tool like Claude Code or Codex nothing, and it hands back what he calls AI slop. > *"You don't even have a folder on what good looks like, your brand colours and stuff. You're not feeding it anything." - Sohun Sanka* His fix is to onboard AI the way you would onboard a person. A new employee gets 90 days, a meticulous onboarding guide, multiple calls a week and time to shadow. Most people give AI two sentences and then conclude it does not work. > *"Think about how much context a human gets when they're signing on for a new job, and then they give nothing to AI." - Sohun Sanka* Working out what good looks like is uncomfortable precisely because it forces you to define processes most of us keep vague. As Matt put it, the reason we do not tell AI what good looks like is often that we are not sure ourselves. ## Win on TikTok Shop Before You Touch AI *15:06* Plenty of the early wins have nothing to do with being AI-forward. Sohun's first moves are the unglamorous ones. - Pull the copy from your highest performing Meta ad angles and landing pages, then feed it into a creative testing brief for affiliates. - Mine your last 30 days of Amazon reviews for one-liners customers actually use, and hand those to creators. - Use social listening to find creators already talking about your product, then reach out personally rather than starting a slow sampling process. - Benchmark against comparable brands. Tools like Kruva scrape and index every video and its KPIs daily, so you can pace yourself against five similar brands and use that as a working definition of good. The context most people miss is scale. There are roughly 4.5 million TikTok Shop sellers in the US, and only about 1.6% are making a couple of hundred dollars in the last 30 days. That gap is an education problem, and it is exactly where AI can help. ## Use AI to Read Your Video, Not Just Your Copy *17:42* TikTok affiliates are not polished UGC creators. They often miss the softer craft, like lighting and camera setup, and on a fast-scrolling feed a slightly off frame can cost you the view before a word is spoken. Sohun uses Gemini to break a video into frames and analyse the first two seconds, pulling out the lighting style so creators can be coached with real examples rather than vague advice. Cost is the catch, because generating a transcript and classifying every hook frame by frame gets expensive across a thousand videos. Tools like Kruva already pass through transcript, hook and lighting analysis, which lets him reduce the Gemini spend by three to five times by only analysing the final slice of each video with his own framework. The payoff is reuse. Inside GMV Max you can pull the two-second and six-second video view rates, find every creative clearing a 33% hook rate, analyse what those openings share, and splice those hooks back into your Meta ads or even Amazon sponsored display. ## Plan Mode, Skills and the Connection Nobody Switches On *28:27* Sohun's view is that prompt engineering is already on the way out. > *"You don't need to learn how to prompt anymore." - Sohun Sanka* Claude Code plan mode interviews you instead. Ask for a report and it works through which brand, which KPIs, what output format, whether it should run as a cron job, and whether it lands in Slack for review or goes straight to a client. When you find yourself doing that 10 to 15 times a week, you turn it into a skill and run it on demand. The bigger miss sits underneath all of it. Most brands and agencies have never connected their TikTok Shop tool to Claude as a connector, even though these tools expose around 110 MCP tools. One connection turns the model into an expert on the platform, yet agencies are still spending 15 to 17 hours a week on manual reporting that could be automated. This ties back to how Sohun runs his own day. He puts a dollar value on his time and sorts tasks into hundred dollar, five hundred dollar and two thousand dollar an hour work, then pushes the cheap tasks to AI or to someone on Upwork. ## Treat Your TikTok Shop Affiliates Like a Sales Team *44:12* On TikTok Shop the follow-through is human. Where a performance marketer lets the ad mechanism do the work, here someone has to sit with creators, coach them, share the data and keep them posting. > *"The top affiliates are posting 15 to 21 times a week per brand." - Sohun Sanka* Volume matters because the algorithm is getting very good at matching niche interests to the right moment, and there simply is not enough content in the US to feed it (China's feed runs at roughly 60% shoppable content by comparison). The brands getting this right build real communities. Javy Coffee grew TikTok Shop revenue from $300K to $1M a month with 15 community managers, each owning a book of affiliates and rewarded on the GMV growth they drove, making sure new affiliates post 30 times in their first 30 days. Then there is the story that anchors the whole episode. One of Sohun's friends is a top-100 affiliate who prints around $1M a year in commissions selling window fans. He runs a dedicated account posting the same fan under different lighting gradients, hundreds of videos, purely to A/B test lighting, then scrapes the top-100 videos each month into a Claude Code testing pipeline. > *"We A/B test copy. And I was like, nah, we need to A/B test lighting, because on TikTok Shop, the lighting matters so much." - Sohun Sanka* Matt recognised the pattern from his own son, a nutritionist with a large Instagram following who tests everything down to the camera angle, and once found that a tea towel over his shoulder lifted engagement. It is the same instinct ecommerce operators already have for split-testing headlines, applied to video. ## Where to Begin with AI on TikTok Shop *55:38* Sohun's single starting move is to connect your Claude or ChatGPT connector to your TikTok Shop outreach tool. That tool holds every affiliate activation you are running, so one connection lets the model learn all of it. From there he asks it to lay the whole flow out on a FigJam board, note what should be switched on and off, and show where the leaks are. You will still need TikTok Shop domain knowledge to sense-check what it says, but it is a fast way to see what is actually going on. Sohun's question for Matt's audience: what is the thing that scares you most about getting started with AI? Matt will be answering it on social, so come and join the conversation. ## Today's Guest **Sohun Sanka** is the founder of Clankers and previously served as head of growth (employee number one) at the TikTok Shop outreach tools Euka and Reacher. With a background in influencer and affiliate marketing, he now helps brands and agencies build custom in-house software with Claude Code and become genuinely AI-enabled across their TikTok Shop and creator commerce operations. - Website: clankersapp.com - LinkedIn: linkedin.com/in/sohun-sanka ## About the eCommerce Podcast The eCommerce Podcast helps you deliver eCommerce WOW. Every Thursday, Matt Edmundson talks with experts and founders who have built the stores and learned the hard way, so you can grow, scale and eventually exit your own business. - Website and newsletter: ecommerce-podcast.com - YouTube: youtube.com/@ecommercepodcast Like and subscribe wherever you listen, and sign up for the weekly newsletter to get the episode notes and AI prompt straight to your inbox. Episode link: https://www.ecommerce-podcast.com/how-to-sell-on-tiktok-shop-with-ai