エピソード

  • The AI That Calls Abandoned Carts for 65 Cents

    The AI That Calls Abandoned Carts for 65 Cents
    2026/11/12

    Andreas Filippov's AI calls customers the moment they abandon a cart, and it's recovering 15 to 20% of them for about 65 cents a call.

    Andreas Filippov joins Matt Edmundson on The eCommerce Podcast to talk about Loop Voice, the AI voice platform he's spent the last six months building. Loop phones customers who abandon their cart, works through whatever's holding them back in a live conversation, then texts a direct checkout link to finish the sale. Across more than 100,000 calls, Andreas and Matt get into the economics of the channel, whether customers mind talking to AI, and where voice fits (and doesn't) in a brand's wider marketing mix.

    Key points:

    06:54 Should Customers Be Told They're Talking to AI

    12:44 What 100,000 Calls Taught Loop Voice About Customers

    13:48 How the Abandoned Cart Call Actually Works

    31:08 Where AI Voice Calling Should Not Be Used

    Should Customers Be Told They're Talking to AI

    [06:54] Matt pushes on the one thing that bothers him about AI voice calling, that he feels duped if he's not told he's talking to a machine. Andreas's answer is that most brands don't lead with it either way.

    "Most of our customers don't really brand it as AI."— Andreas Filippov

    The agent introduces itself with a name, like "Lisa from Brand XYZ", and gets on with solving the customer's problem. Andreas says complaint volume about the calls has stayed low across more than 100,000 calls.

    What 100,000 Calls Taught Loop Voice About Customers

    [12:44] Beyond recovering the sale, Andreas has found a research use nobody built the product for directly. Founders who've shipped millions of units and never once heard a customer's actual voice can now listen back to real conversations.

    "We have a full conversation with the customer on what their pain point is, how they feel, at what point they thought, okay, I need to solve it."— Andreas Filippov

    Those recordings surface specific product, pricing and messaging objections that a tick-box post-purchase survey rarely catches.

    How the Abandoned Cart Call Actually Works

    [13:48] The mechanics are simple. Loop waits around ten minutes after a cart is abandoned, then calls.

    "We've seen you have an open cart, you haven't finished the purchase. Is there anything that came in between you and finalising the purchase?"— Andreas Filippov

    Once the objection is handled (sizing, shipping, a checkout hiccup), the agent texts a direct checkout link so the customer can finish in one tap. At scale, Andreas says 500 abandoned carts a month at a 15 to 20% recovery rate and a $50 average order value works out to roughly $2,500 a month in recovered revenue, for about 65 cents per answered call.

    Where AI Voice Calling Should Not Be Used

    [31:08] Matt asks the obvious follow-up question, could this be pointed at cold outreach, calling people who've never bought anything at all.

    "That would be a big issue in the ecommerce space doing that."— Andreas Filippov

    Andreas draws a hard line between calling someone who's already shown buying intent (an abandoned cart) and cold-calling a list of strangers, which he thinks damages trust in the channel for everyone.

    Today's Guest

    Andreas Filippov is a fractional CMO with seven years in ecommerce, starting as a media buyer in 2019 before moving into head of growth and fractional CMO roles for eight-figure Shopify brands including Moon Magic and Mother's Earth. For the last six months he's been building Loop Voice, an AI voice platform for Shopify brands that recovers abandoned carts, runs post-purchase upsells and handles inbound customer support by phone.

    • Website: loopvoice.ai
    • LinkedIn: Andreas Filippov

    Episode link: https://www.ecommerce-podcast.com/ai-abandoned-cart-recovery-andreas-filippov
    続きを読む 一部表示
    48 分
  • How to Get Your Products Recommended by ChatGPT

    How to Get Your Products Recommended by ChatGPT
    2026/11/05
    Jim Boudreau has studied SEO since 2001, and even he had to ask ChatGPT what was actually in the box.Jim is the founder of CataSEO, an AI tool built in partnership with Moz that optimises ecommerce catalogues for Google and for AI search engines like ChatGPT, Perplexity and AI Overviews. In this conversation with Matt Edmundson, Jim explains why catalogue-scale SEO breaks the old playbook. Once a store has hundreds or thousands of product pages, hand-crafted content stops being realistic, and copying the manufacturer's description starts costing visibility everywhere that matters. Jim walks through what actually moves the needle for AI search. That means tight metadata, structured data, alt text that describes the image rather than the product, and FAQ sections built to answer the questions customers (and AI) are actually asking. He also tells two very ordinary shopping stories, buying a Microsoft Surface and a pair of monitors, that show exactly what goes wrong when that groundwork is missing. Along the way, Jim and Matt get into using AI as a co-writer without losing the original thought that only comes from years spent working inside a category.[09:29] Why SEO Breaks Down Once You've Got Hundreds of Products[39:16] The Keyboard That Wasn't in the Box[15:19] Metadata, Alt Text and the Category Pages Everyone Skips[25:11] Writing With AI Without Losing the Original ThoughtWhy SEO Breaks Down Once You've Got Hundreds of ProductsJim has been in ecommerce since 1998 and studying SEO specifically since 2001, long enough to have heard every guaranteed-page-one SEO pitch going, and to know none of them worked that way. The real problem, he tells Matt, is scale. Writing a strong product description for a handful of SKUs is realistic. Writing one for a few hundred, let alone a few thousand, is not, and that is exactly the level of detail AI search rewards."When you have hundreds or thousands of products, it's impractical to do SEO on a product level." - Jim Boudreau, [09:29]That gap is why Jim built CataSEO. Rather than trying to win from the top down with pillar pages and landing pages, it works from the bottom up, product by product, filling in the metadata most stores leave half populated.The Keyboard That Wasn't in the BoxBuying a couple of monitors from the biggest electronics retailer in the US, Jim could not find which cables were included in the box, so he bought extra ones just to be safe. Researching a Microsoft Surface for a friend, he ran into the same problem on a bigger scale. Every product photo on the manufacturer's site and on the retailer's site showed the Surface with its keyboard attached. Nowhere did either site say the keyboard was sold separately."You can't find anywhere except ChatGPT that this thing doesn't include the keyboard." - Jim Boudreau, [40:32]If the biggest brands in the world cannot get something this basic right on their own websites, most stores still have work to do, and the ones that fix it are the ones AI ends up recommending.Metadata, Alt Text and the Category Pages Everyone SkipsAsked what store owners can do without buying any tool, Jim keeps it practical. Do not copy the manufacturer's description. Even a quick rewrite with ChatGPT is better than lifting the original word for word. Keep URL slugs and page titles tight. And fix the small things almost nobody gets right, like alt text."What an alt tag is supposed to do is actually describe the picture, not the product." - Jim Boudreau, [16:18]Category pages get the same treatment. Jim's advice is to keep products above the fold, since that is what customers came for, while still giving Google, Bing and the AI engines a real description to read rather than an empty template field. Blog posts matter too, as long as they answer a genuine customer question instead of reading like a sales pitch.Writing With AI Without Losing the Original ThoughtJim is drafting two books at the same time, both with AI doing a lot of the heavy lifting. His view is that AI can handle the research and the structure, but not the insight that comes from years spent watching a category change."It can go scrape the internet and come up with the 95% of knowledge. But it's not going to come up with that innovation, that perspective, that foresight that you do as an expert." - Jim Boudreau, [26:43]Used that way, Jim says, AI becomes a co-writer rather than a content generator, drafting from a detailed brief, his sources, his points and his own style."It's not doing it for you. It's doing it with you." - Jim Boudreau, [30:45]Today's GuestToday's guest: Jim BoudreauCompany: Studio 1119, Inc. (maker of CataSEO)Website: cataseo.aiLinkedIn: Connect with Jim on LinkedInEmail: info@cataseo.netAbout The eCommerce PodcastThe eCommerce Podcast helps you deliver eCommerce WOW with real talk about building online stores. Every Thursday, Matt Edmundson talks to experts and founders who have been in the trenches, built the stores, and learned the hard way, so you...
    続きを読む 一部表示
    49 分
  • Are You Running Ads to Products People Can't Buy?

    Are You Running Ads to Products People Can't Buy?
    2026/10/14
    Ecommerce Merchandising, and How to Stop Advertising Sold-Out StockMost stores are quietly leaking sales, and ad spend, to products customers literally cannot buy. In this episode Matt Edmundson talks merchandising with Leon Bart-Williams, co-founder and CEO of Akikumo, a merchandising-intelligence app for Shopify brands. Before Akikumo, Leon built the first shoppable-TV commerce experience at NBCUniversal and helped hundreds of brands scale on Amazon, so this is a proper operator conversation about the part of ecommerce almost nobody owns.SummaryMerchandising is the most overlooked driver of conversion, and it is far more than shop-window mannequins. Leon defines it as putting the right product in the right place at the right time in front of the right audience, with the right inventory to satisfy demand. It is the process of turning inventory into sales.The conversation works through where stores go wrong. Out-of-stock products sitting in the prime slots of a collection. Ads pointing at sold-out or partially sold-out products. A failure to understand which variant is actually driving the sales of a given product. From there Leon and Matt move into curation over volume, why more choice tends to sell less, A/B testing collections rather than just the checkout, and inventory planning that starts with sales velocity at the variant level. Practical throughout, with real brands named and a ten-minute daily habit any store can start tonight.Merchandising Is the Job Nobody in Your Store Is DoingMerchandising usually gets wrapped in grandiose language. Leon's definition is plain."Merchandising is really the process of taking your inventory and converting it to sales." — Leon Bart-WilliamsThe point most brands miss is that merchandising applies to the ecommerce storefront, not just the physical store. The first positions in a collection do the same job as the front window of a shop."Think of the first 10 slots of your website as the moment people walk through the door of a physical store and see a mannequin with nothing on." — Leon Bart-WilliamsOut-of-stock products in the first ten positions of bestsellers or new arrivals are the fastest, most common leak.The problem starts biting at around 400 to 500 SKUs, not tens of thousands, especially in apparel where size, colour and style variants multiply.It is almost always human error and broken team communication, not strategy. As Leon's co-founder Jared (previously a principal engineer at Figs and Chewy) puts it, most of ecommerce is just fixing human errors and self-inflicted wounds.Are You Running Ads to Products People Can't Buy?A costly and common mistake is ad spend hitting stock that is already gone. Leon describes checking a friend's brand, Johnny O, in Meta's Ad Library and spotting a top-selling product with three quarters of its variants sold out, sitting in the top ad slot. One email to their ads team fixed it.The subtler version is the partially sold-out product, and this is where the 80/20 rule matters at the variant level."The dress in blue, size 2, is driving 80% of the sales of this specific product. You need to know that, and you need to at least be fully stocked in that product." — Leon Bart-WilliamsRoughly 20% of a product's variants drive roughly 80% of its sales.If that core variant is out, the listing can look in stock while the ad is effectively pointing at a sold-out product.When a product is gone but the creative still works, point the ad to the collection, or to the category, rather than a dead page. When in doubt, point to the collection.Why More Products Means Fewer SalesLeon and Matt agree that curation beats volume, and the science backs it up. Matt recounts the jam experiment, where a display of six jams sold well and a display of thirty-six exotic jams saw sales fall. More choice created paralysis, not more buying."More does not mean more sales." — Leon Bart-Williams"People want fewer options. People want to be told what to do." — Leon Bart-WilliamsKeep collections tight. Around 50 SKUs or fewer per collection is manageable and tends to convert better.Curation is part art, part science. The science is how many products. The art is telling a story around a hero product, the dress plus the handbag plus the shoes, and writing content around it.Brands Leon rates for this include Cara Cara NYC, and his customer Brooke at Simply Moab and Desert Soul in Moab, Utah, who matches her ecommerce experience to her physical stores.Endless scroll with no curation is the opposite of this, and it kills conversion.Test Everything, and Plan Inventory at the Variant LevelTesting should not stop at the checkout. Leon argues for A/B testing collections, layouts, price points and new products, each with a hypothesis and a finite goal."You can't optimise your way to success all the time." — Leon Bart-WilliamsThe counter-example is Vince at the Texas golf brand Bogey Brothers, who built sale pages by hand from memory, guessed the pricing, ...
    続きを読む 一部表示
    49 分
  • How to Actually Use AI in Your eCommerce Business

    How to Actually Use AI in Your eCommerce Business
    2026/10/22
    AI will not run your ecommerce store on autopilot. Cyril Golub joins Matt Edmundson to explain how to actually put it to work, as a supervised assistant you give a clear goal, not a magician you hand the keys to. It is a practical, often funny conversation about where AI genuinely helps online sellers and where it just burns tokens.SummaryCyril Golub has been building ecommerce software since 2003, so he watches the AI wave with the eye of someone who has seen a few hype cycles come and go. In this episode he and Matt dig into what AI changes for online sellers and what it quietly leaves exactly the same. Cyril's thesis is that "ecommerce" is turning into "AI commerce", where by early 2027 both buyers and sellers have their own agents, and "how do I start selling online?" becomes a single sentence typed to an agent.Along the way they get practical. Most AI startups are building for buyers, leaving the seller side wide open. Marketplaces own your customer data, and on Amazon a buyer is one click from a competitor. Matching the right model to the job matters more than reaching for the biggest one every time. And competitor monitoring at scale lets you act alongside rivals rather than react to them.Key points covered in this episode:Who's building AI for sellers, not just buyers (10:25)Who really owns your customer data (17:41)Which AI model you should actually use (26:55)Using AI to monitor your competitors at scale (31:43)Who's Building AI for Sellers, Not Just Buyers (10:25)Cyril points out something surprising about where the money is going. Roughly three quarters of AI-native ecommerce startups are building buy-side tools, helping shoppers find bargains and streamline their buying. The seller side, by contrast, is far less crowded, which he finds odd given how much easier it usually is to sell to businesses than to chase millions of individual consumers.That gap raises a harder question for any small founder building in the space. If a tool genuinely helps sellers increase their margin, who gets to build it? The giant sitting on terabytes of data, a mid-size player, or a small third party with neither the data nor the audience?"The main question that early-stage investors are asking to startup founders, what if, then put the name of the huge company. What if Amazon will develop it on their own, directly in-house, and then roll out to all of their hundreds of thousands of customers?" — Cyril GolubMatt's hope is that the answer turns out to be the small operator in the garage rather than another platform giant, which is exactly the kind of story this show roots for.Who Really Owns Your Customer Data (17:41)The storefront, Cyril argues, is a habit inherited from the offline world, the digital echo of goods on a shelf. But not every visitor wants to browse a shelf. Some arrive with a clear intention and want to reach checkout in two or three clicks. Others arrive to explore, unsure what they want yet. The problem is that most sellers know almost nothing about which is which, often only the channel or the ad that brought someone in, and rarely anything that ties through to the actual purchase.That blindness is the real cost of not owning your data, and on a marketplace it gets worse."When we are talking about Amazon sales, your buyer is just one click away from buying from your competitor. That's a huge problem." — Cyril GolubHis practical steer is to pull as much data as you can out of whatever platform you sell on, through the richest API access available, so you can connect your own agent and ask it for actionable insight rather than just advice.Which AI Model You Should Actually Use (26:55)This is where the conversation gets genuinely useful for anyone watching their AI bill. Cyril's view is that you should match the model to the job, the same way you would match a hire to a task."It might be like if you are hiring a PhD to just calculate simple numbers in your Microsoft Excel sheet. You must choose the exact type of model that is quite enough, but not more, for your particular goal." — Cyril GolubHe gives a concrete number. Connecting Claude to an Amazon store through the SP-API to process an inventory of around 2,000 SKUs costs roughly 4 million tokens, mostly input. That is far too simple a job to hand a frontier model. His answer is to think in layers:A deterministic layer that prepares the data quickly. If a task is fully deterministic, it is just software, no AI needed.A cheap supervisor model to decide which parts of the data actually matter for the goal at hand.A frontier model reserved only for the work that genuinely needs judgement.Matt notes he is seeing this layered thinking take hold more widely as people wake up to what AI actually costs. Cyril adds a warning about context too, in that the right answer moves over time. What was true a week ago, around a season, a sale or a competitor's campaign, can be wrong the next week.Using AI to Monitor Your Competitors at Scale (31:43)...
    続きを読む 一部表示
    50 分
  • Stop Using Generic AI In Your Ecommerce Business

    Stop Using Generic AI In Your Ecommerce Business
    2026/10/29
    ChatGPT once invented a 20% advertising budget for a founder who hadn't made a single sale yet. Marissa Rodriguez explains why that happens, and what she does differently.Marissa Rodriguez has spent twenty years building and scaling direct-to-consumer brands, first as a fractional CMO and COO, now as the founder of Through Experience, a mentorship platform for ecommerce founders. She talks to Matt Edmundson about the risk of handing serious decisions to generic, industry-agnostic AI tools, and why she thinks the real opportunity lies in training AI on the judgement of people who have actually done the job. She shares how her own email marketing training module grew from six videos to fifty once she tried to capture everything she actually knows, and the moment a founder's P&L came back with a 20% advertising budget that ChatGPT had invented from nothing. Along the way, they cover decision fatigue, shiny object syndrome, and where Marissa tells founders to focus first when they're starting out with AI in their business.08:22 - Why Generic AI Puts First-Time Founders at Risk15:40 - Training AI With Her Own Frameworks29:33 - Decision Fatigue and Shiny Object Syndrome40:06 - ChatGPT Invented a 20% Ad BudgetWhy Generic AI Puts First-Time Founders at Risk (08:22)Matt asks Marissa the question she gets most often, what's going wrong for people right now with AI. Her answer isn't about a specific tool or tactic. It's about who's giving the advice. Most ecommerce founders are first-time founders with no deep domain expertise to fall back on, and a generic AI tool doesn't know what it doesn't know. It won't ask the qualifying questions a human expert would, and it will hand over confident answers regardless.Marissa has watched this happen with financial models and P&Ls built in ChatGPT, where the tool makes assumptions about costs and budgets without ever checking whether they're right for that particular business."Having a generic AI working with you, I think you're gonna end up in a bad place."Marissa RodriguezHer point isn't that AI is the problem. It's that AI without operator judgement built into it is the problem, and that gap is exactly what she thinks founders need to close.Training AI With Her Own Frameworks (15:40)Marissa began mapping her own expertise into an AI the same way Matt describes doing with SAM, Aurion's internal AI layer built on their own playbooks. She assumed it would be straightforward. She already had a structured curriculum from years of building courses and digital products.It wasn't straightforward. One email marketing module alone, originally six videos and forty five minutes, grew into fifty videos once she tried to capture every exception, every variation by AOV (average order value) or customer path to purchase, every moment she'd normally break her own rule because the context called for it."It went from six videos to fifty videos, because there was so much more than I anticipated that I actually had to sit down and give it the context for understanding."Marissa RodriguezThat's the gap most generic AI tools skip. The judgement isn't a setting you switch on, it's decades of decisions that nobody writes down until someone asks them to.Decision Fatigue and Shiny Object Syndrome (29:33)Founders are surrounded by noise, Marissa says, a constant stream of other people's wins and warnings on LinkedIn and Instagram, most of it stripped of the context that would tell you whether it actually applies to your business. Her advice is to narrow the focus deliberately rather than chase every new tactic."There's only three actions a person can take on the internet that matter to an ecommerce business. Click, subscribe, and buy."Marissa RodriguezEverything else, in her view, is in service of those three actions. Paid ads and content drive the click, email marketing earns the subscribe, and the operations behind the scenes, inventory, reporting, unit economics, make the buy possible. She'd rather be great at a handful of things that compound predictably than good at everything and consistent at nothing.ChatGPT Invented a 20% Ad Budget (40:06)A founder came to Marissa with a P&L that ChatGPT had built for them. It looked complete. It wasn't. The advertising line assumed 20% of revenue, with nothing behind the number, for a business that was pre-revenue and hadn't launched yet."ChatGPT just made up the paid advertising line. ChatGPT just decided like 20% is what you need for paid advertising."Marissa RodriguezMarissa's challenge back to the founder was simple, on what basis. A properly trained financial model, in her experience, knows which questions it can answer from publicly available information and which ones it has to ask the founder directly, the weight of a product before it estimates shipping, the likely return rate for that specific category. Generic AI skips the asking and goes straight to the guessing.About Marissa RodriguezMarissa Rodriguez is the Founder & CEO of Through Experience,...
    続きを読む 一部表示
    48 分
  • He Fed 500,000 Emails Into an AI Copy of Himself

    He Fed 500,000 Emails Into an AI Copy of Himself
    2026/10/07
    Drew Morgans spent 15 years and around $50,000 building an AI trained on his own work, and it now helps run his Amazon agency.SummaryDrew Morgans, founder of the Amazon and marketplace agency Marknology, returns to the eCommerce Podcast for a conversation that is all about AI in the real world rather than the hype. Drew explains how he fed 15 years of his own work into an AI system, roughly 500,000 emails, around 300 podcast episodes, blogs, SOPs, playbooks and won and lost proposals, to build something that thinks the way he does. The pair get into the practical build behind it, why context is the hard part, how he keeps his running costs sane by using cheap local models for ingest and expensive frontier models only for client-facing work, and why he chose a single "super agent" over a crowd of specialist bots. Matt and Drew also compare notes on being founders who came up through code, stepped away, and have now picked it back up. It is a generous, concrete look at where AI actually pays off for an ecommerce operator, and where it does not.Turning 15 Years of Work Into an AI That Thinks Like HimDrew describes the shift from using AI for one-off tasks to building a system trained on his own history. He ingested roughly 500,000 of his own emails, around 300 podcast episodes, blogs, SOPs, playbooks and old proposals, turning past work into present value."I really needed a copy of me." - Drew Morgans"The real meat's in the emails." - Drew MorgansThe starting point was helping himself first, then rolling the tools out to his bookkeeper, design team and wider team.Old work he had already paid for in time and energy became a source of ongoing value, including 500 real examples of how he handled objections over the years.A QuickBooks-approved app now auto-drafts invoices from incoming email, always human-gated, never sent automatically.Why Context Is the Real ProblemDrew argues the genuine bottleneck is not the model, it is context. AI works sequentially, checking one source then the next, and never holds everything at once the way a close colleague would. His answer was to build the context in before the model ever runs.Drew IQ is a set of around 350 operator rules, a series of "if this, then this" checks that mirror what he already knows as an operator, plus tags that give each product context such as whether a SKU is a launch.WCS is a context engine that surfaces everything relevant to a brand in about half a second, locally, so the model arrives with the full picture already in front of it.Winston is his always-on system, built to be model-agnostic with fallbacks, so if one provider went down he could switch and keep running. He built it on OpenClaw, an open-source always-on agent framework, rather than tying himself to a single vendor.Cheap Model In, Expensive Model OutOne of Drew's most useful principles is matching the model to the job. He runs cheap local models via Ollama for the overnight grind of ingesting emails, transcripts and Amazon data, and reserves the expensive frontier models for anything client-facing or "trying to be me"."Cheap model in, expensive model out." - Drew MorgansAll the pre-work, the rules, the routing and the context surfacing, happens on his own machine, so it is effectively a free call.By the time the reasoning model is called, it is doing only the final, high-value layer, which keeps costs in check without dumbing down the output.Both Drew and Matt land on the same conclusion about tool-switching, that the opportunity cost of constantly jumping between systems is real, and there is more value in committing to one setup and getting genuinely good at it.One Super Agent Beats Twenty SpecialistsDrew tried the popular approach of building a crowd of named specialist agents, a CFO agent, a design agent and the rest, all chatting and arguing in Discord. It did not work as well as the alternative."I cut down the jungle, my team manicures the lawn." - Drew MorgansInstead of managing twenty different personalities, he invested his time in one super agent that knows where all his information lives and can genuinely help.Drew's view on where things are heading is that mediocre roles are in trouble, trades will grow in value, and strong operators will be fine, with a parallel movement back towards authenticity and in-person work.On Amazon specifically, he sees AI helping him do the deep, non-scalable work an agency normally cannot afford at scale, cross-referencing ranking, margin, inventory, reviews, Reddit and search query gaps to work out why a listing is not converting."It's helping scale the unscalable." - Drew MorgansToday's GuestDrew Morgans (Andrew Morgans) is the founder and CEO of Marknology, a full-service Amazon and marketplace agency based in Kansas City, working across Amazon, TikTok Shop, Walmart and Shopify. He also runs a warehouse and 3PL operation and a portfolio of brands. This is his second appearance on the show.Website: marknology.comLinkedIn: in/...
    続きを読む 一部表示
    54 分
  • Is It Time We Talked About Turning Off Your Ads? (And Yes, You Read That Right)

    Is It Time We Talked About Turning Off Your Ads? (And Yes, You Read That Right)
    2026/09/04
    Your best-ever ROAS month can still lose you money. Brenden Delarua explains why the number on your ad dashboard rarely matches the number in your bank account, and what to measure instead.SummaryBrenden Delarua spent 11 years in paid media, hitting and smashing ROAS targets across Meta, Google and Criteo, until a CFO asked him a simple question at a quarterly review. If the ad numbers are this good, why is revenue down? That question sent him into causal measurement, and he now runs Stella (Growth Intelligence), a marketing-measurement company that helps brands work out what is actually driving sales rather than what the platforms claim.In this conversation Brenden unpacks why platform-reported ROAS is a correlation metric that can overstate how well your ads are working, and how chasing it pushes marketers into retargeting and branded-search quick wins that flatter the dashboard without growing the business. He walks through the practical ladder any small brand can climb, from multi-touch attribution and post-purchase surveys up to full geo-holdouts and media mix models. Along the way there is a YouTube campaign that looked like a disaster, a lesson on why branded search is not always what it seems, and the one post-purchase survey question worth more than the rest.ROAS Is Not the Same Thing as ProfitBrenden's whole approach traces back to one uncomfortable review meeting. As a media buyer he had beaten every ROAS goal he was set, then presented record-breaking results, and still got questioned by the CFO."If you're showing us that ROAS is so high, why am I not seeing that in our actual bank account?" — Brenden Delarua, on the CFO question that started it allThe problem is what ROAS actually measures. It is a correlation metric. Someone saw or clicked an ad and later converted, so the platform claims the sale."ROAS is not correlated with profit, even though a lot of people think it is." — Brenden DelaruaBecause ROAS rewards clicks, it tends to deprioritise channels you cannot click, like connected TV or audio, and it tempts marketers to pile budget into retargeting or branded search for a quick win. The dashboard looks great. The business stops growing. Brenden's point is that what owners ultimately care about is profit, not a platform-reported number.The YouTube Campaign That Looked Like a Money PitOne story makes the gap between platform numbers and reality concrete. Brenden runs ads for a client spending heavily on YouTube, a channel infamous for a poor click-to-convert ratio.In-platform, the YouTube ROAS sat at around 0.2, so the account ROAS looked like it was suffering.The obvious call was to question why so much was going into a channel that appeared to be losing money.Then he pulled the post-purchase survey data from NoCommerce, and around 7% of buyers said they came from YouTube.Set against the spend, that share of revenue worked out at roughly a 3x return, and the channel was actually very profitable.The lesson is not that YouTube always works. It is that a single click-based number can hide the truth, and a second data point, in this case a qualitative survey, can change the whole decision."Sometimes increasing incremental ROAS is the opposite of increasing platform ROAS." — Brenden DelaruaWhat is Causal Measurement?Brenden is keen to stress that incrementality is not just a big-brand concern, and you do not need to turn off ads to benefit from the thinking. He lays out a practical ladder that scales with the business.Start with proper multi-touch attribution. For e-com tools he rates Triple Whale and Northbeam. It maps the customer journey by stitching anonymised touchpoints together until someone converts, then shows first touch, last touch or a fairer linear model.Add post-purchase surveys. Even a brand-new store can ask buyers where they came from and, more importantly, why they bought. Tools like NoCommerce can use conditional logic to ask which specific ad someone remembers.Add causal analysis when you are big enough. Once sales volume is there, usually for brands over roughly 10 million a year running on two or three platforms, holdout testing and media mix models come into play."You don't have to do holdouts to understand that what's causing sales to happen might be different from what you're seeing in ad platforms." — Brenden DelaruaFor a brand doing a million a year, the same issues exist at a smaller scale. The trouble is that as you grow, working out what is really driving growth gets harder, not easier.What are iROAS and Holdouts?For brands ready to test properly, Brenden explains the terms plainly. A holdout study means turning ads off in certain geographical regions and watching your source-of-truth revenue, usually Shopify, against the regions that stayed live. The measure that comes out of it is iROAS."iROAS stands for incremental return on ad spend, every dollar we put in that gives us back money that would not have happened without that investment." — Brenden ...
    続きを読む 一部表示
    53 分
  • How to Sell on TikTok Shop With AI

    How to Sell on TikTok Shop With AI
    2026/09/30
    How do the best brands actually win on TikTok Shop, and where does AI genuinely earn its keep in the process?SummaryMatt 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 does07:36 The biggest mistake we are all making with AI08:33 Why you have to show AI what good looks like15:06 First steps on TikTok Shop before you touch AI17:42 Using AI to analyse video, lighting and hooks25:03 Frontier models versus the agents that connect them28:27 Claude Code plan mode and turning tasks into skills32:38 The hundred dollar versus two thousand dollar task44:12 Why consistency and volume win on TikTok Shop48:25 Treating your affiliates like a sales team50:43 The window fans affiliate making a million a year55:38 Where to begin with TikTok Shop and AIShow AI What Good Looks Like08:33The 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 SankaHis 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 SankaWorking 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 AI15:06Plenty 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 Copy17:42TikTok 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 On28:27Sohun's view is that prompt engineering is already on the way out."You don't need to learn how to prompt anymore." - Sohun SankaClaude Code plan mode interviews you instead. Ask for a report and it works ...
    続きを読む 一部表示
    58 分