AI in Ecommerce Strategy and Marketplace Innovation [New Frontier Podcast]

Max Sinclair: Hello everyone, welcome to another episode of New Frontier. We have with us this week Mike Danford, chief strategy officer at Adverio. He plays a pivotal role in driving innovation strategies for sellers on Amazon, Walmart, and Target. He has helped drive multi-million dollars in ecommerce sales and is renowned for his expertise in, you guessed it, AI. This week I am excited to be talking about lesser-known ways AI can help Amazon sellers. Mike, welcome.

Mike Danford: Max, thanks so much. Thanks for having me.

Max Sinclair: Do you want to add to the introduction at all?

Mike Danford: Sure. I have been in this space about 10 years, went through the Amazon or Amazing Selling Machine, got my feet wet in kitchen appliances and tools, and then realized it was not something I had much of a background in, so I went into supplements and graphic apparel and learned a lot. At the same time I was doing private label, I was also retail arbitraging Nike shoes and apparel, and that is where I started to have fun and understand how to manage larger catalogs with lots of variants and sizes and mitigate and take advantage of customer trends. One of my first internal hires was around advertising, and at that time it was simpler than now, but it was still one of my most expensive expenses, and I was like, this is numbers, I can figure this out. I posted some results on Facebook groups, folks started picking it up, and the rest is history with the agency.

Jo Lambadjieva: Let's dive in, this is what we do every episode, let's talk about AI. My first question is, how do you use AI at Adverio, what have you embedded within your agency processes?

Mike Danford: It is really in almost every department right now, internal and external use. We try to build things in house to answer our own questions and then figure out how to do those at scale. A lot of our goals are how can we do more with what we already have, and process larger amounts of data that we historically could not. We also like to test our bias against it, this is an untrained model, and this is what we have been training internally, and it helps surface those hidden biases and think about it in different ways. For me personally, it is getting involved with big data to understand what is going on, and NLP is an area I really like to leverage. There are fewer and fewer parts of the business that do not have some AI infused. It is about empowering more of the team to use it, have a sandbox. It moves really fast, faster than anything else I have dealt with, but it is great to see everyone going in head first and sharing what they learn.

Max Sinclair: On this speed, everyone compares AI to the internet, and someone told me it is not a fair comparison because AI moves 10X faster than the internet did. Do you have technical talent in the agency, do you see the future of these companies having a technical head of science?

Mike Danford: Yeah, we have some data science. A lot of it is having them focus on one code, one language, one pipeline. Now it is about how do we link two teams together without influencing the other, and understanding when to bring certain models in so we do not develop too fast or the cost is not too high. I have a dedicated full-time data scientist that works directly with me all day, on different projects, and that started earlier this year, I finally needed a full-timer. And we have other teams that work on the BI and reporting for our brands, which is more about managing large data sets, how to visualize it, and bring it into a GPT style where anyone on the team can chat with it as opposed to needing a technical background.

Max Sinclair: How big is the company at the point where you say we are going to get data scientists in the agency?

Mike Danford: Once we had a couple dozen people, that is where we started bringing on more. We are trying to understand the ratio now, we have a catalog manager and a brand manager based on a certain number of brands, so data scientists is probably going to be part of that equation. We are not there yet, but right now I am a bit of a roadblock, so I do not know if I can have more data scientists answering more questions until I clear that.

Jo Lambadjieva: A lot of our listeners are also Amazon agencies, so your case studies will be really helpful. Can you give some concrete use cases of how you have integrated AI with your reporting or analysis function?

Mike Danford: One thing we are working on right now, we call it AMOS, Amazon Marketing Operating System, and basically it is taking a lot of flowcharts of where things are going directionally, and a lot of data. There are a lot of decisions made on data, is my conversion up or down, is my click-through up or down, price change, competitor price changing. Being able to pull all that in, where it would take several hours to dig in and pull those disparate data pieces together, and now communicate with an interface that says, run this diagnostic and give me three or four ideas. For us there are a lot of questions that are less black and white, so we are trying to get the machine to be a little more intuitive, based on what we train it, but also look at it your own lens based on what you have and tell us if it matches or not. It is head to head, us versus the AI, to see if we come up with a similar answer.

Jo Lambadjieva: Who is winning so far, your team or the AI?

Mike Danford: I will tell you where the AI loses, and it is more the hallucinations. My understanding of AI at its basic level is it is a prediction machine, giving you the answer it thinks you want based on the inputs, and it has been told to be really confident and have a lot of bravado. So what we started doing is having it display, to some extent, the confidence that this is the correct answer, and if it does not have a certain level of confidence, tell us, this is what I think but it is a 30 percent chance or 80 percent chance. That has been really fun.

Max Sinclair: That is a brilliant idea, I have not heard of that. And if it does not do it accurately, will it say I am low confidence and then it turns out it is just giving you a hallucinated answer?

Mike Danford: It is a little different. It has probably been about a year since we moved to our own and started developing and bolting it on with our own identity sets. We had to force it to understand that it is okay to tell us if it is not confident, because in the beginning it did not want to tell us it was not confident, like it would be in trouble. Now we have a metric on our end, that is getting more technical than I do, that is what the engineers do, but it is pretty confident. We are trying to figure out the threshold where we just accept it. And you can have it cite or reference where it is getting the source from, which is pretty good. It has given us better stability in where we keep pivoting from when we get a certain answer.

Jo Lambadjieva: Which GPT are you using?

Mike Danford: Honestly I do not have the answer on all that. I know we are on OpenAI for most of our technical, with some stuff built on top. I know our marketing team is really big into different models based on what we are doing for AI image generation, we produce a ton of rhinos. It just depends on what is being trained in the back end and how it can be used.

Max Sinclair: My experience, and I will date this to early September, is AI is very good at many things, but one thing it is not good at is data analysis. When I was doing due diligence for investors, I built these models and tried to put them into ChatGPT to create visualizations, and I found the hallucinations on data analysis really high. Have you experienced that?

Mike Danford: Yeah. For the most part, we have to give it some original direction as to what the column is. The cleaner the data is in the beginning, the better it is. But when the data is old or not structured well, or you have dates and are not telling it how the two sets compare, that seems to be an issue. If you can clean that up, we have seen it be better. For the visualization, there is some stuff it can throw in with Python, but it is pretty limited, though sometimes it is a unique way of looking at it. It is part of the process, not the full process.

Max Sinclair: On the reference point, I found that works well. If you use Perplexity or Gemini and you ask a fact like how many Amazon sellers are there in the US, it gives you a number that could be made up, and then if you say point me to the reference, sometimes it says I cannot find the reference, and sometimes it shows you. So it is a good level-one test for whether it is hallucinated. On the confidence score, how do you know if it is hallucinating the confidence score itself and just saying I am 30 percent confident?

Mike Danford: Great question, and it could very much be gaming us. To your point, send us the references if it is external. And you ask it a fact check on the back end. I do not have a definitive answer, we have not tested that specifically, but we are getting the percentages, and I do think it is pretty good at telling you when it is not confident. When it is a big guess, there is a big delta, and when it is close the confidence is higher. I think that is because it cannot be verbatim, it is putting its own spin on it, so it is learning how close it can be. Maybe our data scientists could answer how they built and trained that model, I will take that back.

Max Sinclair: One more observation. I made a custom GPT to test how well an Amazon listing is optimized for Cosmo, with the disclaimer that it is not connected to Amazon, it is reading the science papers and some data we gave it, and it does a one-to-10 score. And the scoring works, if it has a lifestyle image of a target audience and explains how you use the product, give it a 10, and if it has no lifestyle images and no bullet points, give it a one. So maybe the confidence score does work. Jo, did you have a question?

Jo Lambadjieva: When it comes to data analysis and using ChatGPT, it is all about how concrete you are with the parameters or hypothesis. If you are using it for analytics, you are not really using ChatGPT, the GPT sits on the bottom but it is connected to Python, so you are using the LLM's use of Python to do the analysis. So it is about how specific you are with your prompting and what the flow of steps is that you are instructing. If you are generic and say give me a graph of this, it will probably hallucinate, and it might hallucinate even if you are specific, but if you are very instructional and reference your metrics and what they mean, you get a much more accurate answer.

Max Sinclair: On this point, what I found really useful is it is very good at Excel formulas. Before, I hired a guy on my team because he was super at Excel, and now you can describe in natural language, I want a formula, this is the use case, this is what is in these cells, and it does the formula and it works almost perfectly. That is a really good use case. Any last thoughts?

Mike Danford: We are big into Google Suite and do a lot of macros in Sheets. One of the first plugins I used outside the GPT UI was something for Sheets and Excel. I know enough to be dangerous, so I will go in, here is a function I am thinking of, it is not working the way I want, how can you make it faster or lighter, and it helps you take that next level and write out the macros, which is getting into coding inside the sheet. Where is this broken, what is missing, how do we make this macro lighter for our users. It is always going to give you an answer, coming back to the hallucinations, it will do what you ask regardless of whether it is the right thing.

Jo Lambadjieva: I want to talk about review sentiments, I think you use AI a lot around that.

Mike Danford: Yeah, we look at larger data sets, pulling in the reviews of our products, Amazon has some of their own in the back end of Seller Central now. Understanding the sentiment of your own, what are we missing, how do we improve it, getting ahead of a bad batch, what are the gaps, what are folks talking about, sizes, colors, fit. What I like now is the reviews are being pulled into Rufus on the platform, and we are seeing some folks manipulate reviews by putting certain context in, we have some not fully verifiable proof of bad actors commenting on other people's products. Originally you had to read the reviews and catch it, now you have summaries above the reviews, and if you go into the FAQ or straight into Rufus for the listing, it gives you feedback on what the reviews are saying. For us, if it is correct, let us bolster it in our product images and content, and if it is going in the wrong direction or misrepresenting, we update the listing to answer that specifically. We do live demonstrations with our brands, here is how you play around with your listings, and with large catalogs you cannot do it for every product, but it is great to see what questions people are asking, and how can we answer as many of those across the entire brand with a single update, one image or one bullet.

Jo Lambadjieva: When it comes to listing optimization, if you want to emphasize one use case or benefit, do you replicate the same benefit in both images and bullet points, or keep it to one and maximize your space?

Mike Danford: A lot of the brands we work with have thousands of products and A-plus content they apply across multiple product lines. So you figure out, is this my entire catalog, what are the subsets, and what is the lightest lift. We have been playing around with Amazon Rekognition, understanding it takes the first 100 words available in an image, and anything duplicated goes against your 100, so it is understanding how to stuff, without being too gaudy, things in an image you would not be able to put in the back end. For one apparel brand it is a common t-shirt with jeans question that pops up, but there is not really an answer in a lot of the listings, and there is a gender bias on that. So how fast can we do it, what is their PIM and system, and there is rotation, can we put it in the description, the metadata, where without being overly disruptive. We have brands with tens of thousands of products that have not touched their C1 and C2 tiers, and you go to optimize the listing and have to pull it down and put it back up because it has been so long, so how do we do this without jumping over more hurdles. It is an iterative process.

Max Sinclair: I am loving these tactical tips. So to summarize, you mentioned the Rufus questions when you are on a listing will preempt questions like does this t-shirt work with jeans, and optimizing listings with those in mind helps you rank on Rufus. And the second tip is visual SEO, Rekognition can read the keywords in the images, and you want that present in infographics. Before the podcast you sent me an interesting example of how Rufus hallucinates, do you want to go into that?

Mike Danford: You have to refresh my memory, I know it was an apparel thing I sent to the brands.

Max Sinclair: You were live interacting with Rufus on a Loom video, saying does this t-shirt look good in jeans, and Rufus said plenty of people in the reviews have used this product with jeans, and there were basically no reviews on the ASIN whatsoever, definitely none talking about wearing it with jeans.

Mike Danford: Now I remember, it definitely said in the reviews. My hypothesis is that it goes to your website as well, so maybe the reviews on the website for that product have it, or it is picking up some fringe interpretations, because we have repetitive images of certain products where the model is wearing jeans in the image. So it is picking up some fringe interpretations. In that moment, that is what we are trying to share with brands, go and play, ask Rufus, look at the bubble, see what pops up so you understand the category. Also go to your competitors, see what they are getting asked, and when there is not an answer to something common, give it an answer.

Max Sinclair: I would add that the current AI, GPT-4o, is one of the most advanced, but Amazon will be using something else, probably built their own model for Rufus. Roughly the performance will be as good as GPT-4o or worse. But OpenAI knows hallucinations are a problem, and the next step is agents that break a query into sections and complete those sections rather than just predicting the next word, and then the hallucinations will go down. So this is very much a Q4 2024 thing, and if we did this podcast in mid-2025, no one is going to worry about it anymore.

Jo Lambadjieva: That is the whole point, they let Rufus out so as many people can use it as possible and it learns and gets smarter. If we think about where ChatGPT 3.5 was a year ago, it was still hallucinating like hell, so where we are a little more than a year since then, the more people use it and add context to their listings, the smarter it will get, probably far better in just a quarter.

Max Sinclair: The other topic you mentioned is market basket analysis for product development, do you want to dive into that?

Mike Danford: Absolutely. We work with large catalogs, so we get a lot of data, and it is simply Amazon's market basket analysis to give you the top two other ASINs purchased in combination. We built a dashboard for that and give it out for free, and we layer our AI on top. It is really good at showing this ASIN is commonly purchased with this ASIN. What we like is looking outside the brand at the frequent purchases of other types of products. I gave an example a few months ago, we have a brand selling sage sticks, and one of the top commonly purchased is a sage stick with a Palo Santo stick, and it looks like one brand at 2 percent of total combined sales, but when you look, it is like 30 percent of all the cross-sells are Palo Santo sticks. Then you ask the brand, do you produce this, and the answer is no, and you say you probably should, figure out a virtual bundle, add it to the listing. That brand since released it in the market basket, and the average order value is much higher. It is those little insights, and getting that larger set where a human would find it very difficult to look at all that and pick up those nuances. We pull in Product Opportunity Explorer and Search Query Performance too, layering all that data and having the reports feed off each other, that is the fun part, how can these reports come together to be more informed.

Jo Lambadjieva: Do you feed this all into one centralized dashboard, or into a model like GPT to categorize and mash it up?

Mike Danford: Each brand is a little different, so for now it is separate, it is easier, lighter, and more standardized. It is understanding what the brand can and cannot do, how fast they can produce something, are they hungry to find new products or improve the ones they have. It is the it-depends answer. We are trying to figure out when to pull all of it together, the best way, and if there is a particular order that is better, building that recipe. We love when an apparel brand says we are in this sub-niche and this other design has taken off, and these reports help you see how to put certain styles and designs together, it is not the exact same thing but so close you can piggyback it. It is a lot of bundling and merging and being creative in how you stack the attributes for a parent. And it is combating Amazon, we have a bedding brand constantly getting their parents broken apart, a style with a quilt, comforter, and duvet, and Amazon wants those separately, which kills the basket value, so we try to record those and show Amazon, stop breaking these apart because it is making you less money.

Max Sinclair: So you are using data to understand what bundle listings to create, new or existing. What key metrics and data inform you to say X should be bundled with Y?

Mike Danford: First, are competitors' products commonly being added to your products, and do you offer those products. If so, that is the easiest thing, you already have it live, do we virtual bundle it, merge it, create a new parent, what is the fastest way. Then, 30 percent of your orders are purchasing your product with this competitor product, can you do that. It is the weight of the lift, the speed, and the cost, and can you trial it. We have brands that launch hundreds of products every month, more of a spaghetti-on-the-wall approach, and now we say look at your product opportunities to throw more cooked spaghetti on the wall. Amazon has given us way more data than it used to, but does not do a good job of visualizing or organizing it, so we do that. And we have to be careful not to have shiny object syndrome chasing whatever is out.

Jo Lambadjieva: I want to touch on pricing, which a lot of sellers struggle with. Are you using any AI to analyze pricing, how do you tackle it?

Mike Danford: We work with some vendors right now, a couple of head-to-heads on different catalogs, the term is dynamic pricing. I saw a report that Fortune 500 companies are speaking about dynamic pricing and surge pricing based on demand, like Uber. A lot of brands, when we come in, we ask when did you last update your prices on your bottom tier, and it has been a year or two. Everybody is about advertising, and the next more complicated piece is getting a budget toward optimizing the listing, and I think the Cosmo and Rufus integration is making it easier to pitch. The last one is dynamic pricing. Repricing is about winning the buy box against competitors, but dynamic pricing is about what is going on in the market, your supply and demand, price elasticity. A lot of brands make the deltas too big or do not do it for long enough to have good data, too much recency bias. Little things can go a long way, pennies up or down can make a big impact. The part we are trying to get into, that the service providers have not done yet, is downloading the last couple years of pricing data for your competitors and, this is where the AI comes in, figuring out their cadence, they always have a promotion this week, or every six weeks, or their pricing goes up or down based on inventory. So can we pull that in and decide, should we follow, should we beat it. It is kind of the new PPC, the new game of bidding, your price is the same way as your bids. It is a little early for us, but great data as another way to add to the bottom line when they have maxed out advertising.

Jo Lambadjieva: When I analyze some of my clients' pricing, I sometimes spot that even changing pricing by a couple of pennies down has a dramatic effect, because Amazon factors it into ranking, and it can shoot you up just because of those few pennies, so you do not have to sacrifice on margins. It is almost seeing the sweet spot between consumer psychology and the Amazon algorithm.

Mike Danford: Yeah, getting that lowest price in 30 days badge by flipping it down a few pennies is remarkable. And charm pricing, some brands swear by it, ending in a nine or eight. We have a lot of brands afraid to increase their price, and we say let us do it and see what happens. We just had a CPG brand, really healthy, keto, clean products, considered a premium brand, and they were scared, so they added 15 percent to their price, which had been consistent for three or four years, and it actually increased the purchase rate by showing that validity and the genuine ingredients. It also helped them not be afraid to move the price up or down. A lot of people are concerned that once they lower their price they have to stay there, but it is okay, you can move it back up. How much can you move is a big question they do not understand, so since they do not understand, they do not do anything, which is the worst thing you can do.

Max Sinclair: Mike, this has been one of the most tactical podcasts we have done. To read back the notes I have taken from your tips, number one, start experimenting with confidence scores for your GPT answers. Number two, use Rufus questions to optimize your listing, go see what Rufus is asking, which helps optimize for Cosmo. Number three, add text to images for Rekognition. Number four, use AI to do basket analysis of what competitive products customers are adding, and do that for bundles. And number five, use AI to understand the cadence of competitive pricing. Any other actionable tips beyond those?

Mike Danford: There are so many, just do not be afraid of it, it is not going anywhere. What we are trying to learn is it moves so fast, and I believe Jeff Bezos mentioned that your competitive advantage is not focusing on the shiny new things, but on the things that will be here for five, 10, 15 years and having those foundational blocks. The reason we focus on these points is customers are always going to be sensitive to price, that is not going anywhere, and they are always going to have questions about the product. The more answers you provide before they purchase, the friction reduces and the customer service is better. These are awesome and powerful tools we did not have a year and a half ago. I think a lot of folks fear certain roles will be obsolete, but I also think they will be elevated. A developer's skill set will be elevated, they can do the more complex part the machine cannot, while the machine does the simpler, lower-level tactics. Embrace it, it is not going anywhere.

Max Sinclair: I agree. Knowing what works and what does not gives you an advantage, you can augment certain parts of your work. I would not trust the AI to build my visual graph, but I would trust it to create my Excel formula. Jo, any super actionable tips to add?

Jo Lambadjieva: I would actually contradict this slightly. Use AI to do analytical data, but know how to analyze data. You cannot expect an AI tool to be a great analyst if you are not already a great analyst who knows what they are looking for. How would any AI know what is actually an insightful piece of information? So my number one tip is know what you are asking for, and you shall get what you are asking.

Mike Danford: I will piggyback that. As I mentioned, if you can give it a flowchart, a very clean set of conditions, it will answer those questions pretty confidently, especially if you let it know where and when to look.

Max Sinclair: You are giving ChatGPT a literal flowchart and saying follow this flow?

Mike Danford: Yeah, these are the steps, follow these steps and do it a thousand times, generally in a spreadsheet, we feel that is cleaner. The spreadsheets are easier for us to upload and communicate in the back end through the API.

Max Sinclair: Well, Mike, it has been an absolute pleasure. If people want to reach out to you, how do they find you?

Mike Danford: On LinkedIn, and if you go to our downloads section or homepage, you will see free tools, a lot of the tools we talked about here and our early pilots are available free to download and use.

Max Sinclair: Amazing, very generous. Mike, an absolute pleasure, thank you for coming.

Mike Danford: Likewise, thank you so much.