#606 - There Is No Such Thing As The COSMO Algorithm!

In this episode, our guest is an expert on AI and Amazon Science papers. He'll talk about Rufus, COSMO, Project Amelia, and all other AI advancements from the Amazon side and beyond.

► Instagram: instagram.com/serioussellerspodcast

► Free Amazon Seller Chrome Extension: https://h10.me/extension

► Sign Up For Helium 10: https://h10.me/signup  (Use SSP10 To Save 10% For Life)

► Learn How To Sell on Amazon: https://h10.me/ft

► Watch The Podcasts On Youtube: youtube.com/@Helium10/videos

Join us for an engaging discussion with Kevin Dolan from Pacvue AI Labs as we explore the cutting-edge advancements in AI and Amazon's pivotal role in shaping this dynamic landscape. We'll unravel the mysteries behind intriguing names like Rufus, COSMO, and  Project Amelia, representing Amazon's ongoing AI initiatives. Kevin shares his expertise on the evolution of AI from its early conceptual roots in the 80s to the transformative impact of transformer models around 2019, which paved the way for groundbreaking applications like ChatGPT. Discover how Amazon's increased investment in AI research is manifesting in published papers and sophisticated models that are revolutionizing customer interactions.

We also explore Amazon's integration of AI in tools for sellers, highlighting the launch of advertising AI that optimizes campaigns with precision. The potential of AI in enhancing tools like Helium 10’s Adtomic and Cerebro for more efficient Amazon PPC campaigns and keyword filtering is discussed, along with the impact of Amazon's Rufus on the shopping experience. While Rufus aims to improve customer interactions, we critically assess its current limitations and ponder its potential to shift some search activities directly to Amazon from platforms like Google and Pinterest. Additionally, we dive into Amazon's transition from lexical to semantic search, emphasizing the importance for sellers to align their product listings with customer needs for visibility and success in an AI-driven environment.

Lastly, we examine AI-driven tools like Project Amelia in Amazon's Seller Central and their potential impact on brands and sellers. While chat-oriented interfaces may translate vague intentions into useful actions, skepticism remains regarding their revolutionary potential. We emphasize the importance of exploring third-party tools like Helium 10 for added value and addressing the hype surrounding changes in seller practices, reassuring listeners that successful strategies remain largely unchanged. Kevin's insights and our conversation shed light on the future of AI in e-commerce, leaving us excited for what's to come in this rapidly evolving field.

In episode 606 of the Serious Sellers Podcast, Bradley and Kevin discuss:

• 00:00 - Advancements in AI and Amazon Science
• 00:41 - Decoding the Amazon COSMO Algorithm
• 08:42 - AI Model Cost Efficiency Advancements
• 09:48 - Amazon's AI Innovations and Rufus
• 14:59 - Implementing AI Chatbots Inside Online Marketplaces
• 20:29 - Enhancing Amazon's Semantic Search Capabilities
• 21:12 - Leveraging Rufus and COSMO for Selling Success
• 26:59 - Impact of Science on Amazon Practices
• 28:10 - Enhancing Amazon's Product Understanding With AI
• 30:01 - Customer Preferences for Pregnant Women
• 35:22 - Amazon's Data and Product Listings
• 37:30 - Amazon's Project Amelia in Seller Central
• 38:42 - Amazon's AI Recommendations for Sellers

 

Transcript:

Bradley Sutton:

Today we talk to the person who knows more about AI and Amazon science papers than maybe anyone else in the world, and he's going to talk about all things Rufus, COSMO, Amelia and all other AI advancements from the Amazon side and beyond. How cool is that? Pretty cool, I think. Hello everybody, and welcome to another episode of the Series Sellers Podcast by Helium 10. I'm your host, Bradley Sutton, and this is the show that's completely BS-free, unscripted and unrehearsed organic conversation about serious strategies for serious sellers of any level in the e-commerce world.

 

Bradley Sutton:

I'm not exactly 100% sure what I'm titling this episode, but I might have done something kind of clickbaity and say something. There is no such thing as the COSMO algorithm or something to get people to click on this. But let me just quickly explain that. Now. I don't mean that there's no such thing as Cosmo. There's a lot of documents out there from Amazon that talk about it, but there's nothing that says, hey, Cosmo is the new A9 algorithm, or there's nothing official from Amazon that says, hey, Cosmo is now in full effect across 75% of searches, or anything like that.

 

Contrast that with all the articles from Amazon that talk about Rufus. I mean, Rufus is a thing you can actually see in everything. So I just wanted to do a clickbaity title like that and we'll definitely get into Cosmo and things like that later. But I've got back on the show probably one of the persons who's the highest expert in the world as far as AI and also what Amazon has been doing as far as on the AI front, and that's Kevin from our own Pacvue AI Labs. That's why I'm wearing this. It's actually a Brazilian soccer team, Palmeiras, I think.

 

Bradley Sutton:

I wanted to get something with a P on it. Yeah there you go.

 

Bradley Sutton:

I have a Padres P hat too, but since I'm a Dodgers fan, it hurts every time I even wear that hat. So I was like, no, I'm not going to do it, considering the times that we're in right now. But anyways, Kevin, welcome back. It's been a little over a year since you've been on the show.

 

Kevin Dolan:

Yeah, thanks for having me back. Last year was a lot of fun and we've been seeing a lot of things happen in the last year in AI, especially around Amazon's implementations of AI, so excited to talk about those updates.

 

Bradley Sutton:

Cool. Now let's just talk about AI in general, general. You know, like AI is kind of like, I guess, like about two years, I mean, people have been talking about AI for years but as far as the, the more recent trendy version of the topic, AI, um, it's really been, you know, like you know, ChatGPT and things like that over the last couple of years. And let's just talk about what's happened in general over the last year. You know the improvement

 

Kevin Dolan:

Okay, sure, yeah, I mean, like you said, AI has been around forever. We've been using the term at least since the 80 s in terms of technologies that we can actually use for actual production purposes. As we're using the term today, its meaning has shifted to largely refer to this current generation of models that we're seeing. That began in around 2019 with the introduction of what was called the Transformers model. This led eventually to a variant of that model called Large Language Models, popularized by Open AI's ChatGPT, and we've been seeing a sort of explosion in AI technology and investment into hardware, investment into research as a result of some of these findings. That has become sort of the current modern label of what is AI. We're talking primarily about transformer-based models that perform language or other modalities, including image generation, and we're talking about basically whatever is that front line of research that's happening right now. So you see this explosion happen with the release of the paper around 2018, 2019. And then you see the proliferation of training hardware that led to innovations like ChachGPT, where we're starting to see these emergent behaviors, where these models do start to exhibit something that you can really call intelligence. These models do start to exhibit something that you can really call intelligence.

 

I came on here last year to talk about all of the different papers I had read from the prior four to five years at Amazon Research. You can tell, when you look at the number of papers that Amazon is releasing, that around that time around 2021, 2022, they started to invest a lot more in their research department. When they started releasing papers in Amazon Science in 2018, there were five papers about search. The following year, in 2019, there were 18. By 2021, there were 40. And then the next year there were almost 70 papers. That seems to have leveled off at this point. We saw about 70 papers last year and so far in this year we've seen about 60 papers. So we're probably going to end up in the same realm.

 

So the number of papers that Amazon is releasing isn't really changing. What is changing is the complexity of the models that they're using is much more sophisticated and they're being targeted for much more practical use cases. You're seeing larger A-B tests where they're being run on material percentages of traffic on Amazon. You're seeing Amazon release actual AI features that are customer-facing, like Rufus, and we're seeing investments in hardware that make some of these models that used to be impossible to run in production now very conceivable. So I think we are seeing confirmation that Amazon is taking these technologies seriously. They're implementing it in production and it is starting to impact customer behaviors.

 

Bradley Sutton:

What about non-Amazon AI Like what you know? ChatGPT, imagery you? Know, like a couple of years ago it was just hallucinating nonstop, and then last year a little bit better. You know images. You could not create humans, you know, or products in there without seven fingers and stuff in the general world of AI. How has that come along in the last year?

 

Kevin Dolan:

Yeah, so I mean we are seeing continued investments in research and continued improvements on these models. The transfer model really revolutionized things, but the initial results that we were seeing out of those transformer models were a little disappointing. For the first time, we were starting to see computers understand language, computers being able to generate images, and our initial reaction was holy cow. We didn't know computers could do this, and then, as we started to use it a little bit more, we became really disappointed, because we're like, oh you know, all the people have six fingers. It's making up facts. You know, the things that it's saying don't really make sense. And so there's been a lot of people who have looked at this potential and started to invest material dollars in improving it to basically get to the point where now these technologies produce more reliable, more consistent results. There's still really major shortfalls, there's still issues, and I think you're going to see continued investment in this. The optimistic projections that you're getting from OpenAI. You know I'm personally a little bit cold on those, but who can predict the future? Who could have predicted that this would have happened? Yes, you are seeing improvements in image generation models, where the images that they're producing are now closer to reality. We're starting to see these used widely in industry, especially in fields like advertising, where you need to produce high volume creative. If you look at the features that Photoshop has released related to their Firefly AI image generation model, we're starting to see not only improved models but improved workflows for creatives to actually be using these tools in a way where, instead of just somebody typing some random prompt and getting whatever the system decides to give you now, people are actually able to control the output and get the output that they're looking for. So, between all of these things, you're seeing a lot of development to make these tools more practical to use. I'd say the biggest and most recent news is OpenAI's release of its strawberry model, which they call O1 in their release vernacular. The O1 model from OpenAI is performing thinking steps before it answers the question and hiding that thinking from you, the way that if you're asked a question, you might think about it a little bit before you answer it, and they're seeing really, really impressive results from that. You know we're getting closer to the place where these AI models might be able to do something that's a little bit more functional, a little bit more capable of actually interacting with real life data and real-life processes, you know, but we're still a little bit far away.

 

Another issue that we keep running into is the dollar cost of running these models. Towards the end of last year, at Helium 10, we developed a review sentiment analysis model that basically would read thousands and thousands of reviews for your Amazon products and produce some analysis and produce an analysis of what people are saying about your product. You know Amazon has a similar product. Ours goes a little bit deeper than that but the idea is essentially the same. You know what are people saying about your product, what can you learn about it in order to improve your product, improve your listing, etc. And one of the things that we ran into with that model is just how prohibitively costly these models can be to run on large sets of data, and so we're starting to see investments in making models smaller and more special purpose, and we're also seeing improvements in hardware that make running these models more cost effective. This is really going to start to unlock production capabilities, and that companies will now be able to run AI models profitably.

 

Bradley Sutton:

Interesting, interesting. Now, yeah, we're always looking to add things that can utilize AI that helps Amazon sellers. You know we are launching this week advertising AI on our Atomic side, which is allow somebody to just enter in an ASIN and then our AI engine will kind of just create all the campaigns on its own and optimize them on its own. That's something that we've been using at Pacvue for a while, and we're integrating some AI things into tools like Cerebro, where you could have a prompt that allows you to filter out keywords or say, hey, can you please remove any Spanish keywords from the results? Or, hey, can you remove any branded? You know search terms, you know things that you know you could probably do on your own, but it just takes a lot longer. So, so, definitely, we're, we're keeping track of what AI can do, because anything that is doable. We want to go ahead and bring it into Helium 10.

 

Bradley Sutton:

We know that getting to page one on keyword search results is one of the most important goals that an Amazon seller might have. So track your progress on the way to page one and even get historical keyword ranking information and even see sponsored ad rank placement with Keyword Tracker by Helium 10. For more information, go to h10.me forward. Slash keyword tracker.

 

Bradley Sutton:

Now going back to the main topic, amazon. Before we get into the science more detailed, into whatever science documents have been released and things this year, let's talk about what is 100% already out there or talked about, which is like the Rufus and so Rufus, Cosmo I've got some personal opinions on it and that's all. A lot of this is, you know, until Amazon actually publishes something for sure, like you can't even say that, oh, a science document said this or that, because the great majority of the content of science documents actually doesn't actually get into production on Amazon. You know per se. You know so just because Amazon talked about in a science document. It's just a research paper, you know. But let's first about talk about the stuff that you know Amazon announced at Accelerate or has already rolled out to customers, like Rufus.

 

And then my general thought on that and again I could be wrong and I'll be happy to switch my thinking when Amazon does make some different announcements is that Amazon is always about the customer. Right, they want to give a better result for the customer. And then I don't feel that, like Rufus, for example. Fyi, in my opinion it's terrible as a buyer where I'm like, hey, what did the review say about this product and it gives me an answer. And guess what? There's no reviews on that product. So, as a consumer, being kind of skeptical about some of these AI things, I just can't use it. And now the other part of it is I don't think anytime soon the traditional way of searching on Amazon is going to be improved in that if I know I want to buy and I talked about this in a previous episode recently if I want to buy a coffin shelf, there is no better process than me opening my Amazon app and typing the word towards coffin shelf and looking at the results like there is nothing unless amazon connects my brain to, to the app. That is going to ever be better than that where? In other words, I am not going to go and have a conversation with Rufus with my thumbs, you know, like taking typing in a whole bunch of I used to be a secretary. I type like a hundred words a minute. So like, let's say, I was on the desktop app, I'm still. I'm a lazy person, as all human beings are. I am not going to say what do you think, Rufus, about coffin shelves out there? Like, like, no, I'm going to type in nine letters and then. So that part. I almost don't think Amazon is necessarily trying to change that part, because they know that it's already the most optimized experience for people who know what they're looking for.

 

Now here's the thing, though how did I get to that decision that I wanted a coffin shelf, like maybe I just knew it. But another thing is, maybe I'm just browsing like, hey, I want to uh, search on google what are trending, um, trending gifts in 2024 for teenagers with a gothic inclination, or something like that. Like, right now, I'm not doing that in Amazon, or, historically, I'm doing that like in Google, maybe Pinterest, you know, or maybe these other websites where I'm trying to get ideas. And then, all of a sudden, I read a blog, or I arrive on a TikTok or whatever, and I see, ooh, Coffin Shelf. I didn't even know that existed. Now let me go and type in coffin shelf on Amazon.

 

So I think the potential of, of a fundamental change in the way we shop could be that maybe some of these searches that people would normally start on a Pinterest or on a Google, maybe now you can start in the Amazon app, where what I would have typed for the Google AI or things like it's just going to go ahead and, and, and I can start the Amazon app where what I would have typed for the Google AI or things like it's just going to go ahead and I can start, you know, just browsing, browsing things, and at the end of it, you know like Amazon might, or Rufus might, tell me yeah, you know, like we see some spooky families by coffin shelves, and then here are the coffin shelves Now. Anyways, I normally don't talk very much when I interview somebody, but I'm very passionate about this. But are we on the same page here, or what? Correct me if I'm wrong or if you have different ideas.

 

Kevin Dolan:

I mean totally with Rufus.

 

You know Rufus is out, it's public, it's something that anybody can interact with. So we know it's been implemented and if you've actually used it, I'm sure you found the experience a little bit disappointing. You know it does two main things it helps you to figure out what search you might have wanted to type in if you weren't completely sure, and it answers questions about a product once you're looking at a particular product. I think that those two things could be useful. You know, I think that it's certainly early in the implementation of chatbots to say that these things are fully capable, but I think what you're seeing with Rufus is mainly two things here. The first is there's intense industry pressure to implement AI in a visible way that all companies are feeling. After ChatGPT was released, no major tech company wanted to fall behind on that trend, and so you started to see these types of very visible generative AI features implemented in tech platforms across all industries. If you've got a website, there's a good chance you've got a chatbot at this point, and so it's hard to imagine a world where Amazon was not going to release something like this. They really, really had to because there was so much pressure to at least try it, see if it works, see how customers respond to it. Also, we know that Amazon looks towards other retail experiences to try and understand what ways they can improve the e-commerce experience.

 

It was not always the case that Amazon's primary vehicle for finding a product was a search bar. When Amazon was first released, it was largely node browse based. You would search through a series of categories and get to the product you're looking for, which is much akin to going to a store, looking at the different aisles, walking down the aisle that has your type of product and getting there. It was a major innovation for them to create a search engine that could search through any type of product and understand at some level what a person was looking for, and they've been making continuous improvements to that over the entire development of their company. I think with Rufus, the corollary in real life retail is going to a store and talking to an associate. If you go to a nice store where they have a more curated shopping experience, you might want to go and just talk to a person and ask them questions about the products that they're experts on. I think that's a sort of natural corollary to try to implement in an online context, but when I go to a store, if somebody comes up to me and starts telling me about their products, I'm personally not the type of person to respond to that, and so you know it's natural for me to look at Rufus with a little bit more skepticism than you know somebody who might enjoy that real life experience.

 

I think that there are shortcomings with Rufus. I don't think it's going to materially impact the majority of purchase paths for the majority of customers. I agree with you. There is no easier user interface that I can imagine. When you are looking for something, you want to just go to Amazon, type it in a search box, a brief description of what you're looking for and then yeah, all right, I've got a list of things to look at. I've got some pictures. I can scan some results.

 

I do find some utility with Rufus with respect to answering questions about products. You have to take it with a grain of salt because it can hallucinate. It can produce unactual information. However, I have used it in some context to ask a specific question about you know, can this product be compatible with some other product? And it will give you some kind of information that you can then verify using the listing, using the questions and I think that's helpful in order to use Rufus to come up with search ideas and things like that.

 

I found that those features are a little bit less useful but, like you're saying, if they start to integrate the experience of asking these questions in a more core way, in a way that feels less bolted on and gives you more than just a text output with links if it were to give you, say, a sort of a Pinterest board for product discovery, help you to better understand how to get to the listings that you want to find.

 

I could see a world where those user interfaces become material for less targeted searches, where you aren't really sure exactly what you want to buy off the bat. One of the things that they point out in the blog post about Rufus because they haven't released a scientific paper about it detailing the implementation. But one of the things they point out is, if you are going to involve yourself in some kind of activity like, let's say, ongoing camping in Joshua Tree, I might use a tool like Rufus to answer the question of what types of things do I need? You know the kinds of things that you might talk to a store associate at a camping store about and it can start to give you some ideas about this. But I think we're pretty far from the point where you would give it the same kind of trust as you would give as somebody who has put their body in a camping experience routinely.

 

Bradley Sutton:

I agree. I think Rufus definitely has some potential to help things if the hallucinations stop, because there are things that as consumers, we do that takes time. After I land on a couple of products, I might start looking at the reviews. I might start looking at details of the bullet points and descriptions to see use cases and try and find out material. I might look at the images to see the stats and the ingredients of something, and these are all things that can take a lot of time, especially if I'm not sure where to look.

 

Like I don't know where a seller has put in their listing. You know which material to use, so I can definitely see Rufus helping there. But then, you see, my thing is then you know and this kind of goes now into the Cosmo discussion is I materially do not believe that sellers should be doing anything differently right now. To me, the people who Rufus and Cosmo might help, if anything, is the people. It's kind of like maybe leverage or leveling the playing field a little bit for some of the people maybe who are not doing the best practices.

 

You know, maybe I didn't put all the right keywords in my listing and so I wasn't indexed for it on day one, but then Cosmo or whatever, over time recognizes that the people who are buying my product are actually looking for it for this certain use case. It's kind of like what you and I showed last year on the podcast where noodle camera. Right, you know, noodle camera was not that keyword, was not at the time, I don't know about now, but was not in any listings on Amazon and it didn't have much search volume. So it's not like it was a big loss. But Amazon learned and we don't again. We don't know if this was Cosmo that did it or it's just Amazon algorithm, you know but Amazon learned that, hey, these people who are searching a noodle camera, they're actually looking for this stethoscope kind of camera that looks like a noodle, and so who don't? We don't know how long it took for that to actually become indexed as something, butthat's a benefit you know like. But at the end, if noodle camera was an important keyword, I, if I would have put that keyword in my listing from day one, I would have been the only one searchable. I wouldn't have had to wait for Cosmo or whatever A9, to kind of learn about that. And so again for the person who only keyword stuffs right, you're like, hey, I'm going to pull all my keywords from Cerebro and Magnet and just throw it in my listing and try and get it, each keyword four times.

 

Yeah, you know what? You probably should change your, your methodology, because that's not. That hasn't been the best way of doing things for years. But we've been teaching here at Helium 10 that you have got to talk about pain points to your product solves in your listing. You've got to show it in the images. You know what use cases. If you have collagen peptides, you've got to show people using it in their coffee. Not that they use the keyword coffee to search for collagen peptides, but that's how they are searching for it. They want something that is going to dissolve well in their coffee, and so you've got to be indexed from day one. You've got to talk about what pain points your product solves, and then that's what's going to put you on the radar of these Amazon AI things. And so in that sense, I don't think a seller's you know, most sellers should be changing their methodology at all because of any of these new things. What are your thoughts on that.?

 

Kevin Dolan:

Yeah Well, I mean, I think it'll first be helpful to talk about what Cosmo is and what Cosmo isn't, because I've been reading a lot of the blog articles, watching the videos and I'm seeing something that tends to happen in tech sometimes, where a word or a technology is being used as a stand-in for some broader movement within the space. I'm seeing a lot of people conflating Cosmo, which is a specific research paper, a specific tool that was built and was tested. It's described very specifically in a scientific paper. Cosmo is this tool, but I think it's being used more broadly to capture a shift into focusing more on semantic search and less on lexical search, which is exactly what I had come on last year to talk about.

 

Amazon has been working on this for years and years, improving their search algorithm to not rely on a listing creator to actually put a specific keyword in their listing and then find it based on the existence of that keyword in the listing. Instead, try to understand the meaning of a product, how people use it, what people think about the product and all of these kinds of details, so that when somebody types in a search, it can effectively find the product that they're going to want to buy. That is a shift that's been happening for years. That predates transformer models, but we have started to see for sure an increased ability to actually do these things on Amazon. I think that what you're saying is correct. You know the best practices and what sellers should be doing with their listings hasn't changed. But that really depends on what they were doing, whether they were following the best practices to begin with. You know like you said, if they were keyword stuffing trying to find as many keywords as people might type into a search box and stuff it into their listing in as literal a fashion as possible to make Sammy-looking listings that cover as much search volume as possible yeah, that's a bad practice, and as we move into a more semantically focused search world, that becomes an even worse practice. Semantically focused search world that becomes an even worse practice.

 

What it also tells us is that some of the efforts that are required today to create listings that do involve inserting specific keywords and things like that. You may be able to shift your focus to what would actually be more helpful to customers, which is accurately describing your product, accurately describing how your product will be used and targeting specific customers and specific pain points. The more specific you are and the clearer and more accurate you are, amazon wants you to be in front of the customers who want to buy your product. So that's always going to be a good practice and that's ultimately what Amazon is trying to do when they're doing these types of experiments.

 

Now the Cosmo paper is interesting. The Cosmo paper was tested on a really large chunk of Amazon traffic using a very heavy, large language model. Compared to prior research, which does tell us that Amazon has made investments in the server capabilities to be able to run these models in production and keep searches within their tight latency expectations, so that, I would say, is certainly significant, it tells us that Amazon does have the hardware capacities to run some of these more advanced models and it tells us that we are going to see an increased focus on semantic search. I think that does affect consumer behaviors, it does affect the way that we rank for keywords, but what it doesn't affect is that best practice of describing your products accurately.

 

Bradley Sutton:

Based on those scientific documents. What are some of the things where, again, just because it's in the science document doesn't mean that it's going to be implemented. But, you know, based on the results and sometimes you can kind of tell like, wow, this one had some pretty amazing results, so it's probably for sure going to be implemented. Can you talk a little bit more about the kind of things that maybe you've seen already implemented or you think will be based on all you know? Again, nobody has read more Amazon science documents than Kevin here. So what would you predict as far as the future, the next year or so?

 

Kevin Dolan:

I mean, Cosmo is a specific tool and I think that the function that it performs is valuable to enhancing Amazon's understanding of a listing. So I certainly would not be surprised to see Amazon implementing this in a production capacity on a large swath of searches. That would not be surprising to me, but it's not as massive as the shift that we've seen into semantic focused search. Cosmo in particular discusses essentially a mechanism for enhancing Amazon's understanding of a product by taking into consideration things that aren't expressed in the query and things that aren't expressed in the listing. The example that they use in the paper, the canonical example, is if you're looking for shoes for pregnant women, a listing might not literally say shoes for pregnant women. It might produce a specific type of open-toed shoe that has good support, good comfort. That might not literally be listed as a keyword in the listing, but it might be something that the system can infer based on its knowledge of the universe, about what it's like to be a pregnant woman and the types of products that they might benefit from.

 

Cosmo is essentially a mechanism for enhancing listings with additional information to get closer to the user's intent based on a particular search.

 

If you zoom out and you look at the broader task of semantic search. That's always been the focus. The goal is something might not be said in the same language in a query as it might be when it's written in a listing, when it's answered in a question or when it's written in a review be when it's written in a listing, when it's answered in a question or when it's written in a review, and so the domain of language that's used for these two different ways of expressing thought aren't the same, and so we need to create algorithms that better understand what a user actually means when they type in a search, and what a product actually does and what functions it performs. This idea of understanding deep intent and the actual composition of a product is essentially the goal, and we are seeing for sure that Amazon is making these changes. We're seeing more results come back for listings that do not literally have the keywords typed into search and better match what is a user's real intent on shopping.

 

Bradley Sutton:

But for it to learn that something is a good shoe for pregnant women, it basically would have to have some context, like maybe the reviews. Like somebody said, oh, I was in the second trimester and this was great. It's not going to pull that out of nothing unless, no, I was going to say maybe it knows that. Like, maybe somehow it knows the customer is pregnant and then, without even a review, it's a wow. We see an abnormally large number of pregnant women who are buying this. But I don't, I don't know. I mean, I think I big dad.

 

Kevin Dolan:

I could tell you that, Cosmo, the paper itself does. You're talking about what's usually called avatar personalization, based on your purchase history. I know some things about you. I can kind of put you in this category of person, and I know that these types of people tend to buy these types of products. The Cosmo paper doesn't actually explicitly discuss testing avatar personalization. Doesn't actually explicitly discuss testing avatar personalization. What it does talk about is using recent Search Queries to better contextualize later Search Queries. So like, for example, if I'm searching for camping gear and then I search for mattress after that, there's a good chance that I specifically mean a camping mattress or an inflatable mattress rather than a mattress for a bed in your home that weighs 200 pounds. It can better contextualize a particular search query based on the searches that you've been performing in the recent past.

 

Avatar personalization is another thing that Amazon is always investigating and we have yet to see any really material evidence that it's been implemented. Almost all of the studies that I've read relating to that type of personalization they talk about the potential of it, but in practice they tend to perform pretty poorly. They either reduce sales or they don't materially impact sales, which is a major problem. They don't materially impact sales, which is a major problem, especially considering that cost of performing that personalization. Amazon does a lot to make sure that the searches that come back are within a very tight latency. They need to come back as quickly as possible and that's very important to the shopping experience. The more personalized search results are, the more expensive those search queries are going to be to run and the longer it's going to take, which materially affects your experience as a purchaser. Yes, hardware is improving. Yes, technologies are improving, but if you can just reuse results, it's always going to be a lot faster than if you compute it on the fly.

 

Bradley Sutton:

But then, still, using the same example, I think, if you knew that, hey, your shoes have good cushioning and you designed it actually for pregnant women to be able to use, the best practice still is to put that keyboard in your listing for day one, so that at least you have a. You know, you don't have to wait for the AI to learn based on activity, you know. But then, if it's not something that's readily like, maybe you had no idea that people were using your shoes for gifts for people who are pregnant, like, maybe you had no idea. That's where, like, I think Cosmo, Rufus and stuff is going to help to uncover these sub-niches of people who are getting your product. But again, at the end of the day, this scenario, I don't think there's anything different that the seller needs to do as far as with their listing that we haven't already said. Now, at the same time, maybe they learn. I think this is going to open up some new potentials down the road. Like, let's say, Helium 10 starts seeing what the common Rufus things are being said about the product or what's the common queries. Maybe Amazon will make that available for sellers through some API that says, hey, this persona is buying your product.

 

Well, maybe I would go into my listing and change one of my images to show a pregnant person walking around with these shoes. But again, that's what you should have been doing for years. You know, like when you read your reviews and you notice like I used to sell this or I still do sell this egg tray, and I was reading the reviews one day and people were using this egg tray, this wooden egg tray, to as a serving platter for like sushi and also these chocolates, because you know the holes for an egg tray is very similar I was like I never would have thought that so in that situation, who knows, maybe Rufus would have seen the reviews and saw these images and now, all of a sudden, even though I don't have chocolates or sushi in my egg tray listing, I would be searchable for those keywords. But again, as soon as I would have seen that review or known that people are using my product in a way and this is what I did years before AI. You know cause this was years ago that I did this I went in and I did a reef photo shoot showing other use cases of it and I did one image, or like a quadrant of four images that showed somebody putting sushi in it, somebody putting chocolate in it, somebody putting this and that's, and then I put it in my listing too.

 

So, I was like I didn't want to wait for Amazon to hopefully index me for these keywords. So again, I just go back to the point that what Amazon is doing is not really making things where sellers are going to have to do something completely different, but they they're helping maybe the sellers who haven't been doing the best practices to get indexed for keywords that maybe they weren't smart enough to put in their listing. Yeah, I mean, I think so.

 

Kevin Dolan:           

What you're ultimately seeing with Cosmo is taking information from Amazon's entire catalog, which includes billions of products, billions of product listings, billions of questions, billions of answers, billions of reviews.

 

There's a lot of information contained in all of that data, which starts to build a picture of how the universe works, and so, in a sense, you could think of it as Amazon using the information it's learned from existing listings to enhance all listings and build a more comprehensive picture of their catalog.

 

I totally agree with you that it doesn't change the best practices, and still, I would say it's now even more critical that you are taking into consideration the use cases for your products, the people who might be using it, and accurately describe these in your listings. I think that that is still absolutely the best way to rank for products. I think what it does is it shifts focus from some of those old school techniques that we were probably recommending 10 years ago. It's no longer necessary for you to enumerate all possible customers of a product, but instead focus on the key use cases and the key customers to your products, describe these things as accurately and as naturally as possible. It's not required for you to think of all the ways that you could possibly say pregnant woman. Instead, you can just describe the fact that this is useful for a person who is pregnant.

 

Bradley Sutton:

Outside of Cosmo, Rufus. Obviously, they announced a lot of things at Amazon Accelerate, like Amelia for Amazon sellers. Any comments on other things that Amazon have been working on the AI front? Yeah, I mean I would say Amelia is Amazon sellers. Any comments on other things that Amazon have been working on the AI front.?

 

0:36:59 - Speaker 2

Yeah, I mean I would say Amelia is certainly interesting. Amelia is Amazon's internal chatbot for Seller Central. You know, I've yet to play with it. I've yet to see anybody who's actually had access to it, so I think it's just an early announcement. Maybe some limited people have access, but I would imagine it's going to undergo the hype cycle that we see for most chatbots, including Rufus. There's going to be a lot of excitement. The initial version will be pretty terrible. It will slowly get better over time.

 

The question is whether it will continue to receive enough investment to make it into a chatbot product that is useful for people, and whether chat is as natural an interface.

 

As you know, Seller Central is in and of itself. You know, I think we've spent a lot of time over the past 30, 40 years developing software interaction paradigms, so we have a good idea of what is easy to use software. There is potential that we could be using these more chat oriented interfaces to get to our vague intents that we have in our head a little bit more quickly, but we haven't really proven that out yet, and so I would say Amelia has a very similar potential to Rufus in that it's something that I believe could be useful if it is properly invested in, but the jury's still out on whether or not it's going to be a material impacting to people's workflow as you start to get access to it. I do recommend that sellers give it a try, just like with any of these tools see if it's useful for their workflows, but I'm not really holding my breath on it being revolutionary.

 

Bradley Sutton:

A lot of the recommendations that Amazon gives in Seller Central is. I think a lot of sellers have learned to just ignore them because they're not exactly that useful.

 

And then. So, if this is, it's like putting lipstick on a pig, you know like sure you could put the AI word up, but if it's being based on something that you don't trust in the first place, you know, might be a little bit of time before we can implement it, but I think that Amazon is definitely moving in the right direction and that Amelia has nothing to do with the customer. You know, like we always say, Amazon is all about the customer, which is true, but I think that's just in itself is a step in the right direction, that, hey, Amazon is doing things that are going to try and help the seller, and that's a trend I've been seeing over the last few years. I think it's a very nice step in the right direction.

 

Kevin Dolan:

On that front, we've definitely been seeing Amazon release features in Seller Central using AI that are more seller oriented, that help sellers to understand their products. They've released their own features for review analysis, which does get some basic, surface level summary statistics that could be helpful for people. I think Amazon is making investments there. However, they're always going to be a little bit step removed from the customer. They're always, at the end of the day, competing with sellers to some degree. There are certain things that they can do, certain things that they're limited on in terms of where their interests lie versus where the sellers lie, and so that's where tools like Helium 10 become much more valuable to customers, and so I do recommend that you look at the full suite of tools that you have available to you, because there's going to be things that Amazon will implement and there's going to be things that they're going to be hesitant to implement, for whatever reason.

 

Bradley Sutton:

All right. Well, Kevin, thank you so much for riffing on this with me. It's something I'm passionate about because I'm all about. I'm not like Amazon, I'm all about the sellers, not about the customers, and so anything that affects sellers or you, you know, if there's going to be some big inherent change in the way that sellers need to do things, then I get very passionate about it. And especially when I hear I don't want to, you know, use the word misinformation, you know out there, but almost like scare tactics or just clickbaity stuff, which I just did in this very podcast with the title of it but with at least, if you're in a clickbait, at least let people know that what the real situation is, because I don't want I've had so many sellers come up to me because of just hearing things where it's like, oh, my goodness, I've got to change everything I'm doing for my keyword research.

 

I've got to change everything I'm doing for my listing optimization. And right now, the fact of the matter is, no, I'm still doing the exact same things I did last year. There are some slightly different things because there's new rules at Amazon of what you can and can't do and of course, I've switched, but as far as the way I make my listings and I structure it and how I do my keyword research. Not one iota different am I doing it now, and I have had the exact same success with getting to page one on all my main keywords and getting sales for the keywords I think I'm relevant for.

 

And so I think that's just important to know, guys, that as AI evolves, I'm sure I'm positive there's going to be new things that we might have to do as sellers and stay tuned. We'll let you know what those are, but right now, as long as you've been paying attention to our tutorials the last few years, you're not having to do anything different, in my opinion. So, anyways, thanks, Kevin, let's definitely bring you on in 2025. And you know, who knows, maybe AI will be we'll be driving all of our cars and we're driving like the Jetsons or something. I don't know what's. What's going to happen, but we're going to find out with you next year.

 

Kevin Dolan:

Super excited. Thanks for having me.