Amazon Tools for Sellers Overview

[Music] hey guys welcome back to sell sessions today my guest um I'll bring him on in a moment very very sharp guy um for those who are listening on the podcast I would advise clicking on the link in the description uh because we're gonna have a full-blown geek out session um there's been a lot of chat about comprehend there's been a lot of chat about recognition and of course ruus Cosmo etc etc um so Jeff I'm going to bring on shortly in there Jeff is background is in engineering he uh works with a lot of agencies he's worked with Brands he owns a brand he's exited a brand and one of his biggest skill set is to break down complex things things like recognition and stuff can you know can be kind of complex if you log into awss a lot of it seems very foreign to most people um but without further Ado I'm going to bring on Jeff now and then uh get your pens and pads ready because this one's going to be pretty special Jeff welcome to the show hey how you doing Danny I'm good I'm good um okay anything I missed out in the introduction there I'm although I mentioned uh Cosmo we're not covering that today but we it will be ruus it will be sentiments in terms of Amazon comprehend and we're going to be looking at recognition but we're also going to look at objects as well not just text recognition and we're going to look at how this can be scaled especially if you've got a larger brand with a lot of products right yeah the main thing is is how do we build a repeatable process that could be handed off to a team member and be done you know with the same consistency every time yeah inde especially people with a large catalogs like that's one of the biggest struggles that I see them having yeah indeed so I'll let you kick off here obviously I'll chip in at various different points do you want to give everyone a breakdown of what we got planned now as we go step by step through this process yeah so we start everything off in mural kind of put the you know the main tools that we're going to using be using put any notes that we have on the tools that we're going to be using or any notes that we've read on articles online or Amazon science articles we'll actually put these in here as well yeah and then eventually as we do our testing we'll start to build out the flowchart we'll start to you know the the business improvement process will start to take shape and at the end we'll have a complete start to finish process yeah with every step along the way with the Sops and how the um I guess the idiosyncrasies of how things work yeah and basically by the end of this show we will do some additional ones if required but depends how far we get today but you can take screenshots of this we just anyone who wants them can have these because the idea is to educate you to be able to apply this to your own business or your own agency Etc right yep absolutely yeah so we're going to cover what I call like level one Discovery yeah you know we've done some tests we've kind of got like uh an idea of what's going on still might need might be areas that we need to you know keep testing but we can start to develop some some answers to our questions yeah on what we have cool and so we're gonna start like not like a new listing we're actually going to start with it's an existing listing so this is somebody that wants to optimize our listing the first step I'm you can use any keyword research tool but you know just develop your mkl right here we have data dive for those of you who don't know uh true honey te is a brand that I own part of any type of testing I do public share this account so you know this everything that you're going to see is eventually going to be live on Amazon especially like this royos chai tea and the Hibiscus lemon that we're releasing so we develop our Master keyword list once we're done with that we're then going to put together a graphics brief and send that to the graphic designer so we have an existing listing we have the mkl and then we're going to go to our Graphics here and from the graphics they will then develop the graphics and give it back to us now when they give it back to us what we're going to want to do is run that through recognition and try to start to get questions and answers to like how recognition ties into the Amazon search bar that's going on right now with rofus and with Cosmo so here's the Hibiscus lemon tea that we just developed and the first part of this is going to be the object detection there's two parts of recognition one is object detection which is harder you have to use Python code and the other part is text recognition which is really easy to do and you just upload the limit image and it just gives you the text that it sees now for the object recognition on this first one we can see that the brand name true honey teas hibiscus lemon lift we've got a lemon over here it's really hard to put can you expand it a little bit so it's wider on the screen yeah just so it's bit further up there you go perfect so you know we've got this little tag up here that this is not actually on the packaging but we put it on there um you know you've got the Hibiscus flour some dried hibiscus there and some lemon here now if we look at the detected labels in object we've got the label herbal which we're assuming is coming from this pile of dried leaves we've got herbs plant we've got the flower so that it's pulling this flower in the object we also got petal food fruit produce so they're pick up the lemon they're recognizing that it's produce citrus fruit lemon and you can see the confidence score they know it's food yeah 88% they think it's a citrus fruit and they think it's a lemon you can like watch their confidence score of that go down starts to drop yeah um and then we get the label advertisement and what was really interesting is we got the label hibiscus and I don't and it's even though it's only 56% we did get hibiscus and I was really surprised at that because if I asked you what a hibiscus looked like you know I didn't till we started developing this um typically only returns 10 labels so that's what the object recognition gave on this one right here now the next part would be the text recognition and you can see now it's picking up hibiscus lemon and it's got a huge confidence 99% yep you know teabags 99% so it's really FR through honey which is true honey that just means that you know if you have a creative or artistic brand name it may misread it yeah and then teas it picked that up hibiscus and these are actually really high confidence scores 99 lemon lift so it's picking up that 24 teabags it actually picks up if you look at this total net weight 11 ounces where that is on the product is way down here at the bottom you can hardly even read that it's kind of M it is he's very good at picking up even extremely small text I did some demonstrations in another video where the text was really small and it's still very good at picking it up so it did pick up every part of this now we do have a another one in this sample yeah that's further down here let me get down to it so here's image two and this one we added some icons across it I make this a little bit bigger we added some icons across it we put the lemons over here I mean aesthetically I like this image better personally yeah but as you said in the opening I'm an engineer not a graphic designer so I don't I don't uh try to give a too much of opinion on Graphics but what I see is we got a high confidence on herbal herbs advertisement flower petal poster food fruit produce it didn't pick up lemon and it didn't pick up hibiscus so putting that lemon behind this pile here did that now say the object recognition give it a harder time because it can't distinguish that being a lemon even though the lemon didn't have as high of a confidence score in the first one you bring up a very good point the point I was going to bring up but you've tied it we can elongate into this a lot of sellers will put all their accessories in their images right because they'll go right okay showing the value for money what you find is quite a lot of the time it confuses the OCR so it's not it's not picking up what's in there and it will give them misrepresentation and put different labels on because there's so much going on in the image now there are times it's done well I did an image stack the other day and it captured everything beautifully but I think the people that did the the main image of well classed at what they do they're also running it through recognition to ensure that every attribute is being displayed correctly or near as correct and I think that is some of the problems that the step with the cross referencing I've done an article um called mastering attributes and images is um which relating to uh your keyword ranking timing back to here I mean we're doing this slightly differently I don't want to go off track but the point I was making is that I've seen plenty of times where something is a pencil case and it's shown as the first AQ and where it can't cross reference that attribute with what's going on in the title and other attributes on the listing that will obviously uh work against them in terms of the algorithm understanding what's going on so yeah yeah I back you there maybe that is a possible reason why you're not getting lemon because it's being confused yeah so once we get you know these results we can kind kind of say okay we have two images that are scoring well one of them in my opinion as an engineering mind the one that has hibiscus and lemon I think would be better yeah so we can almost say that image one would be better mhm now going back to our mural board like just how we use it like here's our sop for Amazon recognition doing the objects what we'll do is we'll just take and we'll copy this and put it in our mural board so we have it for you know which we have to authorize it but you know so we'll start to like layer in our Sops on top of here you know so this flowart gets um so we go graphics and then we go AMZ Rec so then you know we have that decision made now I took it one step further inside of Amazon Seller Central they have the generate listing content yeah where you can upload your image and it'll kind of it'll give you an idea and what was interesting about the item name here true honey teas hibiscus lemon lift teabags hibiscus lemon lift was not an any or lift was not in any of the object recognition the object recognition on one of them only pulled hibiscus yeah but when I ran both of them one of them came back through honey te's hibiscus lemon lift 4 o bag the other one came back as true honey tea hibiscus lemon lift tea bags what I think and what I'm kind of I still need to do some testing remember this is all level one Discovery right now mhm is the generate listing content if you have text on your listing we know Amazon is using multimodels but it's going to go with the easiest one first so it's going to say hey this has text we'll use a text recognition over the object recognition and give that a higher score and so now it's giving the text recognition um more weight because these are very close to what the text was bringing versus the object now there's going to be some products like a bookshelf where you just may not have text yeah you know it's just going to be the bookshelf so then it might go to the object recognition so in that multimodal system that they have uh built yeah you know one of them has to win out you know if there's a conflict between two of them between them and I think it's going to be the text recognition that wins out based on these results here yeah I mean I'm just pulling up the as again mentioned that article M the header in there shows you a table the impact of the OCR component on model performance and the OCR detectors in this uh Pim which is the product attribute mot model um paper it shows the o o OCR detectors one of them is mask text spotter which gives an F1 score 71.1 and then you've got the other right which is Amazon recognition which gives you an output of 80.3 which I would say establishes that Amazon does use Amazon recognition on the back end and it's not just for people to use in AWS for third party projects yeah I mean they develop these tools to use in their store yeah I mean that they've come out and said that so you know they just are developing them and they're giving you the raw tool they probably have like some type of language model or you know have trained it some way yeah and I believe that the way that they've trained that the recognition is to take the text first and then the object second yeah you know based on this first test I mean we still have to kind of keep testing this out a little bit Yeah but that's based on this I can you know answer that pretty uh pretty accurately I think yeah so in our mural board what we would do is we would then put our sticky note here and we would say you know text R looks to hold higher value right so we're remember what I said like we're going to build this out and we're also going to have you know we're going to put our notes in here and some Sops that we developed this is how like when you look at a complex process you start to break it down by each individual little component and if you don't have a way to like accurately put your notes or shouldn't say accurately put your notes consistently put your notes yeah you know that's where you can you kind of lose things in the weeds a little bit so once we have Amazon recogition and we've kind we've done that where images all pass we're we're very happy with it yeah the next thing is we're going to take that mkl and we're going to go to our copy and you know whether you use Zang Guru or helium 10 or whatever tool you use or FBA Excel I mean we use data dive I think it's a great tool um so what we're first going to do is we're going to try to get it to score as high as we can yeah and then once we do that now we're going to start running individual components and then the complete title on bullet points in Amazon comprehend and I put the title in here and what a lot of people do is they put their title or they put their listing in there and they say oh I've got a 99 91% sediment yeah and you need to go beyond that because there's a thing called targeted sediment and there's a way to influence your targeted sediment that generates better product attributes in your listing so if we look at the sediment on this you know it's 79% confidence is neutral 20% positive but if you I'm not so worried about that yeah I'd like to see the positive higher but the sediment and if you read the Amazon science article the siment is really about question and answer and review sediment like what do people think about your product what I'm more concerned with is the targeted sediment which that's where it's going to break things down and the other thing it's going to break down it's going to give you an entity type and I don't see a lot of people talking about this or mentioning this and so I you know there's different entity types you know a commercial item is a branded product if you're selling something on Amazon it's a brand branded product I think you want to get as many relevant Search terms or keywords in your listing to be tagged as a commercial item and not be tagged as other you'll notice there's a lot of people that you know when you run their listings it's tagged as other they also have date they have event location organization person quantity that can be in there and then a official title so what we did is I actually I'll blow this up a little bit so the results before we had a 0 44% confidence in neutral 0.55 confidence in positive and then we we tried capitalization like what does that do for us so and actually capitalization actually increased our neutral and decreased our positive so capitalization matters inside of Amazon comprehend so we can say that now here's the before true honey te's royos te we're selling a royos chai right now look at the entity type other yeah with a 90% confidence so it's picking up Roy buus but now if we put an adjective in front of ryos so true honey te is delightful Ros tea we've now flipped it to a commercial item right so by putting that adjective in front and I found that putting adjectives in front is some of the easiest way to flip these items and you can look at your you can go into brand analytics or no the product opportunity Explorer and look at your review siment to find the adjectives that people are using with your product that's the easiest way to do it if you want to take it a step further then you can go to top Search terms to see if people actually searching delightful Roy buus or tasty Roy bus or delicious and you know use the one with the most search on it and then the the other thing the simple change actually increased our positive score to 0.58 so our posit our positivity score actually went up as well y so now here true honey teas is recognized as a commercial item it's it is in the same sentence as royos and T which are also recognized as commercial item now TR and is really the brand but I'm not too worried about that one so you can see it's commercial item now as well yeah you know Ros commercial item with that delightful in front of it and then here's another one slightly tweaking the the name true honey teas and adding the lowers the entity confidence by Point 79 as a commercial item but increases royo's entity confidence by 084 so there again you know how you write things is going to in and even these minor changes like I never would have guessed that capitalization and adjectives would really make that big of a difference but it really does and you can now see like this is a commercial item so we've lowered our neutral all the way to 38 and we've got our positive up to 61 we've also got commercial item tags yeah so even though this one is royos chit it's still tagged as neutral I don't know if we're getting the commercial item tag on royos chit yeah do we care that the sediment is neutral I mean if you look at the definition of what sediment is of what people think of your product I don't think it matters if the sediment is neutral or positive on a commercial item tag on the tag itself or overall no on the tag itself like right here yeah I don't but slightly tweaking the content we got it to a positive right so you know you can see that you know we did get that up to a positive by let's see here it doesn't have the note on there so you know here royos is tagged as other on this one this is where we put satisfying and now it's saying it's a brand h so you can definitely manipulate this yeah at the end this was what we came up with 30 you know 37 confidence and 62 so we know that the adjectives I would say adjectives play a huge role and and capitalization yeah in this just a reminder it was at 62 what did it start at the overall it started at 44 55 okay so yeah so the confidence score has gone up the neutral confidence and the positive confidence have gone up but you've also managed to attain two tags as well as in the entity and the what was the other one entity and commercial commercial item yeah now we got to work out what the values are of those down the road how do we extract the value so you manipulated to get those tags what is going to be the outcome of those tags tax and how do we measure it but it's difficult with sentiments right um there's so much that goes into optimizing listing people have all different methods to doing so and some of it is in the gray area of understanding it's like why not make sure the text is a high sentiment score just for the sake of why not do you see what I mean but is there a way of measuring it as going okay based on getting your list into a a sentiment score above X threshold this brings you a higher ratio of cells and a better conversion rate how do we measure and quantify that do you understand or oh yeah is this is or is this is this is atmospheric stuff that because you've gone into the details and you do a hundred little things like this is what makes the difference I think the biggest difference is going to be in this Comm like making sure you have no negative sent set I wouldn't want any negative but getting these commercial items I think that's going to be the biggest driver yeah so now you know here's a really good line you know of one thing that we did so now we're going to put that in our mural board right put that there because we started with data dive and then we went to Amazon comprehend and when we're done testing this Danny everybody's going to get a picture of this absolutely yeah this is what it is it's just transparent on your brand every step of the way thinking out loud in some places but then also put in something that's in a process in an order that people can evaluate and then add to their own processes yeah so Ian maybe there's a couple other um you know what's another good one that we should put in here is the capitalizing yeah we should put that in here as a note and really this is a day in the life of like the lab that I work in for a lot of these Brands and agencies where it's you know constant testing and R&D back and forth like trying to you know come up with you know this process right here that they can follow yeah um resulted to say positive so the big thing with Amazon comprehend is just don't look at the sediment look at these entity tags I think this is going to hold a lot of weight and so now that we've looked at that now the the next part is going to be Rufus yeah now and I don't want to over complicate I think I in the beginning I talked to uh a very large seller and we were talking about Rufus and I was like really over complicating it I was like well maybe if we take the top 10 asens and we you know if you look at a lot of these questions are very similar across each as in the in the category and we take these and we answer them directly in the alt tags of EC like is that going to like help generate more but I really after kind of testing a little bit talking to some other people I mean this is really where the question and answers used to be so it's a more sophisticated and there used to be a search bar in the question and answers so this is really just more of a um sophisticated search bar for the question and answers that I believe also scrapes the title and bullets to get the answers that's how I am looking at it right now but I still want to know what the questions are so does this tea have caffeine are these bags compostable is this tea gluten-free does this tea have artificial flavors is this te vegan so at least I know that I'm going to be answering these questions in my title and bullets or you know you have somebody asked the question and I can answer it in the question answers so the one thing that I am doing is I still believe that we should take whatever category you're selling in and open up the top 10 products and you know go right to the best sellers page for Herbal te's and you open the top 10 yeah and then if we start to look at it where okay does this have C this is on our listing on one of our listings for True honey tea does this have caffeine and then we can see that in this one it's the first question they ask does this tea have caffeine is this tea gluten-free is that the same is this T Glen free is the third question so these are very similar questions that they're asking in our listing as well um the the the next one does a tea have caffeine is a te gluten-free so you can see that that is a big you know two questions that all we found on three listings already and now on a fourth listing so if you're going to like if your tea is caffeine free and you are gluten-free I would say that you'd have to make sure you highlight that or have somebody ask a question and you can answer it yeah um you know so when we're writing our listing I think like taking note of what these are on the top 10 and if it's relevant to your product we should have those in the title and bullets and try to get people to ask those questions so we can answer them I think that's going to be the biggest influencing Factor on this rofus right so then we can once we do Amazon comprehend then we can do rofus right and then after that we can then uh modify listing copy for r and then after we've done modifying the listing copy we'll go back to data dive put our new listing copy in there and see how we score yeah so it's almost like you know between Amazon comprehend Rufus modifying the listing and data dive you know you may have to go through two or three cycles of this yeah to get the best listing you can so then you're going to come back here and say okay what it doe to my ranking juice am I still answering everything I still have the keywords in there I still think keywords are important yeah but I think there's other layers of complexity that we have to add now oh absolutely I mean look what people talk about all the time when they're absolutely right and I've I've banged this point home for nearly two years now people talk about as many keywords as possible and make sure that all these keywords are indexed in your listing there is nothing wrong with that that is called lexical match which is the foundation of the internet going back two decades right but like you said there is more to it than that and that's not just A9 it's every search engine out there semantics but now we're looking at Cosmo right so you're absolutely right is we do need to make sure at this stage right you want to make sure that you've got all your main keywords and yeah it's great indexing however if you read the science pap but nor semantic match requires anything to be indexed for it to read it it's still going to pick up everything on the listing because it can fully scan it just because you can't see on the front end it's not indexing for you it doesn't mean that the algorithm does not see it does that make sense oh yeah so we have to have a mind shift away from the way it's working now because there will come a day where there will be a big key turn and if you're not prepared for it and you haven't started this process already you are going to get left behind because you're going to spend six months to 12 months trying to catch back up by changing your mindset on this if you if you're very rigid and stuck to that yeah but I say it again before anyone says anything keywords that index are still important absolutely but 100% but that isn't the 100% no more that's the 10 15 20% of everything else that you need to learn that will go towards that which will help enhance things is my take on it yeah so this is how you start to put together these processes you know and you know we've done our level one testing and we can confidently say yeah you know even if you just did this you'd probably be okay like you'd be you know you'd be the top top 5% of Amazon sellers you know doing in your process if you know we can probably refine this down so you're in the top 1% yeah you know by keep testing and go to you know what we've learned in level one take it to level two and keep testing it so that way when we have this completely mapped out you know we know it's a solid process and then it's repeatable like this is something like this process right here you could give to your team your brand manager your graphics and your SEO person to be like follow this process yeah so if you have a clothing brand and you got 200 SKS you know you can follow this on all of them so we're 35 minutes in and it's been pretty heavy session for people to make notes on where do we go next do we Sav the recognition part for the other show because that's a massive sop uh Deep dive if you like to get that right for people to comprehend how to do that based on your sop on scale but what other elements is there anything you want to cover now uh no I just the big thing was to show the level one yes now we're going to go continue in the lab we're going to keep refining the process yeah so the next call we're going to have the you know we've already start to starting to have Sops built out here yeah you know the next call will start to have these built out more this is how we can do it at scale yeah um and that way they'll see this process map become more and more refined mhm and you know we'll have more findings I'm also going to test some recognition on items without text in it y excellent well look let's wrap that there um so we'll come reconvene for part two with Jeff in a couple of weeks next week next Tuesday I got a new monthly Show starting called um main main image monthly and it's a tear down show it's going to be us using our pairwise system for customer objections we've got John from uh piku will'll be a resident and the two Lads over from product opinion and obviously show favorite Sim Mahone and what we're going to be doing is we're going to break down an acing and going to use all of our different Technologies to test and then we're going to come back to the table and do a tear down on all of the images and the improvements that can be made because I think the way things are moving at the moment is that there's not enough testing going on and we need more ideas about that uh for people to have more ideas on what they can test to be surprised of there's so much that you can do you're only limited by your imagination and of course having a good foundation for testing the right things that's next Tuesday and then on Thursday we've got a uh four-part series with Colin Roger as we do a live launch we're going to walk through the whole process with Colin over that per as well so pretty excited about that we've got Sharon Evan will be back in the chair in the next couple of weeks uh Melissa will be back next week she's just had uh an operation so she'll be back in the game next week so there's a lot of stuff coming up Jeff if people want to reach you what's the best way to get in contact they can either do Facebook or we have a general email box Services atth Bedrock agency.com excellent all right guys thank you for joining us today again thank you for giving us your time Jeff I look forward to doing part two and three there's there's a lot more to cover and we'll get that booked in over the next few weeks all right guys we're going to sign off for now take care of yourself and your family much love and I'll see you again soon [Music]

[Music] hey guys welcome back to sell sessions today my guest um I&#39;ll bring him on in a moment very very sharp guy um for those who are listening on the podcast I would advise clicking on the link in the description uh because we&#39;re gonna have a full-blown geek out session um there&#39;s been a lot of chat about comprehend there&#39;s been a lot of chat about recognition and of course ruus Cosmo etc etc um so Jeff I&#39;m going to bring on shortly in there Jeff is background is in engineering he uh works with a lot of agencies he&#39;s worked with Brands he owns a brand he&#39;s exited a brand and one of his biggest skill set is to break down complex things things like recognition and stuff can you know can be kind of complex if you log into awss a lot of it seems very foreign to most people um but without further Ado I&#39;m going to bring on Jeff now and then uh get your pens and pads ready because this one&#39;s going to be pretty special Jeff welcome to the show hey how you doing Danny I&#39;m good I&#39;m good um okay anything I missed out in the introduction there I&#39;m although I mentioned uh Cosmo we&#39;re not covering that today but we it will be ruus it will be sentiments in terms of Amazon comprehend and we&#39;re going to be looking at recognition but we&#39;re also going to look at objects as well not just text recognition and we&#39;re going to look at how this can be scaled especially if you&#39;ve got a larger brand with a lot of products right yeah the main thing is is how do we build a repeatable process that could be handed off to a team member and be done you know with the same consistency every time yeah inde especially people with a large catalogs like that&#39;s one of the biggest struggles that I see them having yeah indeed so I&#39;ll let you kick off here obviously I&#39;ll chip in at various different points do you want to give everyone a breakdown of what we got planned now as we go step by step through this process yeah so we start everything off in mural kind of put the you know the main tools that we&#39;re going to using be using put any notes that we have on the tools that we&#39;re going to be using or any notes that we&#39;ve read on articles online or Amazon science articles we&#39;ll actually put these in here as well yeah and then eventually as we do our testing we&#39;ll start to build out the flowchart we&#39;ll start to you know the the business improvement process will start to take shape and at the end we&#39;ll have a complete start to finish process yeah with every step along the way with the Sops and how the um I guess the idiosyncrasies of how things work yeah and basically by the end of this show we will do some additional ones if required but depends how far we get today but you can take screenshots of this we just anyone who wants them can have these because the idea is to educate you to be able to apply this to your own business or your own agency Etc right yep absolutely yeah so we&#39;re going to cover what I call like level one Discovery yeah you know we&#39;ve done some tests we&#39;ve kind of got like uh an idea of what&#39;s going on still might need might be areas that we need to you know keep testing but we can start to develop some some answers to our questions yeah on what we have cool and so we&#39;re gonna start like not like a new listing we&#39;re actually going to start with it&#39;s an existing listing so this is somebody that wants to optimize our listing the first step I&#39;m you can use any keyword research tool but you know just develop your mkl right here we have data dive for those of you who don&#39;t know uh true honey te is a brand that I own part of any type of testing I do public share this account so you know this everything that you&#39;re going to see is eventually going to be live on Amazon especially like this royos chai tea and the Hibiscus lemon that we&#39;re releasing so we develop our Master keyword list once we&#39;re done with that we&#39;re then going to put together a graphics brief and send that to the graphic designer so we have an existing listing we have the mkl and then we&#39;re going to go to our Graphics here and from the graphics they will then develop the graphics and give it back to us now when they give it back to us what we&#39;re going to want to do is run that through recognition and try to start to get questions and answers to like how recognition ties into the Amazon search bar that&#39;s going on right now with rofus and with Cosmo so here&#39;s the Hibiscus lemon tea that we just developed and the first part of this is going to be the object detection there&#39;s two parts of recognition one is object detection which is harder you have to use Python code and the other part is text recognition which is really easy to do and you just upload the limit image and it just gives you the text that it sees now for the object recognition on this first one we can see that the brand name true honey teas hibiscus lemon lift we&#39;ve got a lemon over here it&#39;s really hard to put can you expand it a little bit so it&#39;s wider on the screen yeah just so it&#39;s bit further up there you go perfect so you know we&#39;ve got this little tag up here that this is not actually on the packaging but we put it on there um you know you&#39;ve got the Hibiscus flour some dried hibiscus there and some lemon here now if we look at the detected labels in object we&#39;ve got the label herbal which we&#39;re assuming is coming from this pile of dried leaves we&#39;ve got herbs plant we&#39;ve got the flower so that it&#39;s pulling this flower in the object we also got petal food fruit produce so they&#39;re pick up the lemon they&#39;re recognizing that it&#39;s produce citrus fruit lemon and you can see the confidence score they know it&#39;s food yeah 88% they think it&#39;s a citrus fruit and they think it&#39;s a lemon you can like watch their confidence score of that go down starts to drop yeah um and then we get the label advertisement and what was really interesting is we got the label hibiscus and I don&#39;t and it&#39;s even though it&#39;s only 56% we did get hibiscus and I was really surprised at that because if I asked you what a hibiscus looked like you know I didn&#39;t till we started developing this um typically only returns 10 labels so that&#39;s what the object recognition gave on this one right here now the next part would be the text recognition and you can see now it&#39;s picking up hibiscus lemon and it&#39;s got a huge confidence 99% yep you know teabags 99% so it&#39;s really FR through honey which is true honey that just means that you know if you have a creative or artistic brand name it may misread it yeah and then teas it picked that up hibiscus and these are actually really high confidence scores 99 lemon lift so it&#39;s picking up that 24 teabags it actually picks up if you look at this total net weight 11 ounces where that is on the product is way down here at the bottom you can hardly even read that it&#39;s kind of M it is he&#39;s very good at picking up even extremely small text I did some demonstrations in another video where the text was really small and it&#39;s still very good at picking it up so it did pick up every part of this now we do have a another one in this sample yeah that&#39;s further down here let me get down to it so here&#39;s image two and this one we added some icons across it I make this a little bit bigger we added some icons across it we put the lemons over here I mean aesthetically I like this image better personally yeah but as you said in the opening I&#39;m an engineer not a graphic designer so I don&#39;t I don&#39;t uh try to give a too much of opinion on Graphics but what I see is we got a high confidence on herbal herbs advertisement flower petal poster food fruit produce it didn&#39;t pick up lemon and it didn&#39;t pick up hibiscus so putting that lemon behind this pile here did that now say the object recognition give it a harder time because it can&#39;t distinguish that being a lemon even though the lemon didn&#39;t have as high of a confidence score in the first one you bring up a very good point the point I was going to bring up but you&#39;ve tied it we can elongate into this a lot of sellers will put all their accessories in their images right because they&#39;ll go right okay showing the value for money what you find is quite a lot of the time it confuses the OCR so it&#39;s not it&#39;s not picking up what&#39;s in there and it will give them misrepresentation and put different labels on because there&#39;s so much going on in the image now there are times it&#39;s done well I did an image stack the other day and it captured everything beautifully but I think the people that did the the main image of well classed at what they do they&#39;re also running it through recognition to ensure that every attribute is being displayed correctly or near as correct and I think that is some of the problems that the step with the cross referencing I&#39;ve done an article um called mastering attributes and images is um which relating to uh your keyword ranking timing back to here I mean we&#39;re doing this slightly differently I don&#39;t want to go off track but the point I was making is that I&#39;ve seen plenty of times where something is a pencil case and it&#39;s shown as the first AQ and where it can&#39;t cross reference that attribute with what&#39;s going on in the title and other attributes on the listing that will obviously uh work against them in terms of the algorithm understanding what&#39;s going on so yeah yeah I back you there maybe that is a possible reason why you&#39;re not getting lemon because it&#39;s being confused yeah so once we get you know these results we can kind kind of say okay we have two images that are scoring well one of them in my opinion as an engineering mind the one that has hibiscus and lemon I think would be better yeah so we can almost say that image one would be better mhm now going back to our mural board like just how we use it like here&#39;s our sop for Amazon recognition doing the objects what we&#39;ll do is we&#39;ll just take and we&#39;ll copy this and put it in our mural board so we have it for you know which we have to authorize it but you know so we&#39;ll start to like layer in our Sops on top of here you know so this flowart gets um so we go graphics and then we go AMZ Rec so then you know we have that decision made now I took it one step further inside of Amazon Seller Central they have the generate listing content yeah where you can upload your image and it&#39;ll kind of it&#39;ll give you an idea and what was interesting about the item name here true honey teas hibiscus lemon lift teabags hibiscus lemon lift was not an any or lift was not in any of the object recognition the object recognition on one of them only pulled hibiscus yeah but when I ran both of them one of them came back through honey te&#39;s hibiscus lemon lift 4 o bag the other one came back as true honey tea hibiscus lemon lift tea bags what I think and what I&#39;m kind of I still need to do some testing remember this is all level one Discovery right now mhm is the generate listing content if you have text on your listing we know Amazon is using multimodels but it&#39;s going to go with the easiest one first so it&#39;s going to say hey this has text we&#39;ll use a text recognition over the object recognition and give that a higher score and so now it&#39;s giving the text recognition um more weight because these are very close to what the text was bringing versus the object now there&#39;s going to be some products like a bookshelf where you just may not have text yeah you know it&#39;s just going to be the bookshelf so then it might go to the object recognition so in that multimodal system that they have uh built yeah you know one of them has to win out you know if there&#39;s a conflict between two of them between them and I think it&#39;s going to be the text recognition that wins out based on these results here yeah I mean I&#39;m just pulling up the as again mentioned that article M the header in there shows you a table the impact of the OCR component on model performance and the OCR detectors in this uh Pim which is the product attribute mot model um paper it shows the o o OCR detectors one of them is mask text spotter which gives an F1 score 71.1 and then you&#39;ve got the other right which is Amazon recognition which gives you an output of 80.3 which I would say establishes that Amazon does use Amazon recognition on the back end and it&#39;s not just for people to use in AWS for third party projects yeah I mean they develop these tools to use in their store yeah I mean that they&#39;ve come out and said that so you know they just are developing them and they&#39;re giving you the raw tool they probably have like some type of language model or you know have trained it some way yeah and I believe that the way that they&#39;ve trained that the recognition is to take the text first and then the object second yeah you know based on this first test I mean we still have to kind of keep testing this out a little bit Yeah but that&#39;s based on this I can you know answer that pretty uh pretty accurately I think yeah so in our mural board what we would do is we would then put our sticky note here and we would say you know text R looks to hold higher value right so we&#39;re remember what I said like we&#39;re going to build this out and we&#39;re also going to have you know we&#39;re going to put our notes in here and some Sops that we developed this is how like when you look at a complex process you start to break it down by each individual little component and if you don&#39;t have a way to like accurately put your notes or shouldn&#39;t say accurately put your notes consistently put your notes yeah you know that&#39;s where you can you kind of lose things in the weeds a little bit so once we have Amazon recogition and we&#39;ve kind we&#39;ve done that where images all pass we&#39;re we&#39;re very happy with it yeah the next thing is we&#39;re going to take that mkl and we&#39;re going to go to our copy and you know whether you use Zang Guru or helium 10 or whatever tool you use or FBA Excel I mean we use data dive I think it&#39;s a great tool um so what we&#39;re first going to do is we&#39;re going to try to get it to score as high as we can yeah and then once we do that now we&#39;re going to start running individual components and then the complete title on bullet points in Amazon comprehend and I put the title in here and what a lot of people do is they put their title or they put their listing in there and they say oh I&#39;ve got a 99 91% sediment yeah and you need to go beyond that because there&#39;s a thing called targeted sediment and there&#39;s a way to influence your targeted sediment that generates better product attributes in your listing so if we look at the sediment on this you know it&#39;s 79% confidence is neutral 20% positive but if you I&#39;m not so worried about that yeah I&#39;d like to see the positive higher but the sediment and if you read the Amazon science article the siment is really about question and answer and review sediment like what do people think about your product what I&#39;m more concerned with is the targeted sediment which that&#39;s where it&#39;s going to break things down and the other thing it&#39;s going to break down it&#39;s going to give you an entity type and I don&#39;t see a lot of people talking about this or mentioning this and so I you know there&#39;s different entity types you know a commercial item is a branded product if you&#39;re selling something on Amazon it&#39;s a brand branded product I think you want to get as many relevant Search terms or keywords in your listing to be tagged as a commercial item and not be tagged as other you&#39;ll notice there&#39;s a lot of people that you know when you run their listings it&#39;s tagged as other they also have date they have event location organization person quantity that can be in there and then a official title so what we did is I actually I&#39;ll blow this up a little bit so the results before we had a 0 44% confidence in neutral 0.55 confidence in positive and then we we tried capitalization like what does that do for us so and actually capitalization actually increased our neutral and decreased our positive so capitalization matters inside of Amazon comprehend so we can say that now here&#39;s the before true honey te&#39;s royos te we&#39;re selling a royos chai right now look at the entity type other yeah with a 90% confidence so it&#39;s picking up Roy buus but now if we put an adjective in front of ryos so true honey te is delightful Ros tea we&#39;ve now flipped it to a commercial item right so by putting that adjective in front and I found that putting adjectives in front is some of the easiest way to flip these items and you can look at your you can go into brand analytics or no the product opportunity Explorer and look at your review siment to find the adjectives that people are using with your product that&#39;s the easiest way to do it if you want to take it a step further then you can go to top Search terms to see if people actually searching delightful Roy buus or tasty Roy bus or delicious and you know use the one with the most search on it and then the the other thing the simple change actually increased our positive score to 0.58 so our posit our positivity score actually went up as well y so now here true honey teas is recognized as a commercial item it&#39;s it is in the same sentence as royos and T which are also recognized as commercial item now TR and is really the brand but I&#39;m not too worried about that one so you can see it&#39;s commercial item now as well yeah you know Ros commercial item with that delightful in front of it and then here&#39;s another one slightly tweaking the the name true honey teas and adding the lowers the entity confidence by Point 79 as a commercial item but increases royo&#39;s entity confidence by 084 so there again you know how you write things is going to in and even these minor changes like I never would have guessed that capitalization and adjectives would really make that big of a difference but it really does and you can now see like this is a commercial item so we&#39;ve lowered our neutral all the way to 38 and we&#39;ve got our positive up to 61 we&#39;ve also got commercial item tags yeah so even though this one is royos chit it&#39;s still tagged as neutral I don&#39;t know if we&#39;re getting the commercial item tag on royos chit yeah do we care that the sediment is neutral I mean if you look at the definition of what sediment is of what people think of your product I don&#39;t think it matters if the sediment is neutral or positive on a commercial item tag on the tag itself or overall no on the tag itself like right here yeah I don&#39;t but slightly tweaking the content we got it to a positive right so you know you can see that you know we did get that up to a positive by let&#39;s see here it doesn&#39;t have the note on there so you know here royos is tagged as other on this one this is where we put satisfying and now it&#39;s saying it&#39;s a brand h so you can definitely manipulate this yeah at the end this was what we came up with 30 you know 37 confidence and 62 so we know that the adjectives I would say adjectives play a huge role and and capitalization yeah in this just a reminder it was at 62 what did it start at the overall it started at 44 55 okay so yeah so the confidence score has gone up the neutral confidence and the positive confidence have gone up but you&#39;ve also managed to attain two tags as well as in the entity and the what was the other one entity and commercial commercial item yeah now we got to work out what the values are of those down the road how do we extract the value so you manipulated to get those tags what is going to be the outcome of those tags tax and how do we measure it but it&#39;s difficult with sentiments right um there&#39;s so much that goes into optimizing listing people have all different methods to doing so and some of it is in the gray area of understanding it&#39;s like why not make sure the text is a high sentiment score just for the sake of why not do you see what I mean but is there a way of measuring it as going okay based on getting your list into a a sentiment score above X threshold this brings you a higher ratio of cells and a better conversion rate how do we measure and quantify that do you understand or oh yeah is this is or is this is this is atmospheric stuff that because you&#39;ve gone into the details and you do a hundred little things like this is what makes the difference I think the biggest difference is going to be in this Comm like making sure you have no negative sent set I wouldn&#39;t want any negative but getting these commercial items I think that&#39;s going to be the biggest driver yeah so now you know here&#39;s a really good line you know of one thing that we did so now we&#39;re going to put that in our mural board right put that there because we started with data dive and then we went to Amazon comprehend and when we&#39;re done testing this Danny everybody&#39;s going to get a picture of this absolutely yeah this is what it is it&#39;s just transparent on your brand every step of the way thinking out loud in some places but then also put in something that&#39;s in a process in an order that people can evaluate and then add to their own processes yeah so Ian maybe there&#39;s a couple other um you know what&#39;s another good one that we should put in here is the capitalizing yeah we should put that in here as a note and really this is a day in the life of like the lab that I work in for a lot of these Brands and agencies where it&#39;s you know constant testing and R&amp;D back and forth like trying to you know come up with you know this process right here that they can follow yeah um resulted to say positive so the big thing with Amazon comprehend is just don&#39;t look at the sediment look at these entity tags I think this is going to hold a lot of weight and so now that we&#39;ve looked at that now the the next part is going to be Rufus yeah now and I don&#39;t want to over complicate I think I in the beginning I talked to uh a very large seller and we were talking about Rufus and I was like really over complicating it I was like well maybe if we take the top 10 asens and we you know if you look at a lot of these questions are very similar across each as in the in the category and we take these and we answer them directly in the alt tags of EC like is that going to like help generate more but I really after kind of testing a little bit talking to some other people I mean this is really where the question and answers used to be so it&#39;s a more sophisticated and there used to be a search bar in the question and answers so this is really just more of a um sophisticated search bar for the question and answers that I believe also scrapes the title and bullets to get the answers that&#39;s how I am looking at it right now but I still want to know what the questions are so does this tea have caffeine are these bags compostable is this tea gluten-free does this tea have artificial flavors is this te vegan so at least I know that I&#39;m going to be answering these questions in my title and bullets or you know you have somebody asked the question and I can answer it in the question answers so the one thing that I am doing is I still believe that we should take whatever category you&#39;re selling in and open up the top 10 products and you know go right to the best sellers page for Herbal te&#39;s and you open the top 10 yeah and then if we start to look at it where okay does this have C this is on our listing on one of our listings for True honey tea does this have caffeine and then we can see that in this one it&#39;s the first question they ask does this tea have caffeine is this tea gluten-free is that the same is this T Glen free is the third question so these are very similar questions that they&#39;re asking in our listing as well um the the the next one does a tea have caffeine is a te gluten-free so you can see that that is a big you know two questions that all we found on three listings already and now on a fourth listing so if you&#39;re going to like if your tea is caffeine free and you are gluten-free I would say that you&#39;d have to make sure you highlight that or have somebody ask a question and you can answer it yeah um you know so when we&#39;re writing our listing I think like taking note of what these are on the top 10 and if it&#39;s relevant to your product we should have those in the title and bullets and try to get people to ask those questions so we can answer them I think that&#39;s going to be the biggest influencing Factor on this rofus right so then we can once we do Amazon comprehend then we can do rofus right and then after that we can then uh modify listing copy for r and then after we&#39;ve done modifying the listing copy we&#39;ll go back to data dive put our new listing copy in there and see how we score yeah so it&#39;s almost like you know between Amazon comprehend Rufus modifying the listing and data dive you know you may have to go through two or three cycles of this yeah to get the best listing you can so then you&#39;re going to come back here and say okay what it doe to my ranking juice am I still answering everything I still have the keywords in there I still think keywords are important yeah but I think there&#39;s other layers of complexity that we have to add now oh absolutely I mean look what people talk about all the time when they&#39;re absolutely right and I&#39;ve I&#39;ve banged this point home for nearly two years now people talk about as many keywords as possible and make sure that all these keywords are indexed in your listing there is nothing wrong with that that is called lexical match which is the foundation of the internet going back two decades right but like you said there is more to it than that and that&#39;s not just A9 it&#39;s every search engine out there semantics but now we&#39;re looking at Cosmo right so you&#39;re absolutely right is we do need to make sure at this stage right you want to make sure that you&#39;ve got all your main keywords and yeah it&#39;s great indexing however if you read the science pap but nor semantic match requires anything to be indexed for it to read it it&#39;s still going to pick up everything on the listing because it can fully scan it just because you can&#39;t see on the front end it&#39;s not indexing for you it doesn&#39;t mean that the algorithm does not see it does that make sense oh yeah so we have to have a mind shift away from the way it&#39;s working now because there will come a day where there will be a big key turn and if you&#39;re not prepared for it and you haven&#39;t started this process already you are going to get left behind because you&#39;re going to spend six months to 12 months trying to catch back up by changing your mindset on this if you if you&#39;re very rigid and stuck to that yeah but I say it again before anyone says anything keywords that index are still important absolutely but 100% but that isn&#39;t the 100% no more that&#39;s the 10 15 20% of everything else that you need to learn that will go towards that which will help enhance things is my take on it yeah so this is how you start to put together these processes you know and you know we&#39;ve done our level one testing and we can confidently say yeah you know even if you just did this you&#39;d probably be okay like you&#39;d be you know you&#39;d be the top top 5% of Amazon sellers you know doing in your process if you know we can probably refine this down so you&#39;re in the top 1% yeah you know by keep testing and go to you know what we&#39;ve learned in level one take it to level two and keep testing it so that way when we have this completely mapped out you know we know it&#39;s a solid process and then it&#39;s repeatable like this is something like this process right here you could give to your team your brand manager your graphics and your SEO person to be like follow this process yeah so if you have a clothing brand and you got 200 SKS you know you can follow this on all of them so we&#39;re 35 minutes in and it&#39;s been pretty heavy session for people to make notes on where do we go next do we Sav the recognition part for the other show because that&#39;s a massive sop uh Deep dive if you like to get that right for people to comprehend how to do that based on your sop on scale but what other elements is there anything you want to cover now uh no I just the big thing was to show the level one yes now we&#39;re going to go continue in the lab we&#39;re going to keep refining the process yeah so the next call we&#39;re going to have the you know we&#39;ve already start to starting to have Sops built out here yeah you know the next call will start to have these built out more this is how we can do it at scale yeah um and that way they&#39;ll see this process map become more and more refined mhm and you know we&#39;ll have more findings I&#39;m also going to test some recognition on items without text in it y excellent well look let&#39;s wrap that there um so we&#39;ll come reconvene for part two with Jeff in a couple of weeks next week next Tuesday I got a new monthly Show starting called um main main image monthly and it&#39;s a tear down show it&#39;s going to be us using our pairwise system for customer objections we&#39;ve got John from uh piku will&#39;ll be a resident and the two Lads over from product opinion and obviously show favorite Sim Mahone and what we&#39;re going to be doing is we&#39;re going to break down an acing and going to use all of our different Technologies to test and then we&#39;re going to come back to the table and do a tear down on all of the images and the improvements that can be made because I think the way things are moving at the moment is that there&#39;s not enough testing going on and we need more ideas about that uh for people to have more ideas on what they can test to be surprised of there&#39;s so much that you can do you&#39;re only limited by your imagination and of course having a good foundation for testing the right things that&#39;s next Tuesday and then on Thursday we&#39;ve got a uh four-part series with Colin Roger as we do a live launch we&#39;re going to walk through the whole process with Colin over that per as well so pretty excited about that we&#39;ve got Sharon Evan will be back in the chair in the next couple of weeks uh Melissa will be back next week she&#39;s just had uh an operation so she&#39;ll be back in the game next week so there&#39;s a lot of stuff coming up Jeff if people want to reach you what&#39;s the best way to get in contact they can either do Facebook or we have a general email box Services atth Bedrock agency.com excellent all right guys thank you for joining us today again thank you for giving us your time Jeff I look forward to doing part two and three there&#39;s there&#39;s a lot more to cover and we&#39;ll get that booked in over the next few weeks all right guys we&#39;re going to sign off for now take care of yourself and your family much love and I&#39;ll see you again soon [Music]

Transcript for:Amazon Tools for Sellers Overview

Transcript for:
Amazon Tools for Sellers Overview