Latest / Inside Commerce: Ecommerce Strategy, CX and Technology Podcast / EP202: The Evolution to Generative AI for Content Personalisation in Ecommerce Sites, with PersonifyXP's Co-founder & CTO
Transcript
- 0:02Frank to listen to the replay from Podcast Sports by Attract
- 0:06and Elastic Path and hosted by me, James Rogers, we'll welcome
- 0:10first episode. We've got a lot of evidence of
- 0:12praying content, bringing you really interesting guests all
- 0:14free of charge. Help us keep it that way.
- 0:17Tell other people about what cross we get.
- 0:19New listers share our posts from LinkedIn and Twitter so you can
- 0:22have a Fireridge. The best thing you can do, if
- 0:24you haven't already, is give us that rating on that response by
- 0:26YouTube or whatever else you might be listed.
- 0:28Watch it. So let me introduce the this
- 0:31week's episode and then I'll hand it back to our guest.
- 0:33The whole topic is the evolution of generative AI for content
- 0:36personalization and ecom sites. So I think this is really
- 0:39interesting area of AI because a lot has been filled around
- 0:42product recommendations, but there's a much wider piece to do
- 0:45on personalization on ecom. And today we're talking to Josh
- 0:49Gosson. He's a cofounder and CTO at one
- 0:53of the leading players in this area, Some Persona Philly, some
- 0:56of you may have come across and some may have already been
- 0:58customers. He's also an editorial board
- 1:00member for Frontiers and AI. And for those of you who go back
- 1:04in the industry, a few more years have worked in fashion
- 1:07people, as might know in these Days of Truth.
- 1:09So I warm up. Josh Haider is her.
- 1:13Thanks, James. Nice to be here.
- 1:15Yeah, I'm doing great. Thank you.
- 1:17Yeah, cool. So do you want to give people?
- 1:22So I was going to say don't give people a bit of an interaction
- 1:25to what your role is and also set the seed what personify is
- 1:28so those I'm across it know exactly where it fits.
- 1:33Yeah, okay. So what I did.
- 1:35So I Co founded personify. I've got a background in AI and
- 1:40user behavior research, and so we started personify when we
- 1:46were working in the fashion industry and we were trying to
- 1:48understand. Why people are buying particular
- 1:51items or why they're browsing particular items mainly to give
- 1:55sort of sort of fashion styling advice to the users in real
- 2:00time. And what we found is that
- 2:03personalization is very at that time and it still is relatively
- 2:08today. It's very about what are you
- 2:10doing right now? Are you looking at product A or
- 2:12product B? You're looking at genes.
- 2:13You're just going to get more genes, but actually.
- 2:16Your behavior across the entire of your session, and also your
- 2:19historic behavior can actually give more information as to the
- 2:24reason why you're coming to a website.
- 2:26You might be looking for a an outfit for a wedding, or you
- 2:32might be just, you know, buying some more jeans because you've
- 2:35just ripped your last pair. You know, there's lots of
- 2:37different reasons for a user to come to a website, and it
- 2:40changes with each visit, and so understanding the context of the
- 2:44visit is incredibly important. To personalizing anything really
- 2:49on the website, it might be not just product recommendations but
- 2:53the content and sometimes the layout of the actual website can
- 2:58be modified to engage the user more and personalized more
- 3:03individually to that user. So that's our goal that
- 3:06personifies to enable that for retailers and brands.
- 3:11Right. And yeah, we'll come on to the
- 3:12content side a lot more later on.
- 3:15So first question, so can you talk us through some of your Rd.
- 3:20activity at the moment around generative AI and how that
- 3:24differs from some of the other solutions on the market and
- 3:26maybe give us a few application examples?
- 3:30Yeah. So what we do, the core of what
- 3:33we do at Personify is identifying different types of
- 3:35behaviors. So looking at all of the traffic
- 3:40on the website and understanding what kinds of reasons do people
- 3:44shop? And once we understand that, we
- 3:46can then start to map that particular content.
- 3:48We can say okay this, this banner really engages well with
- 3:51people who are shopping for holiday and this is for somebody
- 3:56who's giving up their home and buying new furniture or new
- 3:59lighting. And so we can we, we sort of
- 4:02understand the different types of behaviors or shopping goals
- 4:06that user might have. We map content onto that and
- 4:09that then allows us to actually spot where the gaps are.
- 4:13So where you know the particular groups of users aren't being
- 4:17served accurately or effectively and this allows us, you know
- 4:22where we have been in terms of product so far is that we
- 4:27identify those gaps and enable content management systems.
- 4:30The authoring process for CMS is to basically speed up that
- 4:34process and identify where the most effective content
- 4:37generation sort of effort should be placed and when we get into
- 4:40AI generation. So generative AI and sort of
- 4:45content and creation, it's very similar.
- 4:48You identify the gaps and you basically create content for
- 4:52those gaps. So one of those things is you
- 4:55can put in a lot of the information around the context
- 4:57of the behavior. So this you can say this person
- 5:00comes from these types of referrers, they are interested
- 5:04in these types of products and they're going to have these
- 5:06kinds of outcomes and you can. Push that into generative AI and
- 5:11start to generate sort of content examples.
- 5:13So you can say generate me a banner, generate me some copy
- 5:17and so that's that's basically where our Rd. is focused in on
- 5:20is creating that. The benefit of that is obviously
- 5:26you can use content generation to at the moment to do all of
- 5:31that, but it's the context and working out where the most
- 5:35effective use is because. The reality of the situation is
- 5:38that nobody sort of says go away and create content and then
- 5:42slaps it straight on the page because there's a lot of
- 5:45variance in the quality. And so you actually, you not
- 5:48only have to generate it automatically, but you need a
- 5:52workflow which is streamlined to speed up the process of
- 5:57approving that content before it goes and hits the website.
- 6:00And so when you're looking into sort of the rod effort, it's not
- 6:03just. And then to find the gaps, which
- 6:06sort of does, it's not just creating the content to meet
- 6:09those gaps. It's also that approval work
- 6:11stream and making sure the whole process is joined up and simple
- 6:14to use. And that's where you get the
- 6:16best impact because you are taking the whole process end to
- 6:23end and syncing it up and that's where we're focused at the
- 6:27moment. Crown Peak enabling
- 6:30organizations to build meaningful digital experiences
- 6:33for everyone, everywhere. Engage your customers with our
- 6:36content management system, reach your entire audience with our
- 6:40digital accessibility and quality tools, and grow your
- 6:44brand with our product. Discovery solution?
- 6:46Find out more at crownpeak.com. Right.
- 6:49Yeah, super interesting. Yeah, we've got clients that's
- 6:53done quite a lot with some of the generative AI solution to
- 6:57create lifestyle content. And yeah, I think with the right
- 7:00kind of prompts, I think that is very powerful.
- 7:02So yeah, getting that right, I think would be super interesting
- 7:06for brand. The next question to content
- 7:10personalization. I feel like it's often seen as
- 7:13quite a big project, quite a broad thing to kind of get
- 7:16right. What's your view on kind of the
- 7:19best starting point and how do people usually get started with
- 7:22your product? So content personalization.
- 7:27This is like I used to research when I was doing my PhD in
- 7:32education is that we would create individual learning
- 7:36tutorials for individual users and based on their previous
- 7:40knowledge. And it is very similar to what
- 7:42we actually do at Sonify in terms of the content
- 7:44personalization. So I'll explain it a little bit
- 7:45more detail. Effectively, you might be
- 7:48talking about a certain topic and you need to have an
- 7:51introduction. You need to have examples which
- 7:53fit the user's learning styles, and then you need conclusions
- 7:55which wrap it all up. And the problem with doing that
- 7:58and that we found research at Warren University is that you
- 8:05had to create huge amounts of variations in terms of the
- 8:08content, sort of segments. We used to call them and.
- 8:12So you had a huge the content space just exploded in the
- 8:16number of different items and all the metadata that you needed
- 8:19to tag it with. And that was the biggest
- 8:20research problem back then. Going forwards to retail, it's
- 8:26less complicated in terms of how things match together when
- 8:31somebody comes to a landing page and you want to stick a banner
- 8:34on the top of the page. You don't have to make it sync
- 8:38with the rest of the page all the time.
- 8:40It just needs, you know, people understand that there's
- 8:42different blocks, they're going to appeal to different users.
- 8:44The main thing is actually making Zach, which appeals to
- 8:47them in prime position. And so you're more focused on
- 8:51sort of wall content piece at the time when you're looking at
- 8:53landing pages. And then when you get into the
- 8:55rest of the journey, it's more about syncing those content
- 8:59pieces up so they are cohesive across the journey rather than
- 9:02necessarily cohesive across page.
- 9:04Especially early on in the journey where you're sort of
- 9:06trying to expand what you're showing to the users.
- 9:09So what that means practically is the number of content pieces
- 9:14or sort of content assets that you might put into a particular
- 9:17content slot on the page is a lot less than people might
- 9:21expect. So you know, if you've got a
- 9:25landing page and you've got like maybe five or six major groups
- 9:29of users, you're not going to need. 20-30 pieces of content
- 9:34you can make do with like 3 or 4 because what you're trying to do
- 9:38is you're trying to personalize and get basically the most
- 9:41effective option in front of the user, which is better than the
- 9:45default. And that will start increasing
- 9:48your your reducing your bounce rate, increasing your protection
- 9:54immediately. And you could do this with quite
- 9:57a low sort of content asset library.
- 10:03And what you can then do is identify where you're missing
- 10:06particular groups of customers and then you just build the
- 10:08content on there. So it's not like we've got to
- 10:11create you know 40-50 pieces of content before we can even start
- 10:15this personalization. You can actually start quite
- 10:16smaller than scale up. And what we normally do when
- 10:19we're working with clients is we'll say okay, we've got, we've
- 10:23done this behavior review. You know you've got these types
- 10:26of customers, these types of personas which match maybe what
- 10:29you've already know and some that you don't and.
- 10:33You then take that information and say, okay, well, we can,
- 10:37we're gonna hit the major traffic points first.
- 10:40We're going to look at the concept that you already have
- 10:43and repurpose it and basically use the personalization
- 10:47capability to show the right thing to the user, maybe
- 10:50rearrange some of the blocks on the landing pages, that kind of
- 10:52thing. And that enables you to start
- 10:55going straight away and then. What we'll then do is use a test
- 10:59and learn approach. So we basically say okay, how
- 11:00effective is this? Do we want to improve it?
- 11:02Which users need to be targeted next?
- 11:05And that allows you to roll it out incrementally, but still see
- 11:10improvements in your metrics at each stage.
- 11:13And so you can start quite slow that people get very scared when
- 11:17they talk about content personalization and 1:00 to 1:00
- 11:19personalization, because it seems like a really far away
- 11:21sort of goal to what they are right now, but you can actually
- 11:24start to achieve. Really good things with them
- 11:27with almost what you have already if you just have the
- 11:31right smarts behind it. Yeah question relates to that
- 11:35because that sounds definitely think people struggled in the
- 11:38past when they try at least seeing personalization as it's
- 11:41all one to one, it's custom as opposed to actually let's do it
- 11:45based on the behavioral group. So I think that's quite an
- 11:47interesting way of simplifying it initially.
- 11:49How do you get to the level of knowledge of where to focus on?
- 11:53Because the user journey is very massively and you look at large
- 11:56sites and like the user journeys and there are potentially
- 11:59hundreds of thousands of different.
- 12:02How do you get to know which user journeys to focus on?
- 12:05How much of that is automated? I guess in terms of you ingest
- 12:08analytic stage and then you already know what user journeys
- 12:11and which pages. That would be really interesting
- 12:13to understand how you get to that level of low and wet.
- 12:18Yeah. So in terms of how in terms of
- 12:22like the user journeys? That it has been a stumbling
- 12:27block to actually creating sort of automatic segmentation for a
- 12:30very long time. Lots of people, you know, you
- 12:33could create segments automatically all day long using
- 12:36about 100 different sort of algorithms.
- 12:40But actually the hardest part has been to be where the bar is
- 12:44set. You know, where the way you draw
- 12:47the line between this is now irrelevant and this is actually
- 12:51useful. And so a lot of people have gone
- 12:53too, too small. And a lot of people have gone
- 12:56too big and it's always finding that fine line.
- 12:58And what you find is that a lot of the data science capabilities
- 13:01within larger organizations optimize it manually for their
- 13:06particular retailer brand and that works.
- 13:12But when they then try to transport that to sort of
- 13:15translate that across to another brand or basically make it more
- 13:18general purpose, that's where it sort of starts to break down.
- 13:22What we found is we use a custom proprietary sort of algorithm
- 13:29and algorithm that we use to identify the behaviors.
- 13:32That's our sort of secret source.
- 13:34That's what I brought to the table when we started.
- 13:38And what this allows us to do is we're looking at it in terms of.
- 13:43Not necessarily specific behavior.
- 13:45We're not saying, you know, you go to red jeans and then you
- 13:47know your red shirts, then these jeans and then you check out.
- 13:50But more sort of a little bit a step back in terms of grouping
- 13:56like some of those thousands of journeys together and finding
- 13:59the commonality between the journeys.
- 14:01What that enables you to do is you might have people coming in
- 14:04from lots of different referrer places, but parts of the journey
- 14:07are very common. And you might have branching
- 14:11sort of areas in those journeys where they sort of diverge from
- 14:14the door, but that's basically captured within that particular
- 14:17behavior. And what we find is that when
- 14:20you look at this, it might be you're sort of you have some
- 14:27sort of like categories, sometimes they're colors, you
- 14:30know with some customers with women's wear you start to find,
- 14:34OK, this is really bold, you know, red and black colors, all
- 14:37these are pallets. You know, sort of pastel colors
- 14:40or they might be looking for a different reason.
- 14:42They might be looking at combinations of items to buy an
- 14:45outfit and then you're looking at, you know, these people are
- 14:47looking for either jumpsuits or flowing, flowing trousers and
- 14:52the top. Or if you go to electronics, it
- 14:54might be these people are really technical, so they always look
- 14:57at these high end spec and there's a reason why they look
- 15:01at these particular sets of customers in combination.
- 15:04So the algorithm picks that up. And basically you start to see
- 15:08themes, they're not all the same.
- 15:10It's not like you know these are your outfitting journeys.
- 15:12It's like these, this is an outfitting journey, This is a
- 15:15color driven journey and it identifies the driver to the
- 15:19user behind the journey rather than the really granular detail
- 15:24in the journey itself. And that's that's the reason
- 15:29that you can then start to apply it across the user journey
- 15:32because. If you understand the driver for
- 15:35a particular user, you can start to actually show content related
- 15:38to that and then also products which are related to it.
- 15:40And also the outcome which might not be a conversion.
- 15:44It might be a newsletter sign up or it might be getting somebody
- 15:47to interact with a questionnaire or you know download a coupon
- 15:51for some of our cpmg clients in the States.
- 15:54You know, there's lots of different ways that you can then
- 15:56sort of tweak the journey and sort of interact with the user
- 16:00later on depending on that. Sort of their shopping drivers,
- 16:05Elastic Path is on a mission to breakthrough the barriers that
- 16:08prevent commerce leaders from delivering extraordinary
- 16:10shopping experiences. Both business and tech teams are
- 16:14able to innovate on their own terms, fully embrace
- 16:17composability, and in turn, never compromise on what's
- 16:21possible. Learn more at www.elasticpath
- 16:24Calm. Yeah, that's.
- 16:27I mean, I like that thinking as well because there are so many
- 16:29different trapping sources. Often people customize content
- 16:32based purely on source, but actually the behavior can be
- 16:36similar between multiple sources and that we're picking out the
- 16:39the commonality of behaviors regardless of which campaign
- 16:42it's come from. Yeah, that a lot of that's that
- 16:44makes perfect sense. The question that then comes on
- 16:48from that which a lot of people ask is actually if I'm going to
- 16:52use like a content personalization engine, then I
- 16:56need a team of people to make use of this and to make sure we
- 16:59have the right assets and that we're analyzing, understand the
- 17:02behaviors and making the right decisions.
- 17:04And these work on this. Can you talk us through from the
- 17:07customers you work like what sort of resources do they need
- 17:11and how to work alongside sonified to make the most out of
- 17:15the blackboard? What type of roles, how big are
- 17:17the teams use it? Yeah.
- 17:22So in terms of the. It depends on their strategy.
- 17:26Some, some of our clients have taken the, the the sort of
- 17:29landing page optimization and saying OK, we're going to hit
- 17:33banners and that's what we're going to do with content
- 17:35personalization. And what they'll do is they'll
- 17:37take the behavior information that we have.
- 17:39So they'll look at recent behavior and they'll say, OK, we
- 17:42know that you know we have like 20% of customers are are really
- 17:47interested in sports. We know that the majority,
- 17:50though, are going to be just interested in loungewear.
- 17:53And we've got these really specific segments which just due
- 17:56to seasonality, they're really important right now.
- 17:59They'll then look at that behavior and look at the content
- 18:01that they already have created, identify gaps and then create
- 18:05content for that. So it goes into their existing
- 18:07sort of content creation workflow, but they are targeted
- 18:10better. So it's not necessarily
- 18:12increasing the time taken. In fact, you can decrease it,
- 18:18because then they know they're not.
- 18:20They don't need to recreate lots of stuff which.
- 18:22Isn't actually relevant right now, but it makes them work
- 18:25smarter. So that's the that's one way.
- 18:31The other way is that people might say okay, we really want
- 18:36to personalize. So we work with a brand in
- 18:38healthcare with pain medication and obviously content for
- 18:44regulator spaces like that is quite hard because you have to
- 18:47go through legal approvals and other things like this.
- 18:51So that's always going to take some time, but they're working
- 18:55in terms of we are going to target different types of
- 18:59content dependent on whether we think somebody's got ongoing
- 19:03pain or whether it's maybe a sports injury, maybe whether
- 19:06they're an older demographic or younger demographic.
- 19:09And So what they looked at is we we took a project.
- 19:13So they they they had some of their agency resources and some
- 19:16internal resources, not not new resources, but just more of that
- 19:20time. And they they basically said,
- 19:23OK, these are our high traffic points.
- 19:26These are the critical sort of points in the journey from the
- 19:28information that they got from us.
- 19:30I said OK, we want to basically we want to change some of the,
- 19:34we need to create some additional content in terms of
- 19:39articles and to meet these particular goals because they're
- 19:42not served well. And we also need to create
- 19:44different variants of banners. So there might be an article
- 19:47about yoga, and we might show somebody who's.
- 19:50Maybe looks a little bit more sporty for people who have got a
- 19:53sports injury and if you've got arthritis, then we're going to
- 19:56put a picture of maybe an image of a class that might appeal, a
- 20:02yo class that might appeal to those types of users more.
- 20:05And so basically switching out this content as well as putting
- 20:08actually additional copy on the site.
- 20:11So that is a sort of new content added.
- 20:14That's sort of an extra additional project which you can
- 20:16do. So that's where you're sort of
- 20:17going, okay, we're not just. Focusing on improving the
- 20:21intelligence in a workflow, but we're also saying, OK, we're
- 20:26looking at a whole website and we're actually building new
- 20:28things for it. So that's that's that's sort of
- 20:31the additional work on top and that that can be met either by
- 20:36prioritizing things in your workflow and repurposing
- 20:39existing sort of resources or or actually sort of like creating a
- 20:45new project as this customer did with their agency to actually
- 20:49keep this off. And so that's those are the sort
- 20:52of two styles that we've had predominantly with customers so
- 20:55far. And next question, so automatic
- 21:00segmentation is one of your key features.
- 21:04How is the platform kind of defining the segments and kind
- 21:07of keeping them meaningful given the amount of the kind of routes
- 21:12that could be gone downhill? Yeah.
- 21:15So as as I talked before about in terms of the the drivers
- 21:20behind behavior that's that's the main thing in terms of
- 21:23creating them and and working out whether they're improved, we
- 21:27we do a lot, we have, we have a lot of smarts behind the scenes
- 21:30in terms of looking at whether when we create a segment whether
- 21:34it's relevant to the and to the traffic.
- 21:39So we can, we can basically when we create the the segment which
- 21:43is our behavior, so, so the behavior maps to a segment, we
- 21:47can then sort of play it back through traffic, make sure that
- 21:50it's still accurate in terms of predicting that traffic's
- 21:52behavior. And so for every sort of
- 21:55behavior that we recognize, we can also give it a score in
- 21:57terms of accuracy and then track it in real time as as a new
- 22:03traffic comes through. What that allows you to do is
- 22:05basically you have an accuracy error tracker and when it hits a
- 22:09certain threshold, you can say, OK, this actually is no longer
- 22:13relevant. We need to get rid of that and
- 22:15then look at the traffic that was mapped to that and see if it
- 22:18needs to be mapped to either existing behaviors or needs to
- 22:22create new ones again. So you basically have this
- 22:25fairly organic flow in terms of these behaviors, in terms of
- 22:29working out whether they're still valid, working out whether
- 22:31we need to create new ones. And that's that's part of the
- 22:38sort of differentiation with personify in terms of it's not
- 22:41just a one off sort of behavior identification piece and then
- 22:44you have to redo that constantly.
- 22:46But it's it's an always on sort of automatic piece and that
- 22:52doesn't mean that it changes and not a lot of behaviors will stay
- 22:56constant. You know, when you talk about
- 22:58automatic segmentation and it's always sort of identifying
- 23:02different behaviors, people look, people go, oh wait, so
- 23:04I've actually got to do more work to keep on top of these
- 23:07things. And how they sort of translate
- 23:09to my existing personalization strategy is actually less than
- 23:13you'd think. You know, if you'll say
- 23:16electronics or you're in healthcare, your behaviors are
- 23:19going to change dramatically. They're going to adjust to
- 23:24behaviors as they come on that they, the behaviors will, will
- 23:28change and get modified slightly.
- 23:31But the biggest sort of area is actually fashion where you go
- 23:34through seasons where you'll have some behaviors which are
- 23:38what we call Evergreen. They're always on.
- 23:40They're always going to be there, but then you have, you
- 23:43know you go to winter, you suddenly start seeing behaviors
- 23:45around winter coats where you get into summer and you suddenly
- 23:48start to see you know swimwear pick up again.
- 23:50So that's that's generally where you get most change in in the
- 23:55segmentation, but that happens at a manageable pace and we
- 23:59don't find that clients are struggling to sort of stay on
- 24:03top of that. It's it's it's sort of naturally
- 24:06fits in when they sort of change strategic goals anyway because
- 24:09they're they know that behavior is gonna change when you get
- 24:12into summer or when they release a brand new product line.
- 24:16So they're ready for that. They're geared up for that.
- 24:19That makes sense. And how does the, so something
- 24:23that James noted down here was the propensity device scoring
- 24:27and can you just talk us through how that works and how and what
- 24:32you're kind of tracking beyond just PDP views and car events
- 24:35and that kind of stuff? Yeah.
- 24:37So when you look at some of the behaviors that we're
- 24:41recognizing, if you go to retail, you get a lot of
- 24:43behaviors where people are just browsing.
- 24:46So people who might have come from a referrer, which is more
- 24:51of an introductory referrer to a brand, let's say.
- 24:54And it drives behavior where people explore a lot but they
- 24:57don't actually buy. When you dig deeper into this
- 25:00personifies an anonymous tracking platform and
- 25:04personalization platform. We can do this without any data
- 25:06about the users, but when you do add identification of the users
- 25:10across multiple behaviors, you find that those behaviors are
- 25:15exploratory and they give rise to other behaviors later on,
- 25:18maybe a couple of days later or next week, which is actually
- 25:21more driven to purchase. And so these are actually two
- 25:23different behaviors you'll have exploratory where they're
- 25:25looking at like lots of different products, making sure
- 25:28that you have got products in the categories that they want to
- 25:31buy, etc. And then they'll come back like
- 25:332-3 days later and this is when they're actually ready to buy.
- 25:37So when you're talking about dependency to buy score, you're
- 25:39actually looking at a case of the behavior that we're we're
- 25:43we're mapping this user to. Is it actually likely that
- 25:46they're going to convert? And if they are going to
- 25:48convert, what conversion point is it, Is it going to be a
- 25:51checkout? Is it going to be a newsletter
- 25:53sign up? Is it going to be a sort of a,
- 25:55you know, a non financial but still important in terms of
- 25:59engagement thing that we're trying to drive here.
- 26:02And so that allows you to say okay for this user on this
- 26:07behavior, they're probably 0.5% likely to buy another user on
- 26:13another behavior, they might be 10% likely to buy in terms of
- 26:16because they're on that behavior.
- 26:17You can then look at where they are on that journey.
- 26:21So are they actually right at the start of the journey or have
- 26:24they got most of the way through to the conversion point and
- 26:27that's when your propensity storing then can ramp up.
- 26:30So you can say, OK, this user has .5% likely to convert, but
- 26:34they're actually on the leg of this journey, which is actually
- 26:37really close to a conversion point.
- 26:39So their propensity to buy is actually 20%, whereas this other
- 26:42user where the average conversion points is going to be
- 26:47like 10%. They're actually right at the
- 26:49start. So they're actually, you know,
- 26:511% likely to buy in the next couple of steps because they
- 26:54haven't got really got going yet.
- 26:56So that's where the propensity to buy happens because you
- 26:58actually look at the individual behavior and then where they are
- 27:01to actually calculate that score and that allows you to say okay,
- 27:05they've actually added loads of stuff to car, but they're
- 27:07actually doing that because they're basically making a short
- 27:09list. They're nowhere near buying yet.
- 27:11So they're preventative to buy and show and check out office
- 27:13isn't really gonna help. So you can use that kind of
- 27:17smarts to to figure out where somebody is.
- 27:22Yeah, that makes sense. And so essentially you're
- 27:26continuously responding to the behavior and the machine
- 27:29learning of how that behavior corresponds towards.
- 27:32An endpoint purchase to upgrade that percentage rate and I have
- 27:38interest on that. So I know a few updated
- 27:42companies where they I don't listen to.
- 27:44Somebody's have A to purchase which is then surfacing
- 27:48businesses can make marketing automation decisions online.
- 27:51You know, do they follow up with campaigns, Do they follow up
- 27:53with disk guys? Do you have any can businesses?
- 27:57Override or set any controls around the OR is that how,
- 28:01purely algorithmlessly for it? And we have the ability to
- 28:08customize the scoring level. So you've got a couple of levers
- 28:13that you can pull in terms of the sort of weighting that you
- 28:19can give to particular types of action And we do have, we have
- 28:25done a bit of work because that changes per sector.
- 28:28And so people who are sort of you know you're not going to
- 28:34have the same sort of scoring if you're working with a an airline
- 28:40versus somebody who works in retail.
- 28:42And and luxury fashion is going to be very different for fast
- 28:44fashion. So there there are sort of there
- 28:49are the algorithm can work out the best kind of waiting, but
- 28:52then we can go in and override that.
- 28:54We we don't normally recommend that because it it we we
- 28:59recommend taking the output from personifying and pushing it into
- 29:02existing systems to then do your overriding if you wanted to
- 29:08because that allows you to keep personify as a decisioning
- 29:12engine which you can then weight with other inputs.
- 29:16So if you're if you're at the stage where you actually have
- 29:18people who are smart enough to do the weightings then you
- 29:21actually are probably doing a multi algorithm personalization
- 29:24strategy where you're taking in several different things.
- 29:27You might be using personifies some customers use those for the
- 29:31unknown customers. So we go OK this is we don't
- 29:35know this customer, this is personifies suggestion and then
- 29:38they'll have a completely different machine learning
- 29:40algorithm for known customers which then they can mix the two.
- 29:43So that's that's we talk about sort of layered personalization
- 29:47in that way where you sort of take the behavior bit from
- 29:49sci-fi which is like the real time context and then you add in
- 29:52sort of historic data so the user profiling from your CDP and
- 29:57and all of that kind of stuff can be merged together.
- 30:01Yeah, nice. I like that.
- 30:03And the question I've got because we talked about the
- 30:06auto, the generative AI and sort of that auto generation of
- 30:10content based on behaviors. Now obviously a lot of
- 30:12businesses have an existence CMS where they do all that content
- 30:16editing, publishing, and that's where they would have an
- 30:18established work by pushing content into E comics.
- 30:22So you talk, so how does personify work when there's an
- 30:24existing CMS? It might be a native page
- 30:26builder, might be a headless CMS like an average VS or story.
- 30:30What does that integration problem work like?
- 30:34Water generation, getting it into a work play for approval
- 30:37and then being published up to productions?
- 30:40Yeah. I'm sorry.
- 30:42Yeah. So I'll talk about the Ambience
- 30:45integration actually because we work with them directly with an
- 30:49outbox integration. So we personify as a business.
- 30:55Our unique USP is identifying behavior and then being able to
- 31:00leverage that through your digital state.
- 31:03We massively into the headless space with the Mac Alliance and
- 31:07other sort of organizations. And when we come to content
- 31:13management systems, personify has to be able to deliver end to
- 31:15end personalization out-of-the-box because some of
- 31:19our customers don't have all of these different pieces or they
- 31:22might have missing gaps within their CMS or within their CDP
- 31:26that we have to you know we have to plug in and fill.
- 31:28So our strategy from as a business is not necessary to
- 31:33replace a lot of the existing stuff that you have, but
- 31:35basically either supercharged and will integrate and basically
- 31:38make them smarter with ambience. What we do is they have content
- 31:42slots and we have a personify integration which allows us to
- 31:48say okay, we're creating this banner for this content error.
- 31:52We're creating another variant. We're saying, OK, this is
- 31:54targeted at sports. This is targeted at, you know,
- 31:59lifestyle. And what personify will do with
- 32:03the integration into appliance is basically OK for this content
- 32:06slot. You've got 3 variants.
- 32:08They're targeted at different types of behaviors which have
- 32:12been integrated into appliance through Personify.
- 32:14And we'll then say okay, you're going to hit 45% of the people
- 32:17who will actually view this with relevant data.
- 32:20So you actually need to create some more stuff.
- 32:22And so that's the the integration.
- 32:25Interamplus will actually go and say, okay, you're so the users
- 32:30who are going to hit this page, you're going to hit 45% of them
- 32:33with these. But if you create something
- 32:36which is targeted at dogs, you're going to hit another 20%
- 32:39of your traffics. You should really go and do that
- 32:42and that's basically the flow that we have when we integrate
- 32:45into the CMS. Everything else from that point
- 32:47is exactly the same. They'll then push it into shop,
- 32:49you know they'll they'll push it into Shopify or whatever sort of
- 32:55platform that they're pushing into Salesforce etc.
- 32:58And they'll then have a little bit there which then will call
- 33:02out to get the behavior from personify in real time to do the
- 33:05decisioning of what to show. But the actual flow in terms of
- 33:09content creation is that is exactly the same.
- 33:12And so we're looking at sort of how we can integrate into those
- 33:15workflows and sort of create custom sort of integration
- 33:18points which will provide the information.
- 33:21We'll obviously have to create the workflow for ourselves as
- 33:24well, because we'll need some rudimentary sort of flow to go
- 33:29through the approval process. But the idea is not necessarily
- 33:32to replace full CMS. We don't want to become a CMS
- 33:35because that's that's a whole different business.
- 33:38The idea is that we can create those points into content
- 33:41management systems and then enable users to basically use
- 33:46the information for personifier to actually free the web flow
- 33:49approval process or the content generation piece.
- 33:52So we could you know with generative AI is a bit about you
- 33:55know hey you need to create this type of content and we can give
- 34:00you the prompt to use to actually go and create that and
- 34:04that will then speed up the actual interactions with
- 34:07generative AI as well. So that makes sense.
- 34:10I think that, yeah, sounds like a sensible route.
- 34:13And I had a question on reporting and understanding the
- 34:17impact of some of this kind of personalization and particularly
- 34:20kind of as you go more into the kind of generative AI space, how
- 34:25do you measure the kind of positive and negative impact and
- 34:28report back to the end user? So that's challenge always
- 34:34because you you get the whole there's lots if you're showing
- 34:42you know different product recommendations you can track
- 34:45click through and sort of and and you can obviously track the
- 34:48end metrics as well and you can say OK this is working better
- 34:51people are clicking through better.
- 34:54The the issue is, is that your your your goal for metrics
- 34:59should be long term customer engagement and and that's quite
- 35:02a nebulous term that people struggled to measure that anyway
- 35:06and even when they can it's quite hard to you know it takes
- 35:09a long time to measure. So you're generally looking at
- 35:12sort of short term rejectives which generally like conversion
- 35:14rates. We if a lot of the time when
- 35:19we're we're tracking existing personalization that we we will
- 35:23generate the customers. We're looking at A, B testing
- 35:25and basically looking at sort of end and metrics because even
- 35:28with things like product recommendations you can have a
- 35:30really great click through rate but no purchases generated off
- 35:33the back of that. So you always need to look at
- 35:36sort of your final metrics and make sure that that's working
- 35:39well that the in terms of the generative AI that's it's a bit
- 35:47harder because the content asset creation is a lot further
- 35:51removed from the actual impact. So when you're saying okay,
- 35:55we're using personify to increase to identify the gaps.
- 36:01You then create the assets and then several steps later you're
- 36:03putting them in place. The only way that you're
- 36:06effectively going to be able to test directly whether it's
- 36:10working or not is actually to A/B test that and basically okay
- 36:14content, which would have been the usual approach versus
- 36:17content using the personify approach.
- 36:19And most companies don't have the resources to do that.
- 36:22And even if they do, they don't want to because it basically
- 36:25means completely separating the chain of that content and
- 36:29basically having two different streams to go in and saying this
- 36:32is our normal approach, this is personifies approach and then in
- 36:35managing that. So it's quite a challenge to
- 36:38actually directly track the impact.
- 36:42What we normally suggest to do is look at.
- 36:46We can split out users by behavior.
- 36:49We know which ones aren't served by content that you have at the
- 36:52moment. So we can report on individual
- 36:56behaviors and users and how they change relative to others, which
- 37:01is served well. So if you've got somebody who's
- 37:05shopping, if you've got really good content created at people
- 37:08who are shopping for safers on a website and their conversion
- 37:13rate is like say 5%. And then you've got another
- 37:16group which personifies, identified as content is not so
- 37:19good. And you go through the
- 37:20generative content creation approach and you know it
- 37:24increases their conversion rate from 2% to 4%.
- 37:27Then you can start say, OK, Solify is actually making a big
- 37:31difference in this particular area.
- 37:33And so that's where you can start sort of it's not as
- 37:36scientifically correct because you're looking at before, after
- 37:40cohorts, but it's a lot easier to do practically without having
- 37:46a lot of effort to split things up.
- 37:48So you can test cohorts side by side.
- 37:51So that's generally where you'd be looking at sort of identify
- 37:55where the is working because of that, that's sort of a top.
- 37:58Did effect that makes sense. I guess you're the client to do
- 38:01their own reporting allowing engagement metrics and
- 38:04everything else. So I guess that kind of simple,
- 38:06yeah, yeah. If you're doing content for call
- 38:09to actions, you can track that a lot more effectively than if
- 38:13you're just creating copy and slapping it in the middle of the
- 38:15page, because unless you have an interaction opportunity with the
- 38:19content that you're creating then it's a lot harder to.
- 38:25Track, yeah, yeah, that makes sense.
- 38:27And then what verticals are typically best for your business
- 38:31and have you got kind of a really good client example where
- 38:34you've had a really I've proven it.
- 38:39So because son of I can work across the entire journey on the
- 38:44content and the product space. Some of our best customers have
- 38:50have have had some really good sort of results across both the
- 38:55content they're creating that block content sort of authority
- 38:58written pieces that for SEO purposes and then the actual
- 39:04retail store attached to that. It's generally people where it's
- 39:10not necessary specific vertical, it's where people are creating
- 39:13basically a lot of content to drive acquisition into a retail
- 39:16store. So one company which is a great
- 39:19example of this would be MEAN Audio based in States and they
- 39:26are, they sell electronics, they sell really expensive
- 39:29headphones, DAX, so digital sort of digital audio connectors and
- 39:37they're really great at basically doing customized
- 39:40cables which are, you know, you might have a £5000 pair of
- 39:44headphones. If you use a really cheap cable,
- 39:46it's not going to sound any better than a cheap pass.
- 39:48So you have to like make sure your whole flow of, you know,
- 39:52your headphones to your recording desk and all of the
- 39:55cables attached to all its same quality.
- 39:57And that's basically the space that we audio provide.
- 40:00Now they attract customers who are luxury customers who just
- 40:05really want an expensive brand of high quality headphones and
- 40:08that they care more about the the label.
- 40:11They care about the visual aspects.
- 40:13They're very visual shoppers when they come on to me in
- 40:15audio. And the other type of users are
- 40:19more technical users. They're recording artists or
- 40:22people who really have got home cinema setups and really care
- 40:28about the quality of the lesson to and they care a lot about the
- 40:31specifications. They don't really care less.
- 40:34I mean, they care what it looks like, but less so.
- 40:36And so you have these two very different types of users.
- 40:40But then you have the site is very different as well.
- 40:42You have lots of blog and magazine articles.
- 40:46They write reviews and sort of guides for some of their the
- 40:51equipment that they sell, which is better than the
- 40:52manufacturer's content. And so you get lots of users go
- 40:56onto the site to read up. And so they have huge amounts of
- 41:00customers where you'll have people who are exploring the
- 41:04content. And basically the brand is prime
- 41:08position is really well suited in the space as this is an
- 41:11authority on audio equipment. And so when you've got
- 41:15personified these kind of things where people are just browsing
- 41:17to buy and people are browsing to read and find information,
- 41:21you can leverage personified to its full potential because you
- 41:25can start to say, OK, these users aren't going to buy yet,
- 41:27but they are really interested in these types of brands.
- 41:30We're going to push videos, we're going to push content.
- 41:32We're going to push them to bring audio university and sort
- 41:35of their education port for portal and not present that
- 41:40those users will never buy from an audio, but quite a lot of
- 41:43users will, but they might buy later.
- 41:46And so you basically have the opportunity to really engage and
- 41:50then convert when you can see those markers in their journeys
- 41:53where they'll go onto the retail side.
- 41:56And when you get into the retail side, then you've got all of the
- 41:58standard sort of, you know, recommendations, social
- 42:01proofing, all of those kind of messages.
- 42:04But then you have the ability on the product pages to say, okay,
- 42:07we're going to show you lots of visual stuff if you're a luxury
- 42:09customer and we're going to show you all the specification data
- 42:12and push the manuals and the interoperability data for those
- 42:17users who are technical. So you can start to actually
- 42:19change the layout product pages to suit the type of customer.
- 42:24And so that would be a good sort of example of you've got lots of
- 42:28content, you've got really detailed product pages, you've
- 42:30got the standard sort of retail space.
- 42:34Polify can sort of hit every single use case that Polify can
- 42:39sort of implement in that space and that's a good customer.
- 42:43So that would be the sort of sweet spot that you can you know
- 42:50listening to that example you probably can think of quite a
- 42:52few different retailers or brands or use cases which it
- 42:56actually work quite well with outside of that vertical.
- 42:58So it's not really vertical specific, it's more like this is
- 43:01a sort of you know this is you know you've got the content to
- 43:05optimize as well as the retail to optimize as well as the
- 43:08product data. So you're sort of optimizing
- 43:11across the entirety of the not just the individual per session
- 43:16behaviors, but the entire lifetime of the customer you
- 43:19cannot twice cross. I really like the the that.
- 43:24Whole idea of of restructuring content based on people's needs
- 43:27and interests. Because there's so long pages
- 43:30like Pdp's have been, you look at the eye tracking data on it
- 43:35and there are so many like loose spots where stuff is not the red
- 43:39but there's content layered down that the people want, but it's a
- 43:41static temp everybody. And that ability to take logic
- 43:45at a personalization level and feed that in and to dynamically
- 43:49change the page template at the back of it to suit that
- 43:53particular user's behavioral lead, it's just I think it's
- 43:56such a smart way of making pages more relevant to individual
- 43:59users. Yeah.
- 44:04And we can do that quite free form which we do with Mean
- 44:08Audio, but we're actually rearranging a lot of the stuff
- 44:11on the fly and it's quite dynamic.
- 44:13And then we're working with other customers who are sort of
- 44:16more in the, they like a more created approach their content.
- 44:21And so we're looking at personalization around creating
- 44:25different page templates and then just switching templates
- 44:28rather than sort of automatically chopping and
- 44:31changing. So there's sort of lots of
- 44:34implementation practices to sort of suit your sort of your your
- 44:38trust in AI effectively. Yeah, I think the great thing
- 44:43now with the the modern platforms being far easier, even
- 44:46with later paid, they'll just set up different templates to
- 44:48seek different right, be able to have those multiple templates
- 44:53there that then the personalization engine can push
- 44:55the right info to say, right, OK, this is the template you're
- 44:58going to need because this is the customer now serve that to
- 45:00this end of the end, right. Josh has been really interested.
- 45:04We covered all that. We could go raise that because
- 45:06it's a topic that I absolutely love and I'm sure there'll be
- 45:09more more conversations to come, but you can a lot of really
- 45:13useful information about how the black works, the role that
- 45:15generative AI and personalization of content is
- 45:18having any. I'd imagine that's there'll be
- 45:20people with specific questions going I want to do this or how
- 45:23would it work here? If people want to reach out and
- 45:26probe further or even challenge on any of this, how?
- 45:29How it should they connect with Solid, yeah.
- 45:34And if you want to connect then drop the drop the e-mail to us.
- 45:38You can either get in contact for our website which is
- 45:40www.sonifyxp.com or you can drop me a message on
- 45:47josh@sonifyxp.com. Amazing.
- 45:50Thanks so much for your time. Really enjoyed this and I love
- 45:53watching the story of Sonifi's Planet of Growth accelerate.
- 45:57And there's some very interesting clients you've got
- 46:00on your, but it's like love, Honey.
- 46:01I imagine there's quite a lot of interest in personalization, the
- 46:04segmentation stuff in there for sure.
- 46:07Thanks everyone for listening. We hope you've enjoyed it
- 46:09actually. Keep an ear open for the next
- 46:10step and slow and don't forget to us that rating wherever you
- 46:13listen to the elections we go. Thanks very much everyone.
- 46:23For more information on this topic, head over to Replatform
- 46:25dot FM for our audio podcasts to discuss a project, or if you'd
- 46:29like to chat about any of the topics covered in this episode
- 46:31in more detail, please reach out to myself, James Gerd, or my
- 46:34cohost Paul Rogers via LinkedIn and Twitter.
- 46:37Thanks again for listening and keep your ears peeled for the
- 46:39next episode.