How to leverage Target MCP and other major new releases in Adobe Target

Transcript

Hi, everyone. Welcome to this episode of Experience League Live. My name is Daniel Wright, Senior Technical Marketing Engineer. And this episode is how to leverage Target MCP and other new releases in Adobe Target. So we’re going to talk about how Adobe Target enables agentic optimization with Target MCP. We’re going to look at how you can analyze activities faster and explore new capabilities like Bayesian testing, feature flags and CDN experimentation. Plus, we’ll get a quick look at the latest releases with demos of expanded use cases and smarter execution. So why don’t we get to welcoming our guest? Our guest today is Head of Product Marketing for Adobe Target, and he has been doing orange theory for over a year and won the award for muscle gain. Please welcome Drew Burns.

Thank you so much, Daniel.

Wow. That was a generous amount of applause.

What’s that? Yeah. That was a generous amount of applause. Yes, nothing but. So tell us about your newfound workout habit. It’s been going on for a year and bulking you up. Thank you. Daniel’s known me for a long time, and I’ve not been in the best shape, I think, over many years. And so orange theory, actually, I chose that as my fun tip because it shares the same discipline that we need to have in optimization programs. And so orange theory is great. They have classes in major cities throughout the United States. And so you can always find an orange theory near you and it’s a focused hour session and it’s full body. And so I ended up getting into just the consistency of doing it, four to five times per week, these intense hours of full body workout. I’ve gotten into the best shape of my life. And so I think it’s a testament in terms of optimization that if we’re disciplined and we keep up our practice, we will see a lot of exponential results.

Yeah. And how did you get this award for muscle gain? What were your… How did you quantify that? And relating that to Target, as I always do, in the beginning, I think we all know when you start testing or we start personalizing the experience, even for basic category affinity, you start seeing significant results. That’s what happened with me. When I enrolled, they enrolled me into this sort of month test for either losing weight or gaining muscle. And I foolishly, or maybe not so foolishly, signed up for the muscle game and I ended up gaining about a pound of muscle over the course of a month, which is impressive.

Nice. How did they know it was a pound of muscle that you gained? They do one of those full body scans, actually, which is interesting. So they got one of those machines. Yes. They have one of those machines that tells you, unfortunately, your body fat as well. It tells you a lot of information, but they were able to see that I gained about a pound of muscle. So it was a cool experience. Actually, the whole health industry is being sort of revolutionized with these machines that are able to do these body scans. So it’s really fascinating.

Way to improve a key metric there, Drew.

Yes, thank you.

So I was realizing recently that I’ve been using Adobe Target for 18 years, which is a thing to me. So it’s really exciting to hear about the continuous evolution of the product and to see that it keeps gaining muscle as well.

And now in the new agentic world, it sounds like there’s a lot of product development going into that area. So, yeah, tell us about it.

Yeah, I think you can… Sorry, Drew, before we start going, we are taking questions as well. So if you have any questions, you need to be logged into a Google account.

And then you can just type away and we’ll see those questions and I’ll periodically pepper them into our conversation. I see we have Sandeep from Bangalore, Tim from Raleigh, North Carolina, welcome to the session. We have about 80 people with us now live. So, Drew, back to you. Tell us what’s going on with Target and MCC. Yeah, I mean, you mentioned Tim. Tim is a… I know Tim well. He’s been a practitioner for many years, a very progressive user of Target. And he can relate to this very well as you can, Daniel, because the process of optimization has been very manual over the years. Even if we’ve streamlined the workflow to three steps, I think we all know it might take a little bit of development work.

Certainly content creation would take time. It takes time to wait for results. It takes time to analyze those results and sort of surface insights on performance, especially as it relates to specific audiences, and even to then apply those insights to our prioritization of what to do next. And so all of these areas that have taken time over the years and maybe have been a little bit of nebulous, it’s almost like you wanted to plug Target in to some sort of calculator that could do a lot of this analysis for you or at least surface up things that you should focus on. And I think with the rise of agents and AI and Gen AI capabilities, we now see that as an opportunity for those that are familiar, very familiar with Target, as I imagine many of you are. We have 50 plus APIs with Target. So there are things like I remember going into a summit session one time and someone was asking, hey, can we have Target plug into Alexa so we can personalize Alexa based on a commercial running in a specific region? And we integrated Target into Alexa in like 20 minutes. Target is very, very flexible, adaptable, plug and play in many cases. And so we should be thinking about it as something we plug and play into our experiences to test or to use AI for determining the right experience. But wouldn’t it be great if we could do this all through a conversational UI, which I think we’re all becoming familiar with. We’re adopting ChatGBT and Gemini and Clod is actively used within Adobe. And many of these can use an MCP. So actually, if you don’t mind, Daniel, I’d love to show, maybe go to my slides really fast. Yeah. Because I got a definition for what an MCP is just so those that maybe aren’t as familiar are able to learn a little bit more about it. So this is all brand new stuff. We launched the MCP, gosh, it was just a few weeks ago. And we had actually plugged it into Co-Worker, our in-house conversational UI and agent or summit. Learning as we were going, this is all brand new. But basically what an MCP is, it’s a standard interface for AI agents to understand target context. So it, as I was saying, is basically a protocol that can plug into our APIs. And anything that we can do with our APIs, we’ll be able to do with the MCP. We’re building up those skills right now. The MCP is available to you. So you can take it now, again, it’s GA, and you can go ahead and plug it into Clod. If you go into our documentation, especially on Experience League, we’ve got a great article on how to set it up.

And you can plug it into Clod, you can plug it into chat, GBT. Whoop, my notification’s off there.

And here’s, don’t click this link, but this link right here, targetmcp.adobe.io. That’s what you use to plug in target to Clod or to chat GBT. Or again, if you’ve signed the AI writer, which you will need to do to access the MCP, you can use the CX Enterprise coworker, which used to be called Agent Orchestrator. They might have heard of, and that plugs into not just the target agent, but also other agents from other applications. So think about if I’m using target and I want to do some analysis, let’s say with analytics too, the analytics agent, you can use that within coworkers. There’s a lot of this sort of cross use across the agents within the MCP. One thing I also want to call out, and Daniel teased, is the CDN-based delivery. We’ve got a Node.js SDK there to use Akamai edge workers for near latency free performance. So that is already available and so on the list there. But in terms of the MCP, you may want to see like, what is this like? So this is the CX Enterprise coworker, formerly Agent Orchestrator. I’m just going to talk you through a demo here. What we did is at Summit, we plugged it into the Adobe.com account and we radically changed the Adobe.com website. Just kidding. We analyzed their target program and we just asked to summarize activities. So you can just see, okay, well, how many activities are we running? Maybe even to help categorize what these activities are doing or what they’re related to.

You can also get reports. So, you know, we want to get a report on these categories. We want to see it graphically. You know, in many cases, we’re taking what we see in our target reports and we’re creating maybe PowerPoint decks or one page sheets that we’re able to then socialize throughout our organization. You can use that within an agent to help generate that. So it’s going to speed up the process of getting the answers or summarizing information graphically or however you’d like to see it and asking questions even on performance of specific audiences or connections across activities. So there’s a lot of information that we can get out of the target MCP.

And just ask very clear or basic questions. I would say be as specific as you can on where you want to focus it. And so it’s not sort of interpreting or going off into the weeds, but there’s a tremendous amount of drill down and cross-cutting themes here. Lots of interesting things that you can analyze as well as get graphic representations of within the MCP. So again, use that URL there to help plug in, you know, Clod or ChatGBT into the target MCP. Again, you’ll need to have signed the AI writer. So make sure your organization has done that. And then you can begin to do these read-only skills. I will tease, and we’ve been running this internally, this will be released in the next few weeks, is the ability to create, edit, and pause an A-B test. Now, for example, you’ll be able to say, build me an A-B test on this page for this container with these three experiences. I want it to be auto-allocate, so using the multi-armed bandit to find the winner, and I want this to be my primary conversion event. So what it’s going to help you do is build out those activities. So a lot of the manual sort of configuration that was required in target will go away, and you’ll be able to do manual building of A-B tests, I mean, or sorry, MCP, use a natural language interface to build A-B tests, as well as experience targeting. So that’s what we’re releasing in the next few weeks. Now you’ll notice you can’t push this live, and really what we don’t want to do is release that at this point because we want to urge you as you’re opening this up, hopefully this will help you open up more stakeholders to be able to use target who aren’t maybe as familiar with the UI, and also to speed up your process of creating activities.

And then not being able to publish it means that you can go in and then tweak or QA, make sure it’s configured in the right way before pushing it live. We want to make that pushing live or publishing an activity a deliberate process. So I like that safeguard a lot. So again, A-B tests, experience targeting coming shortly, but please, if you can, get your hands on the MCP and take a look at the documentation and our documentation about that. Another thing that you’ll be able to do very soon in the next few weeks is profile scripts.

So this, yeah, Daniel, you’ll remember this. Profile scripts, there’s a powerful, one of the things that distinguishes target I think is the real-time profile and that’s contextual information on what an individual is doing at that given moment. You may know everything about me and certainly with like a unified profile like we have in our CDP, that’s going to be interactions that someone might have in store or on a call center or speaking to a sales team member, let’s say, or a clerk or a front desk. It’s going to have all kinds of information on you and historical information, but the target profile is real-time. So that’s what I’m doing right now. For example, let’s say it’s a sporting good company and I’m a skier, right? And so you might think, oh, let’s show him skiing material, but it’s summer and I’m going away on a summer vacation. So now I’m looking for maybe a tennis racket this time because I want to play some summer sports. So that’s where this real-time profile and target is really, really important, especially at first touch when we don’t know anything else about the individual, that anonymously is really valuable. And one of the sort of hacks, if you will, within target, I would say that experts like Tim know all too well is capturing a behavior on the site. Profile scripts, I would say, is not used as much as it could be. And when you think about it, profile scripts, I think, could be used for any action that a visitor makes on the site that you want to persist as a part of their profile. Maybe they self-select something that then puts them into an audience that you’d like to then personalize the experience for them for. And I think this brings up a really good personalization tactic that I think is sometimes overlooked. It’s a very respectful way to get information from a visitor, which is to ask or to say, hey, do you want to opt in to show me what it is that you’re most interested in? That’s a great opportunity for a profile script. And then that becomes a part of their profile. It used to be that you had to know a little bit of code and we have a library of those codes and it self- it auto-completes actually within the product as you’re building them. But it’d be nice to just ask for that script to be written. That’s teasing out how there’ll be more in terms of the ability to write code and even implement inboxes coming down very soon within the MCP. So just want to tease out some of these ways that we’re developing the MCP and rapidly focused on that. So you’re going to see a lot more skills coming in, but please do get your hands on that. And then one other use case too. So those will be coming out in the near future and they’ll just have access to them as long as they agree to the AI writer? That’s correct. You’ll have that available to you. The idea is for you to get more efficient in your program, more efficient with building out your activities, and also open it up to more stakeholders. People that may not be as familiar because we’ve heard, you know, Target is a simple three-step workflow, but there is a fair amount of sophistication that you can do within the product. How can we streamline that and make it easier for someone to just do it through a conversational UI? And that’s really the point of the MCP.

One other thing too, in terms of flexibility for Target, is brand concierge. So this is an offering from Adobe that you may or may not be aware of, but it’s for companies to create their own conversational UI or conversational agent on their website. So I think we’ve become familiar with this with support and chat, but brand concierge allows you to help people with their own issues help orchestrate that customer experience through a conversation through your product set or through your services or whatever it is that you’re representing on your site. So it becomes like a conversation with the website. And of course, it makes sense for Target to be integrated in there. In fact, I did a summit session a couple years back where we integrated Target with Dynamic Chat and Marketo and using those same principles of being able to integrate Target into this experience, we are now embedded within brand concierge to make recommendations available based on that conversation. So if I’m asking, let’s say about tennis rackets, it could recommend a display of those tennis rackets or a vacation in Italy, a specific area of Italy. It can show me a display of those packages that are available. So opening up all of those recommendations, algorithms, and those containers to be able to better service the conversational customer experience on a website. So something to think about. And then finally, we’ve been talking about Bayesian for a while and Daniel was teasing it earlier. And I know he’s probably heard it over the years from when he was more in the Target UI.

But Bayesian is another statistical methodology that exists. And I think those that could go back to their stats days or maybe are familiar with the statistical methodologies in an optimization products, it’s not more valid or less valid than the students T-test. It’s just another way of approaching the data and getting results. And the students T-test, which is what Target has out of the box and has had over the 18 years of its existence, is very rigorous. And that’s why we landed on it because it’s looking at the entire population. And at reaching a very high bar of 95% confidence, you’re reducing false positives. So you’re more confident that this is repeatable. The results you’re seeing in a test are repeatable. But we know, and this is the reason we went with sequential hypothesis with autoallocate, that you want results faster sometimes, right? That’s what autoallocate is all about, right? Find the winner, give me results right away. Well, Bayesian is another flavor that does something similar to that. Rather than looking at the entire population, it’s looking at samples. You know how you do samples and testing and you can infer results from those samples. That’s what Bayesian is doing at a simplistic level. Now, this will be an option coming in in the July timeframe in Target. So when you’re building an A-B test, it’ll ask, do you want to go Bayesian? Do you want to go students T-test? So your obvious question is going to be, well, when do I choose what? Students T-test, I think we’re familiar with. It is always very rigorous, but you want to wait until that fixed horizon of 95% confidence or near 95% to call a result. With Bayesian, it will give you results faster, as we’re saying. It’ll be more significant over that curve, significant results along the way. But it’s looking at smaller sample sizes. The requirement there is you’ve got an informed hypothesis. So this is one that you feel like, you know, we’re pretty confident this is the way that it’s going to tend. And even we’ve got a sort of rough estimate for where we see those results falling. That’s going to be sort of the litmus test. And then obviously the other litmus test is going to be, we need results right away. So tell us. So auto-allocate, again, another flavor already in the product using sequential hypothesis testing. We’re just finding the winner. Bayesian testing will allow you to look at multiple experiences and compare and contrast, but it’s looking at smaller sample sizes. And so we’re excited to be sharing that with you.

Oh, and sorry, last one. I mentioned feature flagging. So you have been able to do feature flagging in Target. And again, this is another opportunity. Think about, we want to create more utility in the product. We want to obviously improve performance with the MCP and making it faster and easier to run activities. But then we also want to open it up to new stakeholders, right? I was talking about that. And so there might be product teams that you have, right? Or IT team members who are doing feature flagging. Feature flagging is really just where we’ve got a feature on the website. Maybe it’s a cart and we want to roll out a new feature function on the site, but we want to roll it out in a slow fashion, you know, to reduce risk and to make sure it’s operating the way that we want it to operate and that it’s not tanking our metrics so that we can roll it out to everybody or roll it out to an entire group. And so that’s what we have a new workflow. You’ve been able to hack it with Target. You’ve got a new workflow dedicated to feature flagging that you can get access to. It’s in beta, so you have to reach out to us to get access to it. But it does things like experience rollouts. So what I was talking about where you’re ramping up a feature. Another thing it’s really good at is kill switches. I think that really resonates when I’m in front of an IT or product team. They’re like, oh, we’ve got kill switches all over the place when you want to just kill a feature really quickly. This is something that you can use for that. Adobe.com has been using this for several years now. And so it’s tried and true. We’d love feedback on it, though, beyond the Adobe.com team. So that’s why we’re opening it up to you. So reach out to us if you want access to feature flagging. And so that pretty much covers everything that I wanted to sort of show today, show and tell. And so are there maybe questions that I can take in the last few minutes? There is this one question that looks like was answered.

About when you create a B test, the API adds targeting at an experience level, which was not editable via the UI once created. Will the MCP come with improvements to the API or will it have the same? It looks like it’s available in the latest release. Yeah, Santhosh went and it looks like he answered that.

It looks like, yeah, Calvin, it sounds like something that you should test to verify if it’s doing what you wanted to do or if not, let us know.

So the feature flagging, Drew, is that how does that work? I know, you know, you could use, you know, the form composer or, you know, inject code that sets a feature flag. However, your application wants to have it back in the old days. I would set a cookie to indicate, you know, to to the back end, which experience should be delivered to the customer. These days, probably people use like, you know, a JSON configuration. How does what does what does the feature flagging feature technically deliver to the page? And are there any implementation requirements needed to use that so that the application knows? There is. We can provision it for you.

And then it has features, you know, traffic management was not something that our traditional target workflow was very good at sort of allowing you to to toggle or to ramp up traffic to an experience. And so as you’re seeing those results coming in based on a feature flagging activity and it is resonating or it’s not resonating, you can either ramp the traffic up or ramp it down more easily or even push it all the way full to the entire population.

So you can start it out at whatever level that you want.

You know, I think traditionally we would just sort of set I know there were large resort customer would put like 10% or 5% into sort of an extreme version of an experience just to see what that did to the data. It’s sort of an interesting use case there or approach there. But so traditionally that’s how we might do that. But with this, it’s much more easy to react to positive performance over time and to ramp up beyond that sort of a smaller and more conservative population to the full population when ready. So it’s more of the tools that someone might be familiar with the feature flagging tool that are out there. This is just a workflow within Target that allows that to be done easier. And then as I mentioned, configuring kill switches as well. Another use case that we didn’t have a custom workflow for within Target. And so we’ve been hearing that for many years, especially from product teams that were saying, hey, could we maybe replace, you know, we’re only using about 20% or so of this fully featured feature management or feature flagging tool. Could we do some of that in Target or Target take over those responsibilities and now we’re able to do that.

Speaking of ramping up, I know one of the things that, you know, I had seen as a customer and then as a Target consultant is often when you ramp up traffic to one of your experiences, sort of shifts the balance of like the new visitors and the returning visitors. And it would, you’d have to be kind of careful with your data because often the experience suddenly receiving more traffic, it looks like the performance goes down.

But that’s just because of what your, these changes that you’re making to the population in the test. Is that anything that the Bayesian helps with? Like those types of things or is that? Yeah, I think so. You know, that’s a really interesting point, Daniel. I think in fact it does.

And we were so focused on the significant results and comparing and contrasting that to students T and being able to disseminate between that, that we weren’t so focused on that scenario where you’ve got audiences and maybe it’s sort of, you know, we’re always, especially with multiple activities on a page, we’re always worried about cannibalizing, use a whole horrible phrase or term for it, but taking away traffic from another activity. And I do think Bayesian will help with that, especially when we’re looking at smaller sample sizes. Because again, if you’re meeting that sort of higher threshold or meeting the full population for those results, it requires a lot of trafficking. And it can take longer as you could imagine. And so that’s why we have the traffic calculator. And we always say, you know, you can consider that you’re going to be running an activity probably for a couple of weeks before you can get results that you really want to base business decisions on. So it’s, again, there’s nothing against that. When we look at stats, you want to be as rigorous as possible, but oftentimes we want more directional results, but still significant results. And I think looking at sample sizes is very compelling and has less traffic requirements to your point.

We have a new question from Pratma. Are these new features going to improve recommendation engine capabilities? And if yes, can you talk about it? Yes, recommendations.

I know you mentioned brand concierge and that kind of integration between recommendations and that product. What about MCP? What about Bayesian? How does that relate? Thank you, Pratma, for the question. And we are working on that feverishly right now, as you can imagine. One of the things I think Daniel can relate to from his years with Target and certainly a customer, you know, it just takes time, would be feed management and building a feed. And that might be something that we’re taking a look at, something that the MCP could assist with. Other things is obviously the building or configuring of criteria.

So that can be streamlined. And so setting up a recommendations activity has more facets to it than a traditional A-B test. And so I think recommendations will probably come in phases where we’ll start with something that allows you to at least build a recommendations activity, but then requiring you to go in to maybe configure some of the more sophisticated configurations.

But then soon after that, we’ll be able to build out skills that allow you to do the full activity build with recommendations. So we are looking at that, but that’s taking a little bit longer than an A-B and X-T activity. OK, that the creation aspects of the MCP you’re talking about. Yes. And it do.

I mean, I guess personally, when I said I mostly do recommendations these days and but usually using like an A-B activity or an X-T activity and then adding the recommendations in there. So will there be that are there read only? Things that can be done currently with recommendations? I mean, you can ask again, think about what you’d want to it won’t infer. That’s one of the things the MCP won’t do that I’m actually clap for because what we want to do, you know, I think we’re all familiar in using LLMs that sometimes there’s hallucinations. So what we don’t want within reporting, we want to minimize any inferences that’s going on. So what it’s looking to do is organize or categorize or classify from results. So that’s where and Daniel, this question prompt this, I’m grateful for you asking this is we want to be very specific in what we want it to pull out in this activity or across these three activities. I want to see how this audience has performed relative to a specific experience. Same thing can be true within recommendations. You can ask, you know, Daniel, you were talking about putting recommendations within an A-B test. That’s a default actually scenario when you’re building recommendations activities in target because there’s always a default which is no recommendation because we always want to show the improvement to performance the recommendations algorithm provides or in many cases, you have a customer they’ll have several different treatments of recommendations and you’re testing each of those to see what the best one is. And that’s a best practice. We should never just roll out personalization without testing it first. We can ask questions of those results, certainly from the target report. Again, let’s not move into sort of interpreting those results or asking for interpretation of those results. Let’s just from a read-only perspective with the target MCP, just look for surfacing those results so that we can then go ahead and interpret.

I will say that experimentation accelerator, which is an add-on to target which is a Gen-AI first application that does draw conclusions from across activities.

It will take a look at the activities that are currently within the product and draw those connections between audiences and experiences. So it’s got I think 400 to 500 different classifications that it is analyzing audiences, it’s analyzing offers and also behavior. And so it will say, for example, this audience seems to prefer this tone of voice as we’ve seen across these three activities. It lets you know what activities it’s looking at to make that inference. And then from that inference or that insight, it will then make suggestions off of opportunities. So again, the target MCP different from experimentation accelerator where Gen-AI is drawing conclusions based on analysis and comparison across activities.

Okay, got it. All right. Well, I’m not seeing any other questions in the chat.

So perhaps we should skip to, we like to close out these shows with an unrelated cool tip.

So these are unrelated to anything having to do with Adobe Target or work life. So Drew, what do you have for us? Well, we were discussing that those of you that have kids out there, you can turn off their screen time for their phones if you want them to give you a call. So if you’re, that is a good hack that if you’re not hearing from your kids as much, having an effect on their screen time, being able to turn off the screen time for them can urge them to give you a call. So that’s what I found in terms of my kids. I have a high schooler and a college age child and that’s the best way to get them to reach out to me is by doing that. Ouch, yeah. That’s like the Homer Simpson strangling of the present day. It’s like, get that, get certain tension.

Yeah, exactly. Yeah.

Well, thank you so much, Drew. I appreciate hearing about all these exciting Target updates and I look forward to seeing them in my account so I can start using them and get on using the ones that are out now. Thank you so much. Thank you. And we look forward to having you on again soon.

Sounds good. Thanks everybody and get your hands on the MCP and check out Experimentation Accelerator.