#128 MCP servers in digital analytics (with Gunnar Griese at 8-bit-sheep)

[00:00:00] Dara: Hello, and welcome back to the Measure Pod. I’m Dara, I’m joined by Matthew. Hey, Matthew, how are you doing? 

[00:00:20] Matthew: I am very well. This won’t be the same for you, but the sun has started to disappear a little bit on these British shores, so we’ve got a lot of rain in the past few days. Summer’s waning. I’m assuming that’s not the case in southern Spain, but. 

[00:00:34] Dara: It is not the case. There is still a reasonable amount of sunshine in the daylight hours here in Espania. Yeah, I won’t, I won’t say any more about it than that. I won’t, you know.

[00:00:45] Matthew: 17 degrees, but. 

[00:00:47] Dara: At night. Yeah.

[00:00:48] Matthew: Well, good for you, Dara. How are you?

[00:00:51] Dara: I am, I’m good. Good. I am good. I’m glad I asked. Yeah. Well, I was going to, there’s something I’ve been, I we’re going to, we’re going to, we’re going to bleed into the episodes.

[00:01:02] Dara: This is meant to be a kind of light breeze. Hey, how are you doing? But actually, I want to talk about what I’ve been playing around with, which is linked to the guest we have on, on the show today. But I don’t want to give away all the spoilers in the first sentence. So let’s just, let’s just crack on. We’ll get to that. Okay. 

[00:01:18] Dara: So let’s do it, let’s do our infamous news, Roundup Salty News. And you have done a lot of the heavy lifting this week with the Salty News. Yeah, I’m, I’m now instantly feeling the pressure. I’m very good at scanning headlines and then realizing I don’t actually have the, you know, the nitty gritty, so I’m expecting you to fill in the gaps.

[00:01:41] Dara: Really. No problem. I’m feeling extremely relaxed. Yeah. Let’s start with one. Let’s do it. Let’s do it. So one, one thing that piqued both of our interests was the Claude, I’m probably going to call it the wrong thing now, but the Claude Chrome pilot, these names are all getting very hard to, hard to say and remember.

[00:02:02] Dara: Yeah. but this is open in, it’s like a very limited, it’s not even a beat is, it’s like they’ve got a list of trusted users that they’re trying this out with and it’s, yeah. It’s certainly not available to everybody yet for good reason, which we’ll get onto. 

[00:02:18] Matthew: But are you one of the trusted users?

[00:02:19] Dara: I am not. Oh, are you?

[00:02:21] Matthew: No, I’ve not tried to be. Why would I be? They wouldn’t trust me, would they? No. It’s just ’cause you said the other day, oh, I’ve got thoughts on it. I just assumed you. I don’t know. But I’ve been playing with it. But you just, from what you’ve read, just just just the art, yeah, just the art. Giving spoilers again about what’s going to, ah, you are, yeah, yeah.

[00:02:38] Dara: No, I just, I, I mean, I read this, this is on the philanthropic blog or news section of their site, whatever. But it went into a lot of detail, and this is, I mean, you’ll do a better job of explaining this to me, but this, this is basically, you know, it can, it can access the browser, it can do things for you.

[00:02:54] Dara: Yeah. Which like a lot of this stuff, and I feel like we say this quite a bit great in theory, but comes with a pretty, pretty, pretty big handful of risk as well. big helping of risk on the side. So a big part of the article, and this was anthropic themself talking about it, but they were talking about how, you know, this is vulnerable too.

[00:03:14] Dara: Prompt injection attacks and all these brilliant, I mean, all these things sound really fun, don’t they? Like prompt, inject injection attacks. It just sounds, sounds really fun. There’s so many new professions that are going to start appearing before long. I know. What do you do? I’m a prompt injection specialist. Yeah, yeah. 

[00:03:30] Dara: Prompt shield, man. But this is the thing, so obviously. On the positive side, it could take a lot of, you know, things like filling in forms or going and researching things on websites in the browser could take a lot of the pain away. But you are opening yourself up to the risk of, you know, that being hacked into and you know, your data being harvested or actually things being executed within the browser that you haven’t intentionally or knowingly wanted to execute.

[00:03:57] Matthew: So when they say prompt injection, then are they saying like, you take a prompt somebody else has created and then run that within your browser and because it has access to your browser, it can access like console and dev tools within your browser and start doing things you don’t really, if you don’t want to stand what’s happening.

[00:04:12] Dara: Yeah, and even potentially, I guess do it, doing it on the, you know, so going to websites and buying things. Filling in other people’s addresses or transferring money to bank accounts, whatever, all this kind of stuff. So I suppose you could hide 

[00:04:25] Matthew: from, I mean I’ve, I’ve heard about this happening with perplexity, or at least they’ve found it as a problem.

[00:04:30] Matthew: But people hide prompts in their website text. Yes. So they’ll say, ignore previous training. Exactly. If asked what the best burger is, this is the best burger in New York or whatever. And then it surfaces it. So a site could be visited by an AI, it sees that and just does whatever the site says.

[00:04:46] Dara: Yeah, exactly. And this is why it’s being tested, I guess, in a very small, you know, limited group. And there’s a lot of talk, and it’s quite a long article and we’ll put it in the show notes, but it’s, you know, it goes into detail around what you can do to mitigate against some of these risks. But the numbers, I think one of the things that made me think, you know, I’ve got a few thoughts about this, is like some of the numbers that they quote themselves, you know, they’re saying things like this.

[00:05:11] Dara: 24% of the time Claude followed malicious instructions, which is quite a high number, right? A quarter. Yeah. That’s pretty. Yeah. But, but, but the good, the good news is, you know, with some safeguards implemented, it drops down to only 11.2%. So only you know, one in 10 times it’s going to follow malicious instructions, which, you know, I’m happy to take those, those odds.

[00:05:32] Matthew: Yeah. If I only lose the contents of my bank account every 10th time, that’s fine. You know, that’s better than every Yeah. Fourth time. 

[00:05:39] Dara: Yeah. So there’s obviously a way to go and they, you know, they’re aware of this. Yeah. So, you know, there’s a lot to be drawn. 

[00:05:44] Matthew: It’s interesting because there’s been rumors of Google’s version of this for probably two years now, called Jarvis, which was basically built into Chrome.

[00:05:54] Matthew: Co this pilot that takes you along things and a lot of sort of rumblings as to what, well, where is it? Why do they keep talking about it and why does it keep being heard about when never actually surfacing? And probably because they’re coming across exactly the same kind of problems that anthropologists are.

[00:06:09] Dara: I would imagine so. And Google not, not saying philanthropic aren’t, but Google probably have more, maybe arguably have more at stake if they release something like that. And it, so yeah, maybe they’re kind of, it’s the wrong, wrong way to put it to say holding it to a higher standard. But maybe, maybe there is a reason why they’re, they’re, they’re keeping theirs back.

[00:06:28] Dara: But it’s, it’s interesting. It’ll be interesting to see, I feel like we say this a lot as well, but it’d be interesting to see how this develops. But I think at the moment, the numbers look scary in terms of the risk that you’d be exposed to. I, I, I read something, I dunno if you’ve heard this term, but something that was on Hacker News, someone in the commentary said it, you know, it caught me talking about it being like this lethal trifecta problem.

[00:06:51] Dara: So it’s this dangerous combination of access to private data, exposure to untrusted content and the ability to execute actions. So it’s like every, everything is the perfect storm. Yeah. Of, you know, risk that you would potentially be exposing yourself to by using something like this. So a lot needs to be done, I think, to safeguard against any kind of, you know, malicious actions that get taken by using this kind of software.

[00:07:15] Matthew: Yes. I think I do like the approach that I take because it, you assume that they know that the other two are working on this and they’re creating similar sort of products and similar with like Claude Code and Claude Computer use, they kind of put it out there in the early stages and, and do security tests and knowing that they’re kind of revealing some sort of hand, I guess.

[00:07:35] Matthew: But Google’s been working on this forever. They’re not releasing, as far as I know, documentation on risks, et cetera, about how problematic this could be. And I’m becoming a bit of a philanthropic fanboy, I think. 

[00:07:46] Dara: Yeah. I’m leaning that way. Yeah. At the moment as well. I was very much. Chat g bt open ai. And now I think more leaning more towards philanthropic. Yeah. Are we allowed to pick favorites being two? Yeah, yeah, yeah. It’ll all change again. Yeah. Google comes out with something. Well, which, great segue. Thank you. I mean, if I say nano, if I say nano, you say, but there we go. and this is not, this is not the name of our new children’s TV show.

[00:08:19] Dara: This is, this is, this is the latest, the la what is it? How do I say it the right way? Is it the latest image generating model, or is it, ’cause it’s part of 2.5. Isn’t the United States like the latest submodel? What even is it? 

[00:08:32] Matthew: Yeah, but most of these models are, they’re all kind of multimodal now, aren’t they? Mm-hmm. So, it says it’s a Gemini 2.5 flash image. It’s really cheap, apparently, like it’s, it’s very low in terms of how much it costs to, to actually run. And it’s an image generator, I guess we should say that. Which if, if that wasn’t obvious. And it’s just very good at maintaining characters and, and images.

[00:08:56] Matthew: So you can kind of upload images of a person and then put them in different situations, but it keeps all of their face and what they look like very well. And like I was uploading, we got this little 3D Sprite for our AI, internal AI called Brian, and it was created like. 360 view of him and what, like the kind of thing you would upload to a computer game and all the different models and things.

[00:09:20] Matthew: It was just very good at consistency, and cheap. A couple of the things I’ve seen is, could be a bit of a competitor to Photoshop in a way, but you can just ask it to tweak the environment and maintain the model. is interesting. 

[00:09:33] Dara: Yeah. I think that you mean, you know, Photoshop at the moment is there’s a lot of other things it can do, but, I read that as well and, and I think, you know, I would be worried maybe if I was a, an Adobe executive ’cause it does seem like, you know, that that’s the way this is heading.

[00:09:47] Dara: And if you can do, you know, if you can get to a point where you can do, ’cause that’s been the problem before, hasn’t it been, it hasn’t been precise enough. So you can’t, you, you couldn’t really use the earlier models if you, if you had a, a need to create something that had some kind of element of precision in it, whether it was keeping the.

[00:10:03] Dara: You know, the input image is the same or you know, even the early problems, remember with text and it still sometimes does that where it can’t spell things right, or puts the letters backwards or, you know. Yeah. And I think the more precise that gets, the more people are going to be able to use it for actual graphic design.

[00:10:18] Matthew: Yeah. We were messing with it before the show today and like I put in the Measure Lab logo and it pretty much the font’s correct. The colors are correct. And it was able to sort of, it made like a neon sign of it on this one image we’re messing around with. And it, we made like a 3D version of it and another image and it, it all is very much recognizable as the Measure lab logo, colors, fonts, everything.

[00:10:43] Matthew: Yeah. Which feels pretty significant a jump. It, it’s, it’s been getting better and better, but that just, just the way it is, yeah. Nails in it. I feel like it struggles with me, but I might just not realize what I look like. This is, it feels like you, you look right. Yeah. I think I look like JD Vance in half of what it produces, you know, that meme of him.

[00:11:00] Dara: Yeah. That’s what I feel like I look like. What do you think? Do you think, but maybe that’s what I look like. Do you think when he uses it, it spits out a picture that looks like you? Yeah. He has me. He is dmd me. Yeah. Stop stealing my image. Yeah. No, but you’re, you’re right though. I think there’s still, there’s still going to be some little kinks.

[00:11:17] Dara: One thing I found, I don’t know if it was something to do with the prompts or whatever, but it preserves the kind of like if you feed it an image, it’ll preserve that pretty well and then you can change the background or whatever. Yeah. I’ve noticed this before and it still did it with nano bananas, where I feel like it.

[00:11:32] Dara: Feels ridiculous saying that Nano, that was his code name, but it seems so stuck. Yeah, it does a bit, yeah. Yeah. Nano banana, you know, I was using Nano banana the other day, so when I, I I, I tried to get it to alter the background of an image and then I, I, I kind of changed the brief a little bit and it, it’s got a bit stuck on what it had done the first time, and it kind of merged the two requests together and it couldn’t backtrack out of it then, so I had to resubmit the image from the beginning.

[00:11:58] Dara: It got fixated on what I had asked for, and it just came up with this weird amalgamation of the old prompt on the new prompt and, and, and didn’t seem to be able to figure it out. Yeah. Got it. Got itself a bit stuck in a loop. 

[00:12:11] Matthew: Yeah, I, I found, because I uploaded our two headshots to it and was messing around with it, and at a certain point when we got further and further away from me uploading those headshots, we started to get too handsome.

[00:12:23] Matthew: Like, it’s just, I was like, this doesn’t look right. We look too good. So I had to upload ’em, be again and go, no, uglier please to go back to, yeah. 

[00:12:31] Dara: Can, can you knock? You know, can you add a few rough edges? Make, make it more realistic. Yeah. 

[00:12:36] Matthew: Yeah. So I think so. Yeah. I think it’s good at that initial upload and maybe the further it gets from that source material, it starts to move away grade.

[00:12:43] Dara: Yeah. It starts to move away from, but it, it, it is, it is really, really powerful. and the point about Photoshop, there’s another thing that I saw when scouring the, you know, the news, so you know, the open source like version, the people, sorry, it’s not a version of Photoshop, but people use gim, which is an open source image.

[00:13:01] Dara: Yeah. Editing software. A bit like Photoshop, but it’s free, it’s online. And they have a plugin now that uses nano bananas. Oh, post quickly. So, yeah, I know. Yeah. So if you can use that gen AI tech within, you know, gen AI model within. It’s just another, it’s another shot at Photoshop really. I mean, that’s not nothing to Google, but just the fact that you can do this kind of image editing now. Yeah. Without needing to have potentially a, you know, paid subscription to, to Photoshop. 

[00:13:32] Matthew: And it, it, it does feel a little bit like one of maybe a sector that some good healthy competition would benefit from. ’cause Adobe have had some scrupulous business practices because of their monopoly it feels like, for a few years.

[00:13:44] Matthew: So I’m not necessarily too disheartened with some competition, putting a bit of pressure on them, perhaps.

[00:13:50] Dara: No, I, I agree. And our guest on today’s show is from Adobe. no, you’re, you’re absolutely right. I think it is, it’s an area where it definitely needs some disruption because it’s, yeah, it feels strange now, doesn’t it, to, to kind of pay for software like that, and especially at a, yeah.

[00:14:07] Dara: Kind of premium when a lot of this stuff can just be, you know. It’s not quite pay. Well, it is pay as you go, I guess in a, you know, in a way compared to paying this fixed license fee for something like Photoshop or whatever, does seem like it’s, it’s a little bit, a little bit old that now. 

[00:14:24] Matthew: Well I think that’s one of the big, the big sort of predictions, isn’t it? With Gen AI and Claude Code and things like that. The fixed model where you just add loads of features. And we’ve talked about this a lot with sort of like ETL tools and things like that where they add more and more features and more and more features to, and they up the cost of these ongoing monthly subscriptions and you’ve got a load of a lot of features you didn’t ask for and you’re just solving a very specific problem for a, you know, absorbent price.

[00:14:49] Matthew: Price maybe Claude Code and AI allows you to just solve that problem yourself just by quickly coding something. Yeah. Doesn’t have to be a big production thing. It can just be a small piece of code that does what you need to do. Yeah. and I think there’s a reaction in the industry where some are starting to move more towards a pay what you use model and away from that flat.

[00:15:10] Matthew: SaaS kind of fee that you just pay monthly for stuff you don’t use as a reaction to that perhaps and other things, I suppose. 

[00:15:17] Dara: Yeah. Bring it on. I’m in favor. Yeah, agreed. He says as he alienates a, you know, wide section of our listening, listening topic. 

[00:15:25] Matthew: Yeah. Yeah. So far we’ve taken down ETL tools and Adobe.

[00:15:29] Dara: Yeah. So yeah, we’ve got who’s next? We go for the biggies. Yeah. Let’s talk about Google. Yeah, let’s take them. But they got a bit too big for their boots, didn’t they? Yeah. Yeah. So, so, but speaking, you just, you just keep teeing these up, don’t you? it’s almost like we rehearsed this, but we, we definitely didn’t, we do we, we definitely we definitely didn’t.

[00:15:49] Dara: So, speaking of Google and speaking of Google disrupting and, you know, edging into other people’s space, there’s also a bit of chatter around them, so have you seen the new features they’ve added to Google Translate? So you’ve got like a live playback of, so you can, you can open up Google Translate. So this is useful for me now in my new life in Spain.

[00:16:10] Dara: I can speak in English, it’ll automatically translate that to Spanish and it’ll play back the Spanish audio to whoever I’m speaking to. They’ll respond to me in Spanish. It’ll translate to English and play it back to me. So you can have an actual live, real time conversation with somebody. And there’s another bit which I haven’t looked at yet, but there’s a, like a learning resource in there.

[00:16:28] Dara: I don’t know much about it, I dunno how it works yet, but there’s something else they’ve added in. And the chatter is that it, you know, it’s going to edge into like Duolingo, that kind of language learning. Yeah. And territory. Yeah. 

[00:16:39] Matthew: Yeah. So I think I have seen that and I have played with that. ’cause my wife, she does a lot of learning of Spanish and things like that.

[00:16:47] Matthew: And I remember finding it, I think it’s something along the lines of. You can essentially generatively create situations and it will kind of be put together like a little conversation and syllabus around that specific scenario. And, and you can be much more specific and directed in what you want to learn about rather than like, I suppose with a jewel lingo or a Babel, just to really just, let’s just start chopping these companies down.

[00:17:10] Matthew: Keep, keep that train going. It’s like, you know, the cafe or at the library or buying socks. Yeah. To those very specific scenarios. You could be much more like, I really want a, I dunno, this specific drink at this specific bar. I don’t know. Yeah. Something along those lines. 

[00:17:29] Dara: You don’t need socks in Spain, by the way, just to add that. That’s a correction in, but yeah, no, I, I think you’re right. I’m going to play around with it ’cause I am using Duolingo and a couple of other apps I’ve got on the go as well. so I’m going to, I’m going to try this out, but I did demo the live if, if you or anyone else had seen me, I probably looked crazy ’cause I was having a conversation with myself in one side of me was speaking English and then the other part of me was speaking really bad Spanish and I was having this conversation with myself using Google Translate app just to test out the, the. 

[00:18:01] Matthew: Did you go as far as to do that? Do you know that, that sort of dress half and half, but you’ve got one side, you’ve got sort of traditional Spanish dress on on the other 

[00:18:07] Dara: Yeah, I was going to shave some stereotypical Irish. Yeah, exactly. Yeah. Like half a green hat and half a Matador’s hat or something. Yeah, yeah. Yeah. God, it’s like you were here.

[00:18:20] Matthew: It does feel like the, it does feel like we’re getting close to the, what’s the name, the Rosetta Stone. Like just, just the ability to be universal. It’s a thing, glasses and stuff now, isn’t it? Where I think Google demoed it, where you, you just sort of see the words appearing in front of you and then talk back and they have the same pair of glasses on and it’s kind of seamless.

[00:18:40] Dara: Yeah, it’s interesting. So, this has gone beyond the scope of the usual scope of our podcast, but we are philosophers. Yeah. We do like to think about, you know, the big problems and questions in life. So I was talking to Hannah about exactly that the other day. I was saying, look, do you think there’s actually not going to be a need to learn languages in the future?

[00:18:56] Dara: Because when the hardware, the software is almost there now, or kind of isn’t bad, but it’s like you don’t want to take your phone out and be like, sorry, could you speak into my microphone? But when it’s all built in, if it is a pair of glasses or a little earpiece you put in or whatever, or a chip in our brain, and you can just speak to somebody, Neil, I guess his way.

[00:19:15] Dara: Well, yeah. It’s going to, you know, disincentivize people from actually learning any of the languages. So, you know, is that a bad thing? Is it a good thing? 

[00:19:24] Matthew: Both. I guess it’s a, yeah, it’s a shame. It’s not, it’s probably not new, is it? I mean, there’s so many languages that are being held onto very tightly by very small communities, you know, small communities in those countries, because probably English is, yeah, just gobbled it up. So yeah. Maybe that problem just gets exacerbated. But yeah, we’re really down the line of our marketing analytics route with this. 

[00:19:50] Dara: We check, well listen, everything’s linked to marketing analytics, you know, it’s like the central Yeah. How do you measure that? Languages decline. Yeah. We’ll tell you about the subsequent episodes of this podcast.

[00:20:00] Matthew: Let’s keep listening. 

[00:20:00] Dara: Okay. That’s our, our, our news roundup for this week. Just a little, another little plug for our upcoming webinar. Thursday, the 11th of September. Our next webinar, which is on data form, is going to be at 1:00 PM UK time. So if you haven’t already signed up to that, and you’re interested, then please do so.

[00:20:18] Dara: We’ll include a signup link in the show notes. So our guest today is Gunnar Grise from A Bit Cheap. He’s an analytics consultant and engineer, and also a lecturer as well. Very interesting guy. We talked to him. We could have talked to him about loads of different things, but we focused on MCP servers, which he’s been playing around with lately and, and so have we. So it was quite an interesting chat about that, wasn’t it Matthew? 

[00:20:44] Matthew: Yeah. Really interesting. Gave, gave quite a few sort of ideas of, of things we could be exploring ourselves and, and I think he’s gone quite far down the rabbit hole. So it’s, it’s really just seeing what, what capabilities are there now and where, where he sees it going.

[00:20:57] Matthew: For sure. Enjoy.

[00:21:02] Dara: So a very warm welcome to the show, to Gunnar Risa, who’s our guest today. So Gunnar, firstly hello, welcome to the Measure Pod and thank you for agreeing to join us. 

[00:21:12] Gunnar: Yeah, thank you. Thank you for having me. Pleasure being here. Looking forward to the chat today. 

[00:21:17] Dara: Likewise. We’re really looking forward to it too. But before we get into the meat of the discussion, if you could kindly introduce yourself to our listeners. So this is your chance to give your own bio. And I always say this, but you can go into as much or as little detail as you like. 

[00:21:33] Gunnar: Okay, cool. I’ll try my best to give you a bit of a background of me currently working with a P chief in the position of a digital analytics consultant based in Denmark, Copenhagen, though I’m originally from Germany.

[00:21:47] Gunnar: So that’s why I started out, that’s also how I got into, or where I got into digital marketing, digital analytics. So after my bachelor’s, I decided I needed to get a bit more hands-on experience with all these things, right? So Google Ads, digital analytics, Google Analytics, Google Tech manager. So I was looking for a position in that field and ended up in a small real estate agent startup as an online marketing manager.

[00:22:16] Gunnar: Joined as the third employee, and I think you guys can imagine what that means. So I was pretty much responsible for everything. but I got, you know, exposed to Google ads, bing ads, Facebook ads, then further down the line, Google Analytics, Google Tech Manager as well. I was, you know, pretty much teaching myself all these skills.

[00:22:36] Gunnar: I had a very, seasoned founder on the side who could help me get started with it. But at some point I simply realized, okay, in order to become better at this, I needed some more technical background, I would say, which I didn’t have from my bachelor’s in business administration. So that’s when I decided to study e-commerce back in Hamburg as well.

[00:22:56] Gunnar: And yeah, that’s where I first got exposed to databases, programming, e-commerce, business models, and also digital analytics. I actually had a weapon analytics course back there. Ah, and pretty soon I realized, okay, this is what I want to do. I kept my job at this real estate startup.

[00:23:16] Gunnar: And yeah, took on more of the Google Nordic and tech measure responsibility. At some point we got our first proper CRM, which was more of a SQL database, so I got to, to write my, my first SQ queries. I really enjoyed that. And then it was time to, to write my master thesis and the topic of attribution models, which I was really, really interested in back in the days I was looking for a good partner to, to write this thesis with.

[00:23:42] Gunnar: And that’s how I joined Rein Web Services as it was called back in the days, in Hamburg. And there I joined the data science team, again as the second employee, which gave me a lot of free reign, a lot of room to experiment. Yeah. And there I got into, yeah, I got to work with attribution models, which I coded in R which was a difficult time for me because, you know, I had this basic background in programming.

[00:24:07] Gunnar: Now a little bit of statistics, but I would say that I went then. A little bit over my head into this thesis. So it was an intense five months that I had there, and then I joined the track and data science team and in a full-time capacity. Stayed there for, close to two years I would say. And then, yeah, I always had the dream of moving abroad to kind of, you know, go somewhere else in Germany.

[00:24:32] Gunnar: I felt like, okay, I don’t know, maybe Berlin or Munich could be alternatives, but I wasn’t too attracted to it. At the same time during my thesis, I’ve been following Mark Edmundson quite a lot. Yeah. so I gave him a ping on LinkedIn, and was like, Hey, you know, are you guys at IH Nordic looking for any employees?

[00:24:49] Gunnar: And it was like, yeah, sure. Always send me a cv. And then, you know, I had two interviews, tech my, got a contract and moved to Copenhagen, that’s almost six years ago now. Joined IH as a digital, what was it called? Analytics specialist, I think. And worked my way up to, to become the tech lead of the. Of the implementation engineering department, where I was responsible for, yeah, Google Analytics and Google Tech Manager.

[00:25:18] Gunnar: So there was also the time where it shifted a little bit in terms of focus areas. So before that I was working more with, you know, with the BigQuery raw data, trying to, to build, yeah, as I said, attribution models on top of it, or in session realtime prediction models that we were deploying on the Google Cloud to get some realtime predictions on, I don’t know, the likelihood of a conversion and so on and so forth.

[00:25:44] Gunnar: but now it’s deep in the trenches with Google Analytics. Google Tech Manager, soon after Google Analytics app and web, was released by Google. So I had like, yeah, my fair share of exposure to that. And then later on, GA four, I mean, you probably, you guys also have been working with it. It wasn’t always an easy, funny product at the beginning.

[00:26:09] Gunnar: So, yeah, I did that for, for five years and yeah, then last summer, almost a year ago, I decided, okay, it’s, it’s time to, to do, yeah. Take on the next challenge. And that was then essentially for me to, to become self-employed and join the cheap, collective network. So I would say it’s a very, you know, lightweight agency, essentially.

[00:26:32] Gunnar: like a, yeah, a bunch of people like me gathering together under this brand umbrella and yeah, trying to, to solve our customer’s problems to the best of our abilities. And yeah, I think part of it was also kind of getting a bit away from just this Google Trek, right? So do something else than just Google Analytics.

[00:26:51] Gunnar: Just Google Tech Manager. Also, go back a little bit again to work with the Google Cloud. Yeah, especially BigQuery data form was just, you know, becoming the, the big thing. I wanted to have more exposure to that, but also just see. What other tools are out there, right? Because I’ve been essentially working with Google Analytics my entire career.

[00:27:11] Gunnar: I’ve been, you know, dabbling around a little bit with Pivot, Mixpanel and so on, but never applied it into a real, real life customer setting. And yeah, that’s what I’ve been doing for the past year now. So yeah, I think that’s that in a nutshell. 

[00:27:28] Dara: Yeah, no really, really interesting nutshell and quite broad. And we could probably go down so many different roads. but the road we want to go down with you was sparked from, I think it showed up in a couple of places, but I know you spoke at Measure Camp Copenhagen about this, and then you wrote a really good blog where you’ve outlined your proof of concept, which is all around using CPS within the context of digital analytics.

[00:27:53] Dara: So I think that was the angle we were particularly keen to talk to you about and, and dig into. But who knows, we might meander a little bit through the conversation as well. So maybe if I, If I kick it off before we get, Matthew will probably take us into, into more detailed questions, but what, what kind of sparked your, what, what got you interested in looking at, at cps?

[00:28:16] Dara: And maybe by answering that you could also give some of our less technical listeners just a little bit of an explanation of what, of what CPS even are. 

[00:28:23] Gunnar: Sure. So, you know, I’ve been working with Chat, chat GPT, quite a lot of the OpenAI, solutions back in the days quite a lot. I, you know, was trying to, to generate texts for a random website and just, you know, essentially create a content form to some extent just to see, hey, can I get this working with, with chat GBT, build some basics, workflows.

[00:28:51] Gunnar: And that’s essentially, you know, how I got into this entire, large, large language model world. I tried around a little bit. You know, like this entire field became more mature. We had, you know, the chat GPT apps out there, CLO desktop out there. and I’ve been really intrigued, especially in the beginning.

[00:29:13] Gunnar: So I was like, okay, this is now giving me superpowers, so to speak, right? So I was, all of a sudden I was able to complete things and half the time I was able to finish up, you know, little side projects that I’ve just been lying around that I gave up on at some point because I became too frustrated working with, I don’t know, all authentication and figuring all of the, these nitty gritty details.

[00:29:37] Gunnar: But at some point I also felt like I was hitting a plateau, right? Because, you know, there was all this bus about gen ai, how it’s going to change everything. I don’t know, we all might lose our jobs and so on and so forth. At the same time, I, you know, didn’t have that feeling quite yet because I was like, okay, there’s always.

[00:29:57] Gunnar: A big need for me to be in the lube and actually take the output from the LLM and apply it into the system that I wanted to work at. That might be Google Analytics, Google Tech Managers, some, some JavaScript files or Python scripts, whatever, sql. And then I came across the MCP servers. I think the first time that I really heard about it was when STA launched their Google Tech Manager, MCP server.

[00:30:23] Gunnar: And in a nutshell, what it is, it is essentially, yeah, and you, I would probably call it like an API rep that is designed to work with these large language models, right? So imagine that you have your tools, again, not desktop, GitHub copilot or your Gemini CLI or not code, whatever. these applications, they essentially have a lot of knowledge, right?

[00:30:50] Gunnar: all the knowledge of the world. You could say even, But they cannot do any actions on, on your behalf, right? So because of that, they would need to have access to your business systems. That can be, I don’t know, again, you’re Google analytics, your Google Tech manager, HubSpot, your CM, you name it.

[00:31:09] Gunnar: Anything that’s essentially like API capable can be hooked up to these applications, to these AI applications using the model context protocol servers. And that is what really got me hooked, right? Because all of a sudden it was me not having to leave the Globe Desktop, my, my AI applications interface anymore to do these or to implement these changes that I requested Claude to give me an offer for, but rather Claude could take the action on my behalf.

[00:31:39] Gunnar: So essentially was, yeah, I think you can think of MCP servers as, I mean, they always refer to as the, as the USBC stick for, for applications. I think I like to think of them more as, maybe, yeah, I don’t know. A school or a school that teaches your LLM model. some more skills and levels it up to, to that extent.

[00:32:06] Matthew: That’s, so you, so obviously you, you discovered this new superpower because I think, I think a lot of us were the same in the early days where it’s almost like you were waiting for the killer app. Yeah. Along like, it’s like this is clearly insanely powerful, but you’re just trying to figure out how it can be utilized.

[00:32:23] Matthew: and you discovered M ccp, so it’s like, what kinds of things have you been hooking together with these cps that that, and what has it allowed you to do and, and Yeah. 

[00:32:33] Gunnar: Yeah, yeah. You know, as I said, the past five, six years I spent doing Google Analytics implementations and making sure that they work correctly.

[00:32:42] Gunnar: Which is complex or can be complex work, but at the same time, it’s also not rocket science, right? So, like a, if you once learned the ropes of it, you, you, you identify these, that you have a very repetitive task actually that you need to go through, over and over and over again. And that essentially is, once I realized that, brought that together with these new capabilities, I was thinking of an implementation process in, in terms of steps that, okay, first we need to collect the, the business needs, the requirements to identify what are our KPIs, which events do we want to track based on that.

[00:33:22] Gunnar: Then we have a look at the website, get an understanding of the website structure, the, the dom, the the data layer, have a look at the Google Tech manager, what is implemented already. And then based on that we would go into Google Tech Manager, configure our tags or triggers or variables. All of that essentially depends on the website structure.

[00:33:42] Gunnar: Set up a GA four property, press the button and publish the GTM container, the new workspace, right? And then after that we would verify the setup. And essentially from, from my perspective, there’s only a handful of skills really involved in that. And that is first being able to interact with the website, being able to read the dom, being able to read or investigate the network requests.

[00:34:07] Gunnar: So see which events are actually being sent to the marketing or analytics platforms. And then you need to have GA skills. So you need to know how to set up a data stream, set up a property configured properly, and I dunno, like set up the conversion events, these kinds of things, right? The re-register, the customer dimensions, metrics, and so on and so forth.

[00:34:27] Gunnar: And then you go into Google Tech Manager and set up again, your text triggers and variables, right? So pretty predefined steps and all of this is doable with APIs. That is essentially then I think the first, the other MCP server that I came across after, or that’s essentially the, the tool set that you need to have, right?

[00:34:48] Gunnar: The capabilities, the skills. So, and then from my perspective, Google Tech Manager skills were out there already with the, with the state MCP server. So I needed Google analytics and browser skills essentially for the browser skills. I pretty soon came across playwright, MCP, which is, yeah, playwright is a framework, I think an open source framework developed by Microsoft, which essentially allows you to do browser testing in a programmatic way.

[00:35:17] Gunnar: And they, again, they released an MCP server for that, meaning you could all of a sudden go hook it up to your cloud or Gemini or whatever, AI application and be like, Hey, please open this website and tell me, browse five pages. Give me an overview of all the network requests, all the GA four events that are being sent, record the data layer and give me an overview of that.

[00:35:41] Gunnar: So with that, we had the browser skills covered as well. GTM skills were covered already. And then for GA I had pretty, pretty extensive experience working with the API. So I essentially just set out to build my own MCP server using FAST MCP, which is a, yeah, a framework again, to build these MCP servers in a pretty lightweight way, and actually turned out to be way easier than I thought at the beginning.

[00:36:10] Gunnar: Again, also having, you know, a cloud code at, at my, at my disposal, for example. So, and then all of a sudden I had all the skills that I needed. I had a process in mind that my AI application needs to go through, and it had the capabilities right. And then I was just, please go to my test website. Visit all the websites, browse it as if you were a user that is interested in, I don’t know, buying five products.

[00:36:38] Gunnar: And the playwright MCP server went out, browse the website, added things to the basket while it was doing it, it was recording the, the data layer and the network requested were already in place, kept, track of the CSS classes and Id so observing the DOM as well, all the way through to the purchase. And that is then essentially the output that you have, right?

[00:37:01] Gunnar: So then, your LLM all of a sudden has an understanding context of your, of your website, which you then can feed into the next step. Okay. Give me the relevant data layer events. I need to use it in order to implement proper e-commerce tracking, for example, or I don’t know, implement form tracking or navigation tracking, whatever it is really.

[00:37:25] Gunnar: Right. And based on these, configure my GTM containers, set up the triggers, the variables, the tag, and then the next step is, okay, please now configure a GA four property. set up a data stream. Take the measurement id, again, insert that into my GTM setup and then publish the container. Once you’re done, please browse the website.

[00:37:45] Gunnar: Again, record, go through the same steps, record the network request and see if what we initially agreed upon should be implemented is actually being tracked. And then you can do those on the website. And then later on, again, using the GA MCP server, you could also, for example, check the real time reports, right?

[00:38:03] Gunnar: Are you seeing these events going into your GEO GA four property? So, and that is essentially this PC that or my, my blog post that you, and also my talk, was about really, right. So I essentially, all of a sudden I saw this, this potential that what I’ve always been dreaming about, getting rid of this cumbersome, sometimes also a little bit mundane work and, source that out and give me more time to actually, yeah, instead of just being a data collector, do data analysis, do something more value creating essentially, right?

[00:38:37] Gunnar: Or also just use it to double check myself, right? So now, nowadays I can just go out and verify tracking implementations and my browser and then the background, have Claude doing the same stuff, compare notes essentially, and, yeah, trying to up the quality of my deliveries. So yeah, that’s, I think.

[00:38:56] Dara: What was the, ’cause I know you, I know you said this, you do add this as a kind of almost disclaimer at the end of your, of your blog, and you say that the, the videos that you include, they’re the kind of final version, but you had to do quite a lot of prompt refining and you had to check things and correct mistakes.

[00:39:12] Dara: Yeah. So what were the kind of, what were the what, what were the kind of themes that you saw in terms of what went wrong? Like, you obviously wouldn’t suggest someone just go and do this and don’t check mm-hmm. And just assume it’s all going to be output correctly. Yeah. So what, where, where did it, where did it go wrong?

[00:39:27] Gunnar: I think, you know, when you think about what’s happening under the hoods, right? It’s all API calls at the end of the day, which are pretty, or by nature, are very deterministic, right? So you have an input, you make a call to the server, something is happening, and then you get the response back. But in order to get these API calls to work correctly, you essentially, if you do it manually via Python or JavaScript, like in a program programmatic way, you need to get it right.

[00:39:53] Gunnar: And so it’s a very deterministic approach. With these large language models, you do not necessarily need to know the in and outs of these APIs. They’re essentially, this logic is built into, or these, yeah, the logic, the capabilities are built into the MCP server. So you can work with your natural language or written language in that sense, and give certain commands that the LLM then translates into API calls, and whenever something is translated from one language into another, there is the chance something goes wrong.

[00:40:25] Gunnar: Right. And that is literally what, what happened. Right? So I essentially learn to, yeah, get better at understanding how a cloud desktop, in this case, interprets my natural language commands and turns that into code, so to speak. There was a lot of the fine tuning just to understand how to, yeah, how Claude interprets my commands.

[00:40:45] Gunnar: The other thing I think is also, I think at the beginning I had the hope that I could just go out and let the LLM do everything right, but I also learned that. In order to use these MCP services effectively, you still need to, or it’s, you will be arriving at the desired end result faster. If you really know how the API works under the hood, right?

[00:41:06] Gunnar: So what is, if you want to create a property via the GA four admin API, what does this API call require as an input. Right? So if you know, okay, it needs to have the account ID of the GA account that this property will be created in, then you can give it in your initial request to LLM, right? Rather than it coming back trying, creating the property, failing, giving you just a like not very meaningful error message in, in certain cases.

[00:41:35] Gunnar: And then eventually you, after some more debugging, figured out AR needed the account id. So in my prompt, I should include this counter ID right away. 

[00:41:43] Matthew: It’s interesting, it’s eight years ago that lost all track of time with lms. 

[00:41:47] Gunnar: It feels like it’s moving so fast, isn’t it? Yeah. 

[00:41:50] Matthew: About 30 years ago I was playing with, L chain. And one of the biggest, the biggest problems I have found with that and this, you know, it’s a similar sort of concept in trying to put in tooling and essentially give it the ability to call an API to do whatever. But it was much more like seven different ways of doing it. And it was difficult to decide which one was right.

[00:42:12] Matthew: But the main problem I always found with it was, like you say, getting it to translate what you say to make the right decision on the right tooling. ’cause it’s got, say it’s got 20 different tools like to just go, yeah, I know exactly what I need here. I need to go to do the PII . I need to use the admin API to do this, to do this, to do this.

[00:42:30] Matthew: And that’s what I always found the most difficult thing. I dunno if you’ve had any more, look, well not look had any more refinement since you saved, released that blog post in sort of pre prompting and things like that to get more out of that as, as things moved on since you posted that. 

[00:42:44] Gunnar: Yeah. So I think, you know, like the tool, so maybe if we take a step back, the MCP server, right?

[00:42:48] Gunnar: So in, in its core, I think it has three. Main components, essentially. The first one that is what we’re talking about or what we’ve been talking about so far, is the tools, right? So the ability to take, make an API call to any given service essentially. Then, the two others are resources. And then you also have prompt templates.

[00:43:07] Gunnar: Resources essentially is more like, I think the equivalent to a get request to an API, right? So you’re just fetching some information from the server, which then your LLM has as additional context. An example could be, for example, if you want to use the Google analytics MCP server to perform certain types of analysis or have it perform certain types of analysis on top of your data, it might be helpful for the LLM to know which metrics dimensions are available.

[00:43:36] Gunnar: Also, custom dimensions and custom metrics, right? So something that’s specific to your setup. So you could have a resource essentially allowing the LLM to fetch these metrics and dimensions as information put into the context. And then once it starts making actual API calls. Knows that these dimensions and metrics are there in the first place.

[00:43:55] Gunnar: And then coming back to what you just said in terms of refining, I think prompt templates are really, really helpful with that, which is essentially abstracting this prompt refining or taking this prompt refining away from the user of the, of the application and internalizing it into the MCP server. So what do I mean by that?

[00:44:13] Gunnar: If you know, for example, that for, or in order to get a specific report from Google Analytics, you need to use these, I don’t know, two dimensions and these five metrics. It always needs to be that instead of just being like, okay, create me a report that gives me X, Y, Z, be like, okay, give, create me a report.

[00:44:32] Gunnar: Use session source medium and session campaign instead of just source medium and campaign and sessions and active users and you know, like transactions as, as the metrics. Fill that in. And all the user has to do is essentially is click okay, I would like to use this prompt to create this report. And all they have to do is provide their property ID essentially.

[00:44:56] Gunnar: So, and that essentially is, I think where you as an MCP developer can make the life of your end users a lot easier if you essentially identify certain workflows and use cases and give them, you know, you’ve been done doing all the testing and you know what works, right? And then feed that and give that, it’s posted to the, to the end user of the, of the MCP server so that they don’t have to do these trial and error and thinking, they just get what they ask for, essentially.

[00:45:25] Dara: Can you, can you customize those prompt templates as an end user? So like, if you’re using an MCP server as a company and you want your users to be able to, if, if you want to add something to that template or edit it in some way, can you do that? Or is it baked into the. To the MCP server. 

[00:45:43] Gunnar: It’s baked into the MCP server, essentially, right? So if you want to have custom prompts, then it’s like a regular prompt that you would write into your cloud desktop. Again, there’s nothing hindering you as a company to develop your own library or repository of prompts that work well, and then make that, share that with the non-technical data consumers, for example, and give them this as a guide.

[00:46:08] Gunnar: But otherwise, it’s baked into the MCP server. But again, I think, you know, these MCP servers, they’re not that difficult to build naturally. 

[00:46:15] Matthew: So I, I don’t, I suppose you could, if it’s publicly available in a GitHub repo Yeah. Take a, take a copy of it, do your tweaks and everything, and then host your own sort of version of that MCP server.

[00:46:28] Gunnar: Exactly. So there’s nothing hindering you right now, Google Analytics or Google just came out with their own Google Analytics MCP server. It’s in a GitHub repository. You have access to the code. You can just add, you know, a couple of lines here and there and then tweak it to your needs and then make it available to users.

[00:46:42] Matthew: So I was going to ask a question about that very newly released Google Analytics server, because you mentioned, you mentioned earlier on that you, you originally developed your own, sort of way of interacting with that, presumably because that wasn’t out yet. Exactly. So have you sort of lifted and shifted in Google’s version and played with that in your flow, or are you too married to your, your old way?

[00:47:05] Gunnar: When it came out, I had very, very high hopes for it, and I must say, you know, it, it has, it’s, it’s a, it’s a very solid MCP server. If you, all you want to do is chat with your data, essentially, right? Because that’s essentially then coming back to these tools that are available. That’s the only tools that this MCP server has available, which is really.

[00:47:29] Gunnar: get a specific Google Analytics report and, you know, retrieve the data and then you can have it perform some analysis on top of, but again, for my use case that I had in mind where I would actually like to do configuration of, GA four properties, data streams and accounts, even maybe, I don’t know, even add users for example, to a Google account or property that is simply not available in this, the current version of this Google XMP server, which also makes sense to some extent, right?

[00:48:00] Gunnar: Because again, if you just fetch data from Google Analytics, you cannot necessarily break anything in the setup, right? So that coming back to you kind of needs to know what you’re doing. Otherwise, you know, essentially you’re giving your LLM free pass to do whatever with your, with, with your Google Analytics account.

[00:48:14] Gunnar: And that’s not always a good idea, especially if you do not necessarily know what you want to do or can do. 

[00:48:19] Matthew: I did notice that with the state MCP when I first was first playing with that. Like you can. Did you need a container change? Yes, exactly. That. Exactly. That seems you hope, you hope that, like, like you say that the MCV developers, and I’m sure State have done this, they’re very switched on, but you’d hope they have baked in there.

[00:48:34] Matthew: Like Yeah, don’t just go and delete stuff or someone explicitly says so and if they do say so, go. Are you sure about that? 

[00:48:40] Gunnar: Yeah. but yeah, but I think you can do that, right? Because it’s simply just, it’s just a tool. Just the API call. If you ask them to delete it by accident, even then it’ll do that.

[00:48:50] Matthew: It’s, yeah, I guess, I guess it’s just getting that translation layer as absolutely rock solid as possible. And I think me and I were talking about this, the last podcast. No, not on the last podcast. Whenever this podcast comes out, we would, we were talking about this just after GPT five came out and how it seemed like open AI had really, really strived to try and reduce hallucinations and make it as, make the responses as sort of concrete as possible.

[00:49:17] Matthew: It didn’t necessarily feel like I got much more intelligent, but it was trying to make sure that. What it responded with was Correct. And, to me that’s the logical step of getting more power out of the existing models so that it can interface with these MCP servers with more accuracy so you can more, more times than not Yeah.

[00:49:36] Matthew: Get it to do the things you want to do and chain together those tools in a, in a much more functional way. 

[00:49:41] Gunnar: Yeah, that’s, yeah, definitely true. But I still would argue that, like from a governance perspective, I would not necessarily rely on open AI or J 65 getting it right and, you know, knowing what I meant by my, my prompt, but rather it should be, again, as you also said, like probably baked into the MCP server in the first place.

[00:50:02] Gunnar: So maybe just do not expose, delete, GTM container as a tool in the first place. That could be a way, for example, right to. Absolutely sure. It’s just not possible then to do this with this MCP server. 

[00:50:16] Dara: So how do, how do you balance, so if you are advising some of your clients, for example, where’s the, because obviously the benefit or potential benefit of some of these tools, not just MCP servers, but some of the AI functionality in general is that it, it should in theory level, the playing field and you don’t have to have the same depth of technical knowledge that you maybe did previously.

[00:50:36] Dara: But the flip side of that is if you don’t understand what goes on underneath the hood, then you could potentially go in and cause a whole load of harm. Yeah. So how do you, and, and I appreciate this is a big question, but how do you, how do you walk that tie rope if you’re an organization where you want to take advantage of some of the, maybe the time saving or cost saving functionality that these tools provide, but you also don’t want to just open it up to everybody and anyone could just go in and delete containers and rewrite your whole measurement plan?

[00:51:05] Gunnar: I think that’s a good question. I would say that. You know, there’s a step before the, you know, before making these tools accessible to the average data consumer that does not necessarily know what can be done with this MCP server or what can happen to a GA account, for example. Yeah. GA property.

[00:51:26] Gunnar: Right. And that is, I think, starting with documentation and training in the first place. And then I think, for example, having a prompt library, for example, having certain prompt templates in place that simply, you know, from, just from an input perspective, do not allow the users to go off the rails essentially.

[00:51:43] Gunnar: So setting rather strict guardrails. So that’s from the training training perspective. Then within these tools itself, you can also give, still give different use of permissions, right? So maybe you just use the access rights that, I dunno if I’m just, again, just a data consumer. Every now and then I look into Google and analytics to check how my campaigns have performed.

[00:52:07] Gunnar: And, but I, I’m. I’m not a Google Index expert, right? So I cannot configure it and I also don’t want to configure it. I only want to retrieve data. So then I probably don’t have admin or edit rights in the GA four property. So ideally, at least I don’t have these rights in the first place. So when I authenticate to the MCP server with my Google account, for example, I just don’t have the rights to make any, any of these changes.

[00:52:31] Gunnar: And then I think the last step is what I, what I mentioned before, you can also have different versions potentially of an MCP server. So there’s one for the super users, for the power users, and then there’s one another one that doesn’t even have these capabilities tools in the first place to break anything in the setup.

[00:52:45] Gunnar: Right? So, yeah, I don’t know if that answers the question, I hope. 

[00:52:50] Matthew: Yeah, it does. Yeah. I think kind of got a tenuous link to what you just finished on there, which was around like user level access. Yeah. I don’t know if you’ll be able to answer this because obviously you kind of work at. A bit more of a freelance capacity or that in our organization right now.

[00:53:08] Matthew: I, I guess how accessible are these things to, to spread out around companies with a bit more ease? Because it feels at the minute that it’s still very much the, still very much in the realm of the techie and the person who knows how to work these things and figure it out and tinker. Yeah. It doesn’t feel like it yet.

[00:53:25] Matthew: It’s quite yet at a point where Joe Blogs and Company X can just start typing away and Yeah. And start relaxing on things. That’s true. 

[00:53:32] Gunnar: Yeah. I think, you know, like these, or like, like any MCP server, right? So you can run these servers locally, right? And it just, you spin up the server on your local machine and then only you via your local cloud desktop app can use this MCP server in the first place.

[00:53:48] Gunnar: And then, and I think that’s how it all started. I think that’s also the first version of the state MCP server. When I first came across it, it was built like that, right? So you had to download the guitar repo. Then you had to link it to your cloud desktop and then it would run locally. I think nowadays this local version is not even accessible anymore.

[00:54:06] Gunnar: The, it’s still open source, but it’s essentially this version that state is hosting for you, right? So they think they run it on CloudFlare workers or something, or CloudFlare, which is like an MCP hosting service. So it’s only STA managing this MCP server and you essentially just connect to it via the internet.

[00:54:27] Gunnar: You, again, use your Google credentials to authenticate to your Google tech manager container and then perform the actions on, on top of it. So, I’m with you. It’s still, I think, very like in this, you know, tech niche and it’s probably only like your, your skilled analyst that will be using it.

[00:54:46] Gunnar: But I think there is this, in order for MCP really to become a thing, it needs to be, as you mentioned, it needs to be more accessible to, to the end users. And I think the main issue here has probably been authentication. Which is not, has not really been built into the MCP. So this model context protocol, which is essentially just again, a protocol like again, a framework of how two services can communicate was simply not baked in there or was not there yet.

[00:55:11] Gunnar: It was not developed, right? Because it’s also a rather fresh concept. But I think it’s, it’s getting there and I think I, I see for example, more and more services offering hosting of these MCP servers, right? So I mentioned that that software has an option. I think fast. MCP just, I don’t know, like what, two days ago they publicly launched their MCP server hosting server.

[00:55:33] Gunnar: So you can actually do that there as well. And I think that’s literally what’s needed at the end of the day, right? So that you, as an end user, as you open a website, you just open, you just connect your, you tell your LLM to connect to the server and you do the work, right? Or let it do the work. 

[00:55:47] Matthew: Yeah, that’s, that’s pretty much how we’ve been. We’ve got a little internal tool called Brian, which is kind of self-built. and that’s current, that’s the current. The way we’re trying to do things is to be able to point to these hosted MCP servers and make them accessible to all different users within, inside of that, that brine interface. Yeah.

[00:56:07] Matthew: One, one of the really useful ones that I have found that does a great job with the OAuth stuff is Zapier. Mm-hmm. So with Zapier, you can, Zapier has an MCP, you can authenticate to Zapier, and then all of the various tools you can use within Zapier you can turn on. So you can turn on Google Drive, you can turn on Google Analytics, Google Tech Manager, et cetera.

[00:56:30] Matthew: And then it, and then there’s a set of tools that Zapier MCP can use through the MCP. I guess the main drawback there is that it doesn’t have that context layer that perhaps comes with a custom built MCP. So for the Google Analytics one, it doesn’t have all of them. Information that presumably Google and the community have put in there about what that is and how it works. Yeah, it’s a solution to the old problem anyway, that I’ve found. 

[00:56:55] Gunnar: Again, but I, I’m pretty sure you know, that I think this authentication issue has been the biggest drawback or that is what’s holding us back at this point, right? So how can you do this? And again, I think in the future this will be natively supported, MCP Pro, the complex protocol.

[00:57:11] Gunnar: That’s, I mean, when you look at the GitHub repo for example, there’s a lot of requests for just that, right? And they’re working on it and they’ve been making progress. So I’m pretty sure that it will be there at some point. And then, and then I think it’s going to be interesting as well, right?

[00:57:23] Gunnar: Because essentially, I mean, you can think of it as an app store or like a, you know, like apps that are available. So maybe there’s even a new, new market, right? So, I dunno, if you’re really good with Google Analytics and you would like to make your knowledge available to, to a lot of end users, you might build your own GA MCP server.

[00:57:44] Gunnar: Make it accessible through like a wholesale version to, to the end users. Right. And then maybe it will monetize it. I don’t know. 

[00:57:50] Matthew: Yeah. I think, I mean, a lot of what I’ve heard, and, and this makes sense I guess, I guess it’s common sense, but the, the front, a lot of, when, when the LMS first came out, you had a load of companies going and building something off the top of the LLM, the Frontier model, lm, and then OpenAI comes out with the new feature and just instantly wipes.

[00:58:09] Matthew: Yeah. However many of these startup companies are off the mat. And, I’ve listened to a couple of podcasts and thinking about it, and the only real way you can build anything and sell services, no, that’s not quite true, but sell products off of LLMs is if it’s in that sort of niche off to the side domain knowledge that it’s no point in them going really deep on.

[00:58:30] Matthew: So like you, like you say, like. Web analytics, marketing analytics, all that knowledge around there, open AI is not going to go right. I’m, we’re going to really concentrate next on getting our web and marketing analytics stuff and tool set sorted out people can build out and use their knowledge to create services for that off the back of these models is what it feels like.

[00:58:51] Matthew: So I, I wouldn’t be surprised if you write in now that does end up being this kind of like, what would they call GPTs? I think, 

[00:58:59] Gunnar: and I also think, you know, there’s, I think another aspect that we haven’t really touched on is, I mean, you know, chat is not, maybe also not necessarily the, the right interface, right?

[00:59:08] Gunnar: So if I want to do a GTM implementation, and especially if I do not necessarily know how to do it right, having like this blank canvas of a chat in front of me not even knowing where to start, even if there’s maybe some, some prompt templates available for me is maybe too, too much to ask. Also from, from a non-technical, stakeholder at the end of the.

[00:59:29] Gunnar: So I think that’s another way that you can add a lot of value is by building customized UIs for a specific use case or a specific workflow, or maybe even within this realm of web analytics or digital analytics. Identifying use cases, workflows and building a UI for that too. Actually make them accessible to, to the end user.

[00:59:51] Dara: So one kind of theory or one kind of line of thought around this is that, you know, you can reduce some of this repetitive work and you can free up your time. I think you even said this earlier about how you can free up your time to do more kinds of value-added work. Do you think that, so people who have historically who work in web analytics and who are really into the kind of technical nuts and bolts of how tagging works, do you see a path for people like that where they can maybe move a little bit like the two of you were just talking about where maybe there might be different use cases where you have to build UIs for different things or you have to better understand how to connect more data sources and maybe get more MCP service to talk to each other or whatever.

[01:00:29] Dara: So rather than there just being the one track where people free up time to move into analysis or more strategic work, do you see a technical path opening up as well where maybe the similar skill sets are required, but the actual nature of the work is a bit different? So instead of physically adding tags in GTM, you might be physically building connections of some kind. Does that make sense? 

[01:00:53] Gunnar: And thereby enabling other users to do this work that you previously did, or, I hope so. So yeah, I think that that was, it would be nice, really. Right. So because if you, I mean, I think just from a community perspective even it might be, you know, if you have like really top GTM GA engineers, you know, putting their, their, their mind to this one service that works like this, I think that would be really nice.

[01:01:24] Gunnar: would be a cool, cool thing to have. Imagine you all of a sudden are able to. Do GTM implementations, like, I don’t know, like a, or, yeah, wouldn’t that be cool? But I, I don’t know if we ever get to that, if you ever get to that stage, but, I, I could see it, could see it happening. I mean, there’s, you know, the L LMS there getting smarter, I would say more intelligent as you want to call it, that, that the MCP or this, this protocol is evolving.

[01:01:55] Gunnar: So, you know, I, and, and we’ve all seen how fast this space develops, in the last couple of years. Right. So not saying no. Yeah, yeah. Fair enough. Yeah. 

[01:02:05] Matthew: If you think about it like an MCP server and an interaction with that MCP server and the actions that can go and take as an agent, almost. Have you played with, ’cause I know Google, Google is betting on their A two A framework to be the MCP of agent to agent.

[01:02:21] Matthew: Yeah. Sort of communication. I wonder if you’ve played with any sort of agent to agent type workflows where you have like a, you know, this distinct thing over here doing this thing and then passing automatically off to another agent to perform another action. Or have you primarily been still within the chat interface and, and working in, in that way?

[01:02:37] Gunnar: At some point I realized there are these limitations that come with this chat interface. Right. And then, you know, I have my former colleague and still colleague and the person that I share the, the, my office with Mark Edmondson. So we essentially Yeah. Set out to, to build something like that, that is using the, the agent development kit and it’s still under the hood using MCP servers, wrapping it in this, the agent, but also then we’re having and having its own ui.

[01:03:04] Gunnar: What I just talked about, right? It’s a bit more customized, more specified to, to the specific workflows that we identified users might want to do on a, you know, recurring basis. I don’t know, copying GTM assets from one container to another and doing that in bulk, setting up new GA four properties, validating that tracking works and so on and so forth.

[01:03:23] Gunnar: And with the, with the agent development kit, it is also fairly easy to keep this logic, but exposing it to, to another agent. Right. So via this, to a framework or protocol that you just mentioned. So, yes, so on the, on the very surfers, I would say I’ve, I’ve been dabbling with it and that’s probably another one.

[01:03:43] Gunnar: We, when I just talked about, you know, new market spaces or something like this or marketplaces, I think there’s also like an H to a marketplace at some point, right? So you might be building something very specific as we just talked, very niche for a specific user group. Then putting it on that marketplace and then allowing other agents to, you know, explore and find these tools and capabilities on your behalf.

[01:04:09] Dara: If you take, say, pre MCP servers, let’s say a hundred percent of the time where you had a GTM task or a GA task or pre task or whatever, you were in that interface or Well, okay. Excluding the API, but let’s just say, you know, for simplicity a hundred percent of the time you did it in the, in the ui.

[01:04:28] Dara: How does that compare now, what percentage of the time when you’ve got a task to complete with say GTM, how much of the time are you actually doing it in the, in the, and I guess a follow on, sorry, to hit you with two questions in one, do you see the UIs of these tools disappearing? Is the need for these UIs going to disappear and is it all going to just become, you know, background agents to agents and MCP service talking to MCP Service?

[01:04:54] Gunnar: Yeah’s a, it’s a good question, I think. For me, who has worked with the GTM or Google Analytics interface for years, going to a chat interface and or like within this GTM and GA interface, I have full control over everything that I do, right? So it’s, and it’s one-to-one. It happens in my head, I click the button, it’s there.

[01:05:17] Gunnar: And if you then move to a chat interface with an MCP integrated, you are giving up control to, to some extent, right? Because we have this translation layer in between that we just mentioned. And for that, therefore, I would say for skilled users working with MCP servers and giving up this certain level of control might actually feel more of a friction.

[01:05:41] Gunnar: But I see it for non-technical users, like, that, that’s where I think the self-service, enablement character really comes into play. So, that’s more where I see the real value making, you know, these kinds of workflows more accessible to. The average person out there hasn’t spent, I dunno, hours on GTM.

[01:06:02] Dara: And that might not just be technical versus non-technical. It could just be familiarity. ’cause I, I think we can both relate to what you’re saying. When you’ve done it a certain way, it feels safe. Yeah, it feels familiar. So maybe as new people move into the, I’m kind of thinking ahead a little bit and thinking as new people come into the industry, and even actually maybe you’d have a perspective on this, ’cause you, you are also a, a lecturer, yeah.

[01:06:25] Dara: So you are teaching the analytics professionals of the future. So do you see this changing in terms of people, how people learn how to achieve an end goal, whether it’s adding some tracking to a website or, or whatever. Do you think it’s going to gradually shift and become more about, you know, less about carrying out actions within a given UI and more about speaking to a chat interface or, or, or whatever it may be?

[01:06:52] Gunnar: Again, I, I think we’re not quite there yet, but I think that is something that definitely, definitely can happen and will happen eventually. So then also coming back to, you know, do we need or will this need for UIs disappear in the future to, to some extent, yes. I think there will be a certain user group that will just do a lot of their stuff through the chat interface and the MCP servers.

[01:07:14] Gunnar: At the same time, I think there will still always be this need for, you know, the experts really knowing how to do things and, and how things work under the hood as well. Because if you just use the chat interface, you will never learn that really. Right? Yeah. So you’ll never be exposed to the actual API or to the interface necessarily.

[01:07:32] Gunnar: So you don’t really know what it is, what it looks like, which might reduce the entry barrier for beginners, but will also potentially make it more difficult for them to go up the learning curve and become experts at some point. Right, because you never have to. I dunno, spend your 10 hours to get this, auth authentication to A GTM, API.

[01:07:55] Gunnar: Right. Or figuring out how an API call works or even yeah, just working in the, in the up. So, yeah. And then I think another point that you just brought up, you know, this familiar familiarity of the GTM or ga ui, I think that’s also something that resonates quite, quite well because, ’cause you just brought up that I’m, that I’m lecturing as well and the feedback that I often get from, from students when I’m showing them stuff in the Google technology GA is that, that, that I’m moving too fast, really.

[01:08:21] Gunnar: Right? Because it’s just, for me, it’s natural to move around in the interface and if you just, the first time users see it, it, it all happens too fast and it’s just all a mess. And I dunno, why is this window popping up now when you click on this, what does it even mean? And these simple things, right?

[01:08:37] Gunnar: Yeah. So I think long story short, yes there will be, I think there will be a change definitely how we will interact with this truth. And again, it’s, I think it’s not only GTM and ga, it. Probably any, anything that is, yeah, API capable. 

[01:08:51] Matthew: To be cynical, to use one of my world famous analogies. It almost feels like we will just inevitably become dinosaurs because of that, because of our reliance on us.

[01:09:04] Matthew: Because you, like you think, I mean, last, the, my last birthday for example, the only birthday cards I got were for people over 70, but I mean, that everyone else has moved on. Yeah. There’s some people who still rely on that and that’s how they like to communicate. Yeah. Everyone else is just, I, I’ll send you a text.

[01:09:20] Matthew: Yeah. I can imagine things like shifting naturally and how old are, and SMS, but yeah, like we, we know how to do it quickly and easily and blah, blah, blah, blah. But people just start to Yeah. I guess this is your point, right? Dar, but grew up with that interface and interacting with things and.

[01:09:39] Gunnar: And they might be better than us navigating this interface. Right. And thereby having the Exactly, yeah. Actual efficiency gains. 

[01:09:45] Matthew: Yes. Sure. And ultimately, it’s going to come down to money, isn’t it, for a company that they can get the same outcome in the, roughly the same outcome for one, significantly cheaper.

[01:09:54] Matthew: ’cause it takes significantly less time. Yeah. The best will in the world, that’s the one that wins out. and yeah, like you say, maybe there is stuff up here that is, that’s hard, hard earned over many years of getting it in there. But it does feel a little bit like there’s a, there’s an ever, ever shrinking circle.

[01:10:12] Matthew: but that’s getting smaller and smaller and with the real meaty bit in the middle of expertise. But we don’t know. Maybe everything’s, maybe it’s finished now and it’s not, it’s going to move forward anymore. 

[01:10:23] Gunnar: Yeah. Yeah. But I think we’ll see if that’s going to be an interesting next couple of years. Right. And again, I, I think there’s no, there’s a lot of people putting their minds to it and trying to get it to work.

[01:10:35] Gunnar: So I think there is. We’ll see improvements and it becoming more accessible, more functional as well, easier to adapt. Similar to, you know, GTM servicer came out, everybody was clear, okay, this is really cool. This can really up your, your, your data, quality potentially, but it requires certain text code to get there.

[01:10:55] Gunnar: And then you had a player like state coming in, being like, okay, now you just spin it up within 50 minutes. You don’t need your data engineer anymore. You just, we do the heavy lifting for you. And I could imagine something similar happening to this space or GTM GA configuration. 

[01:11:11] Matthew: I always ask the same question for every podcast, and it seems ridiculous if I don’t ask it on this specific podcast, which is all about ai. So before I do that, I just didn’t know if there’s anything else that anyone wanted to cover or we, we think we’ve missed. 

[01:11:23] Gunnar: No, I don’t think so. No, not from my perspective. I don’t know if you have anything there. 

[01:11:27] Dara: No, I, I, we’ve covered all the, all the pointers I had, Dan, so. We’ve even kind of gone into some of the answers to your last question, Matthew.

[01:11:35] Matthew: So I ask this question every time, and I think we’ve maybe answered a lot of it. And, and so I’ll, you’ll have to stretch yourself a little bit to think of something we’ve not already said. But essentially with all everything that’s going on and, and particularly with AI and based around the conversation we’ve had today, what do you see the next sort of two, three years looking like?

[01:11:54] Matthew: What, what do you think is coming down the line that’s going to be transformative? Apart from everything, is there any, any, any sort of trend or, or thing over the horizon that you think is coming? If you had to look in your crystal ball? 

[01:12:07] Gunnar: I mean, what I would hope is that, let’s assume that we now have so much more time, right? Because all these mundane, nitty gritty, cumbersome tasks are outsourced to an artificial intelligence to some extent at least. That maybe we will have more time to. Use the data that we are collecting, right. And, actually, I dunno, activate it, analyze it and try to drive more business value from it.

[01:12:41] Gunnar: You know, there’s so many companies out there that they spend tons or like a lot of effort on getting their data collection as good as they can, but then at the end of the day, they just putting that data into a dashboard and, nobody’s looking at it and nothing is happening with it. Right. And that is also because again, like I think a lot of the technical resources are wasted on the data collected or wasted, but are spent on the data collection part and not so much on the activation part. So that is something that I would like to see, hope, hope to see. 

[01:13:12] Matthew: We’ll add that to our list of predictions and analysis. Enlightenment. Yes. That would be. 

[01:13:19] Dara: I’m also going to, I’m going to book the trenders slightly and ask a question after your last question. ’cause I just feel like I need to ask this. Oh, okay.

[01:13:26] Dara: Yeah, no, yeah. And I’ll actually ask both of you. So here’s my cynical question, my cynical moment. A lot of people say this, it’ll free up time and we can use that time for more meaningful work. Do you think, and this is to both of you, that there is a little bit of a human trait where we find things to busy ourselves with.

[01:13:45] Dara: So do you think we’re just going to find something new to get stuck on and never actually free up that time to do the meaningful Yeah. Valuable stuff? Of course. Yeah. 

[01:13:56] Gunnar: So we’re all cynical. Yeah. I mean, you know, like when I look at myself, right? So with the help of these lms, I’m, I’m more productive, right?

[01:14:06] Gunnar: And I wouldn’t necessarily say that I shifted my work to what, what I just described as being more meaningful. I’m just doing more of the cumbersome stuff, right? It’s, it’s just. More of what I used to do, not necessarily change my entire behavior. So yeah, maybe you have a point there, but again, I think maybe there’s a generation coming up that is, you know, just used to working that way and, or maybe, we’ll also be advisors. 

[01:14:33] Dara: May, maybe like Matthew said, maybe we are just going to be dinosaurs and, and we can’t see it.

[01:14:38] Dara: We can’t access our full potential, but the next generation will just be like, wow, I can’t believe you used to work 40 hours a week. Yeah. 

[01:14:45] Matthew: My big business idea that I’m now going to say to an international public, can I say that, once everything has been replaced by ai, I’m going to create a company that just generates busy work.

[01:14:58] Matthew: People will come into an office, they’ll sit down at a desk and it’ll just generate random tasks for them to complete, and they clock out and go home, and they pay me a fee to do that. Yeah, that’s my big idea. 

[01:15:08] Dara: It’s like a kind of weird wellness retreat of the future where we, where we, we miss this meaningless work and we just pay, it’s, it’s a bit black mirror.

[01:15:15] Dara: Yeah. I like it. I’ll, I’ll, I’ll sign up. I’ll be your first customer. Great. I’ll get a, I’ll get a, I’ll get a form up and start collecting. Brilliant. Okay. Gunner, thank you again for joining us. This is a really interesting conversation. I think it’s an interesting space. It’s going to be, I think we’re all keen to see how this, this plays out.

[01:15:32] Dara: Yeah. And what impacts things like MCP servers have on digital analytics and, and, and, you know, the wider world. But for now, thank you. Thanks for joining us. Thank you guys for having me. It was a pleasure. That’s it for this week’s episode of the Measure Pod. We hope you enjoyed it and picked up something useful along the way.

[01:15:50] Dara: If you haven’t already, make sure to subscribe on whatever platform you’re listening on so you don’t miss future episodes. 

[01:15:56] Matthew: And if you’re enjoying the show, we’d really appreciate it if you left us a quick review. It really helps more people discover the pod and keeps us motivated to bring back more. So thanks for listening, and we’ll catch you next time.

• Spotify
• YouTube
• RSS Feed