--- title: "What Agentic PPC Actually Looks Like in Practic" serpTitle: "What Agentic PPC Actually Looks Like in Practic" description: "Benjamin Wenner takes us through the process of using agentic AI for PPC operations and explains its use cases, challenges, and workflows." date: "2026-08-27" url: "https://www.optmyzr.com/ppctownhall/what-agentic-ppc-actually-looks-like-in-practice/" --- # What Agentic PPC Actually Looks Like in Practic > PPC Town Hall 130 Benjamin Wenner takes us through the process of using agentic AI for PPC operations and explains its use cases, challenges, and workflows. **Published:** August 27, 2026 **Watch:** [YouTube Video](https://www.youtube.com/watch?v=BTJZeabWKhM) **Apple Podcasts:** [Listen](https://podcasts.apple.com/us/podcast/what-agentic-ppc-actually-looks-like-in-practice-ppc-th-130/id1508399985?i=1000786321144) **Spotify:** [Listen](https://open.spotify.com/episode/6Thrzu7hruRkc5gVkezmKD) --- ## Episode Description Fred Vallaeys, CEO & Co-Founder of Optmyzr, sits down with Benjamin Wenner for his first appearance on PPC Town Hall. Benjamin is based in Berlin. He's worked in marketing for about 15 years, across agency and in-house roles, lead gen and e-commerce. For the past three years he's written a monthly column for Search Engine Land, mostly focused on AI, automation, MCPs, and what agentic workflows actually look like once you move past the demo stage. That last part is really what this episode is about. Benjamin has spent the last two to three years building and testing agent-based PPC systems on real accounts, not slide decks. Here's what was discussed in this chat: * Why direct access to Google Ads and Analytics data matters more than people think * Why giving an AI more data can make it less reliable, not more * Why Benjamin insists on read-only access before letting any agent make changes * What data structure and context have to do with whether an agent actually works * Why building agentic PPC systems turns out to be more of a data problem than an AI problem * What Benjamin's own setup looks like today, including what he's abandoned * Why he thinks most marketers only need a handful of core functions, not hundreds of them * Where human judgment still has to stay in the loop --- ## Episode Takeaways Benjamin Wenner has worked in marketing for about 15 years. For the last two or three, he's been deep in agent-based PPC, building systems, testing them on live accounts, and writing about what he finds for Search Engine Land. The opening line of the episode is Benjamin describing the "sitting on a beach while one agent runs a hundred accounts" pitch that circulates on social media, and immediately telling you that's not real. ### Why more data can make an agent worse, not better Benjamin's entry point into this topic wasn't AI models. It was MCP, the protocol that lets an AI connect directly to a data source instead of relying on someone to manually export and upload a file. He calls it a major breakthrough, and for a specific reason. Before MCP, the standard workflow was exporting a CSV, uploading it into a chat, and hoping the AI made sense of it. > "You end up having like 20 CSVs in your downloads folder, and you don't know what's the latest version," he said. The AI gets confused right along with you. MCP solves the access problem. It doesn't solve the next one, which turned out to be the harder problem: once an agent has access to a lot of data, more of it doesn't automatically help. Benjamin described building a system with 10 or 12 functions and watching it start to fall apart almost immediately. He'd built a logging feature specifically so the AI would explain its own reasoning before making a change, expecting that to catch mistakes early. Instead, he found the reasoning and the actual change frequently didn't match. "I'm quite certain this data set is saying this, so I must do that, and then it does something completely different," he said, describing what he saw in the logs. His read on why: an AI without a clear schema, without being told what a given data set is and what it's for, ends up drowning in its own inputs. Access wasn't the bottleneck. Structure was. ### Read-only first, always Benjamin's most concrete piece of advice in the episode is also the simplest. Anyone testing MCP connections or agent workflows on a real account should start in read-only mode, full stop. No exceptions for people who feel confident. > "You never want, especially if you're trying something out, to have an AI that is able to write things," Benjamin mentions. "And then you maybe say something wrong, and oops, your account is gone, or campaign is gone." Fred connected this to a pattern that's older than agentic PPC. Scripts had the same failure mode years ago: a script quietly making a bid change every hour until an account's spend spiraled somewhere nobody intended. Same risk, new wrapper. There's a real-world version of what happens when this discipline slips. Benjamin mentioned Meta banning a batch of accounts that had been using an unofficial third-party MCP connector, only for Meta to launch its own official MCP server shortly after. Fred talked about how Meta doesn't allow AI-generated ads, and heavy MCP usage looked like exactly that to their systems. The lesson isn't "don't use MCP." It's that a platform banning accounts for automated abuse doesn't care whether the automation was well-intentioned. ### The real work is a data problem, not an AI problem This is the part of the conversation that reframes everything else Benjamin talked about. He assumed, going in, that agentic PPC was mostly a matter of stitching some scripts together and letting an AI manage them. What he actually built looks more like a small data engineering project, with AI sitting on top of it. > "When I first started with the whole idea of building some sort of agentic PPC system, I thought it's just putting together some scripts with an AI managing them. But it's not. It's actually more of a database, cloud infrastructure project," Benjamin said. The reason comes back to something called a data schema, which just means telling the AI what a given piece of data is, where it lives, and what it's for, every time, rather than assuming it'll remember. Without that, Benjamin found the same data set, structured identically day to day, would get treated inconsistently by the AI depending on things like model updates he had no visibility into. His fix was building a function whose only job is preparing and explaining data before it ever reaches the agent doing the actual PPC work. He describes it, half-jokingly, as a data warehouse. The word sounds heavier than the reality, but the instinct behind it is correct: an agent that isn't told what it's looking at will eventually make something up instead. ### Ninety percent of the value comes from three or four functions Benjamin's own system has grown to roughly 140 individual functions over its lifetime. By his own estimate, about 30 of those actually work well enough to plug into a live project. The rest are experiments, half-finished ideas, or things built for a use case that came up once and never again. > Benjamin remarks, “There's a lot of dead weight in the whole system, which I think is completely normal if you build and test and try something new, and figure out along the way what's needed and what's not.” That gap is the point, not a flaw. Fred draws a direct line to how Optmyzr itself works: a big suite of tools where a small handful of core functions carry most of the daily usage, while dozens of others exist for the smaller number of people who genuinely need them. His framework for deciding what's core and what's optional is simple. Core functions are the three or four things that are non-negotiable for a given task, like search campaign creation, keyword research, and search term analysis for lead generation. Everything past that is what he calls a cherry on top: useful for some, unnecessary for most, and fine to build only when there's a real reason to. He treats budget pacing and bid management the same way. Both are things an agent could plausibly do, and both are things he keeps a human in the loop on anyway, because the downside of getting either wrong is high enough that speed isn't the priority. ### Why he doesn't think a technical background is optional Fred pushed on whether a solo marketer, without a team, could realistically run something like this. Benjamin didn't pretend otherwise. > "It's definitely not something that you just do on the side," Benjamin said. “You need to have dedicated skills, in the data space, or have at best like a data department, or you need to be willing to invest the time to learn these kinds of skills, also of course with help from AI, but you need to have that amount of time to invest,” he adds. Two or three spare hours a week isn't enough. What actually helps, in his experience, is having at least some technical grounding already: Python, JavaScript, BigQuery, the kind of background that lets someone pick up a framework like CrewAI or Google's Agent SDK without starting from zero. That's not the same as needing a developer background. Benjamin doesn't have one. But he's clear that skipping the learning curve entirely isn't realistic if the goal is a system that holds up on a live account rather than a demo. ### The tradeoff between building it yourself and just using what exists Both Fred and Benjamin admit to the same habit: spending fifteen minutes automating something that would have taken two minutes to just do by hand, because building the automation feels like the more interesting problem. For years, that instinct paid off, because a good custom build kept working for a long time without much upkeep. Benjamin's read on the current moment is that this math has changed. > "What I vibe code, and what I make more efficient, can literally be solved by Claude Cowork," Benjamin said, meaning the tooling is catching up to custom builds fast enough that yesterday's clever workaround is sometimes just a built-in feature today. His answer to where the balance sits isn't a formula. It's an acceptance that some invested time won't pay off, and that this is fine. > "To a certain amount, it's okay to waste a certain amount of time on projects, because that's how you discover new things, learn new things, that's how you build a product. Not every hour spent is spent efficiently when you build something new," Benjamin said. The alternative, waiting until a framework or a feature is perfect before ever building anything custom, isn't really an option in a space moving this fast. --- ## Episode Transcript **Benjamin Wenner:** "Okay, I'm just sitting on a beach and one agent manages 100 accounts." Isn't like something that some people try to sell you often on social media, but if you're really getting into it, and if you're really trying to develop solutions, and also you test these solutions with real accounts, or clients, or projects, you quickly realize that it's not that easy. **Frederick Vallaeys:** Hello and welcome to another episode of PPC Town Hall. My name is Fred Vallaeys. I'm your host. I'm also the CEO and co-founder at Optmyzr, a PPC management software suite that's AI enhanced. So for today's episode, we're bringing in a first-time guest for the show, Benjamin Wenner. He is based in Germany. He's been writing for Search Engine Land for the past three years, one article a month. So, 33 amazing pieces of thought that he's put out there. And a lot of it has to do with, of course, AI, automation, skills, MCPs, all of the things that are kind of top of mind for a lot of PPC practitioners these days. So we wanted to bring him in today, kind of go behind the scenes, and understand how he thinks about all of these technologies, and how it makes him more efficient, and what it can mean for you. So with that, let's get rolling with this episode of PPC Town Hall. Benjamin, welcome to the show. It's good to have you on. **Benjamin Wenner:** Thanks for having me, Fred. It's a pleasure to be here. One of the, I would say, best PPC podcasts. So, yeah, pleasure's on my side. **Frederick Vallaeys:** Yeah. No, thank you for coming on. And so, you also have the great honor of now being a top 10 PPC expert. And I think a lot of that has to do with all of the great stuff that you've been sharing, right? So tell us a little bit about your background, where you're based, and what you mostly work on. **Benjamin Wenner:** Yeah, so right now we're based in Berlin, Germany. I was also born here, and yeah, I live here with my girlfriend now. I've worked in marketing for a long time, around about 15 years now. So basically my whole career, and I've done everything. Classic account work, consulting, growth hacking, supporting other founders growing their companies, lead gen and e-commerce, small clients, big clients. And more recently, I would say for the past two, three years, heavily invested into agents. It's quite a new topic, especially for the PPC community, but that's what makes it so exciting. It's a really fast-paced, fast-moving environment. There's still a lot of potential, and I see personally a lot of possible uplift there for the PPC community. That's why I'm going into it so much lately. We're still not there yet of having the perfect agent system that does everything for you, but I'm quite positive about the development. And also, if you look at other industries, the whole agent space is moving so fast, and it's only, in my opinion, a matter of time until we have some real breakthroughs. **Frederick Vallaeys:** Yeah. No, absolutely. And so it's new to our industry, but I think it's new to everyone, right? So it's really great to have you on and share a little bit about your progress. And now, you've been doing PPC, you said, for 15 years, and then the last couple of years, since that magical day in October of 2022 when ChatGPT was made visible to the world, things have changed, right? But you've been writing for Search Engine Land for about three years, which is fairly close to when that introduction of ChatGPT was. But talk a little bit about maybe how your thinking has changed from that first day of an AI answering bot, to more of the agentic future that we see today. And also keeping in mind that, like you said, we're not quite there yet with the full potential realized, but a lot has changed. So what are the biggest changes that you've had in your way of thinking about all of this in the past, say, three years? **Benjamin Wenner:** So, when ChatGPT came out, I looked more into the SEO space, but also in the developer space, like all these indie hackers, etc. They share a lot of stuff on X and also on LinkedIn, and I looked over to them, and I always saw, okay, the SEOs, they're really, really fast with ChatGPT. Basically a day after it came out, they already had these crazy ideas and strategies, hey, we can do this and that. Same goes for the web developing space, or the indie hacking space. There were all these people, often also going solo, and they had these great ideas how to start new projects or how to fill projects with content, etc. I was like, okay, there's so much going on, but for the PPC space, it's been really slow. The best that we had, from my view, was for the first six months, that we were going crazy about AI writing some ad headlines for us or so. So this was really a little bit underwhelming. And I figured, okay, there must be more to it, but it took some time. I would say other industries, or also other branches in marketing, like maybe content marketing or creative marketing, they were really, really quick in adopting AI solutions. Not only with ChatGPT, but also going forward with Claude or Gemini, or whatever other model you want to test or try out. And my idea was, okay, after a while I figured, okay, I can do more and more, and there must be a way to get more leverage out of it. So it's not only having these headlines, or maybe write a description, but also maybe analyze some more data. And that's where I figured, okay, there must be a way, or also with future development of these products, there must be a way where we can have agents that support our daily work to a certain amount. So it's more like, when you're really good at something, how can you be more efficient? This is where agentic PPC, in my opinion, comes into play. Giving A players even a bigger advantage with agents. **Frederick Vallaeys:** Yeah, I guess I call it the AI-amplified marketer in my book. It's how you take yourself and scale yourself, and do all of the things that you've often been wanting to do but maybe didn't have the particular skill, didn't have the time. So it's quite fascinating. But talk to us a little bit about that struggle between SEOs adopting so quickly. And I think part of that was, and maybe I'm oversimplifying, but in SEO, you basically write landing pages, write content. You don't really need a lot of hidden data for that. You can often look at what is already on a website and use that as a starting point. I think the struggle for PPC managers was that we have a lot of data in GA4 and we have a lot of data in Google Ads, and it's all sort of locked in, we can't get it out. And yes, now we can maybe get a CSV, put that in with the agent, but then the agent runs out of memory context. So maybe that's part of the reason why things evolved more slowly. But talk to us about the whole data spectrum, and what do you need to go beyond just writing better headlines with AI? How do you get it to be an agent that's just as informed as someone who works in your company? **Benjamin Wenner:** Yeah, so about the data thing, this is also, I would say, a really huge topic. I think MCP, being able to connect the data source with AI, this was a major breakthrough in my opinion. Because the whole concept of having to manually send CSVs back and forth, there's no real benefit to it. Then you end up having like 20 CSVs in your downloads folder, and you don't know what's the latest version. AI also gets confused, and in the end, where's the real benefit of using AI, right? So I think many people underestimate the impact of MCP, and I would say MCP is still underused. I think a month ago or so I did a workshop with a couple of people, and out of 90 participants, 10 had ever heard of MCP, and like eight or so had implemented it. Two or three among them were agency founders, so that's also good, because they are multiplicators for many accounts. But I just realized that for so many in the PPC industry, it's still the CSV file upload, and then something back, and you go on from there. But this is not really where AI shines, and it's also not really time-efficient. Maybe you save a couple of minutes, but is this really boosting your workflow, or is this where you really benefit from it? I don't think so. But still, it's something. If you have no starting point, using a CSV upload and saying, hey, this is my business, this is my context, can you analyze this data set for me, it's still better than having nothing at all. So don't get me wrong, I think we all need some sort of starting point. And especially we should not forget, quite often about these AI topics, it's a lot of senior people talking. But marketing is one of the biggest industries that lives from juniors joining the industry, coming, learning, often in agencies, and trying to build knowledge. So we should not forget that these people also need a starting point. But if we look into how AI can really be a benefit for you, you need to have some sort of data infrastructure. MCP is one way of having this kind of data infrastructure, and getting data into AI maybe directly, maybe more in a live fashion, and then having a chat about it, or a more continuous conversation. But there are also other ways. **Frederick Vallaeys:** So let's pause on that for a second, right? MCP, model context protocol. It's kind of like an API for agents. It actually is an API, but with more specification, so the agent can understand what to use when. And the reason a lot of people probably don't really know this term is that Claude just now really calls it connectors, right? So they give it a non-technical name, but on the back end, that's an MCP. Now, that's part of the data infrastructure, right? But what an MCP often allows you to do is go into your existing data infrastructure, but make that available to the agent. So in our case, we would have the Zoho CRM. We have, within Optmyzr, a connector into the Google Analytics data. So these are all the existing data infrastructure sources. So two questions for you. First one, which MCPs do you really love and use on a daily basis? And then second, we should also talk about data infrastructure, and if you've changed your data infrastructure now that you've got agents consuming it rather than humans. So, favorite MCPs first. **Benjamin Wenner:** Favorite MCPs for my day-to-day work is just Google Ads, Analytics, and, yeah— **Frederick Vallaeys:** Also, let me pause you on that too, because the Google Ads and the Google Analytics MCP, not the easiest to install, right? These are really developer MCPs. I looked at the Google Analytics MCP, and I know how to develop, and I was like, I just don't have time for this. So my team built it within the Optmyzr MCP. But talk a little bit through what do you need to be able to do to legitimately set this up and consume it. **Benjamin Wenner:** Yeah, so what I did before having the native integration was, I used the free one from TrueClicks. I think it's a Dutch agency or tool provider or so. Yeah, and they had a really nice way, they were also one of the first to have this free connector, basic connector. It worked quite well, and I used it for quite a while. Only after figuring out, okay, I want things to be more customized, and also I need to get into it, because if I'm talking about a topic, I need to understand how the native one also works, then I got into the native Google MCP connector. That took me a while, I think, maybe one and a half hours overall, after reading everything three times, and creating all the needed tech infrastructure. So it's not that easy, but I would say, from a starting point perspective, there are actually many, many options right now. So I would even say, do not go with the native ones for Google, because the Google one is a little bit more complex, especially if you have no tech background, or if you already struggle with things like ad scripts. There's no need to get even more complicated with something like this. So in that case, I would say go to one of the tool providers. Optmyzr also has an MCP, so there's that. Then there are a couple of low-cost or even free MCP providers, often connected with tools or some sort of agencies, which are perfectly fine to test and try out for basic stuff. The native Google stuff, or analytics stuff, is really for if you want things to be more customized, if you want to understand more of the logic behind it, and if you want to have full control over what the MCP actually does for you, because there are also limits to it, like a simple version or full version, or what gets done or what not. So there's really no need, for a starting point, to have the native integration. But things change when you look at different platforms. I'm not doing social ads every day, but from what I heard, Meta actually banned a couple of accounts who used a third-party MCP, and then days later they published their own official MCP server. So it's a bit tricky, right? Officially there was no official MCP server first from Meta, and then when a couple of solutions got really popular, they launched their own server, and then they banned a couple of accounts. So it's a little bit of a bad taste, I don't know, you never know what. **Frederick Vallaeys:** And I think wasn't the banning in that case mostly driven by the fact that Meta doesn't allow AI generating the ads, and so it was sort of assuming, if you're using an MCP, it's probably doing all this. Because APIs are formally or officially supported, right? And MCP is just a flavor of API. But I think it raises a good point. So if you're using MCPs and agents to create slop on your ads, whether it's Meta or Google, you're going to run into policy violations, and you're going to keep submitting the same disapproved things, and eventually the platform is going to give you a strike, and maybe disapprove or disable your account. So a good word of caution when you start using these automations, because just like with scripts, right? We've seen people set up scripts that make a bid change every hour, and eventually their bids just go crazy too. **Benjamin Wenner:** (laughs) Exactly. I remember also there was something on LinkedIn, a screenshot going around, but I forgot who posted it originally. It turned out it was about an account where this person was massively abusing the API or MCP connector, with like thousands of changes. Obviously that's never in the interest of anybody, or of the platform. So I would say, if you're using the official MCP server with normal human usage, it should be perfectly fine. **Frederick Vallaeys:** And the other angle here is the guardrail. So if you go with the Google native MCP, it's basically a connector into the Google Ads query language, and you can pull reports and get all the data you need, and do some analysis based on that. But I don't think it has a lot of change-making capabilities, at least quite yet. But that's where you also start thinking about what are the protections. A sophisticated MCP will have rules, and will prevent the agent from making stupid changes, or at least tell the agent, hey, you might be doing this too frequently, or this keyword addition looks like it could be conflicting with this, have you considered these scenarios. And that's often where you get the benefit of going with someone who's actually put business logic inside their MCP, rather than it just being a pure, lowest-level data access mechanism. **Benjamin Wenner:** Yeah, there are a lot of pitfalls, actually. First of all, it's a tech product, and PPC is somewhat more on the tech side of marketing, but it's still marketing. So there are many technical pitfalls, as you already mentioned. Definitely you should always proceed with caution. What I always say to people who are new to MCP, or want to play around, always make sure, for example, if you use a third-party provider, it's only in read mode. Because you never want, especially if you're trying something out, to have an AI that's able to write things, and then you say something wrong, and oops, your account is gone, or your campaign is gone. That's the worst case. Obviously it doesn't happen too often, but you can never be, especially with live accounts, cautious enough, because the risk is too high to mess something up. So always play around in read mode, and work with analyzing things, with data reporting. And if you then feel really comfortable, you can go into the write section, if you really want to work on something more advanced. But at that stage, you're also on a level where you know the risk, and can have a better feeling about it. **Frederick Vallaeys:** I would say exactly. It's no different than all of the other types of automations that people do. Take a look at what it suggests first, take a couple of days to look at what the new suggestions are every day, and then if you feel like it's really getting what you want it to do, that's when you set it loose. But then it also runs into, I think, the same issues we had with scripts. So now you have, say, 10 or 15 amazing scripts managing your account, but you need to copy it into all of the accounts that you manage, and maybe have the MCP capability, but then one of the scripts times out, and now you realize this budget-checking script wasn't actually checking budgets. And so it becomes this whole infrastructure question around, how do you stay on top of this? I think it's the same with agents. As you deploy various agents that do different tasks, how do you make sure that they stay active, that they don't start hallucinating, or even the thing where we've seen the memory start to compact, and your initial very important instructions become lost, and all of a sudden it starts behaving differently because it's exceeded its memory allotment. So how do you think about, and now we're talking about the data infrastructure, how do you make sure that this new influx of data doesn't break things? **Benjamin Wenner:** Yeah, it's a fantastic question. Actually, I think it's one of the most important parts in agentic PPC. So maybe let's get back to where I started. I started trying to create little functions, putting them together with a managing agent, and then building more and more functions around it. What I quickly realized is, after adding like 10 or 12 functions, the whole system kind of slips away quite fast. You just look left, and then look into the account again, and the AI is going crazy. Hallucinations, timeouts, and also logging errors are really happening often. For example, what I built quite early on was a logging function, so the AI should write a reasoning log. I see this data, therefore I do that. And then I looked into the change history and compared these things, and what I quickly realized is that the reasoning logic sometimes does not match at all with the changes. I'm quite certain this data set is saying this, so I must do that, and then it does something completely different. Especially in the early days, I was completely overwhelmed. This was also my first big lesson, and I stopped a lot of testing right there and got back to the drawing board, because I realized there's so much data floating around, and AI now has access to so much data, that it's also overwhelming for AI. AI is a powerful tool right now, but if you have no data infrastructure, if you have no data schema, then you're just dumping a lot of trash into AI, and hopefully AI will figure it out. But it's not how it works. **Frederick Vallaeys:** So what does it mean to have a data schema? **Benjamin Wenner:** So data schema, in this sense, means that you need to pre-prepare data for AI, and also explain this data set to AI. What I first thought initially is, you have some sort of connector, maybe MCP, or you have a drive link where you store maybe daily data sets, and AI can just access this and you're good to go. But it's not the case. It might work for some small-scale analysis, but if you want a scaled operation, with many things going on in a day, then you need to explain certain data sets. This is stored here, this is the purpose of this data set, this is how the data set looks. Because what I also figured out, sometimes there was 10 times the same data set, but AI kept working differently on it, although the data had a different value based on the day, but the schema was kind of the same. I was like, okay, why is it treating it differently right now? Maybe it also had to do with models changing, and the big AI providers constantly doing tests, and now this works this way, and now give shorter answers here, and that messes with the whole logic. But I realized you need to implement a function to always prepare this data for AI, to tell it, this is that, this is here, this is used for this purpose, so it always has this context in mind. Otherwise, if you just hope it remembers the context, it loosens over time, and completely drifts away. **Frederick Vallaeys:** Right. So I think here we're now getting into, okay, we've explained MCPs, which is where the data gets into your agent. Now you're starting to get into what are the skills, or the definitions of how the schema is treated, how the data is handled, how you operate on that data. Is that kind of what we're talking about now? **Benjamin Wenner:** Yeah. Yeah, this is quite important, because when I first started with the whole idea of building some sort of agentic PPC system, I thought it's just putting together some scripts with an AI managing them. But it's not. It's actually more of a database, cloud infrastructure project. You have this whole layer before AI. Data warehouse sounds a little bit huge, but it's a good word. You need this, to have kind of like a data warehouse, to have this data prepared and put together for your AI to even make sense of it. Otherwise, if you just let loose 10 MCP servers, it gets overwhelmed. And then you need to have an idea about the data pipeline, or data stream. How do you make sure the data comes to AI in a way that makes sense? You need a certain schema, to keep it as simple as possible, but still as informationally rich as you need, to have all the information to make an analysis. This was maybe in my second year, that I realized this is much more data-warehouse, database-based, cloud-based than I originally thought. My original thought was really, oh, it's just a couple of scripts and AI will manage them, it can't be that hard. This is also, I think, why agentic PPC is being talked about a lot, but if you look at how many people are sharing real insights, you have maybe 10, 20, 30 people maximum. You quickly realize the whole thing is really a massive hype. I know maybe 15 accounts on LinkedIn that are actively sharing real development in the space, and maybe 200 accounts posting every day about some new stuff, but it's not insight. **Frederick Vallaeys:** (laughs) Yeah. No, and that's what sells on social, right, is that vision. **Benjamin Wenner:** (laughs) Yeah. Yeah. It's also true. **Frederick Vallaeys:** And listen, the more I hear you talk about this, it's like, well, because I was in your shoes, I think, when I was doing scripts a while back, before I started Optmyzr. I love writing scripts, I love writing these automations, and all the technology is there to do something truly amazing. But if you need to marry that together with actually managing accounts, and doing the work of talking to your client, and not just constantly updating scripts, there's just not enough time in the day, right? So, kind of curious, from your perspective, can you have a work-life balance? Do you need to hire someone onto a team to manage this? Is this for individuals? What's the real level of effort? Can someone manage accounts more efficiently as a solo freelancer? Does AI make you able to do more, and better? **Benjamin Wenner:** Yes. Yeah, so it's a good question, but it's also one of the biggest pain points, so I have no reason to lie about it. It's more complex. I know that a lot of people even struggle with having more than one script. And then they struggle with connecting maybe BigQuery to Looker Studio. So this is even more complex. It's definitely not something you just do on the side. Let's say you're an in-house marketer, and you have two or three hours to spare, it's not something you just do. You need to have dedicated skills, in the data space, or have at best like a data department, or you need to be willing to invest the time to learn these kinds of skills, also of course with help from AI, but you need to have that amount of time to invest. **Frederick Vallaeys:** And to give people context on what that means, do you have a technical background, or did you just learn this doing so many years of PPC, and being so into the data? **Benjamin Wenner:** Yeah, so I have no technical background, like a developer background, at all. But I have to admit I'm more on the technical side. Years ago I got a little bit into web development, then started with some Python scripts, and then also learning JavaScript, a little bit, alongside other marketing aspects. If you're more on the techy side of PPC, it's much easier than if you're on the creative side. If you're strong with YouTube, upper-funnel formats, creative generation, video ideas, that's a different thing than being on the tech PPC side. I would say tech PPC is also interconnected with tech SEO. They have some tech skills, maybe not a pure developer background, but they can grasp technical concepts much more easily. These sort of people can learn the concepts of BigQuery and so on in a day. For these kinds of people who are more techy, it's kind of easy to jump on it, and then it's only a matter of personal investment to get a full picture of it. And of course there are plenty of tutorials out there, but certain frameworks also make it easier. For example, I started with a framework called CrewAI, if I remember correctly, like three years ago. They made it really easy to connect different sources into one project. Later I jumped on the Google Agent SDK, which is also an agentic framework, and it's not only for PPC, it has many purposes. The thing is, using a framework takes away a lot of programming and customization. You don't need to worry about certain aspects, like connecting BigQuery and all these things. Of course you could kind of vibe code it and get behind it, but it would take hours and hours. Using a framework makes it much easier. Maybe a developer would say, no, I need to customize everything. But you also need to think about it like, okay, I'm trying to build something, hopefully it will make my life easier, or maybe the life of others easier. And then you need to think about how much time you want to invest in it. It's not worth it if you spend a full day on something and see no progress. So you need to keep the balance. But it's hard, because these kinds of projects are really fast-paced. **Frederick Vallaeys:** Yeah, but you need to keep a balance. You need to say, okay, stop for now, I'm not following this part now, I'll keep it more simple. **Benjamin Wenner:** Yeah. And one thing I struggle with is I like to vibe code. I like to build solutions that scale, and I think in the past that worked well, because I could take the time. So instead of something very simple, like answering an email, I could be like, oh, how could AI help me answer this email? And rather than two minutes to answer the email, I spend 15 minutes building that part of my automation. (laughs) **Frederick Vallaeys:** I keep doing that. **Benjamin Wenner:** But the beauty in the past was, this would scale. What I built then would be something we could use for the next 10 years without huge technical change. Now things are changing so quickly that what I vibe code, and what I make more efficient, can literally be solved by Claude Cowork. Figuring out where that balance is, between how much I invest in doing it in a custom way, versus how much I rely on frameworks, versus how much I just do it the old manual way, that question is really difficult for people to answer now. **Frederick Vallaeys:** So do you have an opinion on what that balance is that people should strive to hit? **Benjamin Wenner:** Yeah, so I think you will always hit a point where you say, okay, damn, these last two, three hours were for nothing, it was a dead-end street and time is wasted. But to a certain amount, it's okay to waste a certain amount of time on projects, because that's how you discover new things, learn new things, that's how you build a product. Not every hour spent is spent efficiently when you build something new. I was also talking to a couple of people trying to figure things out and develop solutions in the agentic PPC space, and they had a similar opinion. If it's new for you, and you have so many ideas, you really need to think, okay, let's say I start with an agent, and I try to build a system around it, which functionalities do I want. And then you want a hundred functionalities. Keyword research, negative keywords, shopping segmentation, video creation, whatever, you could name countless functions. But then the question is, how much do you use each function. You know this from Optmyzr, right, it's quite a big suite of tools, but I would say maybe three or four functionalities, 90% of your user base is using them. Then you have maybe 10 functions with a really small user base, but you still need these kinds of functions, to have a competitive edge, or because someone needs to offer these kinds of solutions, even though the majority won't use them, but some will. This is where you need to make a decision. You always have this kind of core functionality, which for lead generation would be search campaign creation, keyword research, search term analysis. These are the core functions. Then there are other functions that are just nice to have. Budget pacing could still be more human in the loop, you don't need an agent for that, but it's nice to have a function for budget pacing or monitoring, though it's not necessarily needed. Or bid management, someone who adjusts the target CPA, it's still good to have a human in the loop there, although maybe agents could also do it. But this is the optional functionality. For the optionals, there's no limit, there are countless optional functions, and this is where you decide how much you want to invest. The core functionalities are usually three, four, five per search, per shopping, per video, per dimension. These are the non-negotiables. Everything else is just a cherry on top. **Frederick Vallaeys:** I would say sometimes you throw away the whole article and just start new, because you think something better will come out of it, right? **Benjamin Wenner:** Exactly. This is how it goes. **Frederick Vallaeys:** You've talked a lot about agents, and I want to make this real for people. Tell me, what is your tech stack? Are you Claude, ChatGPT, different models? And when you say agent, what are you really talking about? **Benjamin Wenner:** Okay, so my tech stack has been, and still is, Claude, because from my perspective, Claude still gives the best results. I think at Hero Conference, when I had a little talking session there, I also said you need to look out, because for tech projects, Claude is the number one right now, and it was also a year or two years ago, I would say. But that doesn't mean Claude will be the number one for the next years. What I'm seeing right now is, especially in the last few weeks, I don't know if it's a European issue or a global issue, but I see Claude working more and more against my interests, I have the feeling. And talking to other people here in Europe, they mention the same issue, but I don't know if it's a European thing, because we have tons of new AI regulations, maybe you've heard of them, and you need to account for that with everything AI-related. Claude is also doing watermarks now, and sometimes Claude mentions things like, no, I can't help you with this topic, but maybe a year ago it was perfectly fine. I figured something changed with Claude. It still gives me the best results, but with a lot more resistance. **Frederick Vallaeys:** Okay, so I'm thinking companies... **Benjamin Wenner:** Yeah, it's, I don't know, I'm still thinking it's maybe a European thing. **Frederick Vallaeys:** So, sorry, when you're using Claude Cowork, Claude Code, do you have like a Mac Mini setup that runs it constantly? What does that look like for you? **Benjamin Wenner:** I'm using Claude Code mainly. I was thinking about this Mac Mini setup, it was really a hype for quite a while, but I decided against it. I'm just using the classic vibe coding way. I think the simple Claude Code setup was more than enough for me. I try to keep things simple, not to have too many sources and tools and applications running at the same time. It just gets more and more complicated if you have so many things on your plate, and then you have to manage subscriptions and functions and connections. I try to simplify as much as possible. If we talk about the agent right now, what am I actually developing, what is agentic PPC for me, it's basically a system. The version I'm working on now is based on the Google Agent SDK, so a framework, and it's basically an orchestrator agent sitting in the middle, which I feed all the context it needs, like the business background, margins, goals. Then I connect a data layer to it, often a Shopify data layer, depending on the project. And then I have certain functionalities around it. Over the lifetime, everything included, it's maybe roughly 140 functionalities. But it's not like each one is built and working fine and we're good to go. I'd say out of these 140, 30 are really working well, and you can plug and play them if you want to use them for a project. The rest is more testing, experimental. Some are just sitting there because there was a use case once or twice, but it never came back again. There's a lot of dead weight in the whole system, which I think is completely normal if you build and test and try something new, and figure out along the way what's needed and what's not. **Frederick Vallaeys:** Great. Hey, well, thank you for taking us behind the curtain and showing the reality of running agentic PPC. It's been fascinating for me, and I think we should probably get you back on in a couple of months and see what the latest is, and maybe explore some of these specific solutions a bit more deeply. Now, you do speak at a number of events, and you write for Search Engine Land on a pretty consistent basis, but where can people come and meet you in person? And which events are you going to be at? **Benjamin Wenner:** I think this year, let me think, at ADworld Experience, in Bologna, I'll be there. I think it's end of October. Besides that, I don't know, this year probably maybe some German events, one or two local events, but for events, I think that's about it, because the year was already pretty intense, and you've got to step back at some point and work on things, and not just go to events. But if people have questions, the fastest way is usually my LinkedIn, and then obviously my website, you can contact me there. LinkedIn is usually, not every day, but almost every day, so for any questions or feedback. Some people also send me concept work and want a review, and I'm more than happy to give my opinion on that. I'm definitely not the expert in this space, because it's just a new space. Anyone who claims he's the expert in this space, I think, is lying, because it's so new and experimental. But I'm more than happy to give feedback or my opinion if someone has an agentic setup they want to spar with me on. **Frederick Vallaeys:** Yeah, I suppose there's not like a full 100% expert, but there's certainly, like you said, 10, 20, 30 people who've really played with this a lot, who can probably help folks out. The website, benjaminwenner.com, is that one? **Benjamin Wenner:** That's it. I'm not really active with my blog posts there, but that's what Search Engine Land is for, they get all my blog content. But you can just contact me there. **Frederick Vallaeys:** Yeah, and LinkedIn is always the best, for myself and for Benjamin. So, great. Hey, thank you so much for being on the show. Thanks everyone for watching. I hope you learned a couple of new things on this one, and are going to experiment more with AI and PPC. If you did enjoy this, please subscribe so you know about the next one. Put anything in the comments so we can have a conversation around it. With that, we'll wrap it up here, and we'll see you for the next episode of PPC Town Hall. --- *Source: [What Agentic PPC Actually Looks Like in Practic](https://www.optmyzr.com/ppctownhall/what-agentic-ppc-actually-looks-like-in-practice/)* *© Optmyzr. All rights reserved.*