How Andrew Lolk Uses AI to Analyze PPC Data, Find Patterns, and Make Better Decisions

Episode Description

Fred Vallaeys, CEO & Co-Founder of Optmyzr and author of The AI Amplified Marketer, sits down with Andrew Lolk, founder of Savvy Revenue, to talk about where AI actually earns its place in a serious PPC agency and where it doesn’t.

Andrew is one of the most respected voices in paid search. His agency runs sophisticated feed-based search campaign structures where keyword management is largely automated. So when it comes to AI, he’s not looking for help managing campaigns. He’s already past that.

What he’s found instead is something more valuable: a personal analyst available at any hour, capable of distilling large datasets in minutes and surfacing patterns that used to require a junior hire, a pile of spreadsheets, and a few hours.

If you want a grounded, experienced perspective on where AI creates real leverage in PPC agency work, this is a good one.

Here’s what was discussed in this chat:

• How Andrew uses AI to analyze PPC data faster
• Why AI is better at analysis than campaign management
• The surprising value of uploading CSV exports into Claude
• Why Andrew tells AI not to infer, conclude, or guess
• How AI is helping Savvy Revenue manage SOPs and internal operations
• The opportunities and risks of AI agents and MCPs
• Andrew’s vision for a future without keyword match types
• Practical ways PPC marketers can use AI today without giving up control

Episode Takeaways

Andrew Lolk is not someone who gets excited about AI for its own sake. He runs Savvy Revenue, one of the more technically sophisticated PPC agencies operating today, and his baseline is already high. Search campaigns are built on feed-based structures. Keywords are added and removed dynamically. New brands and categories roll in automatically.

There isn’t much room for AI to come in and announce that it’s optimizing things, because the optimization is already largely handled.

What Andrew has found is something different and more useful: AI as a thinking partner and analytical layer that replaces hours of spreadsheet work, answers operational questions instantly, and maintains the institutional knowledge of an agency without requiring someone to manually curate it.

The conversation with Fred covers all of that, plus the risks Andrew sees that most people in the industry aren’t talking about loudly enough.

How Andrew uses AI to analyze PPC data faster

The workflow Andrew describes is deliberately unglamorous, which is exactly why it works. Instead of prompting Claude to pull data from Google Ads directly, which breaks down as accounts get large and complex, he downloads a CSV with precisely the columns he wants, then uploads that file and starts asking questions.

Andrew explains, “By uploading the CSV file and asking it to distill the data, find some different visualizations of that data, I’ve sometimes just gotten to a much better point in what I present to clients, but also just a better understanding of what’s going on in an account by doing that.”

Claude writes Python under the hood to aggregate and analyze the data. What used to require 45 minutes of VLOOKUP and pivot table work now takes a few targeted questions. Andrew gave a specific example: a large European account where he needed to understand cross-channel performance. He ingested Google Ads data, Meta Ads data, client revenue figures, and Klaviyo data together.

One chart came back showing that whenever he reduced Google Ads spend, the client increased Meta spend in almost perfect correlation. A media mix model, produced without intentionally building one.

Instead of asking Claude to audit the account, he asks whether the data supports a specific hypothesis, or which segment across ten accounts is dragging on year-over-year performance. That framing produces something useful. An open-ended audit request produces noise.

Why AI is better at analysis than campaign management

Andrew is explicit that he doesn’t use Claude to edit accounts or build campaigns. For the type of work Savvy Revenue does, that’s not where AI adds value. The accounts are already running sophisticated automation. Asking an AI to optimize on top of that creates more risk than opportunity.

The analysis layer is different. It doesn’t touch anything. It just helps you see things faster and more clearly. And it’s available at any hour, which matters when you run a distributed team.

“For me, the challenge almost becomes I’ve got two copies of Claude, a personal and a work one. And then within Claude, you’ve got the three ways of working with it. Which one did I ask this question? Like, where do I need to go to get that answer?” says Andrew,

Managing AI assistants is starting to resemble managing people. You need to know where each one has context, what it knows, and what it doesn’t. That’s a new kind of operational problem, but it’s a good problem to have. The alternative was either hiring someone or doing the work yourself.

The surprising value of uploading CSV exports into Claude

This is the most practical and transferable piece of advice in the conversation, and it runs counter to how most people think about AI and PPC data. The instinct is to connect Claude directly to Google Ads via API or MCP and let it pull whatever it needs. Andrew’s experience is that this falls apart on large accounts.

The CSV approach works because you control exactly what data goes in. You define the columns, the date ranges, the segmentation. Claude doesn’t have to figure out what to pull. It just has to work with what’s there. And on large datasets, only Claude currently handles this well. Other models get confused or hit limits.

The same logic applies to the questions you ask. Specific, hypothesis-driven questions get specific, useful answers. Andrew treats the AI like a sharp analyst who is great at finding patterns but has no business context whatsoever. The business context is the human’s job.

Why Andrew tells AI not to infer, conclude, or guess

This is the instruction that makes everything else work. Without it, Claude will read a dataset showing declining volume and increasing ROAS and conclude that the account’s strategy of tightening up and becoming more profitable is working. If the actual goal was growth, that conclusion is not just wrong but potentially damaging if it shapes a client recommendation.

“I ask it not to conclude anything. All the skills we have in Claude are very specific: do not conclude, do not infer, do not guess,” says Andrew.

The AI has no way of knowing what the account’s strategic goals are unless you tell it. And even then, drawing conclusions from data requires business context that lives in the minds of the people running the account. Andrew wants the data distilled and visualized. He’ll draw the conclusions himself. This is the right division of labor.

It’s a simple instruction but it takes discipline to enforce consistently. The temptation is to let the AI synthesize everything, which saves time right up until it produces a confident-sounding recommendation based on a misreading of intent.

How AI is helping Savvy Revenue manage SOPs and internal operations

Savvy Revenue built its internal operating system on Notion four years ago. Today that system runs with AI agents embedded throughout it, and the results are substantial.

Every SOP has a small AI-specific note below it. Notion agents can read all client emails, Slack conversations, project notes, and meeting records. During quarterly planning, Andrew asked the agent to go through every YouTube video he’d published, every completed client project, and every project outcome, then compile a database of what worked under which circumstances and what didn’t.

That database now gives the team live recommendations: if an account looks like this, this type of project tends to succeed.

The SOP maintenance system is similarly streamlined. Team members used to have to click a button, fill out a form, and submit a request when something in an SOP was unclear or outdated. Now they just comment directly on the page, the same way you’d comment in a Google Doc.

An AI agent runs monthly, reviews all comments, fixes what it can within defined limits, pulls in relevant Slack history for additional context, and packages everything it can’t resolve on its own for Andrew’s operations partner to review.

“The amount of work it takes away from running an agency, not from a Google Ads point of view, is unbelievable,” explains Andrew. “And it’s missing the entire context of what is going on here and how do you like to do things. And I think it’s a huge advantage to do content because you share your thinking, and that can be converted into SOPs.”

The agency operations use case is more mature than the Google Ads use case right now, and Andrew is straightforward about that. The documentation, the scheduling, the institutional knowledge management; that’s where AI is delivering real value today.

The opportunities and risks of AI agents and MCPs

Andrew uses Claude’s ecosystem extensively. He uses Superhuman for email specifically because it has a well-built Claude MCP that connects more cleanly to Gmail than the native Gmail MCP. He has both a personal and a work Claude account. He’s thought carefully about what to connect and what not to.

But he’s also worried. On the question of MCPs more broadly, particularly in the Google Ads space, he’s raising a flag that most people aren’t raising loudly enough.

“Security, cybersecurity, is a really big thing on my end. I think this is going to blow up in everybody’s faces very soon. All the different APIs, all the different MCPs, everything. There are very few trusted MCP solutions, especially within the Google Ads space, and I think that’s something everybody needs to be a lot more concerned about,” Andrew mentions.

His analogy is scripts. In the early days of Google Ads scripts, you could install something written by a stranger and have no idea what it was actually doing to your account. MCPs with write access are a version of the same risk, just with more surface area and more potential for damage. The PPC industry has been burned by this before. The current AI tooling era warrants the same level of scrutiny.

Andrew’s concern about MCPs is part of a larger point about how the industry is rushing toward connectivity without fully thinking through the implications. Read access is one thing. Write access to a client’s ad account, with real budgets attached, is a different level of risk entirely.

Andrew’s vision for a future without keyword match types

Andrew’s wish for Google is to eliminate match types entirely and replace them with a spectrum of smart bidding exploration settings. The current system asks advertisers to signal their intent through keyword formatting: broad match for expansion, exact match for control, phrase somewhere in between. He finds this increasingly redundant.

What he wants instead is a single dial on the smart bidding side. At one end, you can tell the system to explore up to 30% below your ROAS target in exchange for volume. At the other end, you can tighten to 30% above your ROAS target for maximum efficiency. The match type brackets and quotation marks become unnecessary because the bidding system is already handling what they were approximating.

“Why do I have to add a keyword in brackets or quotation marks and argue with people on the internet about whether or not the quotation marks or the brackets are better? It’s ridiculous,” says Andrew.

It’s a coherent vision. Smart bidding exploration already moves in this direction. The argument is that match types are a legacy control mechanism designed for a manual bidding world, and the sooner Google retires them in favor of more direct exploration parameters, the cleaner the advertiser experience becomes.

Practical ways PPC marketers can use AI today without giving up control

The through-line across this entire conversation is control. Andrew is enthusiastic about AI but precise about where it operates. It analyzes, it surfaces patterns, it maintains documentation, it answers operational questions. It doesn’t manage accounts, doesn’t draw strategic conclusions, and doesn’t take actions that can’t be reviewed.

That framing is more useful than the generic “use AI to save time” advice that fills most industry content. The time savings are real, but they come from specific applications with specific guardrails, not from handing over account access and hoping for the best.

The practical starting point he describes is accessible to any PPC team: download a CSV, upload it to Claude, ask a specific question about a specific problem you already have a theory about. See what the visualization shows. Build from there. The agencies that get the most out of AI in the next few years won’t be the ones who connected the most tools. They’ll be the ones who thought carefully about where AI earns trust and gave it room to operate only in those places.

Episode Transcript

Frederick Vallaeys: Hey, my name is Fred Vallaeys. I’m doing a series on AI-amplified marketers, and one of the best people to talk to, one of the most opinionated and smart people in PPC, is Andrew Lolk, who runs Savvy Revenue. So Andrew, good to see you again. Welcome to the studio.

Andrew Lolk: Thank you, Fred. Nice to be here.

Frederick Vallaeys: Yeah. Did Claude tell you to get a haircut?

Andrew Lolk: Actually, I uploaded a picture of me at the gym and got Gemini to just fake it, and I just didn’t like it. But my wife said I should try that. So.

Frederick Vallaeys: Okay. You still went for it. So that’s pretty cool. So you’re trusting of the AI. But let’s talk about AI in business, right? So you asked it for haircut advice. Obviously there are many things we can ask it for our work lives, for PPC. What’s your comfort level these days with using AI in business?

Andrew Lolk: So I think one of the things I’d say, the best way I can say it is when people say that they’re uncomfortable uploading their bank transactions through AI, my first response is, “Why?” I’m in the camp of the more stuff we can give to AI, the better. The more it knows about me, the better. The more it can connect to my email, my bank transactions, everything that we do, the better.

So that’s what we try to do in Savvy. We try to give it as much context as we can. So all our client notes, emails, Slacks, everything, because it just saves me time. We talked earlier in the day about this thing with having people in different time zones. I can just ask AI, “What is this? Why does this work? What is our vacation policy? When is this person working?”

I don’t have to feel bad about inviting somebody to a meeting at 4:00 p.m. when they leave every day at 3:00 because I for some reason cannot remember it. But they leave every day to pick up their kids. I just have that in a document in Notion. I can ask Notion AI, “What is the best time to schedule a meeting with Benjamin?” I’ll do it at 3.

Frederick Vallaeys: Yeah, AI can be super helpful. And really, so you’re one of the leading PPC marketers. I’m assuming you’ve used AI to make yourself stand out, and I just wanted to hear what you’ve been doing and what’s been working for you.

Andrew Lolk: I think for me, like many people, I feel like I’m constantly behind. That AI anxiety is out there where you just see somebody do something and you’re like, “Oh my god, why am I not doing it?” I think I finally found a place where I’m more at ease with what I’m doing and what AI can do.

Frederick Vallaeys: I hear you on that though. We watched the Google IO keynote, and at one point we looked at each other like, “Why are we sitting here watching this when there’s this new stuff that we should be doing something with?” You just always feel overwhelmed.

Andrew Lolk: And I think the stuff that I use AI for today is actually from a Google Ads perspective very uninteresting. I don’t do anything where Claude goes in and edits my accounts or builds campaigns. But I also think it’s important that we look at where we are from a Google perspective. What kind of accounts do we run today, and what benefits do we get out of AI?

Because in Savvy, we already run a lot of stuff more or less fully automated. All our search campaigns are built on feed-based structure. So new keywords are added, old keywords are removed, new brands are added, new categories are added. All that already happens dynamically. Then there’s a little thing with adding keyword variations, and that can be done rather quickly.

But where I found by far the most help with our clientele is basically getting a personal analyst, being able to — and this is going to get very unpopular — downloading the CSV file from Google Ads so you have the columns you exactly want. You don’t prompt it to try to get the data because once the accounts get big enough, Claude or Gemini gets really confused about trying to get the data out. But just uploading the CSV file and then asking it to distill the data, find some different visualizations of that data. And I’ve sometimes just gotten to a much better point in what I present to clients, but also just a better understanding of what’s going on in an account by doing that. That’s been by far the biggest breakthrough for me.

Frederick Vallaeys: And so when you have one of your AIs analyze data, do you find it pretty reliable? Is it writing Python to do the analysis?

Andrew Lolk: So it’ll write a Python script in order to distill the data and aggregate the data and all that stuff. What I’m very careful with is I ask it not to conclude anything. All the skills we have in Claude are very specific: do not conclude, do not infer, do not guess. Because it will read the data and go, “Okay, now your ROAS is increasing and your volume is going down. Your strategy of tightening up and becoming more profitable works.” And I just sit there and go, “The whole purpose is to expand. This is the opposite of what we want.” So it concluding things or inferring your strategy or your reasons for why the data is what it is.

Frederick Vallaeys: So it’s assuming that you’re successful and that you actually meant for that to happen.

Andrew Lolk: Exactly. And I’ve been really hard at saying don’t do that. Don’t infer. Don’t conclude. Don’t do anything. Just help me distill the data.

And that’s something where before, I would have the data in a Google Sheet or an Excel spreadsheet and you would create VLOOKUPs, you would create pivot tables to try to get there. And you would get there after 45 minutes. By uploading the data, even quite large datasets, to Claude — it’s only Claude that can handle this today — I can ask questions about the data. I did analysis from a huge account in Europe a couple months back, and instead of doing all the spreadsheet work, I just asked it and asked it, and it was beautiful.

Frederick Vallaeys: Yeah, exactly. And it can do it fast because it’s not using spreadsheets, which on the surface are actually pretty clunky. It’s just using the underlying Python code to do an actual analysis.

Andrew Lolk: And so I think the magic here is to avoid asking it to do an audit. “Audit this data.” No. You have the data, you already have a theory about what is going on. You go, “Does the data support my thesis about this?” Or, “We are performing more poorly year-over-year across 10 accounts. What segment of the account is it that works?”

And also, you guys had the PPC detective years back. I don’t know if you actually still have that.

Frederick Vallaeys: Yeah, the PPC Investigator.

Andrew Lolk: Investigator. Yeah. I always loved that, having that view where you can see, okay, the CTR of the shopping campaigns has decreased, so therefore you get less volume. Okay, now you know what to go do.

It’s the exact same thing here where you try to find that segment that’s not working and then a visualization of it. I had a client where I couldn’t find out why their Polish account, no matter what we did, we lowered spend and the overall contribution margin or blended ROAS, it never improved. We went from a 250% ROAS to a 450. How can it not improve? And we’d tried it twice, and it didn’t improve.

So I ingested all the Google Ads data, all the Meta Ads data, and the client’s revenue and the Klaviyo data. And the one chart it produced was Meta spend percentage, and it collided exactly with when I would decrease Google Ads spend, then the client would increase Meta. Data point media mix model without knowing it, pretty much.

Frederick Vallaeys: Nice.

Andrew Lolk: And those are the things I just think, for me, the analysis part means I don’t need to have the junior, I don’t need to have somebody on my team do all these reports. I can prompt my way to that report immediately.

Frederick Vallaeys: For me that’s been the unlock of what you’re describing, but also the fact that a lot of my team works in different time zones. And by the time I’m ready to ask them a question, they’re often asleep or they’re with their families. But this assistant is always there and always ready to go whenever I am.

And for me, the challenge almost becomes — I’ve got two copies of Claude, a personal and a work one. And then within Claude, you’ve got the three ways of working with it. Which one did I ask this question? Like, where do I need to go to get that answer?

Andrew Lolk: So just like with managing employees at a company, it’s now I’ve got this workforce of many agents, and how do I — do I need a manager to stay on top of them? And that’s becoming that next level of challenge for me.

Frederick Vallaeys: Yeah. Yeah. In terms of the agents that you work with, Claude Co-work — is that mostly okay.

Andrew Lolk: So we go back and forth. I think at this point in time, Claude is so far ahead that it unlocks so many cool things that you just don’t have in ChatGPT or Gemini today.

Frederick Vallaeys: Yeah, absolutely. Completely agree. It’s the one that works for work. ChatGPT — that’s the one I go to if I need to find something and I don’t want to Google it. But if I need to write an email, why would I go to ChatGPT to get a thing I need to copy and paste into another screen when Claude can just go and do it for me.

Andrew Lolk: It’s the marketplace analogy again. It’s like in order to be a successful marketplace, you need to have a lot of buyers and a lot of sellers. You need buyers to get sellers, you need sellers to get buyers. And the Claude ecosystem is just unbelievable, right? I use Superhuman for my email and they have a Claude MCP. So the Gmail MCP doesn’t work that well inside of Claude compared to Superhuman because they’ve opened up completely and they have a connection into Gmail. So it’s just unbelievable what they’ve built today.

Frederick Vallaeys: Yeah. And so speaking of MCPs, you like the Superhuman one. Have you built MCPs for Savvy, or which ones are you particularly drawn to?

Andrew Lolk: So we haven’t built any, but it’s also because most of all the data that we have — we have all our data storage in BigQuery. Very early on we created a data warehouse for all our client data so we can just consolidate everything in one place.

Security, cybersecurity, is a really big thing on my end. I think that this is going to blow up in everybody’s faces very soon. All the different APIs, all the different MCPs, everything — I’m so worried about all that. And there are very few trusted MCP solutions, especially within the Google Ads space. I think that’s something that everybody needs to be a lot more concerned about. Just letting anybody and everybody create access to an account that’s not just read access but write access.

Frederick Vallaeys: It’s a little bit like scripts back in the day, right? You could install a script that somebody else had written and, unbeknownst to you, just go for it.

Andrew Lolk: Yeah, exactly. You go for it, and it’s now emailing your whole account data to some random person without you knowing.

Frederick Vallaeys: Yeah, exactly. So that’s MCPs. What about skills? Do you build skills in-house at Savvy Revenue?

Andrew Lolk: So we have two layers in Savvy. We built our entire internal operating system on Notion. It was a decision we made four years ago, and it’s always worked really well for us. Today it’s working incredibly well because you can have any page in Notion — we don’t have to overcomplicate it — but every single page in Notion can be AI instructions. So where we have all our SOPs, we just have a small note below an SOP that is specific for AI. So all of a sudden, all the work we’ve ever done, all the notes, all the client emails, all the client conversations, everything we’ve ever had and ever do, from a day-to-day perspective it runs inside of Notion. And agents are running with project management, weekly planning, and all this stuff. It’s all in there.

Frederick Vallaeys: Token maxing.

Andrew Lolk: Oh, really? Like, but you don’t pay for tokens in there.

Frederick Vallaeys: Oh, it’s unbelievable. I bought a Claude subscription, just a regular one, to max it out in two minutes. Okay, great. But I can do whatever I want within Notion. It is a limited version of Claude so it doesn’t max out tokens in the same way.

Andrew Lolk: And that’s one thing that frustrates me sometimes. You buy a software, whether it be Notion or something else, and to make it feasible to give you unlimited AI usage within the plan that you pay for, they don’t give you the best model, right? But sometimes you’re like, “I could really benefit from the best model because then I don’t need to go and review everything.”

Andrew Lolk: Notion has the best model. It has OpenAI 4.6, it has ChatGPT 5.5, but the modality is limited where Claude will generate a nice looking report with different graphs. Notion will not do that. So they’re really limiting it. But now they have autonomous agents inside of the whole workspace. They can just lie there.

So we do quarterly planning in Savvy, and one of the things we did this time for clients was we went, “Okay, go through all the YouTube videos I’ve done, go through all the projects that were done in the last client projects, find the conclusions, and then write up — tally up what were different projects, what projects were successful under what circumstances, and which ones weren’t.” And then we can take all those learnings, and now we have a live database of 12 recommendations to the team that goes: if your account is in this scenario, then this project would work well for you.

The amount of work it takes away from running an agency, not from a Google Ads point of view, is unbelievable.

Frederick Vallaeys: Yeah. And this is important, right? You’ve got an SOP, if this then that, which also applies with skills. But the risk with AI is that you say, “Hey, optimize this account,” and I could optimize it one of a thousand ways. And as you’ve pointed out in some blog posts, it’s risky because it could just muck up your whole account and give you six months of —

Andrew Lolk: Exactly. And it’s missing the entire context of what is going on here and how do you like to do things. And I think it’s a huge advantage to do content because you share your thinking, and that can be converted into SOPs.

And it sounds fancy and stuff, but it’s basically just going, “We believe that you shouldn’t add every single keyword as a keyword variation.” That’s an SOP. But if Claude reads a blog post from somebody from two years ago where you had to add all the keyword variations because match types worked differently, then it will do it and will recommend it and stuff like that.

Frederick Vallaeys: And keeping track of all of these new little tidbits that you learn — just go to your agent and be like, “Hey, here’s another thing we should consider for our SOPs.” It’ll find the right one to put it in, knows how to update it.

Andrew Lolk: A very good example is we used to have a fairly complicated structure with every SOP where, when something was too old or somebody didn’t understand it, they had to click a button. That button created an automation that opened a task which you had to fill out with a request to say why something didn’t work.

Now, all we do is in our SOP database, people can just comment on it. That’s it. I don’t need anything else. It just works just like Google Docs — highlight something, comment it, and say, “This is what works, this is what doesn’t work, or I didn’t understand this.”

And then we have an AI agent that runs every month, goes through it all. What it can fix itself within some limits we’ve given it, it’ll just fix. It’ll pull in the latest information. It will pull in from our Slack and say, “Okay, somebody has asked this before. I can just take that and put it in here.” And what it can’t understand itself, it gathers in a project and gives it to my partner in Savvy who runs operations. He can then go through it.

And it’s just so much more simple. I look at the Google Ads world — I think we’ll get there with doing similar stuff in there, but what it does for operating a team and a company today is unbelievable.

Frederick Vallaeys: Yeah. Now, what’s on the wish list? What’s one big AI thing that would make your life even better?

Andrew Lolk: My big wish is that they kill off match types and leave no match type or just exact match.

So I think that smart bidding exploration can relieve the need for match types entirely. Where today with smart bidding, let’s say we have a 500% ROAS target, I should be able to tell that — smart bidding exploration works where we give it permission to explore more. From my perspective, that’s the broad match version: explore more, go more broad.

Frederick Vallaeys: Right. Maybe you don’t have to meet the stringent targets when you’re exploring.

Andrew Lolk: Exactly. I want the opposite as well. I want it on a scale where I can go, “Okay, you can go 10, 20, 30% lower in ROAS.” I also want the opposite — I want to go 10, 20, 30% higher. So that’s basically what I’m trying to do with match types. If I only run exact matches because I want to be very tight, higher ROAS, no risk, why do I have to add a keyword in brackets or quotation marks and argue with people on the internet about whether or not the quotation marks or the brackets are better? It’s ridiculous.

My big wish is that they get rid of match types.

Frederick Vallaeys: Okay. Maybe one day. Maybe one day soon. We’ll see.

Andrew Lolk: You didn’t see that one coming.

Frederick Vallaeys: Good. Hey, really good to get your perspective on all of these things. Very unique thoughts, and thank you for sharing them, and hope to have you back on another podcast.

Andrew Lolk: Thanks for having me, Fred.