Joel Miller on Google Ads AI, trust, and control. A conversation with host Lisa Raehsler on the AI Ads & Beyond Podcast — “Who Really Benefits From Google Ads AI?”

AI Ads & Beyond: Google Ads AI, Trust, and Control

Lisa Raehsler

My guest today has been in the trenches, and he is not holding back. Joel Miller has been running a boutique marketing agency with his identical twin brother Alan for 17 years, and they’ve never hired a single employee. He works inside live accounts every day using AI tooling, and he has some strong opinions about when the platform’s recommendations are working for you — or when they’re working for Google.

Joel Miller

AI is a capable driver — that’s how I’d think about it. But the question is whether the GPS is pointed in the right direction. If your strategy is clear and your account structure is sound, AI can be a real multiplier. But if you don’t know where you’re pointed, and you have no data because you haven’t made any choices yourself, then AI is just going to accelerate the problem.

Rapid-Fire: Rating Google’s AI Features 1–10

Lisa Raehsler

Let’s do a rapid-fire round. I’ll name some Google features and you rate them one to ten. One means do not trust it, and ten means you can turn it on and walk away — full trust.

Joel Miller

Auto-apply recommendations get a 1, because Google is optimizing for Google’s metrics — not ours, not yours. It’s so aggressively trying to opt you in that unless you’re actively watching it, it’s probably going to get you into it in some fashion.

Budget increases based on Google’s projections — a 2. Their projections are often based on what’s good for spend velocity, not what’s good for the client’s margin. I like to control the increase myself.

Broad match with Smart Bidding — a 3. I think it can work, so I’m going low but not the floor. But “it can work” is not a strategy. It’s just living on hope. And I’d rather have an actual strategy that I can test.

Performance Max with no asset-group segmentation is also a 2 — you lose diagnostic visibility, so you really can’t fix, or communicate to your client, what’s going on if you can’t see it. Google’s campaign objective suggestions get a 5, right down the middle: it feels hopeful, but not all of what it changes is readily apparent, and you can end up getting lost in the noise.

Here’s the full scorecard from the episode — nothing scored above a 6:

• Auto-apply recommendations — 1. Optimizes for Google’s metrics, not yours.
• Rep-pushed keyword consolidation — 1. Quota-driven calls, not strategy.
• Budget increases from Google’s projections — 2. Built for spend velocity, not your margin.
• Performance Max, no asset segmentation — 2. You lose diagnostic visibility.
• Ads in AI Overviews (no controls yet) — 2. Can’t see it → can’t defend it.
• Broad match + Smart Bidding — 3. “It can work” is not a strategy.
• RSAs, full rotation handed to Google — 3. You give up message-to-market fit.
• AI Max for Search — 4. Was a 7 at launch — the claimed stats haven’t shown up.
• Campaign objective suggestions — 5. Directional, but the impact is buried in noise.
• Automated bidding adjustments — 6. With enough conversion data, it earns trust.

The pattern: the less visibility a feature leaves you, the lower it scores. Trust rises only where the machine has real data — and you can still see what it’s doing.

Strategist Calls and Two-Versions-Old Help Docs

Joel Miller

The thing that always happens with Google strategist calls: it’s persistent and annoying. They’re always switching who you talk to, so there’s no continuity. There’s maybe been one time out of — let’s call it 20 over the years — that was useful. I know so much more about the platform than it seems like the person on the other side of the call does, so I’m just navigating circles around their description.

And nearly every suggestion is something I’ve tried already. When I’ve done those edits, it usually ends up spending more, with lateral or worse results. So I’m always skeptical that it’s their number-one button to push right when they want profits and revenue up for the quarterly stock call.

Even the troubleshooting: I’ll click the help article and find out it’s about two versions old. The interface has changed and they haven’t updated the docs — which is always surprising, because ads are their main source of revenue.

What Auto-Apply Does to Your Signals — the Visibility Problem

Joel Miller

A big tipping point for me is just understanding what’s happening. A lot of the AI features are opaque. Right now there’s not a singular place where I can go look at the choices the AI is actively making — what those choices are, and what signals told it to make that choice. I’d love to know.

Auto-apply totally obscures that. There’s an audit log where you can see what it applied, but because I’m proactively doing this for clients, I don’t want to go on a historical mission to figure out when something happened and why. I want to sit at my computer and say: all right, this is my new test. I’ll build an ad group with these experimental theories, run it against everything else I know in the account, and see how it performs. Then I put that into my memory bank to test with other clients.

Keyword Intent, AI Overviews, and the School-Bus Mystery

Joel Miller

As AI Overviews rose last year, we found that when something wasn’t very intent-oriented, it was just getting eaten up by the Overview — and we saw a dip in conversions and even clicks on some campaigns. So we’ve had to get a lot more specific, targeting searchers who are looking exactly for the thing we’re pointing them at.

This reminds me of a pre-AI campaign we ran for a medical group in the Chicagoland area. I repeatedly had to add negative keywords about buses. For some reason Google kept attaching our keywords to something about buses or school buses. I was like, what in the Google world do they think a medical group has to do with buses? No human would look at that and think, yeah, we need to show this ad to people using this keyword combination. Only a human understands that the intent is different.

AI for the Practitioner vs. AI for the Platform

Lisa Raehsler

This question is about AI’s place in PPC — the difference between how AI helps the practitioner versus how it helps the platform.

Joel Miller

AI can be super helpful to iterate faster, to surface anomalies, or to reduce repetitive tasks. And I’m a big proponent of using tools like Claude for idea generation, not just relying on the little AI output Google has built in — especially since, inside the Google sphere, the incentive isn’t necessarily yours. The incentive for the AI should be to help you understand.

But it’s more helpful to the platform when it’s suggesting automated spend increases, pushing broad match over more specific matching, and reducing transparency by taking away the data we need to actually see what’s occurring.

Don’t Compete Against an Invisible Opponent

Joel Miller

We had a client we did ads for who were convinced their next-closest competitor was spending ten times as much as them and getting so many leads. Every meeting: I know it’s going like this for them. Then they closed because of COVID — and the competitor ended up hiring us. About two years later I was in that competitor’s ad account, and I could go back in time and look. They were completely wrong about everything. Our client’s campaigns looked better than theirs at that same time.

What it made me realize — and I think this ties into AI — is: don’t compete against an invisible opponent, whether that’s your competition or what Google’s doing with its AI. The race is against yourself. Optimize based on your own past performance, and keep control of the levers you can actually see.

The Human + AI Play: 30 Hours of Sales Calls

Joel Miller

Always have something you’re in charge of — that you’re actually controlling, monitoring, and iterating on — while you test a new feature. Turn on the AI search assist if you want, but run it as a separate campaign or ad group and pit it against your own management, and see if it’s actually helping instead of assuming it will because of the claims.

Here’s a non-ads example that’s the same idea. For one client, I listened to about 30 hours of their incoming sales calls and made a custom GPT that automatically grades them as they come in. I identified six things their team struggled with. Now every call gets graded automatically, with the problem areas and suggestions on how it could be better — but we kept a very human element, using our own ability to identify patterns in a way the AI might miss. Then we circle it back into the ads: we address an objection in our ad copy before the prospect even brings it up.

Joel Miller

That’s maybe my biggest takeaway from this whole conversation: I want information. I love AI — I think it’s going to help us be more efficient as a society. But I want the information so I can still use my human intelligence to make informed decisions about how I use it.

Where to Find Joel Miller

Joel Miller is the co-founder of The Sky Floor, the boutique agency he runs with his identical twin brother. Read more on the blog, or reach out to start a conversation. Thanks to Lisa Raehsler for having Joel on AI Ads & Beyond.

Key Takeaways

• Nothing scored above a 6. Across ten Google AI features rated live, auto-apply recommendations and rep-pushed keyword consolidation got a 1, budget projections and unsegmented Performance Max a 2 — because they optimize for Google’s metrics and spend velocity, not your margin.
• AI is a capable driver; your strategy is the GPS. With a clear strategy and sound account structure, AI is a multiplier. Without them, it just accelerates the problem — so build your own data by making your own choices first.
• Visibility is the dividing line. There’s no single place to see what choices Google’s AI made or which signals drove them — and if you can’t see it, you can’t defend it to a client. Favor forward tests over forensic hunts through the audit log.
• AI Overviews are absorbing low-intent searches. Campaigns that leaned on looser intent saw clicks and conversions dip — the fix has been remaking campaigns around tighter, more specific targeting.
• Don’t compete against an invisible opponent. The client who was sure a competitor was crushing them was completely wrong — the race is against your own past performance.
• The human + AI play: Joel listened to 30 hours of sales calls, identified six struggle patterns himself, built a custom GPT that auto-grades every new call, and feeds real objections back into ad copy. Test any new AI feature the same way: as a separate campaign pitted against your own management.