As some of you may know, I spoke at SMX in Munich and Friends of Search in Amsterdam during the second week of March. At the end of each presentation, the audience is given a few moments to ask the speaker a few follow up questions. My biggest complaint with this format is that sometimes questions require more time to think about than we’re allotted on stage.
I spoke about a topic that I am incredibly passionate about: automation drift detection and correction. At the end of my sessions, there were so many wonderful questions. Some I answered well and some I kept thinking about on the flight home.
So, I decided that I was going to sit down to give those questions the full answer they deserved. The organizers from SMX were able to provide me with the questions that were asked.
There are eight of them. Let’s get into it!
Question 01
Can you elaborate on Customer Match regarding signals? What does it act as a signal for besides New Customer Acquisition?
Most people know Customer Match in the context of the New Customer Acquisition (NCA) goal. You upload your existing customer list, tell Google to prioritize people who aren’t on it, and theoretically get more net-new revenue. That’s the version that gets talked about most. But it’s only a fraction of what Customer Match is actually doing when you use it as a signal.
Here’s how to think about it more completely. When you upload a Customer Match list, you’re giving the algorithm a concrete picture of who has already converted. Google matches those hashed emails, phone numbers, and addresses to logged-in Google accounts and uses them to build a profile of your converting customer. From that, it does a few distinct things:
Lookalike expansion (Demand Gen). The algorithm finds new users who resemble those on your list and bids more aggressively for them. This is the most powerful use of Customer Match as a signal; you’re not just telling the algorithm what to avoid, you’re telling it what to chase.
Suppression. You can use Customer Match as an exclusion list, telling Google not to serve existing customers at all if your goal is pure acquisition. This is blunt but effective.
Retention targeting. Google now explicitly supports using Customer Match to bid more aggressively for lapsed customers, so people who bought from you once and haven’t returned. If you segment your list properly, you can have the algorithm treat a lapsed customer differently from a brand-new prospect.
High-LTV signalling. If you upload a segmented list of your highest-lifetime-value customers specifically, you’re not just telling the algorithm what a customer looks like. You’re telling it what a good customer looks like. The algorithm then looks for users who resemble that profile and bids accordingly.
None of this works properly if the list is stale. Depending on the business model, 30-90-day list resets are what I would recommend. An 18-month-old Customer Match list can act as a corrupted signal; you’re teaching the algorithm to find people who look like your customers from a year and a half ago, who may have churned, shifted, or become completely different buyers. The algorithm won’t tell you the signal is bad. Your conversion quality could degrade quietly, and you’ll spend months trying to diagnose it from the wrong end.
If you’re using Customer Match only for NCA goals, you’re using maybe 30% of what it can do. The input layer is where the real leverage is.
Question 02
Tell us more about how to block categorically on Google — a real example.
For the folks who weren’t in Munich or Amsterdam: I was talking about negative keywords and how most practitioners think about them reactively. We wait for bad search terms to show up in the search query report (SQR), add them as negatives, and repeat. What I was pushing for instead is categorical blocking before the campaign goes live. You draw the fence first, then let the algorithm run inside it – a flexible boundary.
The example I used on stage in Amsterdam was HVAC: heating, ventilation, and air conditioning companies. To may of us, this is a simple business. But if you run a search campaign for an HVAC client without categorical negatives, here’s what could happen: Google starts matching you to queries from adjacent trades. Plumbers, electricians, roofers, landscapers, etc. These industries share vocabulary with HVAC as they all deal with home services, contractors, emergency callouts, seasonal maintenance. Now, Google won’t send you completely random traffic. But it will send you the irrelevant traffic if it has reason to believe it will convert, based on category proximity, behavioural signals, or past conversion patterns. The problem is that “reasons to believe” and “actually qualified” are not the same thing.
So before a single dollar is spent, I suggest you build your categorical negative list:
HVAC categorical negatives (starting point): plumber, plumbing, electrician, electrical, wiring, roofer, roofing, gutter, landscaper, landscaping, irrigation, pest control, locksmith.
Then go one level deeper: what adjacent services does this client absolutely not offer? Add those categories too.
The principle scales beyond HVAC. A corporate law firm needs to block personal injury, family law, and criminal defence. A B2B SaaS company needs to block job listings, tutorials, and free tools. A premium gym needs to block budget gyms, gym equipment retail, and home workout. Every business has categories it categorically is not, and those categories should be negated before the first impression is served.
Question 03
So you would recommend ignoring the Ad Strength KPI for RSAs? I’ve experienced ads with low “efficiency” having higher CPCs and lower CTR. What do you think?
Yes, I recommend ignoring it as a performance metric.
Ad Strength is a Google-defined score that measures how closely your RSA aligns with what Google’s system prefers: maximum headline variety, keyword inclusion, character-count utilization, and a broad range of messaging. Score high, and Google calls your ad “Excellent.” Score low, and you get suggestions to add more headlines, pin fewer assets, and a “Poor” rating.
What Ad Strength does not measure is whether the ad converts. Those are different things, and conflating them is where the trouble starts.
My response in Munich was: “I would rather pay more for a genuine, semi-qualified lead than shave a few cents off my CPC because Google wants to optimize for attention over intention.”
I say this because a tightly controlled RSA that scores low on Ad Strength allows us to write specific ads for specific people. Someone who knows exactly what they need and immediately filters out generic messaging. That ad has fewer headline variations, possibly pinned assets, and deliberately narrow messaging. Google’s interface sees limited variety and docks your score, but the person who clicks it is genuinely interested.
That’s not underperformance. That’s the system working correctly for your actual goal, and Ad Strength calling it a failure doesn’t matter because it was never designed to measure that goal.
Ginny Marvin, Google’s Ads Liaison, has said publicly that Ad Strength has no weight. It’s a diagnostic tool that is designed to help you write more complete ads, not to evaluate lead quality or conversion performance.
Question 04
What targeting is best for PMax and B2B? Signals, search themes, interests, etc?
Customer Match first, always. Your own first-party data is the most powerful signal you can feed the algorithm. A segmented list (current customers, lapsed customers, high-LTV customers) gives Performance Max a concrete anchor for who it should be finding more of. Without it, you’re asking the system to learn from scratch, which takes longer and costs more.
Search themes second. For B2B, intent is more specific than in e-commerce. Industry terminology, job titles, and specific problem statements are the terms your buyers actually use. Search themes let you guide the algorithm toward that intent rather than letting it discover it through trial and error with your budget.
In-market audiences third, with caveats. Google’s B2B in-market categories are broad. “In-market for enterprise software” could mean anything. Use them, but don’t rely on them as a primary signal; they’re a supporting layer at best.
Interests last. Interest-based targeting is built for consumer behaviour patterns. B2B buyers don’t make purchase decisions based on their interests in the way a consumer does. This signal will do the least work for you in a B2B context.
An important thing to understand about PMax signals in general: they are directional hints, not hard targeting. The system takes your suggestions into account, but is not restricted to them. If it finds converting traffic outside every signal you’ve provided, it will go there. Signals accelerate the learning period; they don’t control the output. That’s the fundamental tension with PMax in low-volume lead gen environments.
Also, none of it works the way you want it to without Offline Conversion Tracking (OCT). Audience signals tell the system who to find, but OCT tells the system what success actually looks like. You can have a perfect Customer Match, well-defined search themes, and a clean in-market layer, yet the system will still drift if the only conversion data going back into the algorithm is a form fill. For B2B, the gap between “form submitted” and “qualified lead” can be enormous. Utilizing OCT helps close that gap. Upload your CRM-qualified leads back into Google Ads, matched by GCLID, and the system will learn which clicks actually turned into real leads, not just which clicks turned into contact form submissions.
Question 05
Google Search Partners are still part of PMax, and in our case, they drive fake leads. How do you fix this if you can’t exclude them?
So, we do have wonderful news on that! We are now able to add placement exclusions in our PMax campaigns. However, there are a few steps to get there:
Go to Tools → Content Suitability → Advanced Settings → Excluded Placements
You must apply these at the account level, which means they will be applied across all campaign types, not just Performance Max. Also, as of March 2024, account-level placement exclusions extend to the Search Partners Network (SPN)! So, if you add a site exclusion at the account level, it will be respected across the SP inventory in PMax as well. This won’t fix the issue entirely, but systematically excluding specific domains that generate bad traffic is a great way to guide the automation.
The great folks at SMEC wrote an article on how to use placement exclusions with a script. You can read it here.
Question 06
How do you control a PMax campaign when it drifts away suddenly?
When this happens, the first thing I want you to do is… nothing. Do not change a setting or slash the budget. What we want to do first is diagnose what happened.
Step One: Find the drift.
Pull three reports before you change a single setting.
First, the Landing Pages report. If PMax has suddenly started sending 85% of your traffic to a blog post, a thank you page, or your careers page, you’ve found your problem. The system found a way to “convert” at scale on a page that has nothing to do with qualified leads, and it’s exploiting it. Final URL Expansion is the likely culprit here. The fix is to add URL exclusions at the campaign level for any non-converting page, and, if the drift is severe, turn Final URL Expansion off entirely until you’ve stabilized.
Second, the Search Terms report. Is the query distribution still where it should be, or has it migrated into lower-intent territory? Cross-reference against your negative keyword list and identify where the gaps were created. Go back to question 2, for how I suggest building out negative keyword lists.
Third, the Asset Performance report. Filter by “Added by: Google AI.” If auto-generated assets have crept in and are now serving heavily, that’s a creative drift signal worth addressing.
Step Two: Make changes incrementally.
Once you’ve identified the source of the drift, I want you to resist the urge to “correct” everything. If you need to pull the budget back, do it in 10–20% increments over several days rather than one large cut. If you’re adding URL exclusions, adding negatives, and refreshing audience signals, don’t do all three on the same day.
When a campaign is misbehaving, and you make twenty changes simultaneously, you lose the ability to read the results. If performance recovers, you don’t know what worked. If it gets worse, you don’t know what broke it further. Every simultaneous change is a variable you can no longer isolate. Fix the most likely cause first. Wait. Then decide if anything else needs adjusting.
Step Three: Give it enough runway to actually tell you if the fix worked.
This is where low-volume lead gen accounts run into problems. The general guidance is to wait until at least 30 new conversions have occurred under the new settings before evaluating whether the change worked. The system needs data to learn, and the less of it you have, the more volatile the results, especially after you’ve made a change and reset part of the learning.
In a B2B account generating 20 conversions a month, that math isn’t great. But the principle holds: judge the fix on conversion data, not on the week-over-week swing you see. Pulling the campaign before the data settles means you’ll never know whether the first fix was actually working.
Sudden drift in PMax is almost always a signal problem that surfaces as a performance problem. The campaign didn’t randomly decide to go rogue, something changed in what it was being rewarded for, or where it was being allowed to go. Diagnose the input layer first.
Question 07
What is your opinion on the conversion metric “store visits”? Is it comparable to a conversion, in your opinion?
No, I don’t think it’s comparable to a conversion.
Store visits are an estimated, modelled metric. Google uses GPS data, WiFi signals, cell tower data, and Maps information, cross-referenced against a panel of opted-in users who share their location history. While the methodology for estimating foot traffic is sophisticated and the accuracy claims are high, I disagree with assigning it the same weight as verified actions, such as a form submission or a completed purchase. Estimated foot traffic should not carry equal weight in your bidding signal.
Google has been auto-enabling store visit conversions on accounts without explicit advertiser consent, assigning them predefined values, and letting them count equally in the conversion signal. Which means Smart Bidding starts optimizing toward a softer, easier-to-achieve target without anyone deliberately deciding that it should.
If foot traffic is genuinely your primary business objective (you’re a retailer, a restaurant, a car dealership, or a gym), then store visits can be a primary conversion. But even then, you need to assign the value separately and lower than that of a verified transaction, so the algorithm weights them appropriately relative to actual business outcomes.
If foot traffic is not your primary objective, store visits shouldn’t be a primary conversion at all. It belongs in your secondary conversions, where they inform your reporting without influencing your bidding.
Check your accounts. If store visits are enabled and counting as primary conversions without you having made that decision deliberately, that’s a signal worth reviewing today.
Question 08
What’s your opinion on broad match? Should we still use it?
Yes, but only under specific conditions.
Broad match works when three things are true simultaneously:
1. You have a clean, high-volume conversion signal
2. You’re running Smart Bidding with a conversion-focused target
3. Your negative keyword list is actively maintained.
Remove any one of those three, and broad match quietly becomes a bottomless pit for your ad spend.
Brad Geddes from Adalysis wrote a great article that gives the nuance that the “broad match is dead” vs. “broad match is everything” debate usually lacks. Broad match can deliver strong results under Max Conversion Value bidding in e-commerce contexts with sufficient volume.
For lead generation specifically, lower conversion volume, higher stakes per lead, and more variation in lead quality, exact match consistently leads in efficiency. Broad match in a low-volume lead gen account gives the algorithm too little signal to calibrate against and too much freedom to expand into irrelevant territory.
There’s also a newer wrinkle worth flagging. Mike Ryan’s research found that AI Max effectively treats all your keywords as broad match, even if you’ve only uploaded exact- and phrase-match keywords, and assigns those impressions back to your existing keywords. Which means that if you’re running AI Max, you may be getting broad-match behaviour without realizing it, and your match-type performance data is no longer clean… yay.
Treat broad match as a signal you audit consistently, not a setting you configure once. If broad match is in your account, it should have a dedicated review cadence: search term report, negative keyword additions, and conversion rate by match type. If it doesn’t have that, you’ll find out it drifted the hard way.
The categorical question — should you still use it — presupposes that at some point it stops being worth the effort. I don’t think that’s quite right.
As for whether we should still use it, broad match is a legitimate tool in specific contexts. The mistake is using it in contexts where those conditions don’t exist and treating the resulting drift as a broad-match problem when it’s actually an account-management problem.
In Closing:
Automation is designed to explore and then exploit. It goes out looking for conversions, finds what works, and doubles down on it. That’s not a flaw, it’s exactly what it’s built to do.
The problem is that “what works” is defined entirely by what the system can measure. And what the system can measure is almost never the same thing as what actually matters to your business. Automation doesn’t understand the difference between relevant and commercial. It doesn’t know that a form filled out by someone who will never qualify as a lead is worthless. It only knows what you told it to count, and it will get extraordinarily good at finding more of that, whether it’s valuable or not.
That’s the thread running through every question in this post. Signal drift, query drift, inventory drift, creative drift; they’re all the same underlying problem wearing different clothes. The system followed its instructions; unfortunately, the instructions were wrong, and because the dashboard stayed green, nobody noticed.
The good news is that this is fixable. Not by fighting the automation, but by being more deliberate about what you feed it, what you tell it to optimize for, and what you check when the numbers start moving in directions that don’t make sense in the real world.
So go look at your account. The conversion actions, the search terms, the landing pages that are actually receiving paid traffic, the Customer Match lists and when they were last refreshed, the gap between Google Ads conversions and CRM-qualified leads over the last 90 days, etc.
Drift is quiet. Oversight has to be deliberate.