ArticleIssue #180 | Time-Tested
I hope you’re enjoying the last few weeks of summer, with back-to-school in full boom, football season around the corner and the final few vacation-acceptable weeks on the horizon.
Back in my younger years (yeah, I feel old even writing that, but it’s true), I loved video games. While my favorites were always MMORPGs, there’s a special place in my heart for turn-based strategy games like Civilization – which is why I still play it every once in a while.
A few weeks ago, I had a free Friday night (kiddos at grammy’s, wife visiting family) and mentioned to someone on my team that I was looking forward to playing Civ7 for a few hours. He looked at me the same way I’d look at a friend who joined a multi-level marketing scheme: like I had lost my mind. I had no idea why – I knew he played Civ. I knew he loved the game. I just had no idea he’d spent the better part of a year refusing to play Civ7.
And why?
Because Firaxis decided to do what tech companies today love to do: make a sudden, disruptive change to the core product, then demand users simply accept it.
In Civ7, that change was the Age system. VII splits a game into 3 Ages, but at the end of each one your empire becomes a different civilization entirely. Most everything – including what you spent hours building – resets, and what carries forward is limited. That’s a tectonic shift for a game whose central premise for 30 years was that if you picked Rome, you won or lost as Rome.
But this isn’t an article about Civ. It’s an article about what happens when platforms – whether they’re ad platforms or game platforms – make massive changes that alienate their users.
In the case of Civ7, it took Firaxis 15 months to design a resolution: they made the disruptive piece optional. You can now carry your selected civilization from start to finish if that’s how you want to play. The Ages still progress. The transitions still occur. But you – the player – have some control over what happens when they do.
There’s a lesson in that for marketing – and it’s one I believe our industry has gotten wrong for 8 years.
We Didn’t Lose Control
What’s striking to me is the parallel between the Civ7 situation and what’s happened with digital ad platforms. At every conference I attend or in every industry blog I read, someone is always lamenting,
“Google took away our control of our accounts.”
News flash: they didn’t, really.
What they removed was granularity. Granularity was never the same thing as control.
Setting a manual CPC on a keyword never controlled anything: you didn’t control who saw the ad, on what device, in what mood, after what, in what position. You controlled one input (how much you were willing to pay) into a system whose output you couldn’t predict. We called it control because it took effort and produced a number….and because the alternative (admitting we were turning a dial wired to a black box) was less comfortable.
What granularity gave us was legibility – the ability to draw a clean line from a decision we made to a result that followed. That’s a real loss for individuals/firms where legibility was the foundation of what they sold. Before all these changes, those people could sit in a room and say: I changed this bid on this keyword, here’s what happened, that’s why. Take that away and an experienced media buyer looks, from the outside, a lot like someone who uploaded a budget.
That’s the anxiety underneath every automation complaint I’ve heard since 2018. It isn’t that the machine makes worse decisions. It’s that the new structure makes it difficult for some people to prove they made any decisions whatsoever.
But – every ledger has two sides. Granularity had a cost marketers never priced (at least not honestly). I still remember the hours upon hours spent every week building bidding sheets in excel. I remember spending every morning going through accounts to identify keywords where our bid was below the first page. Those are decisions automated bidding systems now make in real time, using more signals than we’ll ever see. Changing numbers in excel felt like the job because it was visible and effortful, but it was never the job; the job is, was and always will be the offer, the creative, the measurement and knowing what to count.
So, if we’re all being honest: we traded a lot of busywork, some genuine craft and most of our explicability for better median outcomes and a different kind of craft. That’s a pretty fair trade, if you ask me. Pretending otherwise has made a lot of good marketers sound like they’re mourning a spreadsheet.
What interests me is what follows from that. If legibility is what you lost, legibility should be what you rebuild – just on more stable ground. That’s the actual job change of the last 8 years, but it’s hidden underneath a whole lot of arguing about bidding strategies and optimization scores.
Google Used The Firaxis Method
Performance Max in 2026 isn’t the Performance Max we were handed in 2021.
The original was a half-baked cake: it looked beautiful. It seemed great. But if you ate it, the most likely outcome was salmonella. When Google rolled out PMax, they asked us to hand over budget across Search, Shopping, Display, YouTube, Discover, Gmail and Maps – and in return, we received one line of aggregate performance (plus a not-so-subtle request to increase our budget). The OG PMax campaigns provided no meaningful negatives, no sense of where the money went, no breakdown of who saw your ads. All of that meant there was no way to explain a bad month to a client without (at best) guessing.
Google has spent the years since returning legibility in waves. Channel-level reporting. Search terms. Campaign-level negative keywords, brand exclusions, placement exclusions. Then this year: 1P audience exclusions so you can steer away from customers you already have, network-segmented placement reporting, budget projections, real demographic breakdowns, even channel-level opt-outs for Search Partners + Display Network (this is still in Beta).
Every one of those has a play attached, and almost nobody is running them. Upload your customer/lead list, exclude at campaign level, then watch new-customer CPA vs blended. If the two have converged, PMax has been charging you for your own retention. Pull placement reporting by network and negative out the Display inventory that’s been consuming ~20% of your budget for next-to-no marginal return. Run budget projections and compare them with your own budget modeling BEFORE the client meeting. Check the demographic breakdown against who actually buys rather than who you assumed buys. And if you can get into the Search Partners opt-out beta, take it. Search Partners has been horrific of late.
But notice what Google has not given back: manual channel budget allocation. You still can’t tell it to put 32% into YouTube and 49% into search and 11% into gmail. That particular black box is still sealed.
That’s the same strategy Firaxis used with Civ7: keep the mechanic while restoring the ability to play it on your own terms. Google worked out that advertisers weren’t actually missing the ability to set channel splits by hand. They were missing the ability to see and explain, so that’s what Google has slowly re-introduced.
The result is that PMax has gotten quite good in specific situations. Not tolerable – actually, legitimately good. In a well-structured account with reliable conversion data and a solid creative library, it beats what most of us were building by hand. I’d rather run PMax with strong inputs today than a hand-built multi-campaign structure from 2019. I don’t say that grudgingly – hell, I’ve made the PMax jokes on stage at many events. But, if we can’t update our prior beliefs when provided with new evidence, we have no business running marketing (or anything else, for that matter).
Meta…Didn’t.
Advantage+ went the other way.
Meta made automation the default instead of a product you opt into. Just this week, I was talking to several well-known Meta media buyers who reported Meta was automatically opting them into various “AI enhancements” or “Ad Features” – even when they specifically toggled those settings off.
Treat that as an operational problem, not a grievance. Screenshot your settings before any test starts, re-check them weekly, and log the date any toggle flips back on. A test that ran across a silent re-enrollment isn’t a test – it’s two different campaigns wearing the same name, and nobody is going to tell you which weeks were which.
That’s just the latest example. Detailed targeting still exists, but Meta has done everything in their power to ensure your interest selections are treated as suggestions rather than constraints. They’ve consolidated interests into segments broad enough that the granular audience combinations a lot of accounts were built on don’t function as filters anymore. Budget allocation keeps consolidating upward. The stated destination is a goal-only system: hand over a budget, a goal and creative; Meta will do the rest.
The list of changes Meta has forced down our collective throats is so long I’ve lost count.
Honestly, Meta may be right on the merits.
The performance data has generally favored Advantage+ over hand-built audience structures. Their models have access to more data + more signals than any human could ever hope to process. Gem, Lattice + Brain have turned creative into a legitimate targeting lever.
The easy move here is to say Meta ran a bad migration, but I don’t think that’s right or fair.
Firaxis operates in a market where users can refund, review-bomb, and go back to the previous version on their local machine. Meta doesn’t. There’s no competitor for that inventory and no route back to 2019 targeting. Compliance is enough when there’s no substitute good, and bridges are expensive. Asking Meta to build one anyway is asking a firm to buy conviction it has no use for.
Which tells you where your leverage actually is. On Google you can negotiate with the interface; the controls came back and more will. On Meta you can only negotiate with evidence. Testing there isn’t optional – it’s the only real lever left.
Practically, that means you should run geo holdouts, because they’re the only instrument left that doesn’t depend on anything Meta reports. Split at the DMA or state level, hold out something you can afford to lose for 1.5 sales cycles, and read the result in total revenue rather than platform-attributed conversions. The limiting factor is cell size: at $35-75K/month, you get 2 cells and a fairly definitive answer, or ~6 cells and a story. Write the hypothesis, the cells and the timeline down up front, because the temptation to call a test early in whichever direction you were already leaning is enormous.
A test you run to win an argument with a rep will disappoint you every time. A test you run to decide where next month’s budget goes pays for itself the first time it changes your mind.
Where Legibility Lives Now
1. Start with a conversion hierarchy that maps to money.
Pull the L90 or L180 by conversion action, and for each one find the downstream rate: how many leads became a qualified opportunity, how many closed, what was each worth. The reality is that anything that can’t produce that number isn’t a conversion, it’s a pageview with ambitions. My working floor is 10%: if fewer than 1 in 10 of a given action ever becomes revenue, it’s teaching the algorithm to find you more of the wrong people, and it’s doing it very efficiently. Demote those to secondary and stop bidding toward them. The cost of doing this is small in real terms, but enormous in reporting terms (you’ll have a LOT less conversions).
2. Build a data passback mechanism that runs on a schedule rather than on demand AND pair it with enhanced conversions and/or server-side tracking.
Export closed revenue from the CRM daily, join it to platform-reported conversions on whatever key you have (GCLID, FBCLID, hashed email, phone), and look at the delta by campaign, geo and device. 2 useful things fall out: (1) you learn where the platform’s numbers are directionally fine and where they’re fantasyland, which tells you exactly how much to trust the platform and (2) you get a revenue number you can put in front of a CFO without a platform logo on it.
3. Once the passback exists, calculate the cross-channel overlap.
If you’re running Meta and Google simultaneously, odds are both are claiming credit for the same conversions. Unfortunately, neither platform has any incentive to tell you how many. Start with the crudest possible version, which requires nothing you don’t already have: sum platform-reported conversions across every channel for a fixed window and divide by what’s in the CRM for that same window. If Google claims 400, Meta claims 300, and 500 things actually happened, then at least 200 conversions – 29% of everything you’re being billed against – are contested. The reality is that the true number of “contested” is likely double that, because the odds that you didn’t get a single customer from organic search/social, referrals, existing customers, friends/family, AI or sheer dumb luck is 0.
From there, your next step depends on what your data looks like. If you have high volume with distinctive order values, match on value plus timestamp and let the amounts act as natural keys. If you have low volume and high value (PI, B2B, home services, senior living, dentistry) – stop engineering and just do it by hand; 100 signed deals in a month is a list you can read, marking which platforms claimed each one. Either way you end up with 4 buckets: (1) Google-only, (2) Meta-only, (3) both and (4) neither. The “both” bucket is the one worth examining first, while the “neither” one sizes the demand arriving without any direct touches from paid channels.
Above roughly 20-25% contested, per-platform ROAS targets stop meaning anything and blended CAC becomes the only number worth managing to, because you’re setting independent goals against a shared pool of demand. Add a self-reported “how did you hear about us” field at intake as a tiebreaker on the contested bucket – it’s biased and imperfect, but it’s the only signal in the entire process not produced by a party with a financial interest in the answer. And (when the money justifies it) the optimal next step is a staggered pause: cut one platform in a subset of geos and watch what happens to the other platform’s reported conversions. If Meta’s numbers sag when you turn off Google, Meta was freeriding on Google’s performance.
4. Keep a holdout running permanently.
Not a test you commission when something looks wrong, but a standing dark region/market you rotate every 3-6 months (or 1.5 sales cycles). This one costs real money and a conversation with the client, which is why you really shouldn’t do it until you have the first 3 squared. It’s also the only artifact on the list that answers the question everyone is actually asking, which is whether any of this is incremental or whether you’re paying to reach people who were already coming .None of the 4 live inside an ad platform, which is the entire point. Legibility you rent from Google or Meta is subject to the platform’s ambitions and quarterly earnings calls. Legibility you own survives the next platform change….and there is always a next product announcement.
The Audit
4 things worth doing this week:
Thing #1:
Separate the levers you lost from the ones you abandoned. Make two lists. Most people are grieving controls that still exist and that they stopped using years ago without noticing (i.e. negatives, exclusions, budget pacing, audience signals, asset group structure, search themes). And if your mental model of PMax is more than a year old, it’s (at best) incomplete and (at worst) downright wrong. Go re-read the controls you wrote off in 2022. If you haven’t given PMax a shot since the Beijing Olympics…its probably time to give it another shot.
Thing #2:
Produce 1 page listing every primary conversion action in the account with its downstream revenue rate next to it. That’s it. Actions that don’t have a strong correlation with the thing you actually care about are – more than likely – causing the algorithm to chase the wrong people. Try the 10% rule from above if you’re not sure where that threshold lives.
Thing #3:
Name 1 thing you were wrong about, and 1 belief that’s currently informing how you invest your budget – then put a date on the second. The first is a credibility exercise, and you have to do it out loud, to your team. For me, it’s broad match: for high-volume accounts with large addressable markets AND a high-quality conversion signal, BM outperforms the EM/PM structures most brands (and agencies) used for years. I’ve long been skeptical of broad match (and I still am in some contexts) – but it’s gotten better and now has genuine utility for some brands. The reality is that nobody trusts the person who thinks everything new is worse. More practically, a team that has never watched its leader revise a position in public learns to brinThe second is a discipline exercise with a deliverable attached.
Every marketing team runs on a handful of foundational assumptions nobody has re-examined since somebody said them once upon a time: brand search is incremental, Meta doesn’t work for B2B, YouTube is a brand play, our CAC target is whatever it was in 2023, older customers are more valuable than younger customers. Some of those are true. The ones that aren’t are incredibly expensive to maintain. So, pick the most expensive one and put it to the test: the belief, what would have to be observed for it to be false, who is going to go look, and the date you’ll hear back. You don’t have to run it yourself – you have to commission it and put it on your own calendar. And if you can’t articulate what it would take to change your mind, you have a conviction, not a strategy…and convictions are fantastically expensive to hold.g you evidence that confirms rather than evidence that’s true, and that’s a particularly expensive kind of failure.
Thing #4:
Build 1 measurement asset outside the platform. A server-side event stream, a CRM-matched revenue report, a media mix model. Anything that lets you check the platform’s homework with data it doesn’t own. This is banking your progress before the Age transition. It’s what carries forward when the next mechanic changes underneath you. I don’t know what will change next, but something will.
Coming Around
At the end of the day, my colleague didn’t come around to Civ7 because anyone won the argument (I certainly didn’t win it). He came around, and now genuinely enjoys the game, because he was allowed to arrive.
That’s how every marketer I know who’s actually good with these tools got there: by running the new thing next to the old thing until the evidence was theirs instead of the platform’s. If you’re still holding out, the tools got better while you were arguing. Refusing to engage with a system that now controls the majority of your spend isn’t a principled, heroic last stand – it arrives at exactly the same place as incompetence, just with fireworks along the way.
And if you’re the one rolling something out this month – a new process, a new platform, an AI workflow your team is dreading – remember what Firaxis needed 15 months and a public beating to learn: being right about the mechanic is worth nothing IF you don’t build the bridge, too. So, build the bridge. Let people keep what they already know how to do. They’ll walk across on their own, and when they do they’ll defend it better than you ever could.
Cheers,
Sam
