ArticleIssue #163 | The High Cost of Perfection
The past few weeks have been wildly busy, between launching ChatGPT ads for some clients, to growing the team, to getting ready for a few weeks of conferences. In the midst of all that, I had a conversation with a prospect experiencing a significant amount of fraud in their account, which got me thinking about a similar situation I had years ago, and the non-intuitive result that changed how I think about fraud altogether.
This week, I want to tell you about the client who went to war with ad fraud… and won. And in winning, nearly cut the business in half.
The lesson from this ordeal is one that almost no one talks about, and applies far more broadly than just ad accounts.
The War on Ad Fraud
The client was spending around $500k/month on ads (mostly paid search/paid social) – which was reliably generating $3M+ in profit each month (yes, profit – not revenue, not pipeline, actual dollars at JP Morgan Chase). By almost any measure, things were going quite well.
But there was one thing that kept this from being a “runaway success” account. One thing that the client just could not get past: the fraud in the account.
It took many forms: Bot clicks. Form fills from real homeowners and prospects who, once the call center got in touch, swore they didn’t even have a need for our client, much less had they actually contacted the company. Phantom leads from real (and quite expensive) clicks that went to dead numbers, wasting both the call center + sales team’s time. Massive, out-of-the-blue spikes in impressions in certain zip codes, which resulted in no clicks but skewed expected CTR (which, in turn, harmed quality score + made each click everywhere else more expensive).
The company – from the CEO on down – was beyond frustrated. They (rightfully) believed the fraud was holding back the account and wanted it eliminated, by whatever means necessary. Before we go further, a disclaimer: I’ve seen many people claim that unqualified clicks or bad phone calls or forms driven by legitimate keywords seeking a different product/service are fraud. That’s not what we’re talking about, in this case. This was real fraud, with real costs, targeting a real account.
At some point, everyone was fed up with the situation, so we all agreed to do whatever it took to eliminate it. We deployed every tool in the proverbial toolbox (including inventing some!) to make it happen. No holds barred.
We rolled out low-budget bait campaigns to identify zip codes where impression and fraud spikes were concentrated, then connected them to scripts to exclude those zips on a rolling basis. We analyzed server logs to identify the common behavioral patterns of spoofed leads (forms from real people, with valid contact information, that swore they had never searched for or contacted the company), created custom audiences that matched those patterns, then excluded them across the board. Aggressively-excluded low-quality and logged-out users. Wrote custom scripts to detect and flag suspect traffic patterns in near real-time. Ran probabilistic models to identify if/when/how attack patterns were shifting and adapted to them. Ran holdout tests to detect differences in traffic patterns, then used those gaps to exclude high-probability fraud. Deployed bulk IP blocks to find and remove traffic from server farms and other highly-suspect sources. Added massive numbers of negative keywords.
And…it worked. Within months of implementing the solution, we had eliminated the overwhelming majority of fraud. Spend dropped substantially (just over 35%) as fraud clicks ceased. Spoofed forms vanished. The call center reported entire days without a single fraudulent call. Website traffic quality increased markedly.
By almost every metric most marketers use, this was a runaway success.
Except for one: profit.
Actual money in the bank from paid advertising dropped nearly 50% in the month that followed full deployment. At first, we all (including the client) assumed it was an anomaly. A bit of bad luck, the natural results of Google + Meta recalibrating based on the massive number of new constraints added, a few of the top salespeople at the company being out sick.
Then the second month. Same results. Spend continued to decline – down 40%. Profit still down just under 50%. Week after week, the story was the same. We started digging into the call logs, the form submissions and the sales reports to understand what was going on.
How In The World?
The thing about any kind of “precision” is that it is fundamentally pixelated. Every measurement system – from the ones we use to analyze the physical universe to those we use to manage a Google Ads account – has a minimum resolution: a point below which you cannot subdivide further without destroying the information you’re trying to isolate. Physicists call this the Planck length: the scale at which space itself becomes granular and the very act of measuring distorts what you’re measuring.
Applied to digital advertising (generally) and this situation specifically, the equivalent is the point at which your fraud filters can no longer distinguish a bot from a buyer. The behavioral patterns overlap, the signals converge and your system, designed to remove noise, starts eliminating the very signal it was designed to preserve. Worse, the filters don’t just fail to distinguish – they actively constrain what the algorithm can learn from and optimize toward. When that happens, you don’t just lose resolution; you actively collapse the possibility space.
Put another way: In the process of removing fraud, we unintentionally removed real buyers. The systems we built to identify and exclude bad traffic couldn’t perfectly distinguish between a bot and a legitimate prospect who happened to share an IP range, a zip code, or a behavioral pattern with the fraudulent traffic. And so, every exclusion that caught 100 bots or fraudsters also caught 10 real prospective customers. Multiply that across thousands of exclusions, and the collateral damage compounds fast.
The immediate reaction from another expert was that this fraud wasn’t real – maybe it was made up. Maybe all of us – so convinced it was there – started seeing patterns in the data that weren’t there. So, we went back, pulled the data going back years – and found exactly what we thought we would: the fraud was real. The cost of that fraud was (very, very) real. But in the process of eliminating that very real, very expensive fraud, we also eliminated real buyers. The number of legitimate, high-quality leads had declined since implementing the changes.
The Optimization Cost Curve
This is a pattern that shows up everywhere in performance marketing, but we don’t talk about it enough.
In every system with noise – fraud, unqualified leads, wasted impressions, low-quality calls, return-happy customers, churned subscribers – there’s a temptation to pursue purity and perfection. To eliminate every bad lead, filter out every imperfect click, remove every low-quality user.
Candidly, that pursuit feels both logical and righteous. What CEO or CFO is OK paying for fraudulent clicks? What agency wants to hear about low-quality leads? What brand wants customers who buy, only to immediately return? How many salespeople or call center employees want to spend their days calling spoofed lead after spoofed lead?
But the lesson we learned is that optimization has a cost curve that bends the wrong way past a certain point. We refer to it as the Perfection Premium: the escalating cost of each incremental unit of purity/perfection, which scales exponentially at precisely the moment where it feels most justified.
The first 50% of waste you remove is almost free: obvious bots, clearly fake submissions, data center traffic, job applicants. This is the cleanup that makes everyone feel smart. It costs almost nothing to implement and the savings are immediate and visible. Psychologically, this success is the most dangerous because it validates the thesis (there’s fraud + we can eliminate it at next-to-no cost) and motivates everyone to go further.
The next 30% is where the economics start to shift. Logged-out user exclusions, expansive negative keyword frameworks, geo-level blocks, aggressive audience suppression. Each one is defensible in isolation. Each one removes some real waste. But each one also starts to narrow the aperture the algorithm is working through, and in doing so, reduces the system’s ability to find buyers in unexpected places. The cost here isn’t just the implementation; it’s the opportunity cost of the prospects/buyers/subscribers you’ll never see because the system was never allowed to reach them.
The final 20% is where you kill the patient to cure the disease. At this stage, the Venn diagram between “fraudulent signal” and “legitimate but unconventional buyer behavior” is a circle. The filters can’t tell them apart because, at this resolution, they look identical. You’re no longer removing waste. You’re removing statistical variance… and variance is where outsized returns live.
There’s a misconception among almost everyone in marketing that waste and signal are cleanly separable populations. They’re not. They share behavioral characteristics, IP ranges, geos, device types, session patterns. The further you push toward zero waste, the more real buyers you catch in the net. And unlike the fraud, those lost buyers don’t show up in a report. What happens is something far more nefarious: they just vanish. Fewer qualified calls or leads. Fewer sales. Fewer subscribers. Fraud is loud. It announces itself. It’s memorable. This is the opposite of that – it’s quiet. It is absence.
The question should never have been “Can we eliminate fraud?” – of course we can. It should have been: “At what point does the cost of eliminating fraud exceed the cost of the fraud itself?”
That’s the million-dollar question that almost no one asks.
When we did, we all – us, the client, the rest of their team – agreed to reverse most of the exclusions.
Within days, the call center reported that some of the fraud was back. Spoofed leads returned. The number of hangup calls increased. The quality of the website traffic (time on site, events per session, etc.) started to decline. Not to the levels it had been at before, but far above the levels it was at with the exclusions in place.
Less than a month after making the changes, spend was down 15% vs. previous – but profit was down less than 10% from the original baseline. From an efficiency standpoint, the account is actually more efficient than it was when we started – which is exactly what the Perfection Premium suggests should be the case. We kept only those “initial 50%” and a few of the “Next 30%” restrictions, which had a Pareto impact: they were good enough to eliminate 80% of the fraud.
The client still has that non-eliminated 20% of fraud in the account. They likely always will. That’s a trade we’re all now willing to make, because at the end of the day, the scoreboard that matters most to the client is the one at JP Morgan Chase.
Perfection is a seductive goal, especially when it comes with the bonus of eliminating waste + fraud. But, as we found out, the optimal account isn’t the cleanest one. It’s the most profitable one. And those are rarely the same thing.
Where This Goes Wrong in Every Channel
I’ve watched this same pattern play out in dozens of accounts, across almost every channel:
Search: Negative keyword lists built so aggressively they block real users, with real needs, who happen to not use the exact right words when trying to find a solution. An account I audited last year had excluded 60% of all converting queries, simply because they didn’t conform to what the brand believed were “correct”. Their agency was so focused on “cleaning” the account they cut off the oxygen supply.
Paid social: So many audience + placement exclusions added that the campaign can’t find enough people to spend against. The efficiency is fantastic (their ROAS is over 12; breakeven is a 2.6) – but the scale is gone.
Email: Suppression lists that are so aggressive, re-engagement campaigns can’t reach lapsed customers who want to buy. At one point, half of their list wasn’t getting calendar sends about new product drops or sales/events, simply because they were enrolled in another flow and the account had restrictions suppressing profiles in active flows from getting calendar sends.
Teams & Talent: I did some consulting for a company that used click trackers, screen monitors and productivity scoring to squeeze every minute of idle time from their workforce…only to discover that “idle” time was where their best people did their best thinking. Case in point: this company had an employee who spent an hour of “company time” reading a series of (seemingly) unrelated articles and profiles, found a common thread, and applied it to a client account. The result? An entirely new, non-obvious, wildly profitable audience unlocked…and a policy violation that landed her some quality time with HR. That ended with her leaving the company because they optimized so hard for visible outputs that they eliminated the conditions that yielded disproportionate returns. The same cost curve applies: the last 20% of “wasted” employee time you “recapture” through these tactics often costs more than the total value of savings from the entire endeavor. In this case, the cost was the brand’s single-most-valuable employee. Oops.
The pattern is consistent: somewhere between “good enough” and “perfect,” you cross a line where the cost of the last increment of purity exceeds the actual margin/value you’re protecting. Most teams never even see that line because they’re laser-focused cleanliness, to the point where profitability falls by the wayside.
If you’re curious if this is happening to you, pull up one account – whether it’s search, social, or email – and look at your exclusions. IP blocks, audience exclusions, negative keywords, suppression lists, whatever you’ve got.
Ask yourself one question: are these exclusions reducing cost, or reducing revenue/profit/leads?
If you can’t answer that with data, you don’t actually know whether your “clean” account is working for you or against you.
Some level of waste is the price of scale. The goal isn’t a perfect account. It’s a profitable one.
Cheers,
Sam
