Attribution Is Astrology for Business People

I need to tell you about the worst forty-seven minutes of my professional life. It involves a pair of navy blue dress pants, a hotel iron, an auditorium full of marketers, and the most comprehensive structural failure a garment has ever suffered while its owner was explaining customer journey mapping to a room full of people who were too polite to mention that they could see his underwear.

I'm telling you this because I need you to understand something about attribution modeling. But first you need to understand something about my pants.

The Pants

I bought these pants in 2019. They fit beautifully in 2019. In 2019 I was a younger man who went to the gym occasionally and had what doctors call a "normal" relationship with carbohydrates. I was the kind of person who could buy pants at a store, try them on once, nod at himself in the mirror, and wear them for years without thinking about it.

By the time of the conference in question, it was no longer 2019. Several things had happened between 2019 and the date I stood on that stage. A global pandemic. A prolonged work-from-home period during which I discovered that sweatpants have an elastic waistband and dress pants do not, and that this distinction matters more than most people appreciate. A relationship with sourdough bread that I would describe as intimate. A gradual, imperceptible expansion of the region where my body stores its mistakes.

The pants still fit. Technically. In the same way that a sausage casing technically fits around a sausage. Everything was contained. The button closed. The zipper went up. If you looked at me from the front, you would see a man in dress pants. You would not see a man in dress pants that were engaged in a silent structural negotiation with the forces of physics, a negotiation that the pants were losing by millimeters every time I sat down or bent over or breathed in a way that could be described as "full."

I knew this. I knew the pants were tight. I knew I should buy new pants. I did not buy new pants because buying new pants would require admitting that the old pants no longer fit, and admitting that the old pants no longer fit would require confronting a series of lifestyle choices that I was not prepared to confront. So instead I wore the pants. I wore them and I told myself they fit. This is a decision that will be relevant later.

The Iron

The morning of my presentation, I was in a hotel room in a city I will not name, looking at a pair of pants that had been folded in a suitcase for two days and had the wrinkles to prove it. The wrinkles were primarily concentrated in the seat area, which in retrospect feels like the universe trying to draw my attention to the part of my pants that was about to betray me.

Hotel irons are instruments of destruction designed by people who have never ironed anything. They have two settings: "not hot enough to remove any wrinkle" and "hot enough to melt a space shuttle heat shield." There is no middle ground. The temperature dial is a fiction. You set it to "medium" and it goes directly to "surface of the sun" because the thermostat was calibrated in 2006 and has not been serviced since.

I ironed my pants. I held the iron on the seat area, where the worst wrinkles were, for what I estimated was a reasonable amount of time. What I did not know, because I am not a textile engineer, is that sustained high heat weakens the fibers of cotton-blend fabric. What I was doing, without realizing it, was pre-cooking the structural integrity of the exact square footage of pants that would be under the most stress when I sat down. I was essentially scoring a perforation line across the back of my pants like you'd score a piece of bread dough before putting it in the oven, so it knows where to split.

The wrinkles came out. The pants looked crisp. I put them on. The button closed. The zipper went up. I looked at myself in the hotel mirror and thought, not bad. Not bad at all.

The pants had approximately ninety minutes left to live.

The Stage

The conference was one of those mid-tier marketing events where the coffee is free and the chairs are uncomfortable and everyone has a lanyard and a tote bag they'll throw away at the airport. I was presenting on a panel about digital marketing measurement. Four panelists, a moderator, an audience of about three hundred people, and a set of chairs on stage that I need to describe to you because they are important to what happens next.

The chairs were those modern design chairs. You know the ones. Low to the ground, with a shallow seat and no meaningful back support, designed by someone who values aesthetics over the structural realities of the human pelvis. The kind of chair that requires you to essentially perform a deep squat to sit in and a minor athletic feat to stand up from. The kind of chair that assumes you are a twenty-three-year-old Scandinavian with the hip flexibility of a gymnast, not a middle-aged SEO consultant in pants that are engaged in an ongoing negotiation with entropy.

I did not sit in these chairs immediately. First there was the presentation portion. I stood at the podium. I had slides. I was talking about customer journey mapping and the importance of understanding which touchpoints matter in the conversion funnel, which is ironic in a way that will become apparent shortly. My slides were good. My delivery was solid. I was having, I thought, a good day.

Fifteen minutes into my talk, I dropped my clicker. The slide-advancing remote. It slipped out of my hand and clattered to the floor and rolled slightly under the podium, because of course it did, because small plastic objects are governed by the same malevolent physics that makes toast land butter-side down.

So I squatted to pick it up. A full, deep squat, right there on stage, in front of three hundred people, in a pair of pants that were already stretched to their operational limits by a body that had expanded since 2019 and had been further weakened by a hotel iron that morning.

I heard nothing. I felt nothing. I picked up the clicker, stood back up, and continued my presentation. Everything seemed fine. The audience was attentive. Nobody gasped. Nobody pointed. Nobody made any facial expression that would suggest something catastrophic had just happened to the rear of my pants.

But something had happened. Whether it was a partial tear or the initial failure of the heat-weakened fibers or just the fabric reaching the outer limit of what it could accommodate, something had begun. The structural compromise was underway. The first domino had fallen. I just didn't know it yet.

The Chair

After the presentation came the panel discussion. The moderator invited us to take our seats. The four of us approached the tiny designer chairs, and I lowered myself into mine with the careful deliberation of a man docking a spacecraft, bending my knees, shifting my weight back, descending slowly into a seated position that put maximum stress on the exact part of my pants that had been pre-weakened by the iron and pre-stressed by the squat.

I felt it this time. A sensation. Not a sound, exactly, but a feeling. A release of tension. A shift in the relationship between my body and my clothing. The kind of subtle mechanical event that, if you were an engineer monitoring a bridge, would make you look up from your instruments and say "hm." But I was not monitoring a bridge. I was on a stage answering a question about multi-channel attribution, and the sensation was brief, and I was focused on sounding intelligent, so I filed it away under "the chair is uncomfortable" and moved on.

I was on that stage for another twenty minutes. I crossed my legs. I uncrossed them. I leaned forward to make a point about first-party data. I leaned back to listen to another panelist talk about Google Analytics. I shifted in my tiny chair repeatedly, because the chair was designed for looking at and not for sitting in, and every shift applied new and creative forces to the compromised zone of my pants.

The audience was engaged. People were nodding. Someone in the front row was taking notes. The moderator asked me a follow-up question about attribution windows and I gave what I thought was a quite good answer about the arbitrary nature of lookback periods. At one point I gestured broadly to emphasize a point about the customer journey, and I think this gesture may have involved a slight forward lean and twist that accelerated the ongoing situation in my pants, but I cannot be sure because I was too busy being eloquent about marketing measurement to notice that my pants had given up.

The Photo

The panel ended. I stood up from the tiny chair, which required a kind of rocking, forward-lunging motion that I'm sure did nothing to improve the situation. I shook hands with the moderator. I spoke briefly with an attendee who wanted to ask about UTM parameters. I walked to the networking area. I got a coffee. I stood around for fifteen minutes having conversations with professional contacts, facing them, like a normal person, completely unaware.

My phone buzzed. A notification. Someone had tagged me in a photo on LinkedIn with the caption "Great panel on attribution at [conference name]! Insightful discussion from @amosweiskopf and team."

The photo was taken from the audience, slightly to the side. It showed the four panelists in their tiny chairs. It showed me, in profile, leaning forward to make what was probably my point about lookback periods.

It also showed a tear in my pants that ran from approximately the belt line to approximately my upper thigh. A tear so comprehensive that it was less a tear and more a referendum on the structural viability of the entire garment. Through this tear, visible to anyone with functioning eyes and a line of sight to the stage, was my underwear.

My underwear was plaid. Red and black plaid. The kind of bold pattern choice that is perfectly acceptable when it exists beneath clothing that is intact, and becomes significantly less acceptable when it is displayed to an auditorium of marketing professionals through a gaping wound in the seat of your dress pants.

I stared at the photo. I zoomed in. I zoomed out. I zoomed back in. I performed the analysis you perform when you see something you desperately hope is a shadow or a trick of the light or some artifact of digital photography, and you slowly come to accept that it is none of those things. It is your underwear. On a stage. On LinkedIn.

The Silence

Here is the part that still haunts me. I was on that stage for approximately forty-seven minutes after the initial squat. Forty-seven minutes during which the tear either existed in its full glory or was actively expanding. Forty-seven minutes during which three hundred people had a clear view of my plaid situation.

The moderator sat three feet to my left. Another panelist sat three feet to my right. The front row was maybe eight feet away. The conference had professional videographers. There were photographers. There were people live-tweeting. Three hundred sentient human beings with eyes and opinions and, presumably, some basic instinct toward human solidarity, and not one of them said a word.

Not a tap on the shoulder. Not a whispered "hey, you might want to..." Not a note passed on a napkin. Not a meaningful look. Not a cough. Nothing. Three hundred people watched my underwear make its public debut and collectively decided that the polite thing to do was to say absolutely nothing and let me continue discussing multi-touch attribution while my pants told a different story about what "exposure" means in marketing.

I will never know exactly when the rip reached its full extent. I will never know how many people saw it, or when they saw it, or what they thought when they saw it. I only know that by the time I saw the LinkedIn photo, the rip was total, my underwear was on display, and the only person at the entire conference who didn't know about it was me.

The Forensics

I spent that evening in my hotel room, alone, with the ruined pants on the bed, trying to reconstruct what had happened. I was doing forensics on my own trousers. I was trying to determine, with some precision, the exact moment and cause of the failure. I had become a detective investigating a crime scene where the crime was hubris and the victim was my dignity.

Was it the squat? When I bent down to pick up the clicker, was that the moment the fabric gave way? The squat put maximum stress on the weakened area. It was the most violent single motion I'd made in the pants. If you had to pick one moment, one decisive event, you'd pick the squat.

But was it really the squat? Or was it the iron? The iron had weakened the fibers that morning. Without the iron, the fabric might have held. The squat was the trigger, sure, but the iron created the conditions for the trigger to matter.

But was it really the iron? Or was it the weight gain? The pants were tight. They'd been tight for months. Every day I wore them was another day of sustained stress on every seam. The iron weakened the fabric, sure, but the iron wouldn't have mattered if the fabric hadn't already been under extraordinary load.

But was it really the weight gain? Or was it the sourdough? Or the pandemic? Or the decision to work from home in sweatpants for eighteen months? Or the decision to buy these specific pants in 2019 from a brand that, in hindsight, maybe cuts their inseams a little aggressively? Or the genetic predisposition toward carrying weight in exactly the region that puts maximum stress on exactly the seam that matters?

I was lying on a hotel bed, holding a pair of destroyed pants, running attribution analysis on my own wardrobe malfunction. Every causal chain I followed led to a different root cause. Every root cause led to a different conclusion about what I should have done differently. And every conclusion was simultaneously correct and completely useless.

And that's when I understood attribution modeling.

The Model You Choose Is the Story You Tell

In marketing, attribution is the practice of figuring out which touchpoint in the customer journey deserves credit for the conversion. A customer sees a Facebook ad, clicks a Google search result, reads a blog post, gets an email, and then buys your product. Which of those interactions caused the sale? Who gets credit? Where should you spend your next dollar?

The answer depends entirely on which attribution model you use. And this is where it gets beautiful, in the same way that watching a building collapse in slow motion is beautiful, which is to say: it's horrifying, but you can't look away.

Last-click attribution gives all credit to the last touchpoint before conversion. In my case, this would blame the tiny designer chair. The chair is what I was sitting in when the tear reached its full potential. The chair is the last thing that happened before the catastrophe became total. If you're a last-click person, the chair did it. Fire the chair. Get different chairs. Problem solved.

Most of Google Analytics ran on last-click for years, and most marketers still think in last-click terms because it's simple and it gives you someone to blame. It's also completely wrong, because the chair didn't cause anything. The chair just happened to be there when the inevitable became actual. But it shows up in the data, so it gets the credit.

First-click attribution gives all credit to the first touchpoint. This would blame the 2019 pant purchase. The original sin. I bought pants that would eventually not fit, and everything that followed was a consequence of that first interaction. If you're a first-click person, the solution is obvious: buy better pants. Invest in the top of the funnel. Get the initial touchpoint right and the rest takes care of itself.

This is also completely wrong, because I could have bought those same pants and not gained weight, or gained weight and bought new pants, or kept the pants and not ironed them, or ironed them and not squatted on stage. The first click started the journey but it didn't determine the destination. Millions of people bought pants in 2019 and did not end up on a stage in plaid underwear.

Linear attribution gives equal credit to every touchpoint. The purchase, the weight gain, the iron, the squat, and the chair each get twenty percent of the blame. This feels fair. It feels democratic. It also tells you absolutely nothing about what to do differently, because twenty percent of five different things is a shrug dressed up as a pie chart.

Time-decay attribution gives more credit to touchpoints closer to the conversion. The chair gets the most blame, the squat gets the second most, the iron gets some, the weight gain gets a little, and the 2019 purchase gets almost none. This is like saying that the reason my pants ripped was mostly the chair and a little bit everything else, which manages to be both mathematically sophisticated and profoundly unhelpful at the same time.

Data-driven attribution is what Google now pushes in GA4. It uses machine learning to analyze your specific data and determine which touchpoints actually matter most. It's a black box. You feed it data and it gives you answers and you cannot see the math. You just have to trust it. You have to trust that Google's algorithm, built by Google, running on Google's platform, processing data collected by Google, will give you an unbiased assessment of which channels deserve credit, including Google's own channels.

This is like asking the tiny chair manufacturer to conduct an independent investigation into what caused my pants to rip. I'm sure their findings will be thorough and objective.

The CMO's Horoscope

I have sat in meetings where a CMO presented an attribution report to the board with the confidence of someone delivering a papal decree. Forty-seven slides. Charts in four colors. A waterfall diagram showing exactly how each channel contributed to pipeline. GA4 says paid search drove 40% of conversions. HubSpot says email drove 35%. Salesforce says the sales team drove 60%. The numbers add up to 135%, which should be impossible, but the board nodded, the CFO asked one question, the CMO answered with a number that had two decimal places, and everyone moved on.

The number was made up. Not in the sense that someone fabricated it. In the sense that it was produced by a model that was chosen because it told a story that the CMO wanted to tell, and the story the CMO wanted to tell was that the CMO's marketing strategy was working. The CMO didn't pick the attribution model based on statistical rigor. The CMO picked the attribution model that made the CMO's channels look good. Google's tool says Google is important. Facebook's tool says Facebook is important. The CRM says the sales team is important. What a coincidence that everyone's tool agrees that the tool's owner is the hero. This is what everyone does. This is the entire industry.

A horoscope works the same way. You read your horoscope and it says "A new opportunity is coming" and you think about the email you got yesterday and you think yeah, that's the opportunity, the horoscope was right. You don't think about the nine other things that happened that day that could also be "a new opportunity." You don't think about the fact that "a new opportunity" is so vague that it applies to literally everyone on earth at all times. You just see the match, confirm the bias, and move on with your day feeling like the stars understand you.

Attribution reports are horoscopes for business people. They're vague enough to match whatever narrative you're selling, specific enough to look like science, and wrong in a way that nobody can prove because the counterfactual doesn't exist. You can't rewind the customer journey and remove one touchpoint and see what happens. You can't A/B test the past. You can only look at the data through a model and the model tells you whatever the model was designed to tell you.

And just like astrology, the practitioners have an answer for everything. If the model was wrong, the data was dirty. If the data was clean, the attribution window was too short. If the attribution window was right, there were "offline touchpoints" that weren't captured. There's always an excuse. The model is never wrong. You just didn't read it correctly.

The Moderator Problem

Remember the moderator. Three feet away from me for forty-seven minutes. Clear line of sight to the situation. Said nothing.

Your analytics platform is the moderator. It's sitting right next to the data. It has a clear view of what's happening. And it says nothing useful, because saying something useful would be uncomfortable and might disrupt the panel. The moderator's job is to keep things moving, not to tap you on the shoulder and tell you that your measurement strategy is showing its underwear to the entire organization.

Three hundred people in that audience knew something I didn't. They had information I needed. And they kept it to themselves because telling me would have been awkward, and not telling me was easy, and the social cost of speaking up was higher than the social cost of letting a man walk around a networking event with his plaid boxers on display.

This is the analytics industry in one image. Everyone can see that the models don't work. Everyone knows that the numbers are made up. The attribution vendors know. The platform companies know. The consultants know. The analysts know. Nobody says anything because the entire ecosystem runs on the polite fiction that attribution works, and disrupting that fiction would be awkward, and not disrupting it is easy, and there's money to be made in selling models and dashboards and workshops to people who would rather have a wrong answer than no answer at all.

What You Should Actually Do

The honest answer is that you should stop pretending you know which touchpoint caused the conversion. You don't. Nobody does. The customer journey is a mess of overlapping signals and the idea that you can isolate one touchpoint and assign it credit is about as realistic as isolating one reason my pants ripped and assigning it blame.

But nobody likes the honest answer, so here's the practical one:

Run incrementality tests. Instead of trying to figure out which channel gets credit, turn a channel off for a while and see what happens. If you stop running Facebook ads for two weeks and nothing changes, Facebook wasn't doing what the attribution model said it was doing. This is the equivalent of me wearing different pants to the next conference and seeing if anything rips. It's not sophisticated, but it tells you something real.

Look at holdout groups. Take a segment of your audience and don't show them the ad, or don't send them the email, and compare their behavior to the group that got it. This is actual causal analysis instead of correlational storytelling. It's harder and more expensive and it requires patience, which is why almost nobody does it.

Consider media mix modeling. MMM doesn't try to trace individual user journeys. It looks at aggregate spend and aggregate outcomes over time and uses statistical regression to estimate channel impact. It's what the sophisticated teams are moving to because it doesn't depend on cookies, doesn't need user-level tracking, and doesn't pretend to know what caused any single conversion. It has its own problems -- it needs years of data, it's slow, and it can't tell you what happened last week -- but at least it's honest about what it doesn't know.

Stop optimizing for the model. If your team is making decisions to "improve" attribution numbers rather than to improve actual business outcomes, you've confused the map for the territory. The attribution model is a map. A bad map. Drawn by someone who's never been to the territory. Stop navigating by it.

Accept the uncertainty. The most sophisticated measurement professionals I know are the ones who are most comfortable saying "we don't know." They run experiments. They look at trends. They make educated guesses. They don't build forty-seven-slide decks full of decimal points to make uncertainty look like precision.

The Moral

I bought new pants. Not because I figured out which factor caused the rip, but because I figured out that it didn't matter. The pants were going to rip eventually. The weight gain and the heat-weakened fabric and the aggressive squat and the terrible chair all contributed, and picking one to blame would have been satisfying and useless. The useful thing was to buy pants that fit and stop pretending the old ones were fine.

Your attribution model is a pair of pants that doesn't fit. You know it doesn't fit. You wear it anyway because buying a new one would require admitting that the old one is wrong, and admitting the old one is wrong would require confronting a series of decisions you're not prepared to confront. So you keep wearing it. And you keep presenting the data. And nobody in the room tells you that your measurement strategy is showing its underwear to the whole company, because that would be awkward, and not saying anything is easy.

The next time someone presents an attribution report with two decimal places and a waterfall chart and the quiet confidence of someone who has mistaken correlation for causation, I want you to think of me. Standing on a stage. Talking about the customer journey. In pants that had been compromised fifteen minutes earlier by a squat I didn't think twice about, in fabric that had been weakened that morning by an iron I didn't know how to use, wrapped around a body that had been expanding for three years while I told myself nothing had changed.

Everything in that story was true at the same time. Every causal explanation was valid and incomplete. And the only person who didn't know what was really going on was me, the one standing in front of the room, delivering my insights with great confidence and plaid boxer shorts.

I have since invested in pants with a more forgiving fabric composition and a higher thread count. I recommend a similar approach to your analytics infrastructure.

Although I cannot guarantee the underwear situation will improve.