Marketing — the Atlas · ch.29 · attribution
📣 Chapter 29 · Part V · Growth & measurement

What would have happened anyway?

Attribution hands out credit for the sales it can see. Incrementality asks the ruder question — how many of those sales you'd have gotten for free. This chapter is the war between those two answers, and the armistice that ends it.

Here's the whole chapter in one line: credit is not causation — the only way to know what your spend caused is to compare against a world where you didn't spend, and everything short of that comparison is storytelling with decimals. The rest is how to buy that comparison, and what to do with the dashboards in the meantime.

1The wasted half

John Wanamaker, department-store magnate, is credited with marketing's most durable joke: "Half the money I spend on advertising is wasted; the trouble is, I don't know which half." For a century that line was accepted fate — you bought the page, you rang the register, and the connection between them was a matter of faith and folklore.

Then digital arrived and promised to end the joke. Every impression logged. Every click trailed. Every conversion beaconed back to the ad that "drove" it. Two decades and trillions of logged events later, here is the strange part: the honest answer to "which half" is still contested — inside the most instrumented industry in history.

The reason is that Wanamaker's question was never a data question. It's a counterfactual: what would have happened if you hadn't spent? And no log file, however complete, contains the world where you didn't run the ad. A programmer knows this shape: you can't diff against a branch you never ran. If you want the comparison, you have to construct it — deliberately, expensively, by withholding the treatment from someone.

The click trail tells you who bought after seeing your ads. It cannot tell you who bought because of them. "After" is free and everywhere. "Because" costs a control group. Everything in this chapter follows from the gap between those two words — a gap Chapter 15's coupon-lineage measured with holdouts a century ago and click-land spent twenty years forgetting.

The wasted half as accepted fate.
The line every ad budget was approved under · attributed to John Wanamaker
the stateMedia bought on faith, effect measured in anecdote — the register rings, and everyone claims the ring.
the folkloreMeasurement meant recall surveys and gut feel; the wasted half was a joke because nobody expected an answer.
the ironyThe direct-response wing already had the tool — keyed coupons and split runs — but brand advertising declared itself unmeasurable and moved on.
the legacyA century of budgets defended by charisma. The question survived intact, waiting for a control group.
Turn it off in Denver. Read the ledger.
Geo holdouts as the counterfactual, purchased quarterly · 2026
the movePause (or boost) spend in matched markets, leave the rest alone, and measure the difference in actual sales — the branch you never ran, finally run.
why nowGeo tests need no cookies, no user IDs, no platform pixels — they survived the privacy purge that killed the click trail's credibility.
the findingRun honestly, they keep returning the same verdict: the dashboard flattered the spend. Sometimes by a little. Often by multiples.
the answerWanamaker's half is findable now. The reason it stays unfound is rarely the statistics — it's the meeting where the smaller number has to be said out loud.

2Attribution's family tree

An attribution model is a rule for dividing conversion credit among the touches you observed. A customer saw a display ad, watched a social video, clicked a brand search result, clicked a cart email, and bought. Who gets the sale?

  • Last-click — the default for two decades: full credit to the final click before purchase. Simple, decisive, and structurally biased toward whatever stands nearest the register.
  • First-touch — full credit to the touch that started the trail. The romantic inverse, blind in the other direction.
  • Linear — equal shares to every touch. The diplomatic settlement: nobody's angry, nothing's learned.
  • Time-decay — weight rising toward the purchase, governed by a half-life you choose. A tunable dial pretending to be a discovery: pick the half-life, pick the winner.
  • Position-based ("U-shaped") — openings and closings split the prize, the middle priced at a rounding error.
  • Data-driven multi-touch — Shapley-style credit sharing fitted on the observed journeys. The most sophisticated way to re-divide the same information.

Notice what changing the model does: it redraws the map without adding territory. Same touches, same one purchase — the credit moves, the knowledge doesn't. And every member of the family inherits the same congenital flaw: exposure is not random. Ads are aimed at people judged likely to buy; retargeting literally selects people who carted your product yesterday. When the targeting works, the ad's audience was already converging on the register — and every model in the tree hands the ad credit for gravity.

Programmer's version: attribution is feature importance computed on correlated features. The retargeting pixel isn't a cause of purchase; it's a feature that leaks the label. Any model that only sees the logged journey will reward the leak — elegantly, reproducibly, and wrong.

Try it. One journey, five models — then run the test none of the models can imagine.

Interactive · the credit machine One journey · five ways to divide the same sale
model
Pick a model. The credit re-divides; the sale stays exactly one sale.
💡
An ad that finds people about to buy and an ad that makes people buy produce identical log lines. The click trail cannot tell them apart — not with a better model, not with more data, not ever. Only a withheld group can. That single sentence is most of this chapter; everything else is plumbing.

3Grading their own homework

The second structural problem isn't the models — it's who runs them. Each ad platform measures its own conversions, with its own pixel, under its own rules, and reports its own grade. Three habits of that self-grading are worth knowing by name:

Click windows. A platform typically claims any purchase within some window of a click — seven days, twenty-eight days. Buy a fortnight after clicking anything, and that click's owner books the revenue.

View-through credit. Stricter still: many defaults claim purchases that happen within a day of an ad merely being seen. Scroll past a promoted post at breakfast, buy at lunch because your friend recommended it last month — the platform's dashboard records an ad-driven conversion, no click required.

Overlapping juries. Every platform runs these rules independently, over the same customers. One purchase can sit inside a search click window, a social view window, and an email click window at once — and each dashboard counts it, whole.

The result is an arithmetic scandal hiding in plain sight: sum your platforms' reported conversions and you will routinely have "sold" more than you sold. Each number is defensible under its own rules; the total is fiction. And the incentive gradient underneath is not subtle — the party reporting the grade is the party selling the next semester's tuition.

⚠️
The one-afternoon audit. Pull one quarter's platform-reported conversions across every channel. Add them up. Divide by actual orders from your commerce backend. Numbers well above 100% are common — teams discover they've been shown credit for 130–180% of their own sales. Nothing about that ratio says which platform is lying; it says the sum is not a number about the world. If your budget meeting treats it as one, Chapter 19's dashboard illusion is running the company.

4The control group returns

So how do you buy the word "because"? You reintroduce the thing the click trail skipped: a group that didn't get the ad. Incrementality is just that comparison, and it comes in three practical sizes:

User-level lift tests. The platform randomly withholds your ad from a slice of the target audience. The cleanest versions use ghost ads: the system records the auctions your ad would have won in the control group, without serving it — so both groups have identical targeted intent, and the difference in purchases is your lift, not your targeting. This is the experiment attribution silently pretends to be.

Geo experiments. Randomize markets instead of people: pause or boost spend in matched regions and read the difference in actual sales. Coarser, slower — and unkillable, because they need no user tracking at all. As the tracking substrate erodes (Chapter 33's story), the geo test has quietly become the gold standard again: Hopkins' split run, upgraded with better statistics and worse weather.

Always-on holdouts. A small permanent control — a few percent of users or one quiet market — that never sees a given channel. Less precise than a designed test, but it turns incrementality from an annual event into a dashboard of its own.

And what do these instruments keep finding, decade after decade, platform after platform? The same verdict: attributed ROAS overstates incremental ROAS, often by multiples — and the overstatement is worst exactly where the targeting is sharpest. Retargeting and brand-term search sit at the top of the flattery league every time, for the reason Section 2 predicted: the sharper the targeting, the more the audience was coming anyway. (Chapter 15 told you what eBay found when it finally ran the holdout on brand search; that result was a preview of this entire chapter.)

📈
Evidence check. The published record is unusually one-sided. Platform-run conversion-lift programs (Meta and Google have both published on theirs) keep finding that click-based attribution misallocates: it overcredits bottom-funnel, tightly-targeted tactics — retargeting lift studies commonly measure true incremental effects at a small fraction of the attributed figure — and undercredits upper-funnel exposure that rarely earns the last click. Independent academic audits (the eBay brand-search holdout most famously) found some attributed channels near zero incrementality. Numbers vary by study and category; the direction almost never does. Meanwhile geo experimentation has moved from academic exotica to standard platform tooling — the post-cookie world's measurement of record.
Interactive · the ghost-ad lab 90,000 exposed · 10,000 held out · same intent by construction
2.0%
+15%
Set a baseline and a true lift. The dashboard sees every exposed buyer; the holdout sees what they'd have done anyway.

5What the speedometer is for

At this point the tempting conclusion is: attribution lies, delete the pixel. Wrong lesson. Attribution has three virtues no experiment can match — it's cheap, it's instant, and it's always on. Experiments are the opposite: slow, expensive, and chunky. You get a lift number per quarter, not per creative per hour.

So use each instrument for what it is. Attribution is a speedometer, not an audit. It is genuinely good at questions where its biases hold still: which of two creatives is pulling harder in the same channel, at the same funnel depth, on the same audience; whether a campaign died overnight; whether the tracking itself broke; how spend is pacing against plan. In those comparisons the intent-correlation cancels out, and the cheap, instant number is the right tool.

Where it fails is exactly where the stakes are highest: comparing across channels with different intent profiles. Retargeting versus prospecting video is not a fair race in click-land and never will be — one of the runners started at the finish line. Chapter 19 showed what happens when that race sets the budget: the dashboard illusion, institutionalized.

The operating rule fits in one line: attribution compares like with like; experiments compare against nothing at all — and only the second answers "was it worth it?"

6Triangulation — the armistice

The measurement war ends not with a winner but with a treaty, and the treaty has a name the industry has mostly settled on: triangulation. Three instruments, three questions, three altitudes:

  • Experiments answer "did this cause sales?" — the truth instrument. Coarse, quarterly, and final. They don't scale to every decision; they don't have to. Their job is to be the ground truth everything else is checked against.
  • Marketing-mix modeling answers "how should the budget split?" — the allocation instrument. Top-down econometrics on outcomes, channel-level and privacy-proof. It gets the whole next chapter; for now, know its role: it sets the portfolio.
  • Attribution answers "what do I tweak today?" — the steering instrument. Ad-level, real-time, biased, and fine, because its job is no longer truth. It optimizes within the budgets the other two set.

The triangle's edges matter as much as its corners: experiments calibrate the MMM (a lift test pins the model's curve for a channel to reality), the MMM allocates across channels, and attribution steers inside them. When the three disagree — they will, constantly — the disagreement is information about altitude, not error: the speedometer, the map, and the odometer are describing the same trip at different scales.

Run a campaign through all three lenses and watch them argue.

Interactive · the triangulation table One spend line · three instruments · pick a scenario
scenario
Pick a spend line and compare the three verdicts.
The customer journey, fully credited.
MTA vendors and the perfect-data dream · mid-2010s
the promiseStitch every impression, click, and device into one journey graph, and algorithmically assign every dollar its true parent.
the physicsThe graph was never complete — walled gardens don't share logs — and no completeness could fix it: re-dividing observed credit still isn't a counterfactual.
the endPrivacy regulation and signal loss took the raw material away; the category quietly pivoted or folded.
the lessonMore resolution on the wrong question is still the wrong question.
Three instruments, each in its lane.
The post-signal settlement · 2026
the stackExperiments for truth (quarterly), MMM for allocation (monthly), attribution for steering (daily) — disagreement expected, by design.
the edgesLift tests calibrate the MMM's curves; the MMM sets channel budgets; attribution optimizes creative and pacing inside them.
why it wonIt's the only stack that survives cookie loss: geo tests and econometrics need no user-level data at all.
the cultureThe hard part was never the math — it was demoting the flattering number from verdict to speedometer.

7The practitioner's protocol

Compress the war into a working protocol and it's four moves, run on a calendar:

  • 1 · Rank your spend by flattery risk. The sharper the targeting, the higher the odds the dashboard is crediting gravity. Brand-term search and retargeting go to the top of the suspicion list — not because they're bad tactics, but because their attributed numbers are the least trustworthy in the account.
  • 2 · Buy one counterfactual a quarter. A geo test or conversion-lift study on the biggest suspicious line. One honest experiment per quarter compounds into a measurement culture faster than any tooling purchase.
  • 3 · Keep the wasted-half ledger. Per channel: attributed ROAS next to last-measured incremental ROAS, with a date. The ratio between them is your personal flattery index — it changes slowly, so even stale tests keep correcting fresh dashboards.
  • 4 · Fix the meeting, not just the math. Decide in advance which number is allowed to move budgets. If the answer is "the 8× in the deck" and not "the 1.4× from the holdout," the org has chosen comfort over cash, and no measurement stack can help it.

That last move is the real war. Attribution numbers are big, flattering, and arrive daily; incrementality numbers are small, rude, and arrive quarterly. Every incentive in the building points at the big number — the agency's fee justification, the channel manager's bonus, the platform's next pitch. The organizations that get this right aren't the ones with the fanciest models; they're the ones where the CFO heard the smaller number from marketing first.

One instrument in the triangle still needs its story told: the old econometric workhorse that measures marketing without tracking anyone — dusted off, re-fitted, and suddenly the most fashionable model in the building. Next chapter.

Check yourself

3 questions · instant feedback 0 / 3