Every Attribution Model Is Wrong. Choose the One Wrong in Your Favor.

Attribution modeling is the only discipline in marketing where the textbook chapter and the practitioner's experience have almost nothing in common. The textbook presents a menu, first-touch, last-touch, linear, time-decay, position-based, data-driven, as if choosing well were the skill. Twenty years of watching these models allocate credit, and defend budgets, and quietly reshape entire strategies, has taught me the actual skill: understanding what each model is structurally biased toward, so you can read its output the way you read a witness with a motive. Every attribution model is wrong. The professionals are the ones who know exactly how theirs is wrong.

Be clear about what these models do, mechanically. A customer touches your marketing several times before buying: an Instagram ad, a review site, a branded search, an email. Attribution models are rules for splitting the credit for that purchase across the touches the system could observe. That last clause is the first crack in the foundation: the model divides credit among tracked touchpoints, and tracking is exactly what the past five years dismantled. The podcast mention, the friend's recommendation, the ad seen on another device, the Safari session, none of them appear in the ledger, so the model confidently distributes 100 percent of the credit across the maybe-half of the journey it witnessed. Whatever model you choose, you are precisely dividing an inheritance among the relatives who happened to be in the room.

Now the menu, described by its biases rather than its rules. Last-touch, still the quiet default of most reporting, hands credit to the final tracked click, which systematically enriches the channels closest to checkout: brand search, retargeting, email. Run a company on last-touch and you will conclude that harvesting works and planting is waste, then wonder why the harvest shrinks yearly. First-touch inverts the flattery toward whatever introduced the customer, and inherits the opposite blindness. Linear splits credit evenly, which sounds humble and is really just a confession that you do not know, formatted as a spreadsheet. Time-decay weights recent touches, encoding the same harvest bias with smoother math. Position-based hedges across first and last. And data-driven attribution, the current prestige option, uses the platform's machine learning to infer weights, which would be genuinely superior except that the platform computing it is also selling you the media being credited, an arrangement I would summarize the way I did in the number your CFO actually believes: the dashboard is a witness with a commission.

Run the reconciliation once and the abstraction becomes visceral. A retail client's four platforms claimed, in the same month: Meta $410,000 attributed, Google $380,000, TikTok $95,000, email $210,000. Total claimed: $1,095,000. Actual revenue that month: $640,000. The claims exceeded reality by 71 percent, which does not mean the channels did nothing; it means four witnesses each swore they fired the decisive shot. We did not resolve it by choosing a better model. We resolved it by demoting all four dashboards to channel-management instruments and promoting a quarterly holdout calendar to the position of judge. The models kept their jobs. They just stopped testifying.

Here is the reframe that makes all of this usable: stop asking which model is true and start using models as instruments, each pointed at a decision. Comparing keyword-level performance inside paid search? Last-click is fine; the touches are adjacent to the outcome and the comparison is internal. Judging whether upper-funnel investment is creating customers? No click-based model can answer that, whatever its name; that question belongs to incrementality, geo holdouts and matched markets, the machinery I laid out in you do not need a data science team to run a holdout, with marketing mix modeling layered on when spend and history justify it. Diagnosing journey patterns, which channels tend to open relationships versus close them? Compare first-touch and last-touch views of the same data and mine the disagreement; the gap between the two reports is a map of your funnel's division of labor. The model stops being a verdict and becomes a lens, and lenses are allowed to distort as long as you know their curvature.

The practical calibration loop, for teams ready to run it: once a quarter, take your ledger model's implied CPA for a major channel and test it against a geo holdout on the same channel. The ratio between tested truth and modeled claim becomes that channel's discount factor, and you apply it to the dashboard's numbers for the next quarter's decisions. Channels earn credibility multipliers the way suppliers earn payment terms. It is unglamorous, it fits in a spreadsheet, and it converts attribution from a theology debate into a supplier-quality program.

The governance matters more than the mathematics. Pick one model as the official ledger, last-click or data-driven, it matters less than everyone believing there is one ledger, so that channel owners stop shopping for the model that flatters them, the corporate sport I described in most attribution dashboards get built and then ignored. Reconcile that ledger against reality on a schedule: platform-claimed conversions summed across channels will exceed your actual sales, routinely by 50 to 150 percent in my audits, and the size of that gap is itself a metric worth tracking. Reserve your real conviction for tested truths, and let attribution handle what it is actually good at, which is fast, cheap, directionally useful comparison within channels.

So: which attribution model should you use? The one whose bias you can recite from memory, checked quarterly against experiments that owe nothing to any platform, held by an organization that has agreed on a single ledger. Attribution is not a truth machine and was never going to be one. It is bookkeeping for a world you can only partially see, and the marketers it serves best are the ones who never once mistook the books for the business.