Attribution and its limits
Split a month of orders under last-click, first-click and linear rules, say which channel each flatters, and why none, data-driven included, measures cause.
On this page
Attribution takes the orders that happened and divides the credit among the ads each buyer touched on the way. It is bookkeeping: every order is assigned, the shares always add to 100%, and nothing in the method asks how many of those orders would have happened with no ad at all. That second question is the one a budget decision turns on, the "whether it worked" of the five decisions, and attribution cannot answer it.
One buyer, four touches
A buyer clicks a paid-social ad on day 1. On day 4 a podcast host reads the brand's name, and the buyer hears it but clicks nothing. On day 6 the buyer clicks a retargeting banner, and on day 9 types the brand name into a search engine, clicks the paid result at the top and places a $900 order.
The common rules give four different answers for the same order:
- Last-click gives all $900 to brand search, the final click before the order.
- First-click gives all $900 to paid social, the click that started the path.
- Linear splits it across the clicks: $300 each to social, retargeting and brand search.
- Data-driven learns the split from many paths. It compares how often paths that include a touch end in an order with how often similar paths without it do, and divides the credit in proportion to the difference.
None of them gives the podcast a dollar, because a rule built on clicks cannot see a touch that produced none, even if the podcast is why the buyer searched the name on day 9.
One month, three rules
Scale the buyer up. Assume a brand's 1,000 orders in a month arrive along three paths, with $40,000 spent on paid social, $10,000 on brand-name search ads and $10,000 on retargeting:
- 400 orders: a social click, then a brand search
- 300 orders: a social click, then a retargeting click
- 300 orders: a brand search and nothing else
| Channel | Spend | Last-click | First-click | Linear |
|---|---|---|---|---|
| Paid social | $40,000 | 0 | 700 | 350 |
| Brand search | $10,000 | 700 | 300 | 500 |
| Retargeting | $10,000 | 300 | 0 | 150 |
| Total | $60,000 | 1,000 | 1,000 | 1,000 |
Under last-click, brand search costs $10,000 ÷ 700 ≈ $14 an order, retargeting $33, and social has no orders at all, so the obvious move is to cut it. Under first-click, social costs $40,000 ÷ 700 ≈ $57 and retargeting looks useless. Linear lands in between: social $114, brand search $20, retargeting $67. The rules disagree about the split, and agree about one thing that is false: every column sums to 1,000, so every order was caused by some ad.
Why the rules reward the last touch
Two channels sit at the end of paths by construction. Retargeting shows ads only to people who have already visited the site, so its audience has shown intent before it spends a dollar. Brand search shows ads only to people who already typed the brand's name. Both are close to the order because of who they reach, not because of what they do, and a rule that rewards position rewards them.
eBay tested this directly (Thomas Blake, Chris Nosko and Steven Tadelis, "Consumer Heterogeneity and Paid Search Effectiveness: A Large-Scale Field Experiment", Econometrica, 2015). It stopped bidding on every search containing its own name on two search engines while continuing on a third as the control. Almost all of the click traffic it gave up arrived anyway, through the unpaid link beneath where the ad had been: the paid clicks were people already on their way. The same paper switched off non-brand search ads in a random 30% of eBay's US traffic, chosen by geography, for 60 days. The ads moved new and infrequent buyers, but the frequent buyers who accounted for most of the ad spend bought just as much without them. Regressing sales on search spend put the return at over 1,400%; the experiment put it at −63%, with a 95% interval of −124% to −3%.
The exception is a competitor bidding on your name, where the paid slot defends a click that would otherwise go to them. Test that claim region by region rather than assume it.
Data-driven is better bookkeeping, not cause
A data-driven rule looks more rigorous because it learns from thousands of paths. Its comparison is still between people, not between treatments. The buyers whose paths include a retargeting banner had already visited the site; the buyers whose paths do not include one mostly had not. Their conversion rates differ mainly because they are different people, and the model reads that difference as the banner's effect. It also sees only paths that ended in a recorded touch: the order from someone who heard the podcast and typed the address is credited to "direct", and the orders that would have arrived with no advertising at all are still divided among the ads. It allocates the 1,000 orders more cleverly. Its total is still 1,000.
A second leak sits in the platforms' own reports. Each counts an order if its ad was clicked, or merely shown, within its own window. Assume social claims 700 orders from clicks and views, search claims 700 and retargeting 450: 1,850 claimed for 1,000 orders. A dashboard that adds platform-reported conversions overstates by 85%, and the platform with the longest window wins the argument.
What attribution is good for
Comparing like with like inside one channel. Two creatives in the same campaign, two keyword groups, this week against last: the bias is roughly the same on both sides, so the difference is informative even though neither level is. It also catches a broken tracking tag the day it breaks, and it paces daily spend. That is operations, and attribution is the only instrument fast enough for it.
It cannot compare channels for a budget decision, because the bias differs by channel: largest for brand search and retargeting, smallest for channels whose effect arrives days later through other touches. That comparison needs a holdout or a geo test, and the test's result can then scale the day-to-day numbers, which is how the measurement plan uses all three instruments together.
In practice
To the agency or channel owner reporting a cost per order, three questions: "Under which rule? What share of these buyers had already visited the site or searched our name? What happens to orders in regions where we switch this off for four weeks?" Ask for the credit under two rules side by side. A channel whose orders swing from 0 to 700 between rules is one whose value this report cannot tell you.
To the board, call attributed orders "credited", never "driven by", and show the incremental figure beside them wherever a test has produced one.
The mistake that costs money is moving budget to the channel with the lowest last-click cost per order. That is nearly always brand search or retargeting, which harvest demand other channels created. Cut the channels that create it and, a month later, the harvesters have less to harvest and their cost per order rises too.
Check yourself
1. Retargeting reports 300 orders at $33 each. Assume 60% of those buyers would have ordered anyway; they had already filled a basket. What is its incremental cost per order, and do you keep the $10,000?
300 × 40% = 120 incremental orders, so $10,000 ÷ 120 ≈ $83 an order, two and a half times the reported figure. Keep it only if an order is worth more than $83 to you, and replace the 60% with a measurement: withhold the banner from a random tenth of the retargeting audience for four weeks and count their orders in your own data.
2. The agency proposes moving $5,000 from social to brand search, because brand search shows $14 an order and social shows none. Do you approve it?
No. Brand-search ads already appear on nearly every search for the name, so $5,000 more buys almost nothing new; its spend is capped by how many people search the name. And 400 of its 700 credited orders started with a social click, so cutting social by an eighth shrinks the searches brand search depends on. Expect fewer orders in total, with both channels' reported costs looking worse in a month.
3. A data-driven model moves social from 0% of the credit under last-click to 45%. The vendor says this proves social works. Do you raise social's budget on this evidence?
No. The model re-divided the same 1,000 orders; it did not count how many would vanish without social. A 45% share is consistent with social causing 450 orders or 50. Raise the budget only after a holdout puts a number on the orders social causes, and use the 45% for comparing social's campaigns with each other.