Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

Analytics › Intermediate

Comparing Attribution Models on Sample Conversion Paths

An exercise in applying rule-based attribution models by hand to a small set of invented conversion paths, comparing the credit each channel earns and explaining which channel each model flatters. Doing the arithmetic yourself shows why a model is a lens and not a measurement, and why budget questions need a test as well.
ExerciseIntermediate levelModule 2: Attribution and discrepancies5 minUpdated

Why this matters

Reading that Attribution models disagree is not the same as watching a channel's figure swing when you change the rule. The attribution models comparison describes each rule and where it misleads; the attribution models framework frames the choice as a decision about what you want to see. Doing the sums by hand makes both concrete, and it trains the habit the sample campaign dataset teaches: compute ratios from totals, never by averaging ratios.

Rule-based models are also what you will meet in exported path data and third-party tools, since Google has retired most of them from its own reporting. Knowing how to apply them yourself is a working skill, not history.

The sample paths

Fieldfare Running Shoes is an invented online shoe shop. Its analyst exported six converting paths that together show the shop's typical patterns. Each touch is a recorded click or visit within the lookback window, in order, and the last touch is the session in which the order was placed.

Illustrative example — not a real result

Fieldfare's six converting paths and the spend behind them

PathTouches in orderOrder value
APaid social → Organic search → Brand search£120
BPaid social → Email → Email → Brand search£100
COrganic search → Brand search£80
DAffiliate£90
EPaid social → Organic search → Affiliate£150
FEmail → Brand search£60

Total revenue across the paths: £600. Spend allocated to this sample: paid social £150, brand search £60, affiliate commission £30. Email and organic search carry no media cost in this exercise.

Applying each model

  • Last click: all the order value to the final touch.
  • First click: all the order value to the first touch.
  • Linear: an equal share to every touch, so a channel that appears twice on one path earns two shares.
  • Position-based: the common configuration, with a large share to the first and last touch each and the remainder spread evenly across the middle. A path of two touches splits evenly between them; a path of one touch gives that touch everything. The exact shares are in the worked answer.

Apply the model to each path first, then total by channel.

Task

  1. Credit each path under each of the four models and record the credit per channel per path.
  2. Total the credit by channel for each model and check that every model's total equals the total revenue.
  3. Compute return on ad spend (ROAS) for paid social and brand search under each model, from the totals.
  4. For each model, name the channel it flatters and the channel it penalises, and say why in one sentence.
  5. Write two sentences for the shop's owner on what these figures can and cannot tell her about moving budget from paid social to brand search.

Worked answer

The credit by channel, after applying each model path by path:

Illustrative example — not a real result

Credit by channel under four models

Position-based here gives 40% each to the first and last touch and splits 20% across the middle; path B's middle share of £20 is split across its two email touches.

ChannelLast clickFirst clickLinearPosition-based
Paid social£0£370£115£148
Organic search£0£80£130£94
Email£0£60£80£50
Brand search£360£0£135£158
Affiliate£240£90£140£150
Total£600£600£600£600

Linear on path A gives £40 to each of three touches; on path B, £25 to each of four touches, so email earns £50 there. Position-based on path A gives £48 to paid social, £24 to organic search and £48 to brand search.

Return on ad spend from the totals:

ChannelLast clickFirst clickLinearPosition-based
Paid social (£150 spend)0.002.470.770.99
Brand search (£60 spend)6.000.002.252.63
Affiliate (£30 commission)8.003.004.675.00

What each model flatters:

  • Last click flatters brand search and the affiliate, because they sit at the end of paths that other channels started. Paid social, which opened three of the six paths, earns nothing.
  • First click flatters paid social, because it introduced half the paths, and gives brand search nothing even though it closed four orders.
  • Linear flatters channels that appear often, which here lifts organic search and email from nothing to a real share. It treats the second email on path B as worth as much as the decisive closing click.
  • Position-based flatters the bookends: paid social and brand search both do well, while email, mostly in the middle, ends with less than under linear.

The affiliate is the case to notice. On path E it closed an order that paid social and organic search had built, and last click hands it the whole value. That is the pattern the guide to platform and analytics discrepancies warns about from the other direction: the affiliate network and the paid social platform would each claim path E in their own reports, so their totals together exceed the shop's revenue.

For the owner: the figures show that paid social mostly starts journeys and brand search mostly finishes them, so judging either on a single rule will undervalue it. They cannot show whether brand search orders would have happened without the ad, or whether paid social creates demand rather than meeting it; only an Incrementality test, such as pausing a channel in some regions, answers that. A model is a lens on recorded paths, and every model above is a different lens on the same six orders.

Common mistakes

  • Totalling before splitting. Applying linear shares to channel totals instead of to each path gives different and meaningless figures.
  • Forgetting repeated touches. A channel that appears twice on a path earns two shares under linear.
  • Averaging per-path return on spend. Sum the credit and the spend, then divide.
  • Calling one model correct. Each answers a different question; the attribution models framework asks which decision the number is for.
  • Moving budget on the comparison alone. The comparison points to a question; a test answers it.
  • Treating a sample as the business. Six paths teach the method; real comparisons use the full export.

Next, test the same ideas against discrepancies in the attribution and discrepancies checkpoint, or return to the intermediate analytics level.

Pages this lesson applies

Terms used on this page

Also related

Related