Which Attribution Model Should You Use? What Changes When You Switch

Introduction

Most people reach attribution modeling after an argument. Two channels claim the same order. Changing the model changes channel rankings.

The obvious question is which model is correct. That question has no answer. The better question is which model answers what you are trying to decide.

Read the definitions of all the attribution models at Advanced Attribution Models: Moving Beyond Last-Click. What follows assumes you know roughly what each one does.

What does an attribution model decide?

A model takes the list of touches that led to an order and decides how much credit each touch will get. That is the whole job.

Two things follow from this that people get wrong.

1. A model cannot add information. It can only divide what is already recorded. Say a journey really had five touches and your data caught two. Every model will split the order across those two touches, and every model will be wrong.
   
   

2. A model does not tell you what caused the sale. A touch is a record that something happened before someone bought. First click tells you what was there at the start. It does not tell you the order would have failed without it. That is a different question, and it needs different methods.

What happens to your channel report when you change attribution models?

Here is one month for one brand, run through three models. Same 6,360 orders. Same seven channels. Only the credit rule changed.

Two things stand out.

1. The best channel changes. Google Ads is the clear leader under first click with 1,880 orders. Under last click, it falls to second behind Google Search. Anyone reporting "our biggest channel" is reporting a dropdown setting.
   
   

2. Google Ads loses 660 orders when we change models. Under first click, it holds 1,880 of the 6,360 orders, which is 29.6% of the total. Under last click, it holds 1,220 which is 19.2%. A third of the channel's credit disappears because we changed the models.

That matters if you budget on rules like "Google Ads should drive a quarter of our orders." Under one model, it clears the bar. Under the other, it misses, and someone proposes a cut.

Why your best channel changes depending on the attribution model?

The swing is not random. It shows you where each channel sits in the journey.

Channels that fall when you move from first click to last click are opening journeys. The creator program drops 53%. Google Ads drops 35%.

Channels that rise are closing them. Google Search goes from 860 to 1,390, a rise of 62%. Someone who searches your brand name and buys was sent there by something else. Search caught the order. Search did not create the demand.

Some channels barely move. Meta Ads sits at 800 and 720. Email sits at 680 and 700. A flat channel is either doing both jobs at once, or it shows up on short journeys where the first touch and the last touch are the same touch.

A single model tells you how big a channel is. The gap between two models tells you what that channel does.

Under linear the biggest line is Direct at 1,890 orders, or 29.7% of the total. You cannot buy Direct. It is a leftover bucket that is assigned to any order that can't be cleanly attributed to any channel. Linear spreads credit evenly across every touch, so it always inflates whatever bucket shows up most often. In most stores that bucket is Direct.

Which attribution model to use for which decision?

The attribution model you pick depends on the question you're trying to answer

• Finding new customers. Use first click. You want to know what starts journeys, and every other model undercredits the channels that do this. Any channel that looks weak on last click and strong on first click is bringing people in. Cut it and you break your own discovery.
  
  

• Keeping or killing a closing channel. Use last click. Retargeting, branded search and cart-abandonment email all exist to convert demand that already exists. Last click is the model most generous to them, so if one of them looks weak even under last click, it is genuinely weak. If Direct is swallowing a large share of your last touches, use last non-direct instead, which skips Direct and credits the last channel you can actually name.
  
  

• Splitting next quarter's budget. Use a multi-touch model. Budget decisions involve channels that never get the first or last touch, and single-touch models make those channels disappear. Linear splits credit evenly and is the simplest way to find mid-journey channels you have been ignoring. Position-based weights the two ends and suits brands where discovery and conversion are run by different teams. Data-driven learns the split from your own data and needs a lot of orders before it learns anything. Time decay favours recent touches and only fits short buying cycles, promotions and flash sales.

Whichever you pick, check how much credit lands in buckets you cannot spend against.

• Choosing between two creatives or two campaigns. Use whatever the ad platform uses. Platform numbers are unreliable for comparing across channels. They are fine for comparing two things inside one channel, because the same bias applies to both.
  
  

• Understanding how much your channels overlap. Use any touch. This one is not an allocation model and is not meant to total to your order count. It credits every channel present in a journey.
  
  

• Deciding whether a channel is worth anything at all. No model answers this. Attribution records touches. Whether the order needed the touch is a causal question, and the methods that answer it are covered separately.

Why last click is your default, and why that is not a decision you made?

Most stores run last click or last non-direct without ever choosing it. Shopify, Meta and most ad platforms ship with a last-touch default, and defaults survive.

Last click became the default because it is easy to compute and hard to argue with. It needs one touch and no journey rebuilt. Every other model needs the full path, and most systems do not hold the full path reliably.

The cost is that last click always rewards whatever sits closest to the purchase. On this brand's numbers, Google Search and Direct together take 2,530 orders under last click, or 39.8% of the total. Both are simply converting the demand that something else created.

If you have never changed the setting, you are running the model that was easiest to build. That can still be the right call.

When is a custom attribution model worth building?

Custom models come up once a team has tried the standard set and none of them fit. One of three things usually triggers it.

1. An unusual journey shape. Long consideration cycles, high repeat rates or a big offline component all break assumptions built into off-the-shelf models.
   
   

2. A channel that matters more than its credit. Creator programs, podcasts and referral schemes tend to sit in the middle of journeys, and every standard model underweights the middle.
   
   

3. A decision you make over and over. If you reallocate budget across eight channels every month, a model built for that decision can earn its cost.

Custom is worth less than people expect. Custom weights applied to incomplete journeys gives you a more specific version of the same data. Build the data first, then the model. Most teams skip that order.

How to test attribution models on your own data before you commit

Four checks.

1. Run first click and last click side by side for one month. Sort your channels by the difference between the two. Anything that moves more than 20% has a job in the journey you should know about before you set budgets again.
   
   

2. Count your channels per order. Divide total channel touches by order count. Under 1.3 and model choice barely affects you. Over 2.0 and it shapes your whole report.
   
   

3. Check where Direct lands under each model. If Direct grows as you move toward multi-touch models, your journey data is thin, and the model is pushing credit into a bucket you cannot act on.

Check the totals. Any allocation model should total to your actual order count for the period.

Frequently Asked Questions