Marketing Attribution Models Compared: First-Touch, Last-Touch, and Multi-Touch

A customer sees your Instagram ad on Monday, clicks a Google search result on Wednesday, opens your email on Friday, and buys on Saturday. Which channel gets credit for the sale? The answer depends entirely on which attribution model you use, and each model will give you a different answer that leads to different budget decisions.

This is not a theoretical problem. If you use last-touch attribution, email gets 100% of the credit and you might conclude Instagram ads are not working. If you use first-touch, Instagram gets all the credit and email looks irrelevant. Neither conclusion is true. The customer needed all three touchpoints to convert.

Here is how each attribution model works, when to use it, and what it gets wrong.

The Five Main Attribution Models

1. First-Touch Attribution

How it works: 100% of the conversion credit goes to the first interaction. In our example, Instagram gets full credit.

What it is good for: Understanding which channels drive awareness and bring new people into your funnel. If you are a startup focused on top-of-funnel growth, first-touch tells you where new prospects are discovering you.

What it gets wrong: It completely ignores the middle and bottom of the funnel. Channels that nurture and convert (email, retargeting, branded search) appear to contribute nothing. This leads teams to over-invest in awareness and under-invest in conversion, which is the opposite of what most businesses need.

Best for: Brand awareness campaigns, early-stage companies measuring market reach, content marketing teams measuring discovery.

2. Last-Touch Attribution

How it works: 100% of the conversion credit goes to the last interaction before purchase. Email gets full credit in our example.

What it is good for: Understanding what closes the sale. For short sales cycles (e-commerce impulse purchases), last-touch is often accurate enough because there may only be one or two touchpoints. It is also the default in most analytics tools, which makes it the easiest to implement.

What it gets wrong: It undervalues everything that happened before the final click. Awareness channels (display, social, content) look worthless because they rarely close the sale directly. This leads teams to cut awareness spending, which eventually shrinks the pipeline and makes the closing channels less effective too.

Best for: E-commerce with short purchase cycles, performance marketing teams focused on immediate ROAS, businesses with single-touch purchase paths.

3. Linear Attribution

How it works: Equal credit is split across every touchpoint. In our three-touch example, Instagram, Google, and email each get 33.3% of the credit.

What it is good for: It is the simplest multi-touch model and acknowledges that every interaction matters. It is a reasonable default when you do not have enough data for a data-driven model and want to avoid the extreme bias of first-touch or last-touch.

What it gets wrong: It treats every touchpoint as equally important, which is rarely true. An awareness ad impression is not as influential as a product demo. A retargeting ad is not as influential as a direct referral from a trusted friend. Equal credit is less wrong than single-touch, but it is still inaccurate.

Best for: Businesses with multi-channel strategies and consistent customer journeys, teams moving beyond single-touch for the first time.

4. Time-Decay Attribution

How it works: More credit goes to touchpoints closer to the conversion. In our example, email (last) gets the most credit, Google (middle) gets moderate credit, and Instagram (first) gets the least.

What it is good for: It is a more realistic version of linear attribution. Interactions closer to the purchase decision are generally more influential: a customer who clicks your email on the day they buy was more actively engaged than when they first saw your Instagram ad a week earlier.

What it gets wrong: It still under-credits awareness. The Instagram ad that introduced the customer to your brand may have been the most important single interaction, without it, the other touchpoints would never have happened. Time-decay gives it the least credit. This is fine if your awareness channels are well-established, but dangerous if you are still building brand recognition.

Best for: Businesses with longer sales cycles (B2B, high-consideration purchases), teams that want multi-touch without equal weighting.

5. Data-Driven Attribution

How it works: Uses statistical modeling or machine learning to analyze actual conversion paths and assign credit based on the observed impact of each touchpoint. It compares the paths of customers who converted with the paths of those who did not, and gives more credit to touchpoints that appear more often on converting paths. That is an estimate from path patterns, not a measurement of what each touchpoint caused.

What it is good for: It can be the most accurate of the five when you have enough data. Google Analytics 4 uses data-driven attribution as its default. It adapts to your actual customer behavior rather than applying a fixed formula.

What it gets wrong: It requires significant conversion volume, and the minimum is set by each platform, so check your platform's current documentation. It is also opaque: the model assigns credit, but you cannot easily explain why. And it only sees what it can track, if half your customers opt out of cookies, the model's view of reality is incomplete.

Best for: Businesses with high conversion volume, teams with GA4 properly configured, marketers comfortable with algorithmic models they cannot fully audit.

Side-by-Side Comparison

A Concrete Example

Consider a $200 purchase with this customer journey: Facebook Ad → Blog Post (organic) → Google Ad → Email → Purchase.

Notice how different models would lead to completely different budget decisions. Under last-touch, you might cut Facebook Ads entirely. Under first-touch, you might cut email marketing. Both decisions would hurt your funnel.

Worked Example: Attribution Models Against a Known Truth

The table above shows how models split one purchase. This example shows how they split many, when the true effect of each channel is known. It is simulated, so every number is assumed and invented for teaching. The script runs as written (numpy, pandas, seed 42) and the output below came from it.

In the simulation, 30% of customers already intend to buy and are more likely to be emailed. Facebook and Google ads each truly raise purchase probability (by 2 and 3 percentage points). The blog and email have no effect at all. Touches always occur in the order Facebook, Blog, Google, Email.

import numpy as np
import pandas as pd

# 50,000 simulated customers. Touches always happen in this order:
# Facebook ad -> Blog -> Google ad -> Email. Every number below is assumed.
rng = np.random.default_rng(42)
n = 50_000
intent = rng.random(n) < 0.30                      # hidden: already likely to buy
expo = pd.DataFrame({
    "Facebook": rng.random(n) < 0.50,
    "Blog":     rng.random(n) < 0.30,
    "Google":   rng.random(n) < 0.40,
    "Email":    rng.random(n) < np.where(intent, 0.70, 0.20),   # emailed more often when intent is high
})
lift = {"Facebook": 0.02, "Blog": 0.00, "Google": 0.03, "Email": 0.00}   # true causal lift in P(buy)
p = 0.01 + 0.08 * intent + sum(lift[c] * expo[c] for c in lift)
buy = rng.random(n) < p
journeys = expo[buy]
touched = journeys[journeys.any(axis=1)]           # purchases with at least one tracked touch
print(f"{n} customers, {buy.sum()} purchases, {len(touched)} with a tracked touch")

order = list(lift)
first = touched[order].idxmax(axis=1).value_counts()
last = touched[order[::-1]].idxmax(axis=1).value_counts()
linear = touched.astype(float).div(touched.sum(axis=1), axis=0).sum()
caused = pd.Series({c: lift[c] * expo[c].sum() for c in order})   # expected purchases each channel caused

table = pd.DataFrame({"first-touch": first, "last-touch": last, "linear": linear,
                      "caused (truth)": caused}).reindex(order).fillna(0)
print("share of all purchases, percent:")
print((100 * table / buy.sum()).round(1).to_string())
print("share of purchases that would happen with no marketing:",
      f"{100 * (1 - caused.sum() / buy.sum()):.0f}%")

50000 customers, 2788 purchases, 2661 with a tracked touch
share of all purchases, percent:
          first-touch  last-touch  linear  caused (truth)
Facebook         58.8        10.6    30.6            17.9
Blog             12.8         6.8    13.4             0.0
Google           15.6        29.5    26.9            21.6
Email             8.2        48.5    24.6             0.0
share of purchases that would happen with no marketing: 60%

What the output shows

Last-touch rewards the channel that follows intent. It gives email 48.5% of purchases, yet email caused none of them in this setup: it simply reaches people who were about to buy. Cutting Facebook and Google because email looks best would remove the only channels that moved sales.

First-touch overcredits the channel that comes first. Facebook gets 58.8% of purchases under first-touch against a true 17.9%, because it is the first touch on any journey it appears in, whatever it contributed.

Linear spreads the error around. It gives the blog 13.4% and email 24.6%, where the truth is zero for both, and it credits Facebook 30.6% against 17.9% caused and Google 26.9% against 21.6%. It is less extreme than the single-touch models but is not a measure of cause either.

Most purchases were not caused by marketing here. About 60% would have happened anyway. Every rules-based model hands out all of the credit on the purchases it can see (here 2,661 of 2,788), which is why attribution shares should not be read as incremental impact; a holdout or other experiment is what measures that.

What this does not show. It does not test a data-driven model, and the real world has no known truth to compare against. It shows only that the three rule-based splits can disagree with the cause by a wide margin when intent drives both exposure and purchase.

How to Choose the Right Model

There is no universally correct attribution model. The right choice depends on three things:

1. Your sales cycle length: Short cycle (e-commerce, under 7 days), last-touch is often good enough. Medium cycle (SaaS, 2-8 weeks), time-decay or linear. Long cycle (enterprise B2B, 3-12 months): multi-touch is essential, and you may need to supplement with media mix modeling.
2. Your data volume: With few conversions, as a rule of thumb stick with rules-based models (first, last, linear, time-decay). With more volume a data-driven model becomes an option; check your platform's documentation for its requirements.
3. Your strategic question: "Where should I find new customers?", first-touch. "What closes sales?", last-touch. "How does my whole funnel work?", multi-touch.

The Privacy Problem

All user-level attribution models are getting less accurate. iOS App Tracking Transparency blocks about 75% of app tracking. Third-party cookie deprecation is ongoing. GDPR and similar regulations require consent before tracking. The result: a growing percentage of your customer journeys are invisible to analytics tools.

This is pushing the industry toward two supplementary approaches:

• Media mix modeling (MMM): Uses aggregate channel-level data (total spend vs. total revenue) and statistical regression to estimate each channel's contribution. No user-level tracking required. Works with privacy constraints. MCP Analytics' regression and time-series modules can run this type of analysis on your marketing spend data.
• Incrementality testing: Run controlled experiments: turn a channel off in one market and keep it on in another. Compare the difference in revenue. This is the gold standard for causal measurement but requires enough volume to run meaningful experiments.

The future of attribution is likely a combination: data-driven attribution for what you can track, media mix modeling for the aggregate picture, and incrementality tests to validate both.

Credit is not the same as cause. What incremental impact actually means explains the difference in plain words. If you held out a random group, the free incrementality test tool counts how many customers your campaign caused beyond those who would have converted anyway.

Frequently Asked Questions

Working on this for your business? See the marketing analytics tools.