What attribution has to prove

On April 26, 2021, Apple released iOS 14.5 and made IDFA access opt-in. Within a year, fewer than a quarter of U.S. iPhone users had opted in. Meta later told analysts the change would cost roughly $10 billion in 2022 ad revenue.

The public story was that Apple broke attribution.

The better story is simpler: Apple removed one identifier, and the industry discovered how much of "attribution" depended on that identifier being free.

Most attribution tools do not prove what operators actually need to know. They collect touchpoints, pick a model, and assign credit. That can be useful for reporting. It is not the same as proving that a company action caused a customer to change state: from visitor to buyer, buyer to repeat buyer, subscriber to churned, churned to returned.

The four questions

An attribution claim has to answer four questions in order.

1. Identity. Is this the same person across sessions, devices, emails, orders, and payments? A verified email, login, or billing record is stronger than a cookie or device match. The dashboard should show the confidence of the join.

2. Touchpoint. What did the company actually do? An ad click, page view, email send, sales call, discount offer, landing-page test, or lifecycle flow should be recorded with a source, time, and raw evidence. A UTM parameter is useful, but it is not the whole proof.

3. State. What changed in the customer's relationship with the business? A pixel conversion is not the same as a paid order. A form submit is not the same as a customer. The durable state change usually lives in Shopify, Stripe, the CRM, or the subscription system.

4. Cause. Did the touchpoint cause the state change? Last-click, first-click, linear, time-decay, Shapley, and Markov models distribute credit. They do not create a counterfactual. Causal claims need holdouts, geo tests, audience tests, incrementality tests, or a clearly labeled model.

If a system skips one of these questions, the final number may still be useful. It just should not be treated as proof.

Why model choice is not the main issue

Attribution vendors often compete on model choice. First touch. Last touch. Multi-touch. Shapley. Markov. MMM. Incrementality. Each has a place.

The bigger issue is that the same conversion path can produce very different channel credits depending on the model. Published comparisons have found differences of more than 30% per channel. The model often chooses the answer.

That is why attribution should label the type of claim:

• observed touchpoint
• deterministic revenue join
• modeled credit
• causal lift from an experiment

Those should not be flattened into one number. A budget decision made on a geo-holdout result is different from a budget decision made on modeled last-click revenue.

The budget meeting problem

Attribution matters because someone has to move money.

A growth lead wants to scale a campaign. The media buyer sees platform ROAS above target. The finance lead sees net margin below target. The lifecycle lead says email touched many of the same customers. The agency says the platform is undercounting. The founder asks whether to add another $50,000 to spend this week.

The argument is rarely about whether attribution is useful. It is about which kind of evidence should be trusted for this decision.

For daily campaign management, platform data can be useful. It is fast and directional. For monthly reporting, observed order revenue may be better. For budget reallocation, the team needs spend, net revenue, margin, and a clear window. For board-level claims, the team may need incrementality or a holdout. These are different jobs. One attribution number cannot do all of them.

The mistake is treating attribution as a single answer. It is better to treat it as a ladder:

• Did the touchpoint happen?
• Did the customer later produce revenue?
• Can we join the touchpoint and revenue with enough confidence?
• Did the channel get credit under the selected model?
• Did an experiment show that the channel caused incremental lift?

Each step is more valuable and harder to prove. A good system does not hide the step. It tells the operator what kind of claim they are looking at.

Where ecommerce teams feel this

E-commerce teams usually have the first two layers and struggle with the next two.

They know the ad clicked. They know the UTM. They may know the page viewed. But revenue is in Shopify or Stripe. Refunds may be delayed. Subscription value may come later. Email and paid media may both touch the same buyer. Platform reports model missing conversions. Finance wants net revenue and margin, not pixel events.

So the team gets conflicting answers:

• Meta says the campaign worked.
• GA4 says it was organic or direct.
• Triple Whale says something else.
• Shopify shows orders but not the full journey.
• Finance asks which number is real.

Attribution becomes an argument instead of a decision tool.

The new standard

Attribution should be useful before it is perfect.

That means the system should separate four layers that are often mixed together:

Observed touchpoints. What the customer saw or did: click, page view, email open, SMS click, quiz completion, sales call, checkout start.

Joined revenue. Which orders, charges, refunds, subscriptions, or repeat purchases can be tied back to that customer.

Assigned credit. How the system split credit across channels under a selected model.

Measured lift. Whether a controlled test, holdout, geo split, or audience split showed incremental impact.

Operators can still use modeled credit. They just need to know it is modeled. They can still use last-click. They just need to know what it excludes. They can still use platform ROAS. They just need to know whether finance can reconcile the revenue. The right operating standard is not one perfect attribution model. It is honest labels on every attribution claim.

What Lyberty does

Lyberty stores attribution as a chain of evidence, not just a dashboard number.

First-touch campaign, source, and medium are captured on the tracked user. Touchpoints stay tied to that user. Revenue joins through order and payment data when a durable customer signal exists. Channel metrics like CAC, MER, LTV/CAC, payback, and revenue by campaign only render when the spend and revenue inputs verify. Recommendations can say "not enough signal" instead of forcing a low-confidence answer.

Lyberty also keeps cause separate from correlation. A channel result computed from observed revenue is labeled differently from a result produced by an experiment or holdout. The operator can still use the number. They can also see what kind of number it is.

What to ask any attribution tool

1. Which identity joins are deterministic, and which are modeled?
2. Where is revenue joined: pixel, order, payment, or bank?
3. Can I inspect the raw touchpoints behind a channel number?
4. Which results are causal and which are correlational?
5. What happens when the system cannot prove the number?

If the answer is another model selector, the tool is not solving the real problem.

Sources

1. iOS 14.5 Offers App Tracking Transparency. Apple, April 26, 2021. https://www.apple.com/newsroom/2021/04/ios-14-5-offers-app-tracking-transparency-and-more-on-iphone-and-ipad/
2. Meta Q4 2021 Earnings Call Transcript. Meta Platforms, February 2, 2022. https://s21.q4cdn.com/399680738/files/doc_financials/2021/q4/Meta-12.31.2021-Exhibit-99.1-FINAL.pdf
3. A new path for Privacy Sandbox on the web. Google Privacy Sandbox, July 22, 2024. https://privacysandbox.com/news/privacy-sandbox-update/
4. About Mail Privacy Protection in iOS 15. Apple Support, September 20, 2021. https://support.apple.com/en-us/HT212614
5. iOS 17 Link Tracking Protection. Apple Developer, September 18, 2023. https://developer.apple.com/videos/play/wwdc2023/10053/
6. Sunset of Universal Analytics. Google Analytics Help, July 1, 2023. https://support.google.com/analytics/answer/11583528
7. Twilio to Acquire Segment for $3.2 Billion. Twilio, October 12, 2020. https://investors.twilio.com/news/news-details/2020/Twilio-Signs-Definitive-Agreement-to-Acquire-Segment/default.aspx
8. Amplitude Begins Trading on Nasdaq. Amplitude Investor Relations, September 28, 2021. https://investors.amplitude.com/news-releases/news-release-details/amplitude-begins-trading-nasdaq-following-direct-listing
9. AppLovin Completes Acquisition of Adjust. AppLovin, October 20, 2021. https://www.applovin.com/blog/applovin-completes-acquisition-of-adjust/
10. Triple Whale Raises $52M Series B. Triple Whale, September 13, 2022. https://www.triplewhale.com/blog/triple-whale-series-b
11. Server-side Tagging in Google Tag Manager. Google Tag Manager. https://developers.google.com/tag-platform/tag-manager/server-side
12. About the Conversions API. Meta for Developers. https://developers.facebook.com/docs/marketing-api/conversions-api/
13. Attribution Modeling Increases Efficiency of Bidding in Display Advertising. Diemert et al., KDD 2017. https://arxiv.org/abs/1707.06409
14. Project Robyn: Marketing Mix Modeling Open Source. Meta Open Source. https://facebookexperimental.github.io/Robyn/
15. Meridian: Open-Source Marketing Mix Modeling. Google, 2024. https://developers.google.com/meridian
16. Marketing Attribution Software Market Size & Share Report, 2024-2030. Grand View Research, 2024. https://www.grandviewresearch.com/industry-analysis/marketing-attribution-software-market-report