Performance-Based Compensation

Where an attribution argument turns into a contract argument

Performance compensation looks like the cleanest arrangement in marketing services. The agency is paid on results, the client pays for outcomes, and the incentive is aligned. Then the parties fall out and discover that the contract's operative term — return on ad spend, cost per acquisition, revenue attributed to paid media, qualified leads — is not a measurement of anything that happened in the world. It is the output of a system that assigns credit, and that system is configurable by whichever party holds administrative access to the account.

Two layers sit underneath the number in the cell. The first is the attribution model, which is a rule for dividing credit among interactions on a path. Google's own documentation defines attribution models as determining "how credit for conversions is assigned" — an allocation rule, not a finding of cause. The same clicks, the same conversions and the same revenue produce very different per-channel figures under different rules, and nothing about the underlying events changes when the rule changes. The second layer is that some of the conversions being allocated were never observed at all: platforms model conversions for cross-device journeys, for browsers and consent states where identifiers are unavailable, and for app traffic through platform APIs.

So a compensation clause referencing platform-reported return on ad spend is referencing a modeled estimate passed through a credit-allocation rule, neither of which is disclosed on the face of the export. That is not an argument against performance compensation. It is the reason these clauses need to specify the measurement system with the same care the parties give the percentage.

The mechanism, stated concretely

The forensic question in these matters is narrow: could the reported metric have moved without the underlying business changing, and does the record show that it did? Several documented mechanisms do exactly that.

• Changing the attribution model on a conversion action reallocates credit between channels and campaigns without a single additional sale.
• Changing the conversion window — the period after a click during which a conversion still counts — moves conversions into or out of the measured period.
• Adding or removing conversion actions from the "conversions" column changes the numerator directly. Adding a low-value action such as a page view or a phone-number click can transform a cost-per-acquisition figure overnight.
• Counting "all conversions" rather than "conversions", which includes view-through and secondary actions, produces a materially different number from the same account.
• Time zone. The account time zone and the reporting time zone are separate settings, and a period boundary can move revenue between months.
• Extract date. Platform figures for recent periods restate. Google's analytics documentation states that conversions can be reattributed for up to seven days after the conversion, so the same report run on day one and day ten can legitimately differ.

Every one of those is a configuration change, and configuration changes are logged. In an incentive-compensation dispute the change history of the conversion and attribution settings is a direct record of the mechanism by which the compensation metric could have been moved. That is the point of the exercise, and it is also its limit: the log records the change, the date and the login, and says nothing about who asked for it.

Who controls the number, and why that is the first question

Where an agency's fee depends on a metric measured inside the ad platform's own interface, and the agency holds administrative access to that account, the agency has both the ability and the incentive to change the settings that produce it. That is a structural observation, not an accusation, and it applies equally to a client who holds admin access and later disputes the invoice.

So the first questions in one of these matters are administrative rather than analytical. Which metric does the clause name, and is it defined anywhere in the contract? Measured in whose system — the platform, the client's analytics property, the client's order table, or a third-party measurement vendor? Extracted by whom, and as of what date? Was the configuration frozen for the measurement period, or was it live? Did either party hold a contractual right to inspect the account, and was it exercised?

A clause that names a number without naming the system, the settings and the extract date has not defined a measurable quantity, and I have not yet seen a way to reconstruct one after the fact that both sides accept. What can be reconstructed is what the settings were during the period, and what they were changed to and when — which is usually enough to show whether the invoiced figure was computed on a stable basis.

The 2023 boundary that breaks older contracts

One dated change catches contracts drafted before it. In an announcement dated 20 April 2023, Google stated that from mid-July 2023 developers could no longer choose first click, linear, time decay or position-based attribution for new or existing conversion actions, and that from September 2023 it would switch any conversion action still using those models to data-driven attribution unless the advertiser moved it to last click first. Google Analytics 4 removed the same four models as of November 2023.

Two consequences follow. A contract signed in 2021 or 2022 naming a position-based or time-decay model as the basis for a fee names a model the platform no longer offers, and the conversion actions that used it were changed automatically rather than by either party. And any conversion series crossing September 2023 for a Google Ads account, or November 2023 for a GA4 property, may contain a model change nobody chose. An expert comparing periods across that boundary is obliged to establish which model each period was reported under, which is discoverable from the conversion-action configuration and the change history.

As of August 2026 Google Ads offers two attribution models — last click and data-driven — and GA4 offers three. That narrowing is itself worth knowing when a clause offers the parties a choice the platform no longer provides.

Compensation structures and the dispute each tends to produce

The structure predicts the argument, which is useful early in a matter when the pleadings are still general.

• Percentage of media spend. The fee rises with spend regardless of result, which supplies a ready-made motive theory for over-spending or for recommending expensive inventory. The defense is that the client approved every budget, and the evidentiary test is the approval record rather than the incentive.
• Fixed fee or retainer against a scope. The argument is scope and staffing, tested on the SOW, timesheets, deliverables and the change history.
• Cost-plus against a rate card. The argument is seniority billed against seniority delivered.
• Performance or incentive compensation tied to a metric. The argument is which metric, measured by whom, on whose numbers, as of when — the subject of this page.
• Principal media or arbitrage margin. The argument is the margin itself and whether it was disclosed.

I am not aware of a current, dated primary survey of how common each structure is in US agency-of-record relationships, so no prevalence figure appears here. Anyone offering one in a report should be asked for the source.

How the calculation is actually tested

The work is a recomputation, and it runs in a fixed order.

1. Read the compensation schedule as a formula, and write down every input it requires: metric, system, settings, period, extract date, and any exclusions.
2. Establish the configuration as it stood during the measurement period — conversion actions included, attribution model, conversion windows, time zone, and whether the account counted "conversions" or "all conversions."
3. Pull the change history for those settings across the whole engagement, noting the acting login for each change.
4. Re-run the export on the settings the contract describes, and compare it to the figure that was invoiced.
5. Where the two differ, decompose the difference into definitional, timing and unexplained components, and eliminate the first two in writing before characterizing the third.

A recomputation that reproduces the invoiced figure is a useful result and is often the one the records support. It is worth saying plainly that a defensible answer in this work is frequently that the number was computed the way the parties agreed, on settings that did not move.

What the record does not settle

It does not show intent. A change to a conversion window has the same log entry whether it was made to improve measurement, to follow a platform recommendation applied in bulk, or to move a fee. The defensible formulation is that the setting changed on this date under this login, with this arithmetic effect on the reported metric.

It does not show approval. Clients approve measurement changes in meetings and in email threads that nobody exports. The absence of a written approval is not evidence that none was given.

It does not make a data-driven attribution figure reproducible. Google describes data-driven attribution as a model that compares the paths of customers who convert to the paths of customers who do not, and does not publish the computation. Google also states that in some situations last click and data-driven attribution return the same results, so a data-driven label on a report does not by itself establish that anything other than last-click logic produced the number.

It does not permit conversions reported by different platforms to be added together. Each platform reports conversions attributed to itself, under its own window and its own model, so a total built by summing them double-counts by an amount the reports themselves do not disclose. A fee computed on such a total is computed on a quantity that does not exist, and saying so is usually more useful than trying to correct it.

And it does not answer whether the metric was a sensible proxy for value in the first place. That is a commercial judgment the parties made when they signed, and it is not an expert question.

Frequently Asked Questions

Published 2026-08-12