Fines Based On Algorithmic Revenue Attribution

Introduction

Algorithmic revenue attribution refers to the use of automated models, data systems, artificial intelligence, attribution software, or platform analytics to determine what proportion of a firm's revenue is attributable to a particular product, conduct, transaction, market, customer group, geographic area, or infringement.

In competition law, the concept becomes particularly important when authorities calculate fines or penalties by reference to revenue generated through algorithmically identified conduct. Examples include:

• an algorithm identifying revenue generated through a cartel;
• an AI system allocating sales between affected and unaffected products;
• a platform attributing advertising revenue to a particular anticompetitive ranking system;
• automated models estimating revenue obtained through exclusionary conduct;
• algorithms separating domestic and foreign turnover;
• automated attribution of revenue among jointly sold products or services;
• algorithmic estimation of incremental revenue resulting from an abuse of dominance.

The central legal difficulty is that algorithmic attribution can improve the precision of a fine while simultaneously creating serious problems of causation, transparency, proof, and due process.

1. Meaning of Algorithmic Revenue Attribution

Traditional competition-law fines generally rely on relatively identifiable financial concepts such as:

• turnover;
• value of sales;
• affected sales;
• duration of infringement;
• gravity of infringement;
• geographic scope;
• aggravating and mitigating circumstances.

Algorithmic attribution introduces an additional analytical layer.

For example:

Total platform revenue = ₹1,000 crore
Algorithm estimates that 27% resulted from conduct constituting the infringement
Attributed revenue = ₹270 crore

The authority may then use the ₹270 crore figure as an input into the penalty calculation.

The important question becomes:

How reliable is the algorithm's conclusion that ₹270 crore was actually caused by the infringement?

2. Why Authorities May Use Algorithmic Attribution

Digital markets create situations where conventional accounting cannot easily identify the economic consequences of particular conduct.

For example, a dominant platform may simultaneously:

• rank its own products preferentially;
• sell advertising;
• provide search services;
• operate a marketplace;
• collect commissions;
• provide payment services;
• offer cloud services.

Revenue may therefore arise from several interconnected activities.

An algorithm can attempt to identify:

User → Search/Recommendation → Ranking → Transaction → Commission → Revenue

and estimate the portion of revenue associated with the allegedly anticompetitive intervention.

This can be especially attractive where millions of transactions make manual attribution impractical.

3. Algorithmic Attribution Is Not the Same as Turnover

A critical distinction should be made between:

A. Actual accounting revenue

Revenue appearing in audited financial records.

B. Revenue attributable to a market

Revenue generated from products or services falling within the relevant market.

C. Revenue attributable to affected sales

Revenue from transactions potentially affected by the infringement.

D. Incremental revenue

Additional revenue allegedly obtained because of the unlawful conduct.

E. Algorithmically attributed revenue

Revenue that a computational model estimates to have resulted from particular conduct.

The last category is potentially the most controversial because it involves inference rather than direct accounting observation.

4. Legal Issues Raised by Algorithmic Attribution

A. Legality of the Fine

A competition authority must have statutory authority to impose the fine.

An algorithm cannot itself create jurisdiction.

The legal chain must therefore be:

Statutory power → infringement → legally relevant revenue → prescribed calculation methodology → penalty

If the legislation authorizes fines based on turnover or value of sales, an authority cannot necessarily substitute an entirely new algorithmic concept of "economic benefit" unless the statute permits it.

5. Causation Problem

The most important issue is causal attribution.

Suppose an algorithm concludes:

30% of a platform's revenue resulted from self-preferencing.

That does not automatically establish that self-preferencing caused 30% of the revenue.

Revenue may also have resulted from:

• product quality;
• brand recognition;
• consumer loyalty;
• network effects;
• pricing;
• independent advertising;
• seasonal demand;
• competitors' weaknesses;
• technological superiority.

Thus:

Correlation ≠ causation.

A sophisticated algorithm does not eliminate the legal requirement to establish a sufficiently reliable causal connection.

6. Counterfactual Analysis

Algorithmic attribution frequently relies upon a counterfactual.

The model may ask:

What would revenue have been if the allegedly unlawful conduct had never occurred?

Suppose:

• actual revenue = ₹500 crore;
• predicted counterfactual revenue = ₹380 crore.

The model therefore attributes:

₹120 crore incremental revenue

to the allegedly unlawful conduct.

But the result depends entirely upon the assumptions used to construct the counterfactual.

The model must therefore be tested for:

• selection bias;
• omitted variables;
• model specification;
• data quality;
• alternative explanations;
• sensitivity;
• robustness;
• statistical significance.

7. Evidentiary Transparency

A major problem arises where the authority relies upon a proprietary or highly complex model.

The undertaking should ordinarily be able to understand:

1. what data were used;
2. what variables were selected;
3. how variables were weighted;
4. what assumptions were made;
5. what counterfactual was used;
6. how errors were handled;
7. what confidence interval applies;
8. how the final revenue figure was produced.

Otherwise the undertaking may be unable effectively to challenge the fine.

This creates an important principle:

Computational complexity cannot become a substitute for procedural fairness.

8. Right of Defence

If an authority uses algorithmic attribution to calculate a fine, the undertaking may need access to sufficient information to challenge the calculation.

This creates a tension between:

Algorithmic confidentiality

and

effective exercise of defence rights.

Authorities may legitimately protect:

• trade secrets;
• confidential customer information;
• cybersecurity information;
• proprietary source code.

However, confidentiality should not make the penalty calculation effectively unreviewable.

Possible solutions include:

• disclosure of methodology;
• disclosure of variables;
• expert access;
• confidentiality rings;
• independent technical review;
• disclosure of sensitivity analysis;
• disclosure of relevant datasets in anonymized form.

9. Proportionality

A fine must generally bear a reasonable relationship to the seriousness of the infringement.

Algorithmic attribution can create over-penalisation if the model attributes too much revenue to the unlawful conduct.

For example:

Actual revenue: ₹1,000 crore
Algorithmically attributed revenue: ₹400 crore
Alternative robust models: ₹180–₹250 crore

Using ₹400 crore without adequately explaining why the higher estimate is justified could create proportionality concerns.

The greater the uncertainty, the stronger the justification required for using the estimate.

10. Error Margins

Algorithmic attribution should not necessarily be understood as producing a single perfectly accurate number.

A better approach may be:

Estimated attributable revenue = ₹200 crore
Reasonable interval = ₹170–₹230 crore

The authority must then decide how uncertainty should affect the penalty.

This raises an important legal question:

Who bears the risk of uncertainty?

If uncertainty is created by the authority's methodology, a court may be reluctant to permit the authority simply to choose the highest plausible estimate.

11. Relevant Case Laws

1. Ahlström Osakeyhtiö and Others v Commission — Wood Pulp

The Wood Pulp litigation is important for the principles governing proof and attribution in cartel cases.

The European courts emphasized that competition-law conclusions must be supported by evidence capable of establishing the alleged conduct.

Relevance to algorithmic attribution

An algorithm cannot cure an evidentiary deficiency in the underlying infringement.

There must be a logical chain:

Data → inference → conduct → infringement → financial consequence.

The more inferential the algorithmic calculation, the greater the need for supporting evidence.

2. Baustahlgewebe GmbH v Commission

In Baustahlgewebe, the Court of Justice addressed issues concerning competition-law proceedings and sanctions.

The case is particularly relevant to the broader principle that substantial competition fines engage serious procedural and judicial-review considerations.

Algorithmic significance

Where an algorithm materially affects the amount of a fine, the undertaking should have a meaningful opportunity to challenge the calculation.

A computational model should therefore not become a black box insulated from judicial scrutiny.

3. Dansk Rørindustri and Others v Commission

Dansk Rørindustri is a major authority concerning the European Commission's methodology for imposing competition fines.

The Court recognized the importance of the Commission's Guidelines on fines and the requirement that the methodology be sufficiently foreseeable.

Relevance

Algorithmic revenue attribution must operate within the legally established framework.

An authority should not unexpectedly introduce an entirely new computational methodology that materially changes the basis upon which undertakings could foresee their potential exposure.

The principle of legal certainty therefore becomes particularly important.

4. KME Germany AG and Others v Commission

The KME litigation concerned substantial competition fines and judicial review.

The European courts emphasized the need for effective judicial scrutiny of competition penalties.

Relevance to algorithmic fines

If a fine depends upon:

• econometric assumptions;
• machine-learning outputs;
• data classification;
• causal modelling;

courts must still be capable of examining whether the resulting fine is legally and economically justified.

Technical complexity cannot remove judicial review.

5. Intel Corp. v Commission

The Intel litigation is highly significant because it demonstrates the importance of examining economic evidence when assessing exclusionary conduct.

The Court of Justice required proper consideration of economic evidence concerning whether conduct was capable of producing anticompetitive effects.

Relevance to revenue attribution

The same logic is highly relevant where an authority seeks to attribute revenue to exclusionary conduct.

An algorithm may estimate that a particular practice increased revenue, but the authority should still consider:

• the mechanism of exclusion;
• actual or potential foreclosure;
• counterfactual conditions;
• competitive effects.

Thus, algorithmic attribution should not replace substantive economic analysis.

6. Servier v Commission

The Servier litigation illustrates the importance of careful economic and factual analysis in complex competition cases.

The General Court and Court of Justice scrutinized the Commission's characterization of competitive conditions and economic evidence.

Relevance

Revenue attribution models depend upon assumptions about:

• market structure;
• competitive alternatives;
• substitution;
• counterfactual behavior.

If those underlying assumptions are defective, the algorithmic revenue calculation may also be defective.

7. Deutsche Telekom v Commission

In Deutsche Telekom, the Court of Justice considered the legality of pricing conduct and the Commission's economic analysis.

The case illustrates that complex economic assessments remain subject to legal scrutiny.

Relevance

Where an algorithm estimates revenue allegedly generated by exclusionary pricing, the authority should explain:

• the relevant pricing variables;
• the counterfactual;
• the causal relationship;
• the period of effect;
• the relevant customer base.

The algorithm is evidence supporting the legal conclusion, not a substitute for the legal conclusion.

8. Google Shopping

The Google Shopping litigation is particularly relevant to modern algorithmic markets.

The case concerned Google's use of its search results page and preferential positioning of its comparison-shopping service.

Relevance to algorithmic revenue attribution

A modern penalty assessment could theoretically attempt to calculate revenue generated through:

search query → ranking → click → merchant visit → transaction → advertising/commission revenue.

Such attribution requires distinguishing revenue caused by preferential positioning from revenue that would have arisen through ordinary search visibility.

Google Shopping therefore illustrates why digital conduct can generate highly complex attribution questions.

12. Algorithmic Attribution in Different Competition Violations

13. Algorithmic Revenue Attribution in Digital Platforms

Digital platforms present especially difficult cases.

Consider an online marketplace.

The platform's algorithm:

1. ranks its own products;
2. places competing products lower;
3. increases visibility of its own products;
4. receives commission revenue;
5. collects advertising revenue.

An authority might estimate:

Without preferential ranking: €700 million
With preferential ranking: €900 million
Attributable incremental revenue: €200 million.

The central question becomes whether the €200 million difference is actually attributable to the unlawful ranking practice.

The authority would need to distinguish:

• ranking effects;
• price effects;
• product quality;
• advertising expenditure;
• brand recognition;
• seasonality;
• consumer preferences;
• competitor availability.

14. Machine-Learning Models Create Additional Problems

Machine-learning systems may be particularly difficult to use in penalty calculations because predictive accuracy does not necessarily establish causal accuracy.

A model can accurately predict:

Which consumers purchased the product.

But the legal question may be:

Would those consumers have purchased the product absent the infringement?

These are fundamentally different questions.

Therefore:

Predictive model ≠ causal model.

Competition authorities should preferably distinguish between predictive analytics and causal inference.

15. Dynamic Revenue Attribution

Digital markets also create temporal problems.

An algorithmic intervention today may generate revenue months later.

For example:

2025 ranking manipulation → increased user adoption → stronger network effects → higher 2026 revenue.

Should all 2026 revenue be attributed to the infringement?

Probably not.

The authority would need to identify the legally relevant causal period and avoid attributing unrelated later growth to the original conduct.

This is especially important for:

• network-effect markets;
• social media;
• app stores;
• digital advertising;
• AI platforms;
• cloud ecosystems.

16. Multi-Product Attribution

Large technology companies frequently sell interconnected products.

For example:

Search → advertising → browser → operating system → cloud → payments.

An infringement in one layer may increase revenue in another.

This creates a revenue propagation problem.

The authority must decide whether:

• only directly affected revenue should count;
• indirectly generated revenue should count;
• network-generated revenue should count;
• ecosystem-wide revenue should count.

A broad algorithmic model may substantially increase the calculated penalty.

17. Double Counting

Another serious risk is double counting.

Suppose an algorithm attributes:

• ₹100 crore to increased transactions;
• ₹70 crore to advertising;
• ₹50 crore to commissions.

If all three streams arise from the same underlying transactions, simply adding them may exaggerate the economic benefit.

Therefore the model must establish whether revenue streams are:

independent, complementary, or overlapping.

18. Data Governance Requirements

A defensible attribution system should document:

Data provenance

Where did the data come from?

Data integrity

Has the data been altered or corrupted?

Data completeness

Are relevant transactions missing?

Variable selection

Why were particular variables selected?

Model specification

Why was this model preferred?

Validation

Was the model tested against historical observations?

Sensitivity analysis

Does the conclusion change materially when assumptions change?

Reproducibility

Can another expert reproduce the result?

These requirements become particularly important where the resulting fine is substantial.

19. Burden of Proof

A useful framework is:

Step 1

Authority establishes the infringement.

Step 2

Authority identifies the legally relevant revenue.

Step 3

Authority explains why algorithmic attribution is necessary.

Step 4

Authority establishes the model's methodological reliability.

Step 5

Undertaking receives a meaningful opportunity to challenge it.

Step 6

Authority addresses competing calculations.

Step 7

Authority explains why the final figure is proportionate.

This prevents the algorithm from becoming the starting point and ending point of the penalty decision.

20. Procedural Fairness and Explainability

Algorithmic penalty systems should ideally provide an explanation layer.

For example:

22% of revenue was attributed to the infringement because:

• 14% resulted from increased ranking exposure;
• 5% resulted from increased conversion;
• 3% resulted from reduced competitor visibility.

The authority should then explain the evidentiary basis for each component.

This is much more defensible than simply stating:

"The algorithm calculated 22%."

21. Judicial Review of Algorithmic Fines

Courts could potentially review the calculation at several levels:

Legal review

Was the authority legally entitled to use this measure?

Evidentiary review

Was there adequate evidence?

Methodological review

Was the model scientifically and economically reasonable?

Procedural review

Could the undertaking challenge the calculation?

Proportionality review

Was the resulting penalty excessive?

Error review

Were material computational or factual errors made?

22. Possible Regulatory Safeguards

A robust framework could require:

1. Predefined statutory methodology
2. Documented model architecture
3. Audit trails
4. Human oversight
5. Independent economic validation
6. Sensitivity testing
7. Error margins
8. Disclosure of material assumptions
9. Confidentiality-ring access
10. Right to submit alternative models
11. Periodic model validation
12. Judicial review

23. Key Legal Principle

The most important principle can be expressed as follows:

An algorithm may calculate a legally relevant number, but it cannot independently determine what number the law regards as legally relevant.

The legal framework must come first.

The computational model operates within that framework.

24. Relationship With Proportionality

Algorithmic attribution may actually increase proportionality if properly designed.

Traditional rough estimates may overstate or understate the economic consequences.

A well-validated model could produce a more precise assessment.

But the opposite is also possible.

A highly complex model can create false precision.

For example:

₹243,716,382

may appear scientifically exact even though the underlying model has substantial uncertainty.

The authority should therefore distinguish between:

numerical precision

and

legal/economic certainty.

25. Competition-Law Risks

Algorithmic revenue attribution creates several principal risks:

1. Black-box penalty calculation

The undertaking cannot understand how the fine was calculated.

2. Causal overreach

Revenue is attributed to conduct without adequate causal evidence.

3. False precision

The model produces an exact number despite substantial uncertainty.

4. Data bias

Incomplete or biased data distort the calculation.

5. Double counting

Multiple revenue streams are incorrectly aggregated.

6. Counterfactual manipulation

The authority selects an unrealistic counterfactual.

7. Excessive penalties

Over-attribution increases the fine disproportionately.

8. Reduced judicial review

Courts defer excessively to technical models.

9. Procedural inequality

The authority possesses sophisticated computational resources unavailable to the undertaking.

10. Regulatory automation

Future authorities could increasingly delegate penalty calculations to automated systems without sufficient human judgment.

26. Suggested Legal Test

A court reviewing an algorithmically calculated competition fine could apply the following framework:

1. Statutory authority
↓
2. Identifiable infringement
↓
3. Legally relevant revenue measure
↓
4. Reliable attribution methodology
↓
5. Demonstrable causal connection
↓
6. Disclosure sufficient for defence
↓
7. Independent scrutiny of assumptions
↓
8. Treatment of uncertainty and alternative models
↓
9. No double counting
↓
10. Proportionate final penalty

Failure at a material stage should permit the calculation to be reduced, reconsidered, or rejected.

27. Overall Assessment

Fines based on algorithmic revenue attribution are potentially legitimate and increasingly useful in digital competition enforcement, but they cannot be treated as technologically self-validating.

The strongest legal model is one in which the algorithm functions as an evidentiary and economic tool, while the competition authority retains responsibility for:

• defining the infringement;
• selecting the legally permissible revenue base;
• demonstrating causation;
• explaining the methodology;
• addressing uncertainty;
• respecting defence rights;
• avoiding double counting; and
• ensuring proportionality.

The jurisprudence from Dansk Rørindustri, KME, Intel, Servier, Deutsche Telekom, Google Shopping, Baustahlgewebe and Wood Pulp collectively supports a broader proposition: competition-law penalties involving sophisticated economic calculations remain subject to legality, evidence, transparency, effective judicial review, and proportionality.

Conclusion

Algorithmic revenue attribution will become increasingly important as competition authorities investigate AI platforms, digital advertising, app stores, marketplaces, cloud ecosystems, financial platforms and autonomous pricing systems.

Its greatest advantage is granularity: it can potentially identify the economic effects of complex digital conduct more accurately than traditional accounting.

Its greatest danger is false objectivity: a mathematically sophisticated output can appear legally authoritative even when its assumptions, counterfactuals, or causal foundations are questionable.