Round 97

Marketing Attribution Model Selection

A mid-size e-commerce company spends across paid search, social ads, affiliate links, email, and offline events, and wants to reallocate a $4M annual budget more efficiently. The analytics team has clickstream and CRM data but inconsistent cross-device tracking, partial view-through data, and no reliable holdout regions for causal testing. Leadership wants an attribution methodology that is defensible to finance, actionable for the marketing team within one quarter, and not prohibitively expensive to build or maintain. The team must pick one primary model to drive budget decisions, knowing each approach trades off interpretability, statistical rigor, implementation cost, and vulnerability to data gaps.

Dashboard

Status

DECIDED

Winner

A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes.

Majority

A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes.

Duration

7309 ms

Topic

Options

A Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes.
3.00 of 3 PPV winner
B Build a data-driven multi-touch algorithmic model (e.g., logistic regression or Markov chain removal effects) using existing touchpoint data, accepting higher build complexity for improved credit distribution accuracy.
0.00 of 3
C Implement Shapley value attribution to fairly distribute credit across channel combinations, accepting significant computational cost and reduced interpretability for stakeholders unfamiliar with game theory.
0.00 of 3
D Adopt a first-touch attribution model to emphasize and protect upper-funnel brand and discovery channels that last-touch models systematically undervalue.
0.00 of 3
E Run incremental geo-lift or holdout experiments as the primary decision input, delaying full-scale reallocation until causal evidence is gathered, even though this slows the timeline and limits channel coverage.
0.00 of 3
F Adopt a simple last-touch attribution model, prioritizing ease of implementation and stakeholder interpretability despite known bias toward bottom-funnel channels.
0.00 of 3

Checks

Baselines

PPV no gold answer
Majority no gold answer
Comparison PPV matched majority
k samples8
Agents3
Parsed samples 24 / 24
Quality unanimous

PPV

Top Influences

Claude Sonnet 5 · delegate 1.000
GPT 5.4 mini · delegate 1.000
Gemini 3 Flash Preview · delegate 1.000

Samples

Answer Distribution

Claude Sonnet 5 A: 8
GPT 5.4 mini A: 8
Gemini 3 Flash Preview A: 8

PPV

Influences

Label Role Score
Claude Sonnet 5 delegate 1.000
GPT 5.4 mini delegate 1.000
Gemini 3 Flash Preview delegate 1.000

Voting matrix

Constructed Delegation Matrix

To / From Claude Sonnet 5 GPT 5.4 mini Gemini 3 Flash Preview
Claude Sonnet 5 0.0000.0000.000
GPT 5.4 mini 0.0000.0000.000
Gemini 3 Flash Preview 0.0000.0000.000
A 1.0001.0001.000
B 0.0000.0000.000
C 0.0000.0000.000
D 0.0000.0000.000
E 0.0000.0000.000
F 0.0000.0000.000

Embeddings

Sample Embeddings

24 samples
Claude Sonnet 5 sample #0 · A · 1536 dims A Claude Sonnet 5 sample #1 · A · 1536 dims A Claude Sonnet 5 sample #2 · A · 1536 dims A Claude Sonnet 5 sample #3 · A · 1536 dims A Claude Sonnet 5 sample #4 · A · 1536 dims A Claude Sonnet 5 sample #5 · A · 1536 dims A Claude Sonnet 5 sample #6 · A · 1536 dims A Claude Sonnet 5 sample #7 · A · 1536 dims A GPT 5.4 mini sample #0 · A · 1536 dims A GPT 5.4 mini sample #1 · A · 1536 dims A GPT 5.4 mini sample #2 · A · 1536 dims A GPT 5.4 mini sample #3 · A · 1536 dims A GPT 5.4 mini sample #4 · A · 1536 dims A GPT 5.4 mini sample #5 · A · 1536 dims A GPT 5.4 mini sample #6 · A · 1536 dims A GPT 5.4 mini sample #7 · A · 1536 dims A Gemini 3 Flash Preview sample #0 · A · 1536 dims A Gemini 3 Flash Preview sample #1 · A · 1536 dims A Gemini 3 Flash Preview sample #2 · A · 1536 dims A Gemini 3 Flash Preview sample #3 · A · 1536 dims A Gemini 3 Flash Preview sample #4 · A · 1536 dims A Gemini 3 Flash Preview sample #5 · A · 1536 dims A Gemini 3 Flash Preview sample #6 · A · 1536 dims A Gemini 3 Flash Preview sample #7 · A · 1536 dims A
Claude Sonnet 5 GPT 5.4 mini Gemini 3 Flash Preview

Samples

Agent Responses

Claude Sonnet 5 openai_compatible · anthropic/claude-sonnet-5
#0 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking and no reliable holdout regions, methods requiring granular individual-level accuracy (multi-touch algorithmic, Shapley) or causal experiments (geo-lift) are undermined by the data gaps, while single-touch heuristics (f...

#1 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking and no holdout regions, individual-level multi-touch models (B, C, D, F) would be built on unreliable data, and true causal experiments (E) aren't feasible within the timeline. MMM (A) works with aggregate spend/outcome...

#2 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given the inconsistent cross-device tracking and partial view-through data, individual-level multi-touch or Shapley approaches would be built on unreliable touchpoint data, undermining their statistical rigor benefits. Since no holdout regions exist, true caus...

#3 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking, partial view-through data, and no holdout regions, individual-level multi-touch or Shapley models would be built on unreliable identity data, undermining their statistical rigor, while single-touch heuristics (first/la...

#4 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given the inconsistent cross-device tracking, partial view-through data, and lack of holdout regions, individual-level multi-touch or experimental models are undermined by data gaps, making them less defensible and harder to implement within a quarter. Aggrega...

#5 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking, partial view-through data, and no holdout regions, individual-level multi-touch or Shapley models would be undermined by data gaps, while last/first-touch models are too simplistic and biased for a $4M reallocation dec...

#6 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking and no reliable holdout regions, individual-level multi-touch and experimental approaches (B, C, E) are undermined by data gaps, while single-touch heuristics (D, F) are too simplistic and biased for a $4M reallocation ...

#7 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking and no holdout regions, individual-level multi-touch or Shapley models would be unreliable and hard to defend, while last/first-touch models are too simplistic and biased for a $4M reallocation decision. MMM using aggre...

GPT 5.4 mini openai_compatible · openai/gpt-5.4-mini
#0 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company has incomplete tracking across devices, partial view-through data, and no reliable holdout regions, so a user-level multi-touch or experiment-first approach is hard to defend or execute quickly. MMM is the best fit because it is robust to tracking ...

#1 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company has mixed data quality across devices and channels, making user-level attribution methods less reliable, while also needing something finance can trust and the team can deploy quickly. MMM is designed for exactly this situation: it works from aggre...

#2 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company has inconsistent cross-device tracking, partial view-through data, and no reliable holdout regions, so a user-level or experimental approach would be hard to trust or slow to deploy. They need something defensible to finance, practical within one q...

#3 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given inconsistent cross-device tracking, partial view-through data, and no reliable holdout regions, a user-level multi-touch model or experimentation is hard to trust or execute quickly. Media mix modeling is the most defensible choice for finance because it...

#4 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given the inconsistent cross-device tracking, partial view-through data, and lack of reliable holdout regions, a user-level multi-touch model or experiment-first approach would be harder to trust or execute quickly. MMM is specifically designed to work with ag...

#5 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The team needs a model that works despite cross-device gaps, partial view-through data, and no reliable holdout regions, while still being defensible and practical within one quarter. MMM is specifically designed to be robust to missing user-level tracking and...

#6 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company needs a model that is practical with imperfect cross-device and partial view-through data, can be built within a quarter, and is defensible to finance without expensive experimentation. MMM is specifically designed to work from aggregated historica...

#7 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company has inconsistent cross-device tracking, partial view-through data, and no reliable holdout regions, so highly granular user-level attribution or experiments would be hard to trust or implement quickly. Media mix modeling is the best fit because it ...

Gemini 3 Flash Preview openai_compatible · google/gemini-3-flash-preview
#0 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Marketing Mix Modeling (MMM) is the most defensible choice for finance and leadership because it relies on aggregate historical data, bypassing the company's significant issues with cross-device tracking and partial view-through data. Unlike granular algorithm...

#1 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company requires a model that is defensible to finance, actionable within one quarter, and robust against significant data gaps such as inconsistent cross-device tracking and missing view-through data. Media Mix Modeling (MMM) is the most viable solution b...

#2 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Media Mix Modeling (MMM) is the most defensible choice for finance because it relies on aggregate revenue and spend data, bypassing the company's significant issues with inconsistent cross-device tracking and missing view-through data. Unlike granular algorith...

#3 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Given the constraints of inconsistent cross-device tracking, partial view-through data, and a one-quarter deadline, individual-level algorithmic models (B, C) or experimental holdouts (E) are either data-compromised or too slow. Media Mix Modeling (A) is the m...

#4 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

The company requires a methodology that is defensible to finance, actionable within one quarter, and robust against significant data gaps such as inconsistent cross-device tracking and lack of holdout regions. Media Mix Modeling (MMM) is the most suitable choi...

#5 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Marketing Mix Modeling (A) is the most defensible choice for finance and leadership because it uses aggregate data, bypassing the critical issues of inconsistent cross-device tracking and missing view-through data. Unlike granular algorithmic models (B or C), ...

#6 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Media Mix Modeling (MMM) is the most defensible choice for finance because it relies on aggregate spend and revenue data, bypassing the company's significant cross-device tracking and view-through data gaps. Unlike user-level algorithmic models (B and C), MMM ...

#7 A · Invest instead in aggregate media mix modeling (MMM) using historical spend and outcome time series, sacrificing individual-level granularity for robustness to tracking gaps and privacy changes. 1536 dims

Media Mix Modeling (MMM) is the most defensible choice for finance because it relies on aggregate spend and revenue data, bypassing the company's significant issues with cross-device tracking and missing view-through data. Unlike algorithmic multi-touch models...

Timings

Stages

Sampling6685 ms
Embedding621 ms
Signals3 ms
PPV0 ms

Errors

Agent and Round Errors

No errors recorded.