Choosing an MCP server for Marketing Mix Modeling (MMM) and incrementality testing starts with the decisions you want to make through your AI assistant. Can it explain incremental performance, calculate a new budget plan, and retrieve the experiment results behind a recommendation?
The answer depends on both the measurement platform and the capabilities it exposes through the connection. A platform may support campaign optimization in its own interface while its MCP server only provides access to reporting. Your evaluation needs to establish what your team can actually do from the assistant it plans to use.
This guide presents 48 evaluation criteria across eight categories from our comparison of MCP servers for MMM and incrementality testing. It covers data reporting, causal performance insights, channel planning, campaign and ad set optimization, incrementality testing, interoperability, analytical methods, and enterprise requirements. For vendor evaluation against these criteria, see Top MCP Servers for Marketing Mix Modeling (MMM) and Incrementality Testing.
Table of contents
- What is an MCP server for MMM and incrementality testing?
- Define what you need the MCP server to do
- How the evaluation criteria were developed
- The 48 evaluation criteria across eight categories
- How to use this framework in your evaluation
- Frequently asked questions
- Further reading
What is an MCP server for MMM and incrementality testing?
An MCP server for MMM and Incrementality testing enables marketers to ask measurement and optimization questions within their AI assistant of choice, and get replies that are grounded in MMM and incrementality testing analytics of the connected measurement platform. MCP stands for Model Context Protocol, an open standard for exchanging data and making tools available to AI applications. The MCP architecture documentation describes how servers expose these capabilities to connected clients.
In a marketing workflow, the assistant can request sales data, retrieve MMM results, or call a forecasting tool that the provider has made available. The measurement platform returns the data or calculation, and the assistant presents it in the conversation.
For example, a marketer might ask how much incremental revenue paid social generated last month, then request a forecast for a different budget allocation. An analyst might retrieve a completed geo test and compare its results with the relevant MMM estimate. Each step requires access to the appropriate data or analytical tool.
MCP supplies the connection. The underlying models, experiment methods, and data determine the quality of the measurement. Access to ad-platform spend and attributed revenue alone does not establish that a server can answer questions about incrementality.
Define what you need the MCP server to do
At the start of your MCP evaluation process, write down the questions your team expects to ask, the decisions that follow, and the level of detail needed. A media planner allocating next quarter’s budget has different requirements from a performance marketer adjusting individual campaigns.
- For reporting, specify the sales channels, media platforms, markets, and business outcomes you need to retrieve.
- For planning, specify the budget horizon, optimization objective, and constraints, such as minimum channel spend or committed media bookings.
- For campaign management, decide whether you need measurement, recommendations, or permission to apply budget and bidding changes.
- For experimentation, separate reviewing completed tests from prioritizing, designing, and launching new ones.
How the evaluation criteria were developed
To support marketers in evaluating MCP servers, we developed an evaluation framework spanning 48 criteria. In developing the criteria, our primary research drew on:
- 5,776 prompts submitted by marketers through Sellforte’s MCP connection or in-platform conversational AI.
- 570 discussions with marketers about MCP and conversational AI for MMM and incrementality testing.
- Experience from more than 20 retail and ecommerce RFPs and tender processes containing MCP or AI requirements.
- Desk research and LLM-assisted investigation to identify gaps and check the framework against public information.
The next section presents the evaluated criteria.
The 48 evaluation criteria across eight categories
The first five categories cover the marketing work the connection should support. The remaining categories address interoperability, the measurement system behind the answers, and the requirements for operating the service. The criterion IDs below match the vendor comparison so you can move between the selection guide and the research.
Category 1: Marketing Data Reporting via MCP
Reporting needs to cover the business you manage. An omnichannel retailer may need physical-store sales alongside ecommerce, while a performance team may need orders or new customers by campaign. Check that the returned data uses the dates, dimensions, and outcome definitions you requested.
| ID | Criterion | What it means |
|---|---|---|
| 1.1 | MCP reports sales progress for online sales | MCP reports actual online sales for a specified period. |
| 1.2 | MCP reports sales progress for offline store sales | MCP reports actual physical-store sales for a specified period, using offline store sales data retrieved from customer's data warehouse. |
| 1.3 | MCP reports digital media data (spend, media metrics) | MCP reports digital media data across Meta, Google, TikTok, and other major paid platforms. |
| 1.4 | MCP reports offline media data (spend, media metrics) | MCP reports spend and media metric data for offline media, covering at least TV, out-of-home, radio, print. |
| 1.5 | MCP Reports business outcomes beyond revenue | MCP reports supported business outcomes such as contribution margin, orders or new customers. |
| 1.6 | Filters and groups by business dimensions | MCP applies explicit date, brand, market, product and campaign filters and returns results at the requested supported granularity. |
Category 2: Historical Performance Insights & Causal Explanation via MCP
This category assesses whether the connection can retrieve the model’s estimates of incremental marketing impact and explain performance changes using modeled drivers. Promotion effects, baseline demand, and offline media can be essential to understanding why total sales moved.
| ID | Criterion | What it means |
|---|---|---|
| 2.1 | MCP reports incremental ROAS and incremental revenue for each digital channel | Retrieves MMM-based estimates of incremental revenue and iROAS for each supported digital channel, with a clear distinction from attribution ROAS. |
| 2.2 | MCP reports incremental ROAS and incremental revenue for each offline channel | MCP reports incremental ROAS and revenue for offline channels such as TV, OOH, and radio. |
| 2.3 | MCP reports promotion-driven revenue, in addition to media-driven | MCP surfaces how promotions and pricing changes contributed to sales, not just paid media. |
| 2.4 | MCP's incremental ROAS measurement is updated daily (not weekly or monthly), based on MMM | Provides MMM-based incremental ROAS measurement that updates daily. Verify the measurement refresh cadence separately from the dates shown in reports. |
| 2.5 | Decomposes performance drivers (Base, media, non-media drivers) | MCP exposes modeled contributions from media, baseline and supported non-media drivers to explain changes in business outcomes. |
Category 3: Channel-Level Optimization with MCP
Channel planning requires access to the calculations used to forecast outcomes and allocate budgets. Retrieving a saved plan is useful, but creating a new constrained allocation is a separate capability. Marginal returns help assess the expected return on additional spend at the current budget level.
| ID | Criterion | What it means |
|---|---|---|
| 3.1 | MCP creates an optimized channel allocation | Invokes the provider’s optimization engine through MCP to calculate channel budgets for a specified objective and period. |
| 3.2 | MCP forecasts a specified media plan | Invokes the provider’s forecasting engine through MCP to estimate outcomes for a specified budget allocation. |
| 3.3 | MCP reports marginal returns and response curves | Retrieves marginal returns and spend-response data for supported channels through MCP. |
| 3.4 | MCP simulates a user-specified budget change | Calculates the modeled impact of a specified change to channel spend through MCP. |
| 3.5 | MCP enforces planning constraints | Passes user-defined budget limits, channel restrictions and planning dates to the optimization engine and exposes the resulting constraints. |
| 3.6 | MCP compares scenarios with a baseline plan | Returns comparable spend and outcome estimates for alternative scenarios and an explicit baseline plan through MCP. |
| 3.7 | MCP plans against a business target | Calculates a budget plan for a supported target such as incremental profit, customer acquisition or a specified outcome level. |
| 3.8 | MCP retrieves saved plans and assumptions | Retrieves previously saved scenarios through MCP with their identifiers, inputs and outputs. |
Category 4: Campaign & Ad Set-Level Optimization with MCP
Performance marketers need results at the level where they manage spend. These criteria distinguish campaign and ad set measurement from budget recommendations, bidding recommendations, and execution. A connector can support one part of that workflow without supporting the others.
| ID | Criterion | What it means |
|---|---|---|
| 4.1 | MCP reports incremental ROAS for each campaign and ad set | MCP reports incremental revenue and ROAS at the individual campaign and ad set level, not just at the channel level. |
| 4.2 | For each campaign and ad set, MCP compares incremental ROAS to last-click and ad platform attribution ROAS | Returns comparable incremental, last-click and ad-platform returns for supported campaigns and ad sets. |
| 4.3 | For each campaign and ad set, MCP provides marginal returns (miROAS) | Retrieves modeled marginal returns for individual campaigns and ad sets through MCP. |
| 4.4 | For each campaign and ad set, MCP recommends optimal daily spend/budget | Returns daily budget recommendations through MCP. |
| 4.5 | For each campaign and ad set, MCP recommends optimal bid value (e.g., Target ROAS) | Returns bidding targets (e.g., Target ROAS) through MCP. |
| 4.6 | MCP can push daily spend/budget changes to Meta, Google etc. APIs | MCP can execute daily spend/budget changes directly on major ad platforms via API. |
| 4.7 | MCP can push bidding changes (e.g., Target ROAS) to Meta, Google etc. APIs | MCP can execute bidding parameter changes (e.g. Target ROAS) directly on major ad platforms via API. |
| 4.8 | For each executed bidding change for a campaign or ad set, MCP provides pre/post analysis summarizing revenue and spend impact of the change | Summarizes observed revenue and spend before and after an executed change at campaign or ad set level. |
Category 5: Incrementality Testing with MCP
Experiment workflows begin with access to usable results: what was tested, where, when, and with what outcome. Geo tests, owned-media A/B tests, and platform lift studies need distinct support. Recommending the next experiment and designing a feasible test require additional capabilities.
| ID | Criterion | What it means |
|---|---|---|
| 5.1 | MCP reports results for Geo Tests | Retrieves completed geographic incrementality-test results with the tested intervention and outcome identified. |
| 5.2 | MCP reports results for Own Media A/B tests (e.g. leaflet tests) | Retrieves incrementality results for randomized or controlled owned-media tests such as email or leaflet experiments. |
| 5.3 | MCP reports findings for Meta Conversion Lift tests | Retrieves incrementality results from advertising-platform lift studies such as Meta Conversion Lift. |
| 5.4 | MCP prioritizes experiments | Uses available measurement results and uncertainty to recommend specific hypotheses, channels or markets for testing. |
| 5.5 | MCP designs feasible incrementality tests | Produces a test design specifying treatment and control, power or minimum detectable effect, duration and required spend. |
Category 6: MCP interoperability and workflows
Your team needs a connection it can set up, understand, and reuse. Tool and metric documentation should make the meaning of each result clear. Structured outputs matter when the same analysis must feed a spreadsheet, report, or recurring workflow.
| ID | Criterion | What it means |
|---|---|---|
| 6.1 | Documented instructions for connecting with Claude and ChatGPT | Provides documented MCP connection and authentication instructions for Claude and ChatGPT |
| 6.2 | MCP provides native visual outputs | Returns server-provided tables or chart artifacts that a documented compatible MCP client can display. |
| 6.3 | Exposes tool and metric documentation | Makes available discoverable MCP tool definitions and documentation for interpreting the returned data. |
| 6.4 | Returns reusable structured data | Provides complete structured results or downloadable data files for use in spreadsheets, reports and downstream tools. |
| 6.5 | Supports external agent automation | Documents repeatable MCP invocation from an external agent runtime for recurring analysis workflows. |
Category 7: Analytical Backbone
The analytical backbone determines what the returned numbers mean. This framework includes Bayesian MMM, experiment calibration, model validation, and access to model settings.
| ID | Criterion | What it means |
|---|---|---|
| 7.1 | MCP provides deterministic, model-backed answers with Bayesian MMM as backbone | Grounds numerical answers and recommendations in a Bayesian MMM and its analytical tools, with reproducible outputs for fixed inputs and model versions. |
| 7.2 | The underlying MMM used by the MCP is calibrated with incrementality tests | Uses relevant incrementality-test results to inform MMM calibration, with the relationship between experiments and model assumptions documented. |
| 7.3 | MCP reports model validation and other modelling KPIs | Makes model-validation metrics available, such as holdout performance, MAPE, R-squared, or posterior predictive checks. |
| 7.4 | Model calibration & configuration settings (e.g., priors) are auditable and editable in a self-serve UI | Customers can inspect and configure model priors and other key parameters in a self-serve UI, not just accept the model as a black box. |
| 7.5 | Supports controlled calibration through MCP | Allows authorized agents to propose or apply model-calibration changes through MCP with validation and version history. |
Category 8: Enterprise-Grade Platform
Enterprise evaluation includes reference customers, independent security assurance, deployment options, and authentication. These requirements apply to the wider platform as well as the connection. Select the requirements that match your organization’s procurement and operating needs.
| ID | Criterion | What it means |
|---|---|---|
| 8.1 | At least 10 public reference customers from $1B+ revenue brands | Provides at least ten public reference customers from brands with annual revenue above $1 billion. |
| 8.2 | SOC 2, ISO 27001, or audited IT security by a third-party cyber security auditor | Provides SOC 2 reporting, ISO 27001 certification, or an independent third-party security audit for the relevant service. |
| 8.3 | Data residency: geography option between US and EU | Offers a choice of US or EU data residency. Verify what data and services the residency commitment covers. |
| 8.4 | Multi-cloud: option between AWS, GCP, and Azure | Deployment flexibility to match the customer's existing cloud infrastructure. |
| 8.5 | Supports single sign-on (SSO) for enterprises | Supports enterprise single sign-on. Confirm how identity and permissions apply to users accessing the MCP connection. |
| 8.6 | Hands-on demo or trial of the MCP capabilities is available without sales-call gating | Provides hands-on access to MCP capabilities without requiring a sales call. |
How to use this framework in your evaluation
Prioritize the decisions your team owns
For an omnichannel retailer, give particular attention to Marketing Data Reporting via MCP, Historical Performance Insights & Causal Explanation via MCP, and Enterprise-Grade Platform. Require evidence that physical-store sales, offline media, and the relevant business dimensions are available.
For a team managing channel budgets, focus on Channel-Level Optimization with MCP and Analytical Backbone. The provider should demonstrate new calculations under your constraints and explain the measurement behind the recommended allocation.
For performance marketers, Campaign & Ad Set-Level Optimization with MCP deserves a separate demonstration. Establish whether campaign-level recommendations are actionable at the granularity your media buyers use, and which steps require another interface.
For analytics teams running experiments, prioritize Incrementality Testing with MCP and Analytical Backbone. Confirm whether test results are simply retrievable or actually inform the MMM, and who decides when the evidence is suitable for calibration.
Build a shortlist from evidence, then test it
Mark each criterion as required, useful, or outside your current scope before meeting vendors. Keep essential requirements as pass/fail checks. A high total research score should not compensate for a missing capability that your team needs every week.
For each required criterion, record the supporting documentation, access route, demonstration result, limitations, and any additional cost. A zero in the published comparison means the evaluators found no qualifying public evidence under its method. Use that as a question for the vendor, rather than assuming the capability is unavailable.
Use the same small set of tasks across shortlisted vendors. Include a reporting request you can reconcile, an incremental-performance question, and the planning or experiment workflow most relevant to your team. Ask for missing or unsupported data as well: the assistant should make those limits clear.
Agree responsibilities and total cost
Ask for an itemized quote covering the measurement platform, MCP access, AI-client subscription, implementation, and ongoing support. Confirm any limits on users, markets, tool calls, exports, or model updates. If a workflow depends on an additional API integration, identify who builds and maintains it.
For write-capable workflows, agree who can approve budget, bidding, and model changes, and how those changes are logged and reviewed. Start with the access needed for the agreed use case and establish an owner for credentials, permissions, and failed workflows.
Once you have defined your requirements, use our MCP server comparison to review the supporting research and identify capabilities to demonstrate with each shortlisted vendor.
Frequently asked questions
How do I choose an MCP server for MMM and incrementality testing?
Start with the decisions and data your team needs, then evaluate the server across the eight categories in this guide. Verify your required workflows in the intended AI assistant, trace numerical answers to model or experiment outputs, and confirm the access route, permissions, and cost.
What is the difference between a marketing data MCP server and an MMM MCP server?
A marketing data server may retrieve spend, sales, and attributed conversions. An MMM server needs access to model outputs or tools that estimate incremental impact and support the relevant planning tasks. Check which measurement method produced each returned metric.
Does a capability available through an API also work through MCP?
That requires verification. The source comparison gives API capabilities equivalent credit, but using an API may require separate integration work. Ask the provider to demonstrate the exact workflow through your chosen MCP connection and identify any additional components.
Can an MCP server run incrementality tests?
Capabilities vary. Retrieving a completed result, recommending an experiment, designing it, and launching it are separate functions. The five incrementality-testing criteria in this framework cover result retrieval, prioritization, and design; confirm execution and monitoring separately if you need them.
Can an MCP server change advertising budgets and bids?
Only if the required write tools, advertising-platform connections, and permissions are available. A budget recommendation does not establish execution capability. Ask for a controlled demonstration that shows authorization, the change applied, and the recorded result.
Does connecting an AI assistant improve MMM accuracy?
The connection can make results and analytical tools easier to use. Measurement quality still depends on the data, model specification, validation, and calibration. Evaluate those separately from the fluency of the assistant’s explanation.
Do I need every capability in the framework?
No. A team reviewing monthly channel performance may have little need for automated bid changes. Use the full framework to understand scope, then define the requirements that match your decisions, operating model, and procurement needs.
Further reading
- Top MCP Servers for Marketing Mix Modeling and Incrementality Testing
- How to Choose a Conversational AI Tool for MMM and Incrementality Testing
- How to Choose an Incrementality Testing Tool
- How to choose a Marketing Mix Modeling solution in Retail
Author
Lauri Potka is the Chief Operating Officer at Sellforte, with over 15 years of experience in Marketing Mix Modeling, marketing measurement, and media spend optimization. Before joining Sellforte, he worked as a management consultant at the Boston Consulting Group, advising some of the world’s largest advertisers on data-driven marketing optimization. Follow Lauri on LinkedIn.
Top MCP Servers for Marketing Mix Modeling (MMM) and Incrementality Testing (2026 Q3)
How to choose a Marketing Mix Modeling solution in Retail: 69 Criteria
