Your CFO Is About to Ask You a Question You Can't Answer

It's Tuesday. You're in a conference room you didn't pick. Your CFO is holding a printout of last quarter's paid media spend. She taps the Meta line - $2.1M - and says, without looking up:

"How much of this would have happened anyway?"

You know the answer your attribution dashboard gives. You also know, in the back of your head, that the answer is wrong. Meta is claiming credit for 47,000 signups. Your gut says the real number is maybe half that. Maybe less. You don't actually know, because nobody on your team has ever run a real incrementality test, and the last time you asked a vendor what it would cost, they quoted $75,000 and a six-week "scoping engagement" before they'd even commit to whether the test was possible.

So you say what every VP of Marketing says in that moment:

"We're looking into it."

And you feel the temperature in the room drop by two degrees.

This Is the Meeting Every Growth Leader Is About to Have

If you haven't had this conversation yet in 2026, you will. CAC is up. Attribution is softer than it used to be. iOS 14 didn't get better, it got normalized. Every CFO in the country has learned the phrase "incrementality testing" from the same podcast, and they're all going to ask you the same question in the same quarter. This is exactly the CFO-CMO partnership gap that makes or breaks marketing budgets.

The problem isn't that you can't run an incrementality test. The problem is that nobody can tell you whether you can run one in under a month of scoping, and by then the board meeting is over.

This is the thing that's changed.

What the Old Workflow Looked Like - and What It Looks Like Now

Here's the part that should make you sit up. Six months ago, getting from "raw conversion data" to "a defensible incrementality test plan you can walk into a boardroom with" looked like this:

That's not an incremental improvement. That's a category shift. The thing that used to require a vendor, a statement of work, and a quarter of patience now fits inside a coffee break.

The Ten-Second Version

A few weeks ago I got access to a new planning tool - an MCP server (Model Context Protocol - an open standard that lets AI assistants call external tools and run structured analyses) that plugs into Claude - built by the team at BlueAlpha specifically for this moment. The pitch is deceptively boring: *"it tells you whether your data can support an incrementality test."*

That doesn't sound like much. Then I ran it on a realistic dataset and watched it do this:

10 seconds in: "Yes, your data is ready. 52 weeks of history, 20 geos, ~1,500 conversions per geo per week, clean pipeline, geos track each other well. Proceed."

20 seconds in: "Here's your treatment/control split. 10 states each. Similarity score 0.785, volume ratio 1.02. Good match, balanced."

30 seconds in: "Recommended test: Meta holdout, Causal Impact methodology, 4 weeks, pause spend in GA/FL/AZ/NY/MO/IN/WI/NC/WA/CA, maintain in the rest."

40 seconds in: "Warning: At your expected 5% lift, you're only powered at 9%. You'd need a 23% lift for this design to detect anything reliably. Add geos, extend duration, or expect a bigger effect."

That last line is the whole game.

Read it again. The tool just told me - a non-statistician, with zero setup, in under a minute - that the test I was about to run was going to produce a null result I couldn't trust. That's the finding that usually costs $75k and six weeks to discover. And it's the finding that has killed more measurement programs than bad data ever has.

Why This Matters for You, Specifically

Put yourself back in the conference room. The CFO has just asked her question.

Version A (today): "We're looking into it. I'll come back in six weeks." Everyone in the room quietly updates their belief that you don't actually know what your budget is doing.

Version B (with this tool, 20 minutes of prep): "Good question. My analytics lead pulled our geo-level conversion history yesterday and ran it through Claude with the BlueAlpha incrementality MCP. Here's what we have. We can run a Meta holdout test in 10 states starting May 1st, 4 weeks long, with a $240,000 opportunity cost. Here's the thing though - at our current spend level we'd only be powered to detect a 23% lift or larger. So what we're proposing is: run the test, and if the lift is under 23% we treat it as 'Meta is doing something but we can't precisely size it,' and we plan a larger, longer test in Q3 to get the precise number. Either way, you'll know more in six weeks than you know right now."

Notice what just happened in that monologue. You didn't do this alone. Your analytics lead ran the tool. They walked you through the output this morning. They're the one who produced the artifact you're holding. In Version A, your team is invisible - or worse, a source of delays. In Version B, your team just shipped a measurement capability your company didn't have yesterday, in one afternoon, and you're the one presenting it. That's not just your career getting better. That's your analytics lead becoming the person every other VP in the company wants to borrow.

Version B is not a better answer. Version B is a different career. Version B is the version where the CFO relaxes, the CEO nods, and someone at the table quietly thinks *"okay, this person knows what they're doing."*

The difference between Version A and Version B is not talent or budget or even data. It's $75,000 and six weeks, or twenty minutes and a CSV. Same data. Same marketer. Same channel. The tool is the entire delta.

The Part I Wasn't Expecting

Here's what surprised me. I went in expecting a planning tool. What I got was closer to a translator. Every output comes back in English - "good match," "balanced," "low power, increase duration" - not in statistical notation. Which means for the first time I can remember, a marketer can walk out of a twenty-minute working session with their analyst and explain the whole test to a CFO without anyone in the room pretending to understand a p-value.

That's not a measurement upgrade. That's a political upgrade.

The Bigger Story Nobody's Telling Yet

Here's the thing I didn't expect to write, but I have to, because the incrementality tool is actually the smaller of the two stories I want to tell you.

The bigger story is what Claude just became.

Until recently, AI assistants were text generators. You'd ask one a question, it would write you a paragraph, and that was the product. Useful, sometimes impressive, but fundamentally limited to "things you could read." That era is over. What Claude is now - and specifically what the MCP pattern unlocks - is an assistant that can operate inside specialized software on your behalf. The incrementality MCP gives Claude the ability to run feasibility checks, power analyses, and test designs. But the same pattern applies to almost everything else in your marketing stack.

There's already an MCP for Meta Ads - Claude can pull your campaign performance, analyze creative fatigue, spin up audiences. There's one for Google Ads - Claude can audit your account structure, diagnose underspend, apply recommendations. There's one for marketing mix modeling - Claude can load your MMM, pull contribution breakdowns, simulate budget reallocations. There's one for your Granola meeting notes, one for your Slack, one for your Notion docs, one for your CRM. Each one takes Claude from "chatbot" to "operator" inside a specific tool.

Incrementality is just where I'd start, because it's the tool where the before/after gap is most dramatic and where the output is most immediately defensible in a boardroom. But the real shift isn't that you can now plan an incrementality test in 20 minutes. The real shift is that a VP of Marketing with a well-configured Claude sidebar has, today, more operational leverage across their stack than they had with a full analytics team six months ago. That sentence sounds like hyperbole. It isn't. It's a description of where the tooling has quietly moved while everyone was arguing about ChatGPT.

Who This Is For (Be Honest with Yourself)

You should try this today if any of these is true:

• You have a $2M+ channel line item you can't confidently defend.

• Someone above you has used the word "incrementality" in the last 90 days.

• Your attribution dashboard and your gut disagree about Meta, TikTok, or YouTube.

• You've asked a vendor for a test and they started a conversation about scoping instead of answering you.

• You're heading into QBR season and you need one concrete number that isn't platform-reported.

You can skip this if you already have a PhD statistician on staff whose whole job is incrementality - they have better tools. Or if your total paid spend is under about $500k a year, in which case the test isn't your bottleneck, scale is.

Getting Set Up: The 10-Minute Install

Before you can run any of this, you (or more realistically, your analytics lead) needs to get Claude Desktop plus the BlueAlpha incrementality MCP installed. This is genuinely a 10-minute job. No IT ticket, no procurement, no vendor contract. If you can paste text into a file, you can do it.

There are three steps.

Step 1 - Install uv. uv is a small Python tool manager that the incrementality MCP uses under the hood. It handles Python installation automatically so you don't have to think about versions. Open Terminal (on Mac) and paste:

curl -LsSf https://astral.sh/uv/install.sh | sh

On Windows, the equivalent instructions live at astral.sh/uv. One command, thirty seconds.

Step 2 - Add the MCP to your Claude Desktop config. Claude Desktop reads a JSON config file that tells it which MCPs to load. Open it in any text editor:

• Mac: ~/Library/Application Support/Claude/claude_desktop_config.json

• Windows: %APPDATA%\Claude\claude_desktop_config.json

If the file doesn't exist yet, create it. Paste in exactly this:

{
  "mcpServers": {
    "incrementality": {
      "command": "uvx",
      "args": ["bluealpha-incrementality-mcp"]
    }
  }
}

Save and close. (If you already have other MCPs configured, just add the "incrementality": {...} block inside your existing "mcpServers" object - don't replace the whole file.)

Step 3 - Restart Claude Desktop. Fully quit (not just close) and reopen. On first launch, Claude will pull the incrementality MCP package via uvx and wire it up. This takes maybe 20 seconds the first time.

Confirm it's live. Open a new Claude conversation and type: *"What incrementality tools do you have available?"* Claude should list the ~15 tools in the MCP - feasibility check, power analysis, treatment/control recommendation, test design, budget estimate, and so on. If you see them, you're done. If you don't, check that the config file is valid JSON and that Claude Desktop was fully restarted.

Total time: about 10 minutes, most of it waiting for downloads. Total cost: $20/month for Claude Pro (Claude Desktop works with a free account too, but Pro removes rate limits that matter for real workflows). What you get: a full incrementality planning capability bolted into your Claude sidebar, permanently.

While you're in there, consider adding the other BlueAlpha MCPs to the same config - the MMM MCP, the Google Ads MCP, and the Meta Ads MCP all install the same way (add another block inside "mcpServers"). If you're going to make this investment, this is the moment to do it - you want all four connected before you start running the real workflows.

What to Do Before Friday

1. Complete the 15-minute installation above, or have your analytics lead do it.

2. Pull your last 12 months of conversions by geo by week out of GA4, your warehouse, or whichever tool you trust most. One CSV. Three columns: geo, date, conversions.

3. Drop it into Claude with the incrementality MCP connected.

4. Ask: *"Can I run an incrementality test on Meta this quarter?"*

5. Read what comes back.

That's it. You'll either get a green light and a plan you can walk into a boardroom with, or you'll get a list of three specific things to fix before you can measure anything credibly. Either way - for the first time in this category - the answer takes ten minutes instead of six weeks.

And the next time your CFO asks how much of that Meta spend would have happened anyway, you won't have to say "we're looking into it."

You'll have an answer. Or at least a plan to get one that nobody in the room can poke a hole in.

Which, in this job, is basically the same thing.

The VPs of Marketing Who Will Be Hard to Replace

Six months from now, the VPs of Marketing who are hard to replace will be the ones whose teams know how to operate Claude against their marketing stack. Not the ones with the biggest agencies or the fanciest attribution dashboards - the ones whose analytics leads can stand up a defensible measurement program in an afternoon because the tools finally caught up. Incrementality is just where I'd start. It's the most dramatic before/after delta, it's the most politically useful output, and it's the one where the phrase "we're looking into it" stops being a valid answer this quarter.

Install Claude. Install the MCP. Give your analytics lead thirty minutes. See what comes back.

You'll know within an hour whether I was right.

Real-World Results from Incrementality Testing

If you're wondering what these tests actually find when run on real budgets, here are some results from companies that went from planning to execution:

• Pettable: 2.12M in annualized savings from eliminating non-incremental Google Search spend. Reddit killed entirely (zero incrementality, 150K redeployed). TikTok validated at 10.5x ROI. The team ran 5+ tests in 9 months using a weekly action cadence.
  
  

• Klover: Cut Meta iOS spend 50% with no conversion loss. Scaled Apple Search Ads 10x. 35% iCAC improvement. The MMM showed what was truly incremental; the incrementality tests proved it.
  
  

• Cann: Tested whether AppLovin's reported ROAS was real using a geo holdout. The test revealed the truth about incremental contribution and saved half a million in potential wasted spend.

The pattern across all of these: the test plan came first (often using the same readiness and power analysis methodology built into this MCP), and the savings followed. The planning phase is the unlock.

Frequently Asked Questions

What is incrementality testing?

Incrementality testing is a causal measurement method that compares marketing exposed to a treatment group against an unexposed control group to isolate the conversions a channel actually caused - as opposed to conversions that would have happened anyway. It's the gold-standard way to validate whether a paid channel is actually working, separate from what the platform's own attribution reports claim. For the fundamentals, see our guide: What Does "Incremental" Mean in Marketing?

What is the BlueAlpha incrementality MCP?

The BlueAlpha incrementality MCP is a Model Context Protocol server that plugs into Claude and gives it a specialized toolbelt for incrementality test planning. It runs feasibility checks, treatment/control geo recommendations, power analyses, test designs, budget estimates, and full handoff-ready test plans - all from a single CSV of geo-level conversion history. It replaces the traditional 50,000-75,000 scoping engagement with a ~30-minute workflow.

What is an MCP in Claude?

An MCP - Model Context Protocol - is an open standard that lets AI assistants call external tools. Instead of Claude only generating text, an MCP turns Claude into an operator that can pull data, run analyses, and take actions inside a specific system (Google Ads, Meta, an MMM, an incrementality engine). Think of it as handing Claude a domain-specific toolbelt.

How do I install the BlueAlpha incrementality MCP in Claude?

Three steps, about 10 minutes. (1) Install uv by running curl -LsSf https://astral.sh/uv/install.sh | sh in Terminal. (2) Add the MCP to your Claude Desktop config file at ~/Library/Application Support/Claude/claude_desktop_config.json on Mac (or %APPDATA%\Claude\claude_desktop_config.json on Windows) by pasting a six-line JSON block with "command": "uvx" and "args": ["bluealpha-incrementality-mcp"]. (3) Restart Claude Desktop. Then type *"What incrementality tools do you have available?"* in a new conversation to confirm it's live. Total cost: $20/month for Claude Pro.

How long does it take to plan an incrementality test with this tool?

About 30 minutes end-to-end: 10 minutes to pull a CSV of conversions by geo by week out of GA4 or your warehouse, and 20 minutes to run feasibility, treatment/control recommendation, test design, power analysis, and budget estimation inside Claude. The traditional scoping-engagement equivalent took 4-6 weeks.

How much does it cost?

20 per month for Claude Pro, plus the cost of the underlying incrementality MCP (contact BlueAlpha for access). Compared to a traditional 50,000-$75,000 measurement scoping engagement, the planning phase is effectively free.

Who should use the incrementality MCP?

VPs of Marketing, Heads of Growth, and analytics leads at mid-market consumer companies with 5M-50M in annual paid media spend. It's especially valuable for teams facing CFO scrutiny on specific channel line items, teams without an in-house statistician, and agency or measurement-vendor CS teams running feasibility qualification with clients.

Does this replace a data science team?

No. It replaces the 4-6 week scoping engagement that used to come before the data science work. When the tool greenlights a test, you still want a data science team - internal or at BlueAlpha - to execute the final Causal Impact model on the in-market data. The tool makes the handoff faster, cheaper, and much more informed on both sides.

What if the tool tells me my data isn't ready?

That's a feature, not a failure. A "not_yet" verdict means the tool just saved you from a 50,000-75,000 scoping engagement that would have told you the same thing in six weeks. It explains exactly which checks failed - not enough history, volume too low, geos too dissimilar, data quality issues - so you know what to fix before trying again.

Can the same Claude + MCP pattern work for other marketing tools?

Yes. MCPs already exist for Meta Ads, Google Ads, marketing mix modeling (MMM), meeting notes (Granola), Slack, CRMs, and more. Incrementality is a particularly dramatic example of the pattern because the old workflow was so expensive, but the broader shift is that Claude is becoming an operator across the entire marketing stack, not just a text generator. To see how incrementality testing and MMM work together as a complete measurement framework, see our playbook.

If you try it and it tells you "not yet" - that's not the tool failing. That's the tool saving you from a $75k engagement that was going to tell you the same thing in six weeks. Screenshot it, send it to your CFO, and use it to buy yourself a quarter.