Ask Sellforte MCP: Do our priors match our lift tests?
Welcome to the Sellforte MCP Use Case Library. Each post takes one question that marketing teams really ask their AI assistant through Sellforte MCP, which is now in general availability and accessible to all marketers worldwide. It strips out anything customer-specific and walks through what the assistant does with it, step by step. We run every question on our demo account so the numbers you see are illustrative, not anyone's real results. The idea is simple: steal the prompts, try them on your own model, and see what you get back. We hope you find this series inspiring!
-Juha, CEO at Sellforte
Putting a channel's prior next to its lift tests
Every marketing mix model starts with assumptions about what each channel is likely to return. Those priors should move when new experiment evidence arrives, but checking them usually means opening the calibration settings in one place, the lift study results in another, and doing the matching by hand. It's a job for whoever owns the model's calibration: on the Sellforte side, that's typically a data scientist or customer success manager working through a calibration round.
What you ask
For our demo account, read the calibration settings and prior ROI for Meta Prospecting in Germany. Compare them with the Meta Prospecting conversion lift studies. If they disagree, propose a new prior value and show the change, but don't apply anything until I confirm.
Which of those tests actually feed the prior, and with what weight? Why is one of them left out?
What Sellforte MCP does
The assistant opens Sellforte's calibration tools and does three reads.
1. The prior ROI table
First it pulls the prior ROI row for Meta Prospecting in Germany. That row shows where the prior comes from: what the platform's own reported revenue implies, what the pooled experiments imply, and the prior the model actually uses once the calibration rules are applied. Here's what came back on the demo account (illustrative demo data):
| Meta Prospecting, Germany | ROI |
|---|---|
| Prior implied by platform-reported revenue | 1.04 |
| Prior implied by pooled experiments (2 tests) | 1.91 |
| Prior mean the model uses, after rules | 2.08 |
So the platform's own numbers would put the channel at roughly half of what the experiments say. The model's prior sits close to the experiment evidence, not the platform figure. That's the point of calibrating with lift tests in the first place.
2. The experiment evidence
Next it reads the experiment calibration evidence for Meta and filters it to the same channel and country. Three conversion lift studies are linked to Meta Prospecting in Germany on the demo account:
| Study window | Test iROAS | Confidence interval | Significance | Used in prior |
|---|---|---|---|---|
| 2026-02-02 to 2026-03-01 | 3.05 | 2.68 to 3.42 | 0.96 | Yes, 92% of weight |
| 2026-06-01 to 2026-06-28 | 2.70 | 2.11 to 3.29 | 0.89 | Yes, 8% of weight |
| 2026-08-03 to 2026-08-23 (full-funnel) | 0.87 | 0.76 to 0.97 | 0.95 | No, excluded by user |
Illustrative demo data.
3. Why the weights look the way they do
The follow-up is where it gets useful. Before pooling, Sellforte corrects each test's iROAS so it's measured on the same basis as the model. After that correction, the February test lands at 1.75 and the June test at 3.66. The pooled figure of 1.91 is the weighted blend of those two, with the February test carrying almost all of the weight.
The August full-funnel study isn't in the pool at all. The evidence table marks it as excluded by a user, so someone made a deliberate call to leave it out. That's exactly the kind of thing you want surfaced before you touch a prior: a highly significant test with a much lower result, sitting outside the calibration.
On the demo data, the run stopped at this comparison. Nothing was proposed or written. When you do ask for a change, the assistant should show the exact field with its old and new value and wait for your confirmation before it saves anything.
Why it matters
A lift test and a model shouldn't be two separate truths. The test result goes into the model as a prior, and the model is then held up against the test. When the two disagree, that's where you learn something about the channel, the test design, or both.
What this question gives you is the full chain in one place: platform view, experiment view, the weights behind the pooled number, and which tests were left out and why. That's what you need before deciding whether to re-include a test, change a weight or leave the prior alone. And because every write needs your confirmation, nothing moves until you've seen the before and after.
One practical note: calibration tools in Sellforte MCP are admin-gated. They're available to Sellforte staff working on a model, not to customer users.
Try these follow-ups
- Show the same prior and experiment comparison for every Meta channel in Germany, in one table.
- Which channels have no experiment evidence behind their prior at all?
- Show the uninformative prior diagnostics for Germany. Which channels' results are driven mostly by the prior?
- Do the same comparison for Meta Prospecting in the US. Do the tests and the prior agree there?
Related
- Sellforte MCP
- Model calibration
- Model validation
- Ask Sellforte MCP: Is our new model run safe to release?
Try it yourself
Book a demo to see Sellforte MCP on a model like yours, or read how to connect Sellforte MCP to your LLM client. New use cases land in this library regularly.
