Proofseal

Performance validation for marketing teams

Can your ad numbersbe trusted with a budget?

Upload ad spend and sales by channel. Proofseal checks the data and methods behind your numbers and tells you how far the evidence goes, and why.

Start free →See an example

Free during the research preview · no credit card

01You provide

  • Weekly ad spend by channel
  • Weekly sales, orders or conversions
  • Or connect GA4, Google Sheets or BigQuery, or import ad platform reports
  • Demand signals such as brand search (optional)
  • Regional data (optional, for regional experiments)

Proofseal does

  • Checks whether the data can answer your question
  • Estimates each channel's effect and its likely range
  • Tests the model on simulated data built to resemble your data's conditions
  • Applies written verdict rules to reach a verdict
  • Analyzes regional experiments: regions where ads paused or grew, compared with regions that stayed the same

03You get

  • A verdict, such as Supported, Conditional, or More evidence needed
  • The reasons and limitations behind it
  • Which data would help most next
  • A report you can share with a manager or client
  • Effect estimates from regional experiments, to compare with the model's estimates

No code needed, and every step comes with a plain-language explanation.

An example: a channel with ROAS 4.2Example

The number you have

ROAS 4.2

The ad platform's report says every $1 spent on this channel brought back $4.20 in sales.

Proofseal checks first

  • How many weeks of data? Enough to see a pattern?
  • Does each channel's spend vary enough to learn from?
  • Does spend tend to follow demand? If so, a model can credit ads with sales that demand brought in.
  • On simulated data with similar conditions and known answers, how often do the model's ranges contain the answer?
  • Does this channel always scale up with others, so credit is hard to split?

What you get

Conditional

4.2 is an attribution figure the platform calculates under its own rules, not necessarily the extra sales the ads caused. There are enough weeks of data, but this channel is usually increased together with Google, so their separate effects cannot be told apart. The verdict supports only a modest shift and names the next evidence that would help: a regional experiment.

What a verdict looks like

Every verdict states its reasons, its limitations and what would help most next.

Supported

Under the data, method and assumptions checked, the evidence supports this conclusion.

Example: "The overall marketing return can inform annual planning." On simulated data with similar conditions, the model's 90% ranges contained the known answer 84% of the time; the provisional threshold in rules v0.1.0 is 70%.

Conditional

The evidence broadly supports it, but a few things need attention, and the verdict lists each one.

Example: "The model's evidence is enough to back a budget shift of up to 10%." The direction has support, but two channels always rise together, so their separate effects cannot be told apart; the verdict lists this as a limitation.

More evidence needed

The evidence is not yet enough to conclude. The verdict says what is missing and which piece would help most.

E.g. "Overhaul the budget split." The model's evidence cannot carry a move that large; one regional experiment would help most.

When is it useful?

Before changing an ad budget

Before you scale up, cut back or move budget between channels, check whether the numbers in hand can carry that decision.

When you explain results to a client or manager

When an agency, consultant or in-house team reports on performance, attach a verdict and its reasons: whether the data is enough and whether the model is accurate under data like this, so the number has evidence behind it.

Before bringing numbers into a meeting

Before presenting to a manager or client, get clear on what the number rests on and what is still uncertain.

After new data comes in or the market shifts

When a new quarter of data arrives, check again and see whether the earlier conclusion still holds.

How is this different from asking an AI?

An AI's answer can change from one ask to the next, and its basis is hard to pin down. Proofseal uses fixed steps and written rules: the same evidence under the same rule version always gets the same verdict.

  • The same steps every time

    Questions of the same kind go through the same data format and the same checks, so wording a question differently does not change the result.

  • The method is written down

    Each verdict records which model and which version of the rules it used, so it can be checked later.

  • Every verdict comes with reasons

    The numbers a verdict relied on and its limitations sit right next to it.

  • Rerun it when data changes

    Rerun after new data arrives and compare directly with last time.

AI is good at reading reports and organizing your thinking. A verdict that decides a budget needs every step to be open to someone else's check.

How the checking works

Five steps, each one recorded.

  1. 1

    Data checks

    First, is there enough data: how many weeks, which weeks are missing, whether each channel's spend moves up and down. With too little data, everything after is unreliable.

  2. 2

    Estimate effects

    Estimate how much each channel contributes to sales, with a likely range around every number, not a single point.

  3. 3

    Test accuracy

    We build simulated data that mirrors your data's conditions, with each channel's answer set by us in advance, then check how often the model's 90% ranges contain that answer.

  4. 4

    Reach a verdict

    Written rules map the check results to a verdict, with the reasons and limitations spelled out.

  5. 5

    Keep the record

    Data version, method, rules and verdict are stored, so you can rerun and compare later.

Check it yourself with a regional experiment

Besides the model, you can pause or raise spend in a few regions and compare them with regions left unchanged. This is closer to direct measurement than reading past data, but the result is still an estimate: it covers those regions and that period, and it comes with a margin of error.

1. Plan: pick the right regions

From past data, find the regions best suited to testing, the regions most like them, and how large an effect this design can detect. Before the experiment starts, you know the smallest change it can pick up.

2. Analyze: compare with the comparison group

After the test, build a comparison group from the unchanged regions, see how far the test region moved, whether that exceeds natural drift, and what each 1 of ad spend did.

A hypothetical example: Ohio pauses Meta ads for 8 weeks and revenue comes in about 13% below the comparison group, an estimated $2.30 less revenue for every $1 not spent. You can set this estimate against the platform's ROAS, keeping in mind that the two are defined and timed differently.

Where the method comes from: synthetic control was published by Abadie, Diamond and Hainmueller in 2010 and is widely used in economics and policy evaluation; Google's and Meta's regional experiment tools use the same family of methods. See the full references

Needs weekly data by region. After signing up there is a regional experiment demo to try first.

Sign up free and see the demo →
Test period
Test region actual Comparison group Test period
Illustration: during the test, the test region's actual revenue (solid) drops below the comparison group (dashed). The gap between the lines is the estimate of the ads' effect.

Said up front: what Proofseal cannot do

Too little data, no forced answer

With less than a year of data, uncertainty is high. The verdict will usually be "more evidence needed" rather than an attractive number you cannot rely on.

A model is not an experiment

A model estimates effects from past ups and downs in spend. An experiment changes spend on purpose and compares, which gives a different and usually more direct kind of evidence. The verdict points out when an experiment would strengthen the evidence most, and Proofseal can help pick test regions and analyze the results.

A verdict does not decide for you

How to split a budget also depends on your goals, stock and calendar. Proofseal answers one question: how far the evidence in hand reaches.

The method is still improving

This is a research preview, and checks and thresholds will keep being adjusted. Every change gets a version number, and earlier verdicts are kept.

Origin

A question that started with a thesis

Proofseal began as a prototype I built for my thesis research in the Master of Science in Decision Making and Applied Analytics at Minerva University. From evidence-based decision making, statistical modeling and formal reasoning to the marketing measurement questions of daily work at CloudAD and the complex operational decisions at plantārium, I kept running into the same question: once a measurement result has been computed, how far does its evidence actually reach? Proofseal starts from that question.

Read the origin of Proofseal →

The path of instruments

Before statistics existed, people used these instruments to measure what they could not see, and to leave a record of what their judgment stood on. Proofseal is the modern version of the same work.

  1. 1 · Connect data

    Core sample

    A drilled column of rock: layer upon layer, time keeping its own record

  2. 2 · Data checks

    Plumb line

    A weighted string, the oldest standard: only with a true line can you speak of tilt

  3. 3 · Measurement

    Gnomon

    A rod and its shadow: the sun's unseen height, worked out from what can be seen

  4. 4 · Validation

    Armillary sphere

    The observatory's rings: hold the model against the sky, and correct it where they disagree

  5. 5 · Verdict

    Seal

    Checked and approved, then stamped: someone answering for the result

Frequently asked questions

What is Proofseal?+

A validation tool for marketing teams. Upload weekly ad spend and results by channel, and Proofseal checks the data and methods behind the performance numbers, then states how far the current evidence goes, with its reasons and limits.

What data do I need?+

Weekly ad spend by channel, plus weekly sales, orders or conversions. Upload a CSV, import ad platform reports, or connect GA4, Google Sheets or BigQuery. A year or more of data is ideal.

How is this different from ad platform reports or GA4?+

Platform reports and GA4 tell you what a number is. Proofseal checks what the evidence behind it can support: whether there is enough data, how accurate the model is on simulated data with similar conditions, and whether spend tends to follow demand, which can lead a model to credit ads with sales that demand brought in.

What verdicts are there?+

Supported, Conditional, Not supported, and More evidence needed. Every verdict comes with reasons, limitations and next steps. When data is updated you can recompute; earlier verdicts are kept so you can compare them.

Can I verify with my own experiment?+

Yes. Pause or raise spend in some regions: Proofseal first helps pick suitable test regions, and after the experiment it estimates the effect for those regions and that period with synthetic control, then gives a verdict. Synthetic control was published by Abadie and colleagues in 2010; the regional experiment tools from Google and Meta take a similar approach.

How much does it cost?+

Free during the research preview, at least through December 31, 2026, with no credit card. Existing users get 30 days' notice before any billing starts.

Is my data safe?+

Data is stored in Google Cloud's Taiwan region and encrypted in transit and at rest. We do not sell data or use it to train AI, and the site has no third-party tracking.

Try it once with your own data

Free during the research preview, no credit card. After signing up there is a demo project where you can read a complete verdict first.

Start free →