CMO Touchstone

CMO Touchstone Marketing Management

Attribution and incrementality for the CMO — the chief marketing officer — and the marketing team. A touchstone tests whether the gold is real.

“Which channel do we cut?” The same journeys under six models, side by side. The channel that looks worst on last click is usually the one doing the introducing.

“Revenue is up. Was that us?” Test against control with the confidence interval attached. When a lift cannot be told apart from zero, it says so.

“How big does the test need to be?” The traffic required to detect the lift you would actually act on — before you run it, rather than after it fails.

Free. Your figures never leave your browser — no signup, nothing you enter is uploaded. The chat assistant is the one part that talks to a server, and only what you type into it is sent.

Who gets the credit

Showing a sample path set so you can see how it works. Paste your own over it.

What to paste

One journey per line: the channels in the order they were touched, then how many converted, then how many did not.

THE CHANNELS In order, separated by >. Use whatever names your reporting uses — they are just labels, and the tool never assumes it knows what they mean.

CONVERSIONS How many journeys took exactly that path and converted. This is what every model shares out.

DID NOT CONVERT How many took the same path and did not. Optional, but it is what unlocks the causal model — without it, every journey in the data converted and removing any channel appears to cost you everything.

Where to get it, and what each source gets you

Be warned: the third number is harder to obtain than the first two, and it is the one that unlocks the causal model.

FIVE MINUTES — FIVE MODELS GA4: Advertising → Attribution → Conversion paths. Most other platforms call it a path or multi-touch report. This gives you CONVERTING paths only, so paste it with two numbers per line and you get the five rule-based models.

A DATA REQUEST — ALL SIX The non-converting journeys are not in that report, and GA4 does not export the conversion paths report to BigQuery at all. They have to be rebuilt from raw event data with SQL, or pulled from your CDP or warehouse. Ask for sessions grouped by channel sequence, with converted and not-converted counts.

EITHER WAY Roll the long tail into an Other channel rather than pasting thousands of one-off paths. The models do not need volume, they need the shape of the journeys.

Five models is still a real answer — the gap between first click and last click alone tells you how much of the journey your reporting is ignoring. But the Markov removal effect is the only column here that asks what you would actually lose, so it is worth the data request if the argument you are settling is a budget one.

Rule	Why
Order matters	Five of the six models score purely on position, so `Email > Search` and `Search > Email` are different journeys with different answers.
Keep Direct in	It is usually the biggest channel on last click and it starts almost nothing. Leaving it out hides the problem; leaving it in is how you see how much of your reporting is really a tracking gap.
Group the tiny channels	Anything under about 2% of conversions cannot be told apart from noise, so ranking those against each other is reading tea leaves.
Rough beats blank	A path export with the long tail grouped tells you far more than a perfect one you never got round to producing.

JourneyConvertedDid not convert

Journeys

Channels in the order they were touched, separated by >. The number beside each row is how many touches were read — if it says 1 where you meant 2, a > is missing. Did not convert is optional and is what unlocks the causal model.

One journey per line: channels, converted, did not convert. Edits here rewrite the rows above, and the rows above rewrite this, so the two can never disagree.

Model settings

Only two of the six models take a setting. They are here so the numbers can be matched to whatever your reporting already uses, not because they should be tuned until the answer looks better.

Time decay half-life touches

Position: first touch %

Position: last touch %

Did it actually work

Attribution divides up conversions that already happened. This is the only page here that produces evidence: a group that got the treatment, a group that did not, and an honest interval around the difference.

The test

Counts, not rates — the size of each group is what decides whether the result means anything.

Test group people

Test conversions

Control group people

Control conversions

Confidence %

Power %

Size the next test

The question to answer before you run it, not after it comes back inconclusive.

Baseline conversion rate %

Lift worth acting on %

Confidence %

Power %

What to do about it

The Attribution tab shows where two defensible models disagree. This asks the next question: which disagreement is worth spending a test on, and what is standing in the way of trusting any of it? Everything here is read from your own data rather than assumed.

Every measurement model rests on assumptions. These are the ones behind these numbers, so you can judge where they hold and where they do not.

What is assumed	Why it matters
Your path data is complete	Every model divides up what it can see. Cookie loss, cross-device journeys, walled-garden platforms and consent refusals all remove touches, and the touches that go missing are disproportionately the early ones. A large Direct channel is usually the symptom.
The order is right	Five of the six models score purely on position. If your export orders touches by anything other than time, the answers are confidently wrong.
Channels are the right unit	Credit is assigned to a channel, not a campaign, creative or audience. Two campaigns doing opposite things inside one channel cancel each other out here.
The Markov chain is first order	The next touch depends only on the current one, not on the whole history. Real journeys have memory, so treat the removal effect as a strong signal rather than a precise quantity.
Attribution is not causation	Even the removal effect is causal only within the model of your data. It cannot see the customer who would have bought anyway, and no attribution model can. That is what the holdout is for.
The holdout was random	The two groups must differ only in the treatment. A geo test where the treated cities are also your best cities measures the cities. The maths cannot see that.
One test, one conclusion	Run twenty tests at 95% confidence and roughly one will look significant by chance. Nothing here corrects for testing many things at once, so decide what you are testing before you look.

None of this makes the tool useless — it makes it honest about which of its numbers are evidence and which are argument. The Attribution tab produces argument. The Incrementality tab produces evidence. Do not confuse them.

CMO Touchstone by CMK Sons Labs. Six attribution models including a Markov removal effect, and two-proportion holdout testing with confidence intervals and sample sizing. Every figure is computed on your own device and nothing is sent to a server. A touchstone is the stone used to test whether gold is real.