Outbound Plays

ICP testing outbound experiments

Pit verticals, personas, and angles against each other in small controlled batches, and learn which slice actually converts.

Kill the losers early. Pour the budget into the slice that wins.

By Kshitij · Updated June 2026 · 8 min read

Book a Fit Check

Play Snapshot

Learning-led

Outcome

A proven ICP slice, plus the meetings the test booked along the way

When to run it

Your ICP is a hypothesis, or reply quality is uneven across segments

Signals it uses

A few candidate slices, controlled batches, clean tracking

Channel mix

Email and LinkedIn, one variable held steady per batch

3 to 5 slices at a time Read on replies and meetings

ICP testing outbound is the motion where you run small controlled batches across different slices of your market, a vertical crossed with a persona crossed with an angle, to learn which slice actually converts. You read positive replies and booked meetings, not opens, then kill the slices that miss and scale the one that bites.

When to run it

Run it at one of three moments

This is the learning play. It earns its keep when you are not yet sure who you should be selling to, or which slice deserves the next dollar.

1

Early, ICP still a guess

You have a hunch about who buys, but no evidence yet. Test the hunch in small batches before you commit a quarter to it.

2

Scaling, picking where to pour

Volume is going up and you have to choose a slice to fund. Doubling down on the wrong one is the expensive mistake. Test first.

3

Reply quality is uneven

Some segments reply warm, others go cold, and you cannot tell if it is the message or the market. A controlled test separates the two.

The point is not to send more. It is to learn faster, so the budget follows evidence instead of a guess. If you already know your winning slice cold, skip this play and go straight to expansion.

What the play uses

This is not a signal play, it is the learning layer that sits above the signal plays. It needs three inputs, and it falls apart if any one is missing.

Input 1

A few candidate slices

Three to five slices defined as vertical crossed with persona crossed with angle. Enough to compare, few enough to read. A slice can hang off a signal you can detect, like a recent job change.

Input 2

Controlled batches

Equal-size batches per slice, sent in the same window, with one variable changed at a time. Hold the rest steady so the result points at the slice, not at the noise around it.

Input 3

Clean tracking

Replies tagged positive or not, meetings attributed to the right slice, and a scorecard you actually read. Without clean tracking you are guessing, just with more steps.

The read

The slices can be built from anything you can target cleanly: a vertical, a company size band, a persona, or a buying signal you already track. If you want help deciding which signals predict your buyers, and therefore which slices are worth testing first, that is what signal mapping does, and the signal and intent tools guide covers the software that feeds it.

Not sure which slices are worth testing first?

Book a Fit Check

The experiment

How we design and read it

Every slice gets one row. You write the hypothesis, the batch size, the metric, and the decision rule before a single email goes out. Decide what would make you kill it in advance, so the result reads itself.

Slice	Hypothesis	Batch	What we measure	Decision rule
Vertical A, VP persona	They feel the pain acutely and own the budget	Equal, sized to read	Positive replies, then meetings booked	Clears the bar twice, expand it
Vertical A, Director persona	Same pain, but the buyer sits one level down	Equal, sized to read	Positive replies, then meetings booked	Below the bar once, hold and watch
Vertical B, same persona	A neighbouring market with the same role	Equal, sized to read	Positive replies, then meetings booked	Two cold batches, kill it
Vertical A, pain angle	Same slice, opener leads on the cost of the problem	Equal, sized to read	Positive-reply rate vs the control angle	Beats control, make it the default
Vertical A, outcome angle	Same slice, opener leads on the result you drive	Equal, sized to read	Positive-reply rate vs the pain angle	Loses, retire it, free the volume

The loop around the table

The table is the design. This is how a round runs, start to finish, then repeats with the survivors.

1

Define the slices

Pick three to five slices as vertical crossed with persona crossed with angle. Write the hypothesis for each one, plainly, before you build a list.

2

Send controlled batches

Equal batches, same window, one variable changed at a time. Hold cadence, channel, and offer steady so the slice is the only thing moving.

3

Read replies and meetings

Score on positive-reply rate and meeting rate, never opens. Treat the first small batch as a hint, then confirm the promising slices on a larger second batch.

4

Kill losers, expand winners

Cut the slices that miss the bar twice and free the volume. Pour it into the slice that bit, then start the next round on a fresh question.

The one rule

Change one thing at a time. The moment you swap the slice and the opener and the channel together, you have learned nothing, because you cannot tell which change moved the number. One variable per batch, every time, or the test is just sending.

Where it wins, and when it fails

A test is only useful if you trust the read. Here is the honest case for and against.

Where it wins

✓Turns your ICP from a guess into evidence
✓Stops you funding the wrong slice at scale
✓Books meetings while it learns, the test is not free spend
✓Compounds, every round narrows the next one

When it fails

!Reads noise as signal when batches are too small
!Tells you nothing if you move many variables at once
!Slow if you need pipeline this week, not a read
!Wasted if you never act on the result and kill nothing

Common mistakes

What ruins the read

Four ways teams run a test that looks rigorous and teaches them nothing. Each one is common, and each one is avoidable.

✕

Changing too many variables

Swap the slice, the opener, and the channel in one batch and you cannot attribute the result to any of them. Hold everything but the one thing you are testing.

✕

Samples too small to read

At normal reply rates a tiny batch gives you two or three replies. The jump from two to three looks like a big lift and means almost nothing. Treat small batches as directional and confirm before you scale.

✕

Optimising for opens and clicks

A high open rate tells you the subject line landed in the inbox, not that the slice cares. Score on positive replies and meetings, the metrics that become pipeline.

✕

Never killing the losers

Keeping every slice alive out of hope is how a test becomes a permanent spread bet. Set the kill rule up front, then honour it. A test that never cuts anything is not a test.

Want the slice experiments designed and run for you?

Book a Fit Check

How we would run it

The experiment in motion

An illustrative walkthrough of the method, not a specific client result. We report real numbers only when they are real.

1

The slices

Four hypotheses

A new founder believes ops leaders in two verticals will bite. We write four slices: two verticals, two personas, with one shared angle to start.
2

The batches

Equal and steady

Equal batches go out in the same window, same cadence, same offer. Only the slice changes, so the read points at the audience, not the setup.
3

The read

One slice bites

One slice replies warm and books meetings, one is flat, two are quiet. A larger second batch confirms the warm slice held, so it was not a fluke.
4

The call

Kill, then scale

The two quiet slices are cut, the volume moves to the winner, and the next round tests two angles inside it. The guess is now a proven slice.

Plays like this fed pipeline we've built inside these companies

Palm.ai

Alcméon

Mindflow

CEF.AI

Boolee

CoachHub

Inrō

Buster.AI

Palm.ai

Alcméon

Mindflow

CEF.AI

Boolee

CoachHub

Inrō

Buster.AI

FAQ

Questions founders ask

What is ICP testing in outbound?

ICP testing in outbound is running small controlled batches across different slices of your market, a vertical crossed with a persona crossed with an angle, to learn which slice actually converts. You hold the variables steady, send to each slice, and read positive replies and booked meetings rather than opens. The slice that earns real conversations is the one you scale. It turns who you target from a hypothesis into something you have evidence for.

How is ICP testing outbound different from a normal A/B test?

An A/B test usually pits two versions of one variable against each other, a subject line or an opener, on the same audience. ICP slice experiments test the audience itself: which vertical, which persona, which angle. The discipline is the same, change one thing at a time so you can read the result, but the question is who you should be selling to, not which wording wins. You can A/B test copy inside a winning slice once you have found it.

How big does each slice batch need to be to read anything?

Big enough that the result is not noise. At typical reply rates a batch of fifty or seventy-five contacts gives you two or three replies, and the gap between two replies and three looks like a fifty percent swing while meaning almost nothing. Treat small batches as directional, not as winners. We read the early batch as a hint, then send a larger second batch to the promising slices to confirm before we pour budget in. If the number drops on the bigger sample, the fit was narrower than it looked.

Which metric tells you a slice is working?

Positive-reply rate and meeting rate, in that order. Opens tell you about inbox placement and a subject line, not about whether the slice cares. A high open rate with no replies is a slice ignoring you politely. We judge a slice on how many positive conversations and booked meetings it produces per batch, then on whether those meetings turn into real pipeline. Volume metrics flatter losing slices, so we keep them off the scorecard.

When should you run ICP slice experiments?

Run it early, when your ICP is still a hypothesis and you need to find the slice that bites before you commit a quarter to it. Run it when you are scaling and have to decide where to pour budget, because doubling down on the wrong slice is expensive. And run it when reply quality is uneven across segments and you cannot tell whether the message or the market is the problem. If you already know your winning slice cold, this is not the play, expansion is.

What happens after a slice wins?

You stop testing it and start scaling it. The learning job is done, so the motion becomes finding more accounts that look like the winner, which is the best-customer expansion play, lookalikes by shared traits and signals. You keep a small share of volume in test mode so the next slice is always being found, but the bulk of the budget follows the proven slice. Testing finds the winner, expansion compounds it.

Keep going