How to Test an ICP Hypothesis With Outbound

An ideal customer profile (ICP) is a claim about who will buy from you. If you write it once, send a lot of email, and then look at the results, it is hard to say whether they prove anything. This guide takes a different route. You treat the profile as a set of hypotheses, test one at a time with a small batch, and write down what would change your mind before you send anything.

A hypothesis names a segment, a reason, and a result that would count as support. Here is one:

"Companies of this type, with this trigger, reply to this offer and book meetings."

Compare that with a profile such as "mid-market SaaS companies." That describes a population. It does not predict anything, so no result can contradict it.

Outlevo's own tooling uses the same framing. Guesses are labelled as hypotheses, never presented as findings [[1]](#sources). Segments proposed by AI are labelled experimental until a person approves them [[2]](#sources). You can borrow that discipline even if you never use the product: keep a short list, mark each item "hypothesis" or "observed," and move items between the two only when evidence moves them.

If you want the background first, read How to Define an Ideal Customer Profile and the glossary entry on the ideal customer profile (ICP). This guide starts after you have a draft.

Who this is for

This guide is for founders, sales leads and marketers at B2B companies who have a draft ICP and want evidence before they commit budget, mailboxes and time to it.

It fits you if:

You have two or three plausible segments and cannot tell which one deserves a full campaign.
You have run outbound before and the results were mixed, and you cannot say which part of the profile was wrong.
You have past deals in a CRM or a spreadsheet and have not yet compared them with the profile you use today.

It does not fit if you need an answer in a day. A real test has to wait for replies and meetings.

Steps

1. Collect candidate hypotheses from evidence you already have

Start with what happened, not what you hope will happen. Your past deals are the cheapest evidence you own.

If you use Outlevo, you can import past deals from a CSV or your CRM and get a win/loss analysis that proposes ICP hypotheses [[3]](#sources). The CSV import takes won or lost deals with fields such as stage, amount, outcome and loss reason [[4]](#sources). HubSpot and Salesforce connections read companies, deals or opportunities for the same analysis [[5]](#sources) [[6]](#sources).

Read the confidence label before you trust any proposal. The analysis counts companies, not deals, and under 10 won companies is labelled very low confidence [[2]](#sources). Treat a segment drawn from so few wins as a guess to test, not a conclusion.

If you have no deal history, the free ICP assessment reads up to 5 public pages of a website and returns up to 5 experimental segment hypotheses [[7]](#sources). It labels its facts as verified observations or AI inferences [[7]](#sources). Either way, the output is a list of candidates, not an answer.

For more on how teams describe segments, see ICP Examples for B2B Teams.

2. Pick one segment

Choose one. If you test three segments in one batch and meetings come in, you will not know which segment produced them.

Narrow the segment until you can describe it in one sentence a colleague could apply to a list. Use criteria that a person can check from public information, such as what the company sells, who it sells to, or a visible trigger.

Markets are your decision. Outlevo asks which countries you sell into and never proposes markets [[2]](#sources). Pick the countries from your own business before you build the list, and keep them fixed for the whole test. Changing them midway turns one experiment into two.

3. Set the success measure and the sample size before you send

Do not skip this step: it keeps the rest of the test honest. Write these down before the first email goes out:

The measure. Pick one primary measure. Booked meetings with people in the segment is the strongest. Positive replies is the next. Total replies is the weakest, because a reply can be "remove me."
The denominator. Define reply rate in a way you can repeat. One public example is Belkins' 2026 cold email study, which covers 7.5M B2B emails sent in 2025 and defines reply rate as replies ÷ emails sent [[8]](#sources). Whatever definition you choose, use it for every batch. Do not compare it with other people's numbers unless their definition matches yours.
The sample size. Decide how many companies the batch will cover. Count companies, not emails, the same way Outlevo's win/loss analysis counts companies, not deals [[2]](#sources). A company that gets four follow-ups is still one data point.
The decision rule. Write three lines: what result means keep, what means narrow, what means drop.

For example: "If at least X companies out of Y book a meeting, keep. If replies come but few are positive, narrow. If fewer than Z reply at all, drop." You choose X, Y and Z. They should reflect how much a wrong answer would cost you and how many conversations your team can actually handle.

A caution on sample size. A small batch is unlikely to detect a small difference. If your decision rule needs a precise result, the batch is too small, and you should say so now rather than after the replies arrive. Be willing to call a result "inconclusive" and run a second batch on the same segment.

Do not set the bar after you have seen the replies. That is how a weak result gets described as encouraging.

4. Get sending ready

A test on a broken setup tests your setup, not your segment.

If you send through Smartlead, every mailbox must be connected and warmed up before a campaign sends, and its domain must have SPF, DKIM and DMARC checked in the last 7 days [[9]](#sources). If Outlevo manages your mailboxes, a new mailbox sends nothing until it is warm. That takes at least 14 days and usually about 3 weeks [[10]](#sources). Plan the test around that wait. Build the list, write the hypotheses and set the decision rule while the mailboxes warm up.

You can also rehearse the loop without sending. Test API keys run against a sandbox with 100 fake companies, so you can build and check a workflow without sending anything or spending credits [[11]](#sources).

Check the basics of compliance now, not after the first complaint:

The law makes no exception for business-to-business email, according to the FTC's CAN-SPAM compliance guide [[12]](#sources).
Google's sender guidelines say: "Marketing messages and subscribed messages must support one-click unsubscribe, and include a clearly visible unsubscribe link in the message body." [[13]](#sources)
If you rely on legitimate interest under GDPR, the text of Article 6(1)(f) covers cases where "processing is necessary for the purposes of the legitimate interests pursued by the controller or by a third party." [[14]](#sources)
The UK regulator says: "You can email or text any corporate body (a company, Scottish partnership, limited liability partnership or government body)." [[15]](#sources)

These are quotes, not legal advice. Ask a lawyer about your own case.

If LinkedIn is part of the test, read the platform's rules first. Its user agreement prohibits: "Use bots or other unauthorized automated methods to access the Services, add or download contacts, send or redirect messages" [[16]](#sources).

5. Approve the batch: where the human step fits

Before the batch runs, a person reads it. This is the checkpoint that protects the experiment from drift.

In Outlevo, nothing is sent outside a policy that a person approved [[17]](#sources). Policies set who can be contacted, on which channels, how often and with what budget [[17]](#sources). Every external action is checked against the approved policy before it runs and recorded in a ledger [[17]](#sources). For a hypothesis test, that gives you a written record of what the batch was allowed to do.

What the approver should check:

The list matches the segment. Open a dozen companies at random. If a third do not fit the one-sentence description, fix the list, not the copy.
The evidence behind each account. A Readiness Card carries fit, intent and data quality scores, and needs at least three observations, each with a source and a date [[1]](#sources). Facts read from a website are kept only when their quote is found on the page [[1]](#sources). Look at the observations on a few cards and check that the fit claim rests on them.
The message tests the hypothesis. The offer should speak to the reason you picked the segment. If the copy is generic, a weak result tells you about the copy.
The limits. Confirm the channels, the cadence and the budget. Kill switches pause a campaign, a channel or the whole workspace at once [[17]](#sources). Know who can press them before you start.
The credit cost. Qualification costs 3 credits, light qualification 1 credit and contact discovery 4 credits, and the Builder plan includes 320 credits a month [[18]](#sources). Sending and follow-ups through Smartlead use no credits [[9]](#sources). Check the full list on the pricing page before you size the batch.

Save the approved version. Your later reading depends on knowing exactly what was sent.

6. Run the batch and leave it alone

Resist editing mid-test. If you change the offer after two days, you have two tests with half the sample each.

Smartlead sends sent, reply, bounce, unsubscribe and category-change events back to Outlevo by webhook [[9]](#sources). That lets you read replies against the hypothesis in one place. If you want the same events elsewhere, Outlevo can send signed webhooks to an HTTPS address, at 1 credit per delivered event; failed attempts and test pings are free [[19]](#sources).

Meetings are the cleanest outcome to track. Each Calendly or Cal.com booking with one of your contacts counts as a meeting booked and is written to your CRM when one is connected [[20]](#sources) [[21]](#sources). For an interested reply, the answer goes at once on autopilot. Otherwise it waits in Replies for you to send [[20]](#sources) [[21]](#sources). For a first test, consider leaving autopilot off, so a person sees each interested reply before it goes. That is a choice, not a requirement.

If you use HubSpot, Outlevo writes status, fit score, ICP version and next step to its own fields only, and nothing is written until an owner or admin approves the field mapping [[5]](#sources). The ICP version field gives you a place to tie each result to the version of the profile that produced it.

7. Read the results

Read in this order:

Delivery first. Check bounces and unsubscribes. If delivery was poor, the result says little about the segment.
Replies by type. Sort replies into interested, not now, wrong person, not relevant and remove me. The mix matters more than the total.
Meetings. Count companies that booked, against the companies in the batch.
Reasons. Read the actual replies. "We already use a competitor" is a different finding from "we don't have this problem." If competitors come up, Competitor tracking software for B2B covers one way to follow them.

Then compare the result with the decision rule you wrote in step 3. Do not rewrite the rule now.

8. Keep, narrow or drop

Each hypothesis ends in one of three states.

Keep. The result met your rule. Do not scale yet. Run a second, larger batch on the same segment to see whether it holds. Change the label from "hypothesis" to "supported," and keep the sample size next to it.

Narrow. There was a signal, but it came from part of the segment. Look for what the companies that replied have in common: size band, role, trigger, tooling. Write that as a new, tighter hypothesis and test it as a new batch. Do not pull a sub-segment out of a small batch and call it proven. It is a fresh guess.

Drop. The result missed your rule and the replies give no pattern to narrow on. Record it, with the date and the batch size, so nobody re-tests it by accident in six months. A dropped hypothesis is a useful result. It saved you a larger campaign.

Inconclusive is a fourth outcome, and it is allowed. If the sample was too small to separate the outcomes, say so and run another batch.

A person should make this call. In Outlevo, AI-proposed segments stay experimental until a person approves them [[2]](#sources). Use the same rule for your own list: nothing moves out of "hypothesis" without a named person signing it off.

What to check

Use this list before you call the test done.

You tested one segment, with fixed countries.
The success measure, sample size and keep/narrow/drop rule were written down before sending.
Samples were counted in companies, not emails.
If you sent through Smartlead, mailboxes were connected and warmed up, and SPF, DKIM and DMARC were checked within the last 7 days [[9]](#sources).
Every message had a working unsubscribe, in line with Google's guidance [[13]](#sources).
A person approved the list, the copy, the channels, the cadence and the budget before the batch ran.
Nothing about the offer changed during the test.
Replies were sorted by type, not just counted.
The decision was recorded with the date, the batch size and the profile version.
Inconclusive results were labelled as such.

Common mistakes:

Testing a segment and a new offer in the same batch.
Reading a reply count as success when most replies were refusals.
Treating a very small number of won companies as proof. Under 10 is labelled very low confidence in Outlevo's analysis [[2]](#sources).
Comparing your reply rate with someone else's without checking how each was defined.
Dropping a segment because of a deliverability problem.

Next step

Pick one segment from your draft ICP and write its hypothesis in one sentence. Then write the measure, the sample size and the keep/narrow/drop rule next to it.

If you want a starting list of hypotheses from your public site, run the free ICP assessment. If you have past deals, import them as described on the CSV and Excel page, or connect HubSpot or Salesforce. To see how approval, policies and the ledger work, read How Outlevo works. To rehearse the loop without sending, the developer hub covers the sandbox.

If you are choosing data tools to build the list, see Data and enrichment tools: pricing compared and Outlevo vs Clay. For the wording of a campaign step, the glossary entry on sequence is a short reference.