Sergio Iacobucci
·
Every B2B SaaS company has an ICP document. Most of them read the same way: mid-market, 50 to 500 employees, “digitally mature”. Ask the sales team when they last opened it and you’ll get a blank look. Which is fair enough, because it’s a description of who the company would like to sell to, written in an afternoon and never tested since.
An ICP needs to be data-driven and heavily validated, it's not a quick an easy task. Here’s how to find out whether yours has gone through the correct process.
1. Start with the customers you already have
Name your three best customers. Not the biggest logos: the ones who bought fast, churn least, expand the most and refer others. Now work out what they share.
To do this you need to push past the firmographics. Industry and headcount are the easy answers and they’re usually the least interesting ones. The better question is why each of them bought: the specific moment they decided to go looking for a solution. A new VP arrived. A compliance deadline landed. The spreadsheet finally broke. Triggers repeat across your best customers far more reliably than industries do, and sales can act on a trigger. At least at that point, the pain is real.
If you have fewer than five paying customers, be honest that you’re working with a hypothesis, not evidence. Write it down anyway. Just don’t defend it like it’s data.
2. Treat churn as evidence
Your churned accounts are the most honest dataset you own. Nobody churns politely to spare your feelings.
Every account that leaves is marking the edge of your ICP for you. Line up your lost and churned accounts and look for the pattern: what did they have in common on the way in? Often you’ll find the warning was visible at the first call and everyone ignored it because the logo was nice and shiny.
This is where your disqualifiers come from. A disqualifier that came from three real churns (“no internal owner”, “bought us to fix a problem the product doesn’t solve”) is worth more than a page of fit criteria, because it stops bad deals entering the pipeline instead of explaining them after they’ve died.
3. Watch where the value actually lands
Knowing a customer’s industry and headcount won’t tell you why they renew. For that you need to see the product through their eyes, and the best place to do it already exists in your business: the QBR/weekly calls/product usage data.
I’ always go deep in QBR call recordings for clients, listening to how their customers describe the product when nobody’s selling to them. (You develop a strange affection for other people’s customer success teams after a week of this.) What a customer says they’d miss in a renewal conversation is the sharpest ICP signal there is, and it’s frequently NOT the thing on the homepage.
Read the QBR notes. Listen to the calls. Compare which modules your best accounts use against which ones they bought. When the accounts that renew and expand all lean on the same corner of the product, the companies who need that corner are your market. Build the ICP around them.
4. Get your data into a state you can trust
Everything from here on involves running numbers against your account list, and the numbers are only as good as the list. Raw CRM exports are always worse than anyone admits. The industry field is half-blank free text with ten spellings of “manufacturing”. One global customer is scattered across ten country rows, which means your biggest account might be masquerading as ten mediocre ones. And sometimes each client can be paying in different currencies.
Three fixes before you test anything. Roll subsidiaries up into parent accounts, conservatively: same brand is not always same company, and merging two independent co-ops because they share a name will cause you pain further down the line. Classify every account against a fixed list of sub-industries instead of whatever the rep-typed in 2019. And put every value in one currency.
None of this is glamorous. However it's extremely important because the segmentation you're doing gets built on the raw export and every number downstream inherits the mess. It needs to be right.
5. Enrich the base with AI agents
The cleanup in step 4 used to be the real reason nobody validated their ICP: classifying and verifying a few hundred accounts by hand is weeks of work nobody has. It isn’t anymore.
I run enrichment with a fleet of AI agents working in parallel. Each one takes a batch of accounts and the fixed taxonomy, checks the less obvious companies against their actual websites and writes back a label, a confidence level and a one-line evidence note. The fixed list matters more than the agents do: ten agents left to pick their own words will invent ten different labels for “car dealership”, the same way ten analysts would. Closed vocabulary in, consistent data in the columns out.
Run a pilot batch of 40 accounts first. It’ll expose the gaps in your taxonomy while they’re easy to fix, and then you can let the rest loose. A few hundred accounts, classified and web-verified in an afternoon, and every judgment stays auditable because the evidence travels with the label.
The part I like most: the engine is reusable. Once it exists, testing a new hypothesis against the whole customer base (“how many of our customers run a distributed brand model?”) stops being a quarter-long project and becomes a question you can ask on a Tuesday.
6. Test whether each dimension earns its place
Your ICP probably scores accounts on several dimensions: region, vertical, company type, product depth, maturity. They look like separate cuts. Run the numbers and they usually aren’t. One client had four dimensions in their segmentation: geography, platform, customer cohort and module depth. Statistically they were one axis. The four cuts moved together so tightly that segmenting on all of them meant counting the same thing four times, while looking four times more sophisticated.
The test is simple to describe: does each dimension still predict account value after you control for the others? Most don’t. They’re proxies for company size or for each other. A dimension that survives the controls earns a place in your ICP. A dimension that doesn’t gets demoted to metadata. Fewer, cleaner axes beat an impressive-looking matrix every time.
7. Separate gates from scores
A segment can be worth more on average AND be useless for ranking any individual account. Those are different claims, and mixing them up is probably the most common ICP mistake I see.
Think of it as a guessing game: if I tell you an account has the fit criterion, how much does your guess about its value improve? For one client, accounts matching their fit profile were worth 33% more after controls. It sounds like a scoring variable however the overlap between fitting and non-fitting accounts was so large that fit explained only around 6% of the variance. Knowing an account fit barely narrowed down the guess.
So treat a dimension like that as a gate. Fit decides whether an account belongs in the pipeline at all; the dimensions that survived step 6 decide which account to call first.
8. Clean the base before you quote a number
Somewhere in your deck is a line like “ICP accounts are worth 2.5x more”. Before you repeat it in a board meeting, check who’s in the denominator.
Dead and dormant accounts sit in every CRM: cancelled customers never marked as churned, zero-revenue stubs, duplicate entries from an old billing migration.
Pull the dead accounts into their own bucket, recompute, and only then quote the multiple. If a number is going to decide where your sales team spends the year, it should survive ten minutes of hygiene.
9. Write it down so sales can use it
The output is one page.
One line at the top: we sell [what] to [who] when [trigger]. Your two or three real segments, each with a named flagship customer and the structural reason it fits. The disqualifiers, with their receipts. A scoring rubric small enough to use on a live call: fit, timing, access, intent, three levels each. And the deal economics, because an ICP with no revenue attached is a persona.
Then revisit it quarterly against actual conversions: which segments closed, which disqualifiers held, which signals predicted anything. An ICP is a working document. The one from six months ago is a historical artefact.
If you want help with it (the churn interviews and QBR digging are the parts most teams never get to): that’s what I do. sergio@thedialog.co.uk
