Is Your ICP Still a Portrait? Six Stages to a Dynamic, Data-Driven ICP
By Arpit Srivastava & Hitika Saini, Growth Natives ·
When did you last update your ICP? Most teams say they have a clear, confident answer. Arpit Srivastava's experience is that if you dig two or three layers deeper, something always turns up that the official definition missed. In this session, Arpit and Hitika Saini of Growth Natives walked through how they build an ideal customer profile backward from a company's own customer data, score the prospect database against it, and push the result into outreach, expansion and churn work. This recap covers the six-stage model, how their hybrid scoring works, what the demo showed on synthetic data, and why they think the people running the model matter as much as the model.
Why does a static ICP stop working?
Arpit described it as a portrait versus a mirror. A portrait painted five years ago shows who your customers were back then. A mirror shows who they are now. Plenty of ICP definitions were set by someone in the C-suite five or ten years ago and never checked against the data since.
His other point is that an ideal customer is different from an average one. Some customers simply pay you. Others drive your profitability and growth, and the job is to separate those two groups. If you already have somewhere between 1,000 and 10,000 customers, you're sitting on enough data to do that. The questions to answer are which segments bring in the most revenue, which grow fastest and which are most profitable.
A CRM like Salesforce will tell you total revenue by account, pipeline, closed-won deals and contact activity. It won't tell you which segments are profitable, which accounts expanded year over year, or which cohorts those accounts belong to. Getting there takes some work, and that work is what the ICP gets built on.
What are the six stages of a dynamic ICP?
| Stage | What happens |
|---|---|
| 1. Grade the data | Score CRM data for completeness and harmonization, produce a health card, then enrich the blanks |
| 2. Rank customers by value | Look past revenue to growth, margin, retention, expansion, product adoption and how fast deals closed |
| 3. Discover shared traits | Find the industry, sub-industry, revenue range and employee count combinations with the most revenue and growth concentration |
| 4. Score the prospect database | Run lookalike modeling against customer and prospect databases to find accounts that resemble the growth accounts |
| 5. Tier and activate | Grade accounts into Tier 1, 2 and 3 (or A, B, C) and use the tiers to drive ABM programs |
| 6. Protect and expand | Spot expansion opportunities and churn risk in existing accounts |
What makes a customer a "growth" customer?
Revenue is only one input. Growth Natives also looks at whether a customer bought several products up front, how quickly the deal closed, and whether the customer renewed two or three times. One purchase could be a coincidence. Two or three renewals is evidence. The top cohort that comes out of this ranking is the evidence base for everything after it.
In their experience the Pareto principle usually holds: 20 to 25% of customer accounts produce somewhere between 70 and 90% of revenue. If the first three stages show you that pattern, Arpit said, you're headed in the right direction.
Why does stage six matter?
Churn among accounts that look like your growth accounts should get top priority. When one of those leaves, you lose more than its current revenue. You also lose the growth you should have been able to count on from it.
How does the hybrid scoring model work?
The scoring blends two approaches. The first is rule-based. During discovery, Growth Natives learns a client's target geos, priority industries and existing ICP definitions. Most clients already have some kind of scoring set up in Marketo, HubSpot or an ABM platform like 6sense, and those rules get carried over into the model.
The second is a machine learning model that works across all of the data. The reason for using both is bias. Rules carry business context, but they also carry whatever assumptions people already hold about their market. The ML side can surface accounts with high potential that no rule would have picked out.
That leads to an obvious risk: what if the existing rules are wrong and the client is attached to them? Arpit's answer is to make the model a "glass box":
- Contributing factors are visible. If a cohort is graded Tier 1, you can see why.
- Side-by-side comparison. The AI scores sit next to the rule-based scores, so you can spot-check where they disagree.
- Validate before production. The analysis only goes into production once the contributing factors make sense to the client.
He also said this kind of analysis can't be fully productized. Someone has to get into the client's CRM, customer success and activity data and work closely with the client. The goal is a monthly meeting with sales, the marketing head and the CRO where they review which accounts to go after, then collect feedback the following month and tune the model. "The biggest currency is the trust," Arpit said, and a tool by itself can't earn that.
How does the enrichment waterfall work?
Stage one depends on enrichment, and Growth Natives uses a waterfall model. It connects to commercial data vendors like Apollo, ContactOut and ZoomInfo, plus the company's own in-house marketing database.
Every vendor's data goes stale quickly, their own included, especially company-level details and signals like funding or revenue. For refreshing those fields, Arpit has found AI to be relatively more accurate, so it feeds into the account enrichment process too.
The waterfall is configurable because both data subscriptions and AI tokens cost money:
- Cost control. If Apollo is the cheapest source, you can put it first and set a cost-per-lead threshold the list shouldn't go over.
- Trust order. In discovery, the team agrees on which source wins when vendors disagree. If a client trusts Apollo, Apollo goes first. If not, the team spot-checks and recommends a priority order.
- Use what you already pay for. Many clients pay for ZoomInfo or Apollo and don't fully use them. The waterfall puts those subscriptions to work alongside the in-house data and AI.
Demand has changed too. Clients used to ask for enrichment every six or nine months. Now some want it monthly or twice a month, because things move too fast to wait.
What did the demo show?
Hitika ran the demo on synthetic data, so no client data appeared on screen. It followed the six stages in order.
Data quality grader and enrichment
The grader takes a CRM export, or connects directly to Salesforce, HubSpot or another CRM, and color-codes the current state of the data in green, amber and red. Missing industry, headcount, revenue and employee size fields showed up in red. Without those fields, nothing can be scored.
Account enrichment fixes four firmographic fields first: country, employee size, industry and revenue. Contact enrichment then works in two ways. It fills in job titles, business emails and LinkedIn profiles for existing contacts, and it sources new contacts by persona. You choose a job title and a seniority level such as C-level, and you set how many contacts you want per account (two, in the demo).
ICP discovery model
The discovery model builds the ICP backward from customers you've already won. In the demo dataset:
| Metric | Value |
|---|---|
| Accounts in the database | ~173,000 |
| Customers | Close to 60,000 |
| ICP-fit customers | 20,000 |
| Share of revenue from ICP-fit customers | ~87.3% |
The model reports growth factors for each segment, including average customer lifetime value, average win rate and number of deals won. It then suggests segment combinations. The top one in the demo was software and tech as the industry and SaaS platforms as the sub-industry, with set employee, revenue and region ranges. Each segment gets an AI score and counts of customers, prospects and open prospects.
An industry-level rollup shows the same view one level up. Arpit said this is often an eye-opener for mid-market and entry-level enterprise clients deciding whether to fund a dedicated vertical team for, say, manufacturing, financial services or hospitality. A vertical team is expensive, so the decision should rest on data.
Prospect scoring
Each prospect account gets three scores, which combine into a blended weighted average:
| Score | What it measures |
|---|---|
| Model fit (AI/ML) score | A classification model that learns from your historical data which factors predict a good account. Nobody sets the weights by hand. |
| Rule engine score | Business rules set explicitly, for example ranking a given industry or geo higher |
| Engagement score | How active the account has been in the last 90 days |
The AI and rule scores sometimes disagree, and the console shows those cases instead of hiding them. Engagement data comes from web activity (through the marketing automation platform, or GA4 with identity stitching), email opens and clicks, and display campaigns run from an ABM platform. An account that is both active and a strong model fit ranks higher.
Contacts get scored too, on ICP fit, seniority or decision-maker status, engagement and recency. A separate Pareto view of the demo data showed the top 20% of accounts producing about 81.7% of revenue.
What happens when a score changes?
Scores are tied to workflows. If an account starts at Grade C and engagement picks up, it moves to Grade B. If it moves from B to A, it's considered sales-ready. A workflow kicks in, the account is assigned, and it becomes a priority for the sales team over the general pool.
Every score syncs back to the CRM and marketing automation system through reverse ETL, whether that's HubSpot, Salesforce or 6sense. Arpit was clear that the console isn't meant to duplicate those systems. It supplies the intelligence and data, and the teams keep working in the tools they already use.
Seasonality
The seasonality view shows deals won and revenue by industry across one year, or all time for three years or more. In the demo, software and tech did well in August and dropped off in February. Arpit said this view exists because customers requested it. Most teams already know Q4 is strong and Q1 is slow, and this puts numbers behind those calls.
Competitor analysis
You can't see a competitor's customer list, so Growth Natives builds an approximation from public sources and AI research. The result is a competitor profile covering the industries, employee counts, revenue ranges and geos they serve. The console then shows two lists:
- Competitor customers not in your CRM. These are net-new targets.
- Competitor customers already in your prospect list. These are the ones to look at first, because they're already in your CRM.
This helps in two ways. Marketing gets the input for competitor displacement campaigns, and the competitor overlap goes back into the scoring model as a market signal. An account that looks like a competitor's customer should score higher, which brings in information your first-party engagement data can't provide.
Expansion and renewal
Hitika called existing customers "the cheaper half of the growth." The expansion view shows each customer's products, adoption and renewal status, then suggests a next-best product based on what similar customers use. A renewal runway view tracks overdue accounts, the current renewal rate and the potential uplift.
How does the AI-written outreach work?
The last step turns the scores into email. You build a send list from the scored contacts (the demo used Grade A contacts), enter the sender's name, role and company, and pick a message type. Each message type comes with a prompt you can edit.
Brand guidelines are set once and stay in place until you change them: phrases to use, phrases to avoid, and anything every email should include. The AI-written emails all follow them. There are three personalization levels:
| Level | What it means |
|---|---|
| One-to-many | The same email goes to everyone, with no personalization |
| One-to-few | Personalized at the company level |
| One-to-one | Hyper-personalized to the person's role and persona and to their company |
In the demo, one click on "Generate Emails with AI" produced a personalized email for each of the 24 selected contacts.
Arpit brought up token cost. If you're writing one-to-one emails for a 50,000-contact list, the bill adds up. For a large Grade C segment, industry-level personalization is less complex and cheaper to generate. One-to-one is worth it where the account deserves it, and there the AI gets everything known about the person, the company and their past activity.
Growth Natives doesn't want to be the email engine either. The copy goes back to the client's automation platform. With HubSpot Enterprise, messages can be pushed into HubSpot to run as campaigns, and engagement metrics come back so the prompt can be adjusted over time. Other vendors can be connected the same way.
Was this built with Claude?
No. The console runs on a platform that started as a B2B marketing attribution tool. Attribution is hard in B2B because of buying committees, long sales cycles, many channels and many marketing systems, so the team had already built integrations with Salesforce, HubSpot, Marketo, Pardot, GA4 and Google Search Console to bring marketing data together.
That base made the ICP work faster to build. The integrations, the data schema and the grader already existed. Arpit described it as four or five years of work. His view: you can prototype quickly, but production needs a system that is repeatable and that people trust.
Key Takeaways for GTM Teams
- Check your ICP against current data. A definition set years ago describes who your customers were, not who they are.
- Rank customers by more than revenue. Renewals, multi-product purchases, expansion, margin and close speed separate growth customers from average ones.
- Fix the data before you model anything. Grade completeness, enrich the gaps, and refresh more often than every six months.
- Blend rules with machine learning and show the reasoning. Rules carry business context, ML corrects for bias, and visible contributing factors are how the model earns trust.
- Tie scores to workflows and sync them back. A grade change should trigger action inside the CRM and marketing automation tools your teams already use.
- Use competitor overlap as a signal. Accounts that look like a competitor's customers are both displacement targets and a scoring input.
- Match personalization to the tier. Save one-to-one AI emails for top accounts and use industry-level copy for large, lower-grade segments to keep token costs down.
Want to Attend the Next Session?
GTM Clawllective is a community of go-to-market operators sharing how they actually use AI in their work. The live discussion stays off the record, and every recap includes the speaker's demo video. Come see the next build for yourself and join the community.
About the Speaker
Arpit Srivastava is VP Growth Marketing and Analytics at Growth Natives and head of product for the company's growth platform, which began as a B2B marketing attribution tool and now includes the ICP and scoring work shown in this session.
Hitika Saini is a Senior Data Scientist at Growth Natives and works on the same platform. She ran the live demo. Learn more about their work at Growth Natives.