Skip to content
← Articles & field notesArticlePlaybook15 min readRevised

Use customer evidence to revise your ideal-customer hypothesis

Start with who buys, why they buy and whether they are profitable; the persona follows the evidence.

A customer segment found in source records is an observation; an AI-generated persona is a hypothesis. Validate demand and contribution before treating either as the next profitable audience.

Your ICP Is Probably Wrong

Your ICP is fiction. Your customer data tells a different story entirely.

That's not a provocation. It's what we see every time we onboard a new e-commerce client.

We have used enrichment across selected e-commerce brands to investigate segments absent from their planning. The article does not retain a comparable dataset establishing that every brand will reveal two or three previously unknown segments.

Treat the clusters as candidate patterns until their sample quality, linkage and commercial behavior have been checked. A visible cluster alone is not evidence that chance or selection has been ruled out.

The gap between who brands think their customer is and who their customer actually is doesn't just waste ad budget. It means entire profitable audiences are sitting there, untouched, while competitors fight over the same obvious demographics.

This article breaks down what happens when you stop guessing your ICP and start enriching your actual customer data. We'll cover the specific process, what we've found across client engagements, and why the hidden segments consistently outperform the assumed ones.

If you've ever built a persona in a meeting and then built campaigns around it, this is for you.


The Persona Deck Problem

Every brand we work with starts the engagement the same way. They walk in with a persona deck.

Sometimes it's a detailed 15-page document from a brand workshop. Sometimes it's a one-page profile sketched by the founder. Either way, it says something like:

"Our ideal customer is a 35-year-old female, household income $120k+, interested in wellness and clean living. She shops online 3-4 times per month and values sustainability."

That sounds reasonable. It might even be partially correct. But here's where it breaks down.

Persona decks are built from assumptions, not purchase data. They're constructed in rooms by marketing teams working from intuition, competitive research, and best guesses. Sometimes from surveys - which measure what people say they do, not what they actually do.

The result is a fictional character that the entire marketing strategy orbits around.

Campaigns get built targeting that demographic. Ad copy speaks to that persona's pain points. Landing pages mirror that persona's priorities.

And nobody checks whether the real buyer data confirms any of it.

This isn't a minor oversight. When your ICP is based on assumptions, every downstream decision inherits that assumption. Bidding strategies, audience targeting, creative angles, keyword selection - all of it points at who you think is buying instead of who is actually buying.

The compounding effect is brutal. You're not just missing one segment. You're optimizing an entire campaign architecture around an incomplete picture of your customer.


What Customer Data Actually Reveals

When we say "enrich the customer list," we mean something specific.

Raw customer data - emails, purchase history, transaction records - tells you what someone bought and when. Useful, but shallow. You know the behavior. You don't know the person behind it.

Data enrichment adds layers that raw purchase data can't give you. It pulls company descriptions, job roles, company size, industry categories, technology usage, and behavioral signals from external data sources. The raw list gets filled in with context.

Here's what changes when you run enrichment on an actual customer list:

You stop seeing transactions and start seeing patterns.

A list of 5,000 purchases looks like 5,000 individual orders. Enriched, it looks like clusters. Groups of buyers who share characteristics you'd never spot from purchase data alone.

Think about what you know from a raw transaction: email, order date, products purchased, order value. Maybe a shipping address. That tells you what happened, not why.

Now add enrichment data: the buyer works at a 30-person marketing agency. Their company serves DTC brands. They've been in business for 7 years. They use Shopify Plus and Klaviyo in their tech stack. They've visited your pricing page 4 times before purchasing.

Suddenly that transaction has context. And when you run the same enrichment across thousands of transactions, you start seeing buyer profiles repeat in patterns that no amount of persona workshopping would have surfaced.

The enrichment data feeds into AI-powered clustering analysis. Instead of grouping customers by what they bought or how much they spent, you're grouping them by who they are - their role, their company size, their industry, their competitive landscape.

This is where the hidden segments show up.

Not in the purchase data. Not in the Google Analytics reports. In the enriched profiles that reveal shared characteristics across buyers who seemed unrelated on the surface.

The common enrichment dimensions we pull:

  • Company data: Name, size, revenue range, industry, years in business
  • Role data: Job title, department, seniority level
  • Technology: Tech stack the company uses (CMS, marketing tools, ad platforms)
  • Behavioral signals: Content engagement, page visits, purchase frequency
  • Competitive landscape: Who else they buy from, what alternatives they've evaluated

Each dimension adds a layer of understanding. Stacked together, they create a buyer profile rich enough to drive targeting, messaging, and campaign architecture decisions.

Selected engagements surfaced candidate segments the brand was not explicitly targeting. Validate linkage, sample quality and buyer behavior before treating a cluster as a new profitable market.

The segments weren't small, niche afterthoughts. They were substantial groups with consistent buying patterns and distinct needs.


The Segments That Stay Untargeted

Let me be specific about what "hidden segments" actually look like in practice.

Case Study 1: The Agency Blind Spot

One client - mid-size e-commerce brand selling B2B software tools - was convinced their buyer was mid-market DTC brands. Their persona deck said so. Their ad copy spoke to DTC operators. Their landing pages featured DTC case studies.

When we enriched their customer list, the data told a different story.

The author-reported tool-business example identified agencies among valuable buyers. The note does not establish the sample, LTV definition or period needed to transfer the earlier 40% share to another business.

Zero ad spend targeting that segment. Zero landing pages speaking to agency pain points. Zero ad copy mentioning multi-client use cases.

The opportunity is to test whether current acquisition work addresses a supported segment. A discovered cluster is a hypothesis until its identity, economics and buying motive are validated.

Case Study 2: The Enterprise Mismatch

Another client was spending $15k/month going after enterprise companies. Big logos, long sales cycles, expensive keyword bids. Their persona deck positioned them as an enterprise solution.

The enriched data showed their actual best customers - highest repeat purchase rate, lowest return rate, strongest LTV - were companies with 10-50 employees.

Not enterprise. Not even mid-market. Small teams that used the product differently than enterprise buyers, had different pain points, and responded to completely different messaging.

$15k/month targeting the wrong company size. For over a year.

Case Study 3: The Category Surprise

A third client selling premium home goods assumed their buyer was the end consumer - homeowners with disposable income. Standard DTC persona.

Enrichment revealed a significant cluster of interior designers and home staging professionals buying in bulk. Different purchasing behavior, different product preferences, different price sensitivity.

(The brand had no idea this segment existed. They'd never asked.)

These aren't hypothetical examples. This is what happens when you look at the data instead of the whiteboard.


The Data-First ICP Build

After seeing these patterns repeat across multiple client engagements, we built a structured process around it. We call it the Data-First ICP Build - and we run it before creating a single Google Ads campaign.

Here's how it works at a high level:

Stage 1: Enrich

Start with the actual customer list. Not a survey. Not assumptions. Transaction data from real purchases.

Run enrichment to add company data, role information, company size, industry categories, and behavioral signals. The goal is to turn a list of transactions into a list of people with context.

The enrichment layer is what makes everything downstream possible. Without it, you're clustering on purchase behavior alone - which tells you what people buy, not who they are or why they buy.

Choose an appropriate sample and permitted data scope for the question. More records can reveal additional patterns, but full-list enrichment is not automatically necessary, proportionate or accurate. Validate missingness, linkage and potential sensitive inferences.

Stage 2: Cluster

Feed the enriched data into AI-powered analysis. The AI identifies natural groupings based on shared characteristics across the enriched dimensions.

This is where segments emerge that no human would spot manually. The clusters form around combinations of attributes - role + company size + industry category + purchase frequency - that create distinct buyer profiles.

Typically, we see 4-6 distinct clusters emerge. Of those, 2-3 are segments the brand was already targeting (usually confirmed by the persona deck). The remaining 2-3 are the hidden ones.

The key insight from this stage: the clusters don't always align with demographic assumptions. You might expect company size to be the primary differentiator. Instead, the AI finds that role + tech stack combination is a stronger predictor of purchase behavior than any single demographic dimension.

That's why human-built personas miss these segments. Humans think in categories. AI finds combinations.

Stage 3: Gap

For each newly discovered segment, run a competitor gap analysis. Who else is targeting this audience? What messaging are they using? Where are the openings?

This involves scraping competitor landing pages, analyzing their ad copy in auction insights, and mapping their positioning per segment. The output is a competitive landscape map per discovered segment.

A discovered segment can suggest a new research or messaging question. It does not reveal industry-wide competitor targeting or prove that the audience faces little competition.

We've seen cases where a discovered segment had literally zero competitors running targeted ads. Not low competition - zero. That's free attention waiting to be captured with the right message.

Stage 4: Angle

Based on the segment characteristics and the competitive gaps, identify the messaging angles that would resonate with each segment.

Different segments need different messages. An agency buying your product for their clients has completely different motivations than a DTC brand buying it for themselves.

The pain points are different. The value proposition is different. The objection handling is different.

This stage produces specific positioning, copy angles, and offer structures per segment.

Stage 5: Architect

Build the campaign architecture mapped to real segments instead of assumed ones. Each segment gets its own:

  • Audience targeting parameters
  • Ad copy angles
  • Landing page messaging
  • Bid strategy calibrated to segment LTV
  • Keyword sets aligned to segment search behavior

The campaigns are built on evidence, not opinion. Every targeting decision traces back to enriched customer data, not a persona workshop.


Why Hidden Segments Outperform

Here's the part that surprises most clients.

Evaluate the discovered segment against a compatible baseline. Enrichment helps organize evidence; it does not establish that the resulting segments will outperform.

Why?

The mechanism is straightforward: competition density.

The segments your persona deck identifies are the same segments your competitors' persona decks identify. Everyone in your category is building campaigns targeting the same obvious audience. That drives up CPCs, increases competition for attention, and compresses margins.

The hidden segments have a structural advantage. There's minimal competition for that attention because these segments haven't been found through traditional persona work.

When you're the only brand (or one of few) running targeted campaigns for a specific audience segment, several things happen:

The author's agency-keyword example reported a lower CPC than its earlier DTC-focused campaigns. Different queries and buyer groups are not a controlled cost comparison, and lower CPC alone does not establish lower acquisition cost.

Relevance scores improve. Your messaging actually speaks to this audience's specific needs - because you built it from their data, not generic personas. When an agency sees ad copy about "managing campaigns for multiple clients" instead of "growing your DTC brand," the click-through behavior changes immediately.

Conversion is a question to test. Better supported buyer context can inform ad, offer and destination choices, but alignment alone does not guarantee that rates rise or explain a measured change.

Check actual cohort contribution. A discovered segment may have different repeat behavior, but modeled LTV and the fact that it already purchased do not prove higher lifetime value or product-market fit.

This isn't a theory. It's the consistent result of building campaigns around data-validated segments instead of assumption-based personas.

The irony is that the "obvious" audience - the one every brand targets - often has the worst unit economics precisely because it's obvious. You're paying a premium to reach people in the most crowded auctions.


What Changes When You Build Campaigns on Real Data

The shift from assumption-based to data-based ICP changes more than just your targeting. It restructures how you think about your entire Google Ads architecture.

Campaign structure changes. Instead of organizing by product category or match type, you organize by audience segment. Each segment gets campaigns that speak their language, address their pain points, and match their buying behavior.

Creative angles multiply. When you discover 2-3 new segments, you don't just get 2-3 new audiences. You get 2-3 entirely new messaging angles.

The agency segment needs different ad copy than the DTC segment. The small business segment responds to different proof points than the enterprise segment.

Allocate a test from the segment's actual reach, contribution and uncertainty. A share of historical high-value customers does not mechanically prescribe a share of future budget.

Value-based decisions: use validated, comparable collected customer contribution and permitted signals. A cluster's modeled LTV is uncertain and does not automatically prescribe a higher bid.

Keyword strategy sharpens. Different segments search differently. The small business buyer searches "affordable [product] for small teams." The enterprise buyer searches "[product] enterprise pricing." When you know which segments matter most, you build keyword sets that match their actual search behavior - not the generic terms that attract everyone and convert no one particularly well.

And the compound effect works in your favor. Each campaign optimization builds on real data instead of assumptions. The feedback loop tightens. You're not guessing which audiences respond - you're measuring it against enriched baselines.

The most underrated benefit: your reporting becomes meaningful. Instead of looking at blended ROAS across a generic audience, you're tracking performance per segment. You know which segment is growing, which is plateauing, and which needs a messaging refresh. Budget decisions get made on segment-level data instead of gut feel.

Here's the practical impact: you stop spending money convincing the wrong people and start spending money reaching the right ones.

That sounds simple. It's the execution that most brands skip.


The Cost of Guessing

Let me frame what's at stake.

Every month you run campaigns on assumed ICPs, you're paying for two things: the campaigns you're running and the campaigns you're not running.

The visible cost is the ad spend going to audiences that may or may not be your best customers. If your persona deck is partially right (it usually is), some of that spend generates returns. You see ROAS. It looks functional.

The opportunity is uncertain. A segment that lacked a dedicated campaign may already be buying through broad targeting, other channels or organic routes. Missing segmentation cannot be counted as zero acquisition or a known revenue loss.

The author-reported agency-buyer example is a reason to inspect customer composition. Without the retained cohort definition and period, its earlier 40% figure cannot quantify revenue acquired by accident or transferred opportunity.

Now multiply that by every month the enrichment step was skipped. Every month without campaigns targeting the discovered segments. Every month paying premium CPCs in crowded auctions while low-competition audiences sit untouched.

A hypothetical 20-30% segment share does not establish missed revenue for the previous twelve months. Confirm existing acquisition, audience overlap, costs and a useful test before sizing an opportunity.

The first brand to identify and target a hidden segment gets a head start that compounds. Every month of data, every optimization cycle, every creative iteration builds an advantage that late entrants have to overcome.

The cost of assumption-based ICPs isn't a single bad campaign. It's a compounding opportunity cost that grows every month the data goes unexamined.

We used to skip this step too. (For years, if we're being honest.) Built campaigns on persona decks, optimized from there, reported on the metrics that looked good.

Then we started enriching the data. And the gap between what brands assumed and what the data showed was consistent enough that we made it the first step in every engagement.

Not because it's fancy. Because the numbers demanded it.


Precision Beats Volume

The brands that win in paid acquisition aren't necessarily the ones with the biggest budgets.

They're the ones who know exactly who's buying and why - because they looked at the data instead of the whiteboard.

The Data-First ICP Build isn't complicated. Enrich the customer list. Let AI find the clusters. Identify the gaps. Build the angles. Architect the campaigns.

Five stages. The output is a campaign architecture built on evidence instead of opinion.

The brands that skip this step aren't lazy. They just haven't seen what the data reveals when you actually look. Once you see 2-3 segments with real buying behavior sitting there completely untargeted, the return to assumption-based personas becomes difficult to justify.

Your customer data already contains the answers. The question is whether you've asked it the right questions.

Revisit an assumed ICP against permitted, reliable customer evidence. The article does not establish a universal missing-audience percentage or guarantee that discovered segments face lower competition.

A pattern we see often: teams read this and think, "We should probably do that sometime." The ones that act on it discover what we've seen consistently: the hidden segments are real, they're substantial, and they're waiting.

Precision beats volume. Data beats opinion. And the brands that figure this out first get the advantage of low-competition segments before the rest of the market catches on.

Put the framework to work

Validate one unexpected segment

Give your growth
an operating team.

Google Ads, Meta Ads and Klaviyo Email/SMS. Buying, creative, copy, pages, testing and optimization, handled by the team that built the AI systems.

Typically live within 48 hours. The clock starts when we have account access and the assets we need.

Discuss your growth goalCompare services and fees

Keeping execution in your team?
Book a paid working session or explore licensed systems.