ICP Research with AI: How We Find the Ideal Customer and the MVP Test Market

ICP research with AI means using automated search tools to find out, with sources, who buys a product, why they buy it and where those buyers are. ICP stands for ideal customer profile. At Headcore, we run this research before every launch. It runs in Claude Code, uses two search tools, Tavily and Exa, and ends with a table that compares cities and regions so we can choose where the MVP will be tested first.

This is the low-cost method we use for a preliminary measurement. It relies on public data and two search tools that have free or inexpensive plans, and it delivers a first read of the customer and the markets in a few hours. More expensive research, such as interviews, consumer panels and field studies, comes later, when the project needs to confirm what this first read pointed to.

This article walks through the full process in plain language: what each tool does, how to install it, which questions the research answers, how we size a market and the prompt we use. At the end, as a bonus, we show how to check search intent in each market with Google Ads.

The process in six steps

1. Write down what you already know about the customer.
2. Turn the topic into 3 to 5 questions.
3. Look for the answers in Tavily (numbers and news) and in Exa (similar companies and the way the audience talks).
4. Read 3 to 5 sources all the way through.
5. Compare cities by their audience: size, maturity and profile.
6. Choose where to test the MVP and set the 90-day goal.

What ICP and market mean here

The ideal customer profile (ICP) describes the customer who gets the most value from the product and whom the company can serve at a profit.

In B2B sales, the ICP states the industry, the company size, who makes the buying decision and what makes the company start looking for a solution. In consumer sales, it states income range, habits, the problem the person wants to solve and where they look for help.

ICP, target audience and persona describe the customer at different levels. The target audience is the broad group of people who might buy, such as “women aged 25 to 40 who run”. The ICP narrows that group to the people who get the most from the product and bring profit, using criteria you can count and filter.

A persona puts a face on the ICP: a character with a name, a routine and objections, used to write the messaging. This research works with the ICP, because of the three it is the only one that can be counted in public data.

A market, in this article, is the place where the product will be tested: a city, a metropolitan area or a state. Market sizing means using public data to estimate how many people or companies that fit the ICP live or operate there, how well they already understand the problem and whether they match the profile of a buyer.

The ICP comes first, because it tells you what to count in each market. The market comes second, because an MVP launched in several places at once produces results that are hard to compare.

The tools and what each one does

Claude Code is Anthropic’s AI assistant that runs in your computer’s terminal. It connects to outside tools through a standard called the Model Context Protocol (MCP). MCP is what lets you ask Claude to “research this” and have it use Tavily or Exa on its own to search.

The third piece is a deep research skill. A skill is a set of working instructions that Claude follows. This one says how many questions to ask, how many sources to read and how to write the report.

Tool	What it is	Available commands	What we use it for
Deep research skill	A research script: questions, searches, reading and a sourced report	It has no commands of its own; it calls Tavily and Exa	Making sure every claim in the report has a link and that data from a single source is flagged
Tavily	Web search built for AI, with filters for topic, time range and country	tavily-search, tavily-extract, tavily-map, tavily-crawl	Market figures, recent news, official data and full reads of competitor websites
Exa	Search based on the meaning of the whole sentence	web_search_exa, web_fetch_exa, agent_run and, optionally, web_search_advanced_exa	Companies similar to the ICP and conversations where the audience describes the problem in their own words

Tavily: numbers, news and competitor websites

Tavily works well when a question has a factual answer. The search documentation lists the filters it accepts: topic (general, news or finance), time range (last day, week, month or year), country and a list of allowed sites. Each search returns up to 20 results.

One example: “how many physical therapy clinics opened in Santa Catarina in 2026”, with the search limited to government and news sites. Another: using tavily-crawl to read a competitor’s entire website and note its prices, offer and the cities it serves.

Exa: similar companies and the audience’s own words

Exa works better with questions written as full sentences. According to its search documentation, it has categories such as company, person and news, plus a deeper research mode that runs several searches in a row.

When we ask for “Brazilian supplement companies that sell subscriptions to amateur runners”, Exa returns websites of companies with that profile. We also use Exa to find forums, groups and reviews. That is where the audience explains the problem the way they talk about it every day, and those words become raw material for the product’s messaging.

How to install the tools in Claude Code

Tavily and Exa run on their own servers. To use them, you register each server’s address in Claude Code with one terminal command. The --scope user option makes the tool available in all your projects. Without it, the tool is available only in the folder where you ran the command. The details are in the Claude Code MCP documentation.

# Tavily: you sign in through the browser the first time
claude mcp add --transport http --scope user tavily https://mcp.tavily.com/mcp/

# Exa: works without a key, with a usage limit
claude mcp add --transport http --scope user exa https://mcp.exa.ai/mcp

# Exa with advanced search turned on
claude mcp add --transport http --scope user exa "https://mcp.exa.ai/mcp?tools=web_search_exa,web_fetch_exa,web_search_advanced_exa,agent_run"

Next, type /mcp inside Claude Code. This screen shows whether both tools are connected, and it is where you sign in.

Tavily also accepts the access key written into the address itself (?tavilyApiKey=), but then the key is saved as plain text on your computer. That is why we prefer signing in. The official instructions are on the Tavily page and the Exa page.

Step-by-step ICP research

1. Write down what you already know

Every company knows more about its customers than it thinks: who has bought so far, who dropped out in the middle of a deal, which question comes up in every sales meeting. We write this down in a few lines and put it in the prompt. Without this starting point, the AI returns a generic description of the whole market, and that does not help you choose anything.

2. Turn the topic into questions

“Who is my customer?” is too broad for any search. The skill splits the topic into 3 to 5 smaller questions:

• Who buys, and who makes the buying decision?
• What problem makes the person or company look for a solution?
• How much does the audience pay today for alternatives, and to whom?
• Where does this audience get information and gather?
• What signals show that someone is ready to buy?

Each question becomes two or three searches, phrased in different ways, so the results do not depend on a single wording.

3. Give each tool the right kind of question

Questions about numbers, the market and news go to Tavily, with time range and country filters. Questions about behavior, the audience’s language and similar companies go to Exa. This split avoids two common mistakes: asking a meaning-based search for official data, and asking a factual search to understand how people talk about a problem.

4. Read the full sources

The snippet shown in a search result is short and often out of context. So we map 15 to 30 sources and read 3 to 5 of them from start to finish, using tavily-extract or web_fetch_exa. Our order of preference is official sources, academic studies and well-known news outlets, and only then blogs and forums. We also prefer material from the last 12 months.

5. Require a source for every claim

The report only accepts claims that have a link. When a data point appears in just one source, it is marked “unverified.” Estimates are kept apart from facts.

When there is no good answer to a question, the report says “insufficient data.” This last rule matters most. It shows what still needs to be found out, instead of letting the AI fill the gap with a guess that later turns into an assumption in the plan.

How we size each market

Once the ICP is described, the question becomes: in which city or region does this customer exist in large enough numbers, already understand the problem and match the profile of a buyer? We measure each market by looking at its audience along three dimensions: size, maturity and profile.

Studies on test markets set two conditions for the chosen place. The population needs to resemble the target audience, and the test needs to stay inside that area without spilling over into other regions (Test market entry).

These studies also show that looking for an “average city” that represents the whole country usually fails, because regions differ too much from one another. It works better to group similar cities and choose, within the group, the one that matches the ICP (European Journal of Marketing). For a digital MVP, the second condition is easier to meet, because ads can be shown only to people in the chosen market.

Size: how many potential customers the market has

The math starts from the customer: the number of people or companies that fit the ICP, multiplied by what each one would pay in a year. This method is called “bottom-up,” and it is the most common way to estimate the market a product can serve (the SAM, or serviceable available market) and the share it can win within a set time (the SOM, or serviceable obtainable market).

Its advantage is that every number in the calculation can be checked (guide to TAM, SAM and SOM).

We built this method for projects in Brazil, so the data sources below are Brazilian.

To count people, we use the 2022 Census from IBGE, Brazil’s national statistics institute. It divides the country into census tracts, areas the size of a few city blocks, and reports the population and number of households in each one. Since April 2025, IBGE has also published the average income of the head of household in each tract (IBGE methodology notes). With this, you can estimate the audience neighborhood by neighborhood.

To count companies, we use the open CNPJ data. CNPJ is Brazil’s national business registry. We filter it by CNAE, the code that identifies each company’s line of business (the list of codes is on the IBGE website).

When the product depends on household spending, we also check IPC Maps, a paid study that estimates how much each of Brazil’s 5,570 municipalities spends across 22 consumer categories (Varejo Brasil report).

Maturity: how well the audience already knows the problem and the solution

Two markets of the same size can call for very different amounts of work. In one, the audience is already looking for a solution. In the other, people have not yet noticed they have the problem, and the MVP will have to spend time and money explaining it. We measure maturity in three ways:

• Audience awareness. In his book Breakthrough Advertising (1966), the copywriter Eugene Schwartz described the stages a person goes through before buying. They range from not knowing they have the problem to already wanting the product and only waiting for an offer (summary of the stages). To measure this, we look at what people in the market search for. Searches for a symptom (“knee pain after running”) point to people who know the problem. Searches for a type of solution (“sports physical therapy”) point to people who know a fix exists. Searches for brand names point to people who are already comparing products.
• How many competitors say the same thing. Schwartz also noticed that the more companies repeat the same promise, the less the audience believes it. So we map the competitors in each market and what they promise. If ten brands say the same thing, a lower price will not be enough. The MVP needs to show a different way to solve the problem.
• Presence of people who try new things early. According to the adoption curve of sociologist Everett Rogers, 2.5% of people adopt a new product first and another 13.5% follow soon after. These figures are a convention and vary from product to product (Journal of Marketing Research, 1990). Since an MVP’s first customers come from these two groups, we look for markets that show signs of them: active communities, industry events and professionals who already talk about the topic.

Google Trends helps measure maturity, but it cannot count the audience.

Each Trends number is the share of searches for a topic out of all searches in that region and time period, on a scale of 0 to 100. Because of this, two regions with the same index can have very different search volumes (Google Trends help). Trends shows where a topic carries more weight. The audience count still comes from IBGE and the CNPJ registry.

Profile: age, income, social class and lifestyle

Age, income, education and household makeup come from the Census. Economic class comes from the Critério Brasil, a standard way of classifying households by economic class published by ABEP, Brazil’s association of research companies. ABEP publishes how the classes are distributed in each metropolitan area.

The 2026 figures show how the national average can mislead. Class A makes up 3.6% of households in Brazil. In the Brasília metropolitan area, it is 17.4%. In São Paulo, 5.1%. In Recife, 2.8% (ABEP, Critério Brasil 2026). For a high-priced product, this difference alone can decide the market.

Beyond demographics, we look at the psychographic profile: values, lifestyle and motivations. This is what explains why a person buys, and it works better added to demographics than in place of them (guide to psychographic segmentation).

In the United States and Canada, companies such as Experian, with Mosaic, and Environics, with PRIZM, sell this combined data ready to use, neighborhood by neighborhood. In Brazil, we did not find an equivalent public database. We build the profile by combining Census tracts, the Critério Brasil and what Exa finds in local communities and reviews.

Dimension	Question	Where we look for the answer
Size	How many people or companies that fit the ICP are in the market, and how much would they pay per year?	2022 Census by census tract, CNPJ filtered by CNAE, IPC Maps
Maturity	Does the audience already know the problem, the solution and the competitors?	Symptom, solution and brand searches; Google Trends by region; competitors mapped with Tavily; communities found with Exa
Profile	Does the market’s audience have the age, class and lifestyle of a buyer?	2022 Census, Critério Brasil by metropolitan area, local communities and reviews

Each market gets a score from 1 to 5 on each dimension. Each dimension gets a weight, which changes with the product. For a high-priced service, profile usually weighs more than size. In a new category that the audience does not know yet, maturity weighs more than the other two. The weights are written in the table so the team can question each choice.

From the market table to the MVP

The market with the highest score is not picked automatically. We discuss the result with the client, who knows things no public database shows, such as where they already have staff, partners, inventory or contacts. In the end, the research delivers a recommendation with three numbers:

• the test market for the first test;
• the size of the test, meaning how many people or companies need to be reached for the result to be meaningful;
• the goal that defines whether the test worked within 90 days.

This is the same reasoning behind our Growth Science projects: make decisions with data before investing, measure the test, and only then take the product to the second market on the list. When the MVP reaches the paid media stage, the market list sets the geographic targeting for the Growth Marketing campaigns.

Testing in one market before expanding also protects cash. With high interest rates, a wrong expansion costs more, as we show in how to make company decisions when money costs 14% a year.

And if the MVP is a new product inside an existing company, it is worth deciding early whether it will carry the company brand or a name of its own, the subject of our article on brand architecture.

The prompt we use

Replace the text in square brackets with your project’s details. Keep the rest as it is. These instructions are what make the AI use each tool for what it does best, read the full sources and tell you when it could not find a data point.

Context: [COMPANY] will launch [PRODUCT/MVP] in the [SEGMENT] segment, in [COUNTRY/REGION].
What we already know about the ideal customer: [ICP HYPOTHESIS: who buys, company size, job title, problem].
Goal: define the ICP with evidence and choose the test market where the MVP will be tested first.

Use the two tools for different jobs:
- Tavily (search/extract/crawl): market figures, news from the last 12 months, official
  data (IBGE, CNPJ by CNAE, industry associations) and full reads of competitor websites.
- Exa (meaning-based search): companies and brands similar to the profile described, communities,
  forums and reviews where the audience describes the problem in their own words.

Part 1, ICP. Answer each question with sources:
1. Who buys and who decides? (profile, company size, job title, what triggers the purchase)
2. What problem or desire leads to the purchase, in the audience's own words?
3. How much do they pay today for alternatives, and who do they buy from?
4. Where do these people get information and gather?
5. What signals show they are ready to buy now?

Part 2, markets. Compare [LIST OF CITIES/REGIONS] or propose the 5 most promising ones.
Measure each market by its audience, with sources:
- Size: people (2022 Census by census tract) or companies (CNPJ by CNAE) that
  fit the ICP, times the annual value per customer.
- Maturity: symptom, solution and brand searches; Google Trends by region, read as a
  relative index and never as volume; competitors and what they promise; signs of people who
  try new things first (communities, events, professionals who talk about the topic).
- Profile: age, income, education (Census), economic class (ABEP's Critério Brasil by
  metropolitan area) and lifestyle signals in local communities and reviews.
Build a table with a score from 1 to 5 for each dimension, explain the weight of each one and recommend
the test market for the MVP, the size of the test and the 90-day goal.

Bonus, search intent. List 10 to 20 search terms in your audience's language, split into three
groups: symptom (the problem in the audience's words), solution (the type of product or service)
and brand (competitors that serve the markets). Use the language found in communities and
reviews. This list goes into the Google Ads Keyword Planner.

Rules:
- Read 3 to 5 main sources in full; do not draw conclusions from search snippets alone.
- Every claim needs a link. Anything supported by only one source is marked "unverified".
- Keep facts apart from estimates. Where data is missing, write "insufficient data".
- Prefer sources from the last 12 months.

Deliverables: executive summary, one-page ICP, market table, MVP recommendation,
list of terms for Google Ads, list of sources and methodology (how many searches and how many
sources were read).

Bonus: check search intent in Google Ads

Google Trends shows where a topic carries more weight, but it does not say how many people search for it. Google Ads answers that. Keyword Planner shows search volume and cost by city or state, and a small campaign reveals what the audience in each market actually types. We use both steps to check the market table before deciding where the MVP will run.

1. Split the terms into three groups

The list of terms comes from the research itself. Exa brings the words the audience uses in communities and reviews, and Tavily brings the competitors’ names. Then the terms are sorted by the stage of awareness they point to:

Group	Example for a muscle recovery product	What the search volume tells you
Symptom	“muscle pain after workout”, “how to recover faster”	People who feel the problem and have not chosen a solution yet
Solution	“muscle recovery supplement”, “cryotherapy for athletes”	People who already know the type of product or service
Brand	names of the brands that already serve the market	An established market with competition

2. Measure volume and cost in Keyword Planner

Keyword Planner is part of Google Ads, and you can use it without a live campaign. For each term and each location you choose, it shows how many times per month, on average, the term was searched over the last 12 months.

It also shows how much advertisers usually pay to appear at the top of the page, as a range with a low and a high value (Keyword Planner forecasts). Accounts with low ad spend often see volume in approximate ranges.

Once the three groups are measured in each market, the table gets real numbers:

• search volume for solution terms goes into the size dimension, as demand that already exists;
• the ratio between symptom, solution and brand searches goes into maturity. Many symptom searches mean the MVP will need to explain the problem. Many brand searches mean competition is already in place;
• the amount paid to appear at the top shows how much it will cost to reach the audience.

3. Run a test campaign in the top markets

When two or three markets end up with similar scores, a search campaign with a small budget helps break the tie. The ad leads to a simple page with a waitlist or contact form, and each market runs separately from the others. The test shows three things:

• What people searched for. The search terms report lists the searches that triggered the ad. Searches made by very few people are left out for privacy reasons, but search terms insights group all of them by theme.
• How the audience reacts to the message. The click-through rate and the cost per click in each market show where the MVP’s promise draws the most attention.
• Who wants the product. People who join the waitlist before the product exists give the most reliable demand signal you can measure at this stage.

Google Ads only sees people who search on Google. In a new category, low volume may simply mean the audience has not noticed the problem yet. That is why we use Google Ads to check the market table alongside the other sources. When the MVP reaches the paid media stage, the same term groups can structure the Performance campaigns.

The limits of this method

Because it is a preliminary, low-cost measurement, the method has clear limits. Public data arrives late. The 2022 Census is still the most complete count of Brazil’s population, and a growing city changes a lot in four years.

An active CNPJ also does not guarantee an operating business, because the registry records when a company was opened and does not report revenue. So the market table is the starting point of the decision. The MVP test is what confirms the choice.

AI also makes mistakes. It can misread a table, mix up the year of a data point or cite a page that changed later. The rule of putting a link on every claim exists to make review fast: the reviewer opens the link and checks it, without redoing the research. To learn more about how these assistants work and where they tend to fail, read our article on agentic AI in business.

Frequently asked questions about ICP research with AI

Sources and method note