This is a 90-day plan for building a GTM (go-to-market) engine from zero, week by week, plus the four things I’d leave out.
A GTM engine here means a repeatable way to find the right companies, reach the right people at them, and learn from every reply.
The plan has five phases. Every phase ends with a done-test, a check that tells you the phase is finished. The next phase starts only when it passes.
Say you run GTM at Hopwright, a company that sells production and inventory software to independent craft breweries in the US.
After 90 days you have a market you can count and a list where every row shows its source and its reason. You also have replies from your first waves, batches of companies that get the same emails in the same weeks. And you have a written routine someone else can run.
This guide is for founders and first GTM hires at B2B companies that have a product, a handful of customers and no repeatable way to find the next fifty customers.
Read it once in full, then work through it phase by phase. Half of the 90 days goes into the first three phases, before a single email is sent, and that’s on purpose.
Weeks 1 to 2: Describe your market as yes-or-no tests, and pick the opening line
The first two weeks produce one document, the market card, and no list. This is the cheapest point in the whole build to be wrong, so think hardest here.
Write the market as tests. A test is a yes-or-no question about a single company, and it names where you’d check the answer. Words like “growing” or “mid-sized” don’t work as tests, because two people will judge the same company differently.
Take the tests from your best customers. Read your won deals, up to ten. If you have fewer, add your best open deals. The tests come from what these companies have in common. The disqualifiers, the reasons a company is out, come from deals you lost or shouldn’t have taken. They catch companies that look like they pass but don’t fit.
For Hopwright, the market card looks like this:
MARKET CARD v1 Hopwright
Must pass every test:
1. Brews beer on its own premises evidence: website, brewery listing
2. Not majority-owned by a larger evidence: about page, press
beverage company
3. Sells beyond its own taproom evidence: "find our beer" page,
distributor announcement
4. 5 to 60 employees evidence: headcount field, team page
Disqualifiers (any one excludes the company):
- Bar or restaurant that sells local beer but doesn't brew
- Homebrew supply shop
- Equipment or ingredient supplier with "brewing" in its name
- Closed, or no sign of activity in 12 months
Opening line the first email must be able to prove, per company:
"You started working with [distributor] in [month]."
Proof: URL and date of the announcement
A taproom is the brewery’s own bar. Selling beyond it means stores, bars or distributors (wholesalers that move beer to stores and bars) carry the beer.
Pick the opening line before you build the list. The opening line of your first email decides which list you build, because every company on the list has to back it up. Hopwright could start its first email with a new distributor, a job post for a second brewer, or a new taproom.
All three point to the same problem: inventory spread over more places than a spreadsheet handles. But each one gives a different list. I’d pick the one whose proof is easiest to find and keep the other two for later waves.
Done-test: two people sort the same 20 breweries into in or out, separately, and disagree on two or fewer. For the chosen opening line, you can link to the proof for at least five real companies.
Keep those 20 breweries with your answers. They’re your answer key: in week 3, the AI check has to agree with them before it runs on the whole market.
Weeks 3 to 4: Map the market, then count the companies that show your signal
Weeks 3 and 4 turn the card into one table with every company in the market, plus a monthly count of how many show your signal. A signal is a sign that a company may need you now, like a new distributor for Hopwright’s breweries.
Build the table the way my TAM playbook for local businesses describes. A TAM (total addressable market) is every company that could buy from you. The playbook shows how to collect every company that could fit, from company databases and, for local markets like breweries, Google Maps. It also shows how to check each company with AI against your tests, how to search again with lookalike tools (they find companies similar to the ones you give them) until a round adds under 3% new companies that pass, and how to find the people and their emails. This plan adds when to do each part and for which companies.
Three things in this phase are specific to the 90-day plan:
- Give every company one status. In or out, always with a reason. Out means a test failed, a disqualifier applies or the data to prove a test is missing, and the reason names which one. The status table holds every company with its status and its reason.
- Check your count against a public number. The Census counts 5,218 US brewery locations with paid employees (2023), and 3,579 of them have 5 to 49 employees. If your “in” count lands far above that, a test is too loose.
- Count the signal over three time windows. Count how many companies showed the signal in the last 30, 90 and 365 days. One month can be unusual, and the longer windows show what’s normal. Use the 90-day count divided by three as your monthly number.
Say Hopwright’s table ends with 3,200 breweries in and 4,200 out. Say 144 of the 3,200 started working with a new distributor in the last 90 days, so 48 a month. That monthly number decides how often you can send a new wave (weeks 7 to 8 show how).
Start the sending setup now, in the background. Send cold email from new domains, separate from your main domain, so a problem with cold email doesn’t hurt your normal email. New domains and mailboxes need a few weeks of small, slowly rising volume before they can send a full wave. This is called warming up. I’d set them up in week 3, so they’re ready in week 7.
Gmail checks a few records on every domain that sends to it. Google requires SPF or DKIM from every sender, two records that prove an email comes from your domain. Every sender also has to keep the spam rate (the share of emails people mark as spam) below 0.3% in Postmaster Tools, Google’s free dashboard for senders.
If you send 5,000 or more messages a day to Gmail accounts, you need SPF and DKIM together, plus DMARC. DMARC is a third record that tells mail servers what to do with email that fails the first two. You’ll stay far below that volume, but set up all three now, so nothing has to change later.
Done-test: every company in the table has exactly one status, and every company that’s out has a reason. The monthly signal count is written down, and the new mailboxes are warming up with SPF, DKIM and DMARC set.
Weeks 5 to 6: Find the contacts, then check the list
Find people only for the companies you’ll contact in the next two waves. If you find contacts for the whole market in week 5, you pay for contacts that change jobs before you reach them.
- Name the buyer roles. For Hopwright: owner or founder, head brewer, and operations or production manager. Find every person with those titles at each brewery, with no limit per company.
- Find the people and emails the way my TAM playbook describes. Flat-rate tools like GetLeads and BlitzAPI come first, and Prospeo comes last, because it charges per contact. For owners who aren’t on LinkedIn, a web search for “who owns [brewery] in [city]” often works.
- Keep only confirmed emails. Send to addresses a verifier marks valid. Send to catch-all addresses (mail servers that accept any address) only after LeadMagic or BounceBan confirms them, and drop every address that comes back “unknown”.
- Write the source and the date into every row. When someone asks in week 11 where an email came from, the row shows it.
Then come the checks, before anyone else sees a row. Claude Code, Anthropic’s AI agent that runs in your terminal, can write a check script (there’s a separate guide on setting it up) that stops when one of the first three fails:
- The counts add up from the table to the waves.
- Email domains match company websites, or the difference is explained.
- Shared inboxes like info@ sit in their own column and never count as people.
- You read ten rows by hand. If a row is wrong, you fix the rule that let it through and pick ten new rows. The list is ready when three rounds of ten in a row come back clean.
Say the first two waves need 100 breweries with a person you can reach. Of the 144 breweries with the signal, say 14 are customers, open deals or asked not to be contacted. That leaves 130 to search.
Of those, say 100 come back with a verified owner or head brewer, 18 with only a shared inbox and 12 with nobody. The 18 can get a later email at the shared inbox, addressed to the owner by name if the website names one. The 12 stay in the table with that status, ready for a manual look later.
Done-test: all four checks pass, and every email in the two waves has a source and a date.
Weeks 7 to 8: Send on one channel, in small waves
Send on one channel first, in waves small enough that each one teaches you one thing. For most B2B companies that channel is email, sent through a tool like Instantly or Smartlead, because it gives you the fastest answer for the least money.
Each wave tests one segment, a group of companies with the same signal, and one angle, the problem the email leads with. Wave 1 is the baseline, the wave the others are compared with.
| Wave | Companies | Segment | Angle | What changes |
|---|---|---|---|---|
| 1 | 50 | New distributor, last 90 days | Inventory across distributors | Baseline |
| 2 | 50 | New distributor, last 90 days | Taproom and wholesale in one count | Angle only |
| 3 (week 9) | 50 | Hiring a second brewer | Winner of waves 1 and 2 | Segment only |
Waves 1 and 2 use the 100 breweries from the last 90 days. After that, 48 new breweries join this segment each month, close to one wave of 50 a month. Wave 3 tests a second segment with the winning angle.
- Change one thing per wave. If you change two, you won’t know which one caused the change.
- Write lines that stay true until the last follow-up. A sequence is the first email and its follow-ups, often spread over three weeks. “You started working with [distributor] in March” stays true. “You’re expanding right now” may not be true by the third email.
- Create every campaign paused. In paused mode, nothing goes out until someone presses start. Someone reads five finished emails, with the name and company fields filled in, before anyone presses it.
- Set a bounce limit before the first send. A bounce is an email that comes back because the address doesn’t exist. Pick a limit, for example 2% of sent emails, and pause the wave when it’s crossed. If you pick the number in advance, nobody argues about it in the middle of a wave.
- Log every reply with its wave, segment, angle and outcome (for example interested, not now or wrong person), in the same table that holds the companies.
Done-test: two waves have been sent, and every reply sits in the table, tagged with segment, angle and outcome.
Weeks 9 to 12: Measure, score, and hand it over
The last month turns the first waves into a weekly routine someone else can run.
Measure per segment and angle. I’d track interested replies, from people who want to talk, per 100 companies contacted, split by wave. Say wave 1 got 3 interested replies from 50 breweries and wave 2 got 1. With numbers this small, you have a direction for the next test. You don’t have a rate you can plan the next three months with.
After every wave, make one decision and write it down. Keep the angle and try it on the next segment, keep the segment and try a new angle, or drop both. Put the decision in the wave log, a short record per wave, next to the numbers. Then whoever plans wave 4 can see why wave 3 looks the way it does.
Score every company with written rules and a script. A score is a number of points per company that decides who gets the next wave first. Write a first version of the rules with an AI model’s help: which signals, which sizes and which roles answered. Update them after every wave. Claude Code can write the script that applies them to the whole table. Read ten scored rows by hand and check that the order makes sense to you.
Hand it over. Write the runbook, the written routine for each week, while you still remember why each step exists. For Hopwright, the routine could be Monday for counting new signals and running the checks, Tuesday for the launch when a wave is due, and Friday for reading replies and updating the card.
Write down who approves spending on data, who reads the sample and who launches. Then watch someone else run a full weekly cycle from the runbook without helping.
HANDOVER CHECKLIST
[ ] Market card, current version and date, every change logged
[ ] Status table: one row per company with status, reason, source, date
[ ] Signal definition, time windows and monthly count
[ ] Contact order (which tool first) and verification rules
[ ] Check script that stops on errors, plus three clean rounds of ten rows
[ ] Wave log per wave: segment, angle, companies, replies, interested
[ ] Scoring rules in a file, applied by a script
[ ] Owners: who approves spending, who reads the sample, who launches
[ ] Weekly calendar: signal day, launch day, review day
[ ] One named owner for the whole engine
Done-test: someone other than the builder runs one full weekly cycle from the runbook, and the counts at the end add up.
The four things I leave out, and why
Each of these is useful at some point. In the first 90 days, I think each one costs more than it teaches.
1. An AI-written line for each person. This means a model writes a custom line for every person from their posts or profile. The first waves test segment and angle. A different line for every person makes it impossible to tell which one worked, and it adds one more thing that can be wrong in front of a buyer, in every email. The relevance comes from the list and from the opening line you picked in week 1. I’d add it once one segment got the most interested replies in two waves and the angle is settled.
2. An AI SDR that sends and replies on its own. An AI SDR (sales development rep) is software that writes, sends and answers sales emails without a person. Every reply in the first 90 days teaches you something about the market, like which role answers or which words buyers use for the problem. A system that answers for you skips that lesson, and nothing it sends can be taken back. An AI model can draft every reply and pull up the company’s history, and a person sends it. I’d look at this again after you’ve read a few hundred replies yourself and know what a good answer looks like.
3. Intent data you can’t show the buyer. Intent data is a guess that a company is researching a topic, based on website visits or reading behavior elsewhere. It can help rank companies, but early on it has two problems. You can’t start an email with it, because the buyer has no way to check the claim. And without a reply rate from your own waves, you can’t tell whether companies flagged as interested reply better than the rest. I’d add it once the waves give you a rate to compare against.
4. A CRM rebuild and dashboards. The CRM is the database where your sales team tracks deals. A rebuild means new objects, custom fields, scoring inside the CRM and a reporting dashboard. In week 3 you don’t know yet which fields matter, and every field built on a guess has to be moved or rebuilt later. Until the routine has run a month unchanged, the status table holds the market, and only real conversations and deals go into the CRM. I’d build the CRM side and the dashboard after that.
Where to start today
You can start the 90-day plan today in three steps:
- Write the market card: four tests with the evidence for each, four disqualifiers and one opening line.
- Map the market and count the companies that show your signal.
- Find the contacts and send the first wave on one channel.


