An AI agent for cold email research and drafting for startups does two jobs, and the research is the one to judge it by. First it has to find the founder's address on a public page and write down which page, since an address guessed from a name pattern or bought from a database is the one that gets a new sending domain marked as spam. Then it has to find one true fact about that founder's own site to open with, and write the email in 80 words or fewer with a free report as the only ask. Ours is two agents that share a ledger. The finder looked at 35 companies between October 23 and October 31 and recorded an address for 26 of them. The drafting agent sent its first five emails on October 31. As of November 1 nobody has replied, and the rest of this post is what the two ledgers hold.
What GPT-6 Sol's search returns for this question
On the morning of November 1 we put the question through GPT-6 Sol's web search and kept the list: 30 results, none of them ours. Ten were Reddit threads ("Anyone using AI agents for cold outreach & lead research?", "Built an AI cold email tool that finds verified emails + writes personalized copy", two posts about an agent that cold-emails professors). Twelve were product pages: Taskade's free cold email drafter, AgentCeres, SimpleSend twice, Xevanti ("Research-First AI Cold Email Agent"), Nex ("Personalized Cold Email From Real Research"), Tailwrite, DraftFly and the rest. We read the titles only. In those titles, research means reading the prospect's page to personalize the copy. Ours means something narrower and, we think, more useful to a founder sending the emails under their own name: where the address came from, and one fact the check actually saw.
The research half: 35 companies, 26 addresses, each with its page
The finder reads seven public sources in a fixed order and records nothing that is not printed somewhere: the YC company page for a YC company, the company's own site, a Show HN author's profile, matching GitHub profiles, the X bio, web results that name the founder, and, when all of that leaves nothing, a deep pass over the site's legal, team and press pages and the founder's own site. It never derives an address from a name pattern, queries no third-party database, and never opens LinkedIn. Each address is classed as personal (the founder's name on the company domain), generic (hello@, team@, founders@) or other.
| Finder's ledger, October 23 to October 31 | Count |
|---|---|
| Companies looked at | 35 (37 rows; two companies were tried twice) |
| An address recorded | 26: 13 personal, 13 generic |
| Nothing a cold email could use | 9 |
| Where the 26 came from | The company site 18, the YC page 4, a web result naming the founder 3, GitHub 1 |
| The week's Product Hunt and Show HN launches | 4 looked at, 0 found: the poster was not named on the site in two cases |
Two rows explain the 9. One company was found on the second try, a day later, three pages deeper into its site. One was tried twice and came back empty both times. Of the four launches the collectors added from Product Hunt and Show HN on October 31, none yielded an address, and the notes say why: the person who posted the launch is not named on the product's site, so there is no founder to match an address to. A YC page names its founders; a launch post often does not. That is the kind of thing a founder learns from the ledger, and the reason the ledger records the page beside the address.
An AI agent for cold email research and drafting for startups: what the first five emails said
The drafting agent takes the finder's addresses, drops companies already written to or on the do-not-contact list, and runs our free site audit, five checks, on each domain for its most concrete finding; the audit allows five runs a day, which is where the daily cap of five comes from. Then a session writes one email per founder to a fixed shape, and a checklist refuses a draft that breaks it.
| The rule | What it means in the email |
|---|---|
| Length | 80 words or fewer, not counting the greeting, the opt-out line and the signature; four short paragraphs at most |
| First | The greeting, then one sentence saying who is writing and why they looked at this site |
| Second | The audit finding, about their site, with the specific thing the check saw: the missing file, the crawler name, the page it read |
| Third | One sentence on what it costs them, using only mechanisms we have measured ourselves; no number we did not measure |
| Last | Whether they want the full report, free, sent the same day they reply. No call, no demo |
| Always | The opt-out line, word for word; no link in the body; the engine named by name, never "AEO" or "GEO"; nothing about other clients, nothing guaranteed |
| Pace | Five a day at most, inside 09:00 to 11:30 US Eastern, a minute or more apart; one follow-up after 14 days, never a third |
The five that went out on October 31 opened with three findings. Three said the site has no sitemap, so Bing, which GPT-6 Sol's search reads, and Brave, which Claude's reads, find its pages by links alone. One said the site carries no structured data about the company. One said the site passes all five checks we run, which the rules treat as a fact worth sending rather than a reason to invent a problem. Four of the five were written that morning; the fifth had been drafted on October 17 and waited two weeks for the sending address to exist, then went out re-signed.
The sending is the agent's as well. With the switch set to auto it sends from a dedicated outreach mailbox with a one-image card of the five checks attached, stamps the ledger row with the message id, and reads every sent row's thread at the start of each run, the five of October 31 included: a reply marks the row replied, an opt-out puts the address on the do-not-contact list, a bounce marks it bounced. When a founder writes back, a session drafts the answer and a person approves it by replying "send" to the note the run emails; the answer goes out in the same thread. Nothing is answered without that word.
What it cannot do, from our own numbers
It cannot find an address nobody published. 9 of 35 companies, and 4 of 4 launches, came back empty, and the right response is a new public source a person approves, which the finder proposes in a backlog when a week runs thin, and never a guessed address.
It cannot promise replies. The five emails are one day old and have none. The closest record we have is a different channel: since September 21 a person sent 32 LinkedIn messages to founders who had accepted a connection request, and by October 30 none had replied. Cold messages to strangers who have not asked for anything got nothing from us on that channel, and we expect the email agent to need weeks of ledger before it says anything either way. The cold email agent's log shows each row's status as the reconcile settles it.
It cannot judge fit on its own. A consumer app, a hardware company or a company with no website is skipped by the writing session with a reason, as its rules have said since October 17, and an email written to make the daily count of five is the one email that gets a domain marked as spam.
How to judge one before you buy or run it
Ask for the page each address came from, and read five. An agent that reports "verified" addresses without a source page has verified that the mail server accepts them, which is not the same as the founder having published them. Then ask for the fact each email opens with, and check it against the site yourself; ours is one of five checks the recipient can repeat in a browser in a minute, a missing sitemap being the most common of the five sent on October 31. Then ask what happens to a reply. If the answer is that the agent answers it, decide whether you want that; ours drafts and a person sends.
How we checked
The finder's counts are from its look-up ledger, 37 rows dated October 23 to October 31, read on November 1: the result, the kind, the source and the page of each. The emails are from the cold email agent's ledger, five rows with their subjects, findings and send times, and its configuration notes, which hold the daily cap, the window and the follow-up rule; the writing rules are from its voice guide. The LinkedIn figures are from our weekly performance record of October 30. The search results are from our GPT visibility round of November 1, one search for this question with the list kept; we did not open any product page in it, so no vendor's price or claim is quoted.
Disclosure
ClawWorld runs AI growth agents for early-stage startups and small businesses: they carry out the growth routine a team already planned (posting, replying in communities, outreach, launches and measurement) on a schedule, with every run on a public log, and AEO (answer engine optimization) is one of the services. The two agents in this post are among them, free to copy from their pages or priced per agent to have us run them, and the free site audit the emails open with is ours, so we have an interest in this question. Every number above is from our own ledgers, and the empty rows are left in.