Was it written
to you, or at you?
Paste an email. In one second you'll see the structural fingerprints every AI sales sequence leaves behind, quoted with the exact lines that gave it away.
We score the structure, not the sender. You'll see every line that fired and what it's worth.
The original Trish scores 0.77. Red starts at 0.55. Honest email scores 0.
Show the math
No model, no lookup. Each fingerprint is a regular expression with a fixed weight. Weights combine with noisy-OR: 1 − Π(1 − wᵢ), so ten weak tells never add up past certainty, then real human context is subtracted.
- +0.35 Bulk-mail footer
- +0.30 Fake bump
- +0.30 Database personalization
- +0.25 Soft opt-out
- +0.25 Vague value tease
- +0.25 Fact-to-pain pivot
- +0.25 Phantom referral
- +0.20 Meeting-slot ask
- +0.20 Self-quoted thread
- +0.20 Brand/domain mismatch
- +0.15 The ‘I help’ formula
- +0.10 Burner domain
- −0.30 First real-context cue
- −0.15 Each additional cue (max −0.55)
One more gate: if nothing in the email is actually selling, the automation evidence is discounted to a fraction of its weight. A templated note from your accountant shouldn't score like a sequence. Thresholds: 0.55+ red, 0.30–0.54 amber, below 0.30 green.
Who is Trish?
Trish R. apologized for the delay, referred back to their last conversation, and mentioned $600k. There had been no last conversation, there was no $600k, and there was no Trish, just an AI sequence in a friendly mask, sent from a domain bought that month.
The test is named after the mask. Millions of Trishes go out every day, and they all leave the same fingerprints: the bump, the two weekdays, the polite escape hatch, the number with nothing behind it. Here's the trick: those fingerprints aren't bad writing. They're structural. The bump is the sequence. The opt-out is the compliance line. The vagueness is all they know about you. The only way to remove them is to write a specific, honest email. Which is the point.
Sender name and domain fictionalized. The email is real.
What we check
Fake bump
"Just wanted to make sure this didn't get buried" is step 2 of a sequence, firing on a timer.
Meeting-slot ask
Two weekdays and fifteen minutes. The email exists only to fill a calendar.
Soft opt-out
A polite escape hatch bolted on because the sender knows it's bulk.
Vague value tease
"The kinds of results we're producing." Big number, zero specifics.
Self-quoted thread
It quotes its own earlier step to fake a conversation you never had.
Burner domain
Disposable .info / .xyz / .club domains, bought by the hundred.
Database personalization
"I saw you've been family-owned since 1967." That's not research, that's a Clay column.
Brand/domain mismatch
Says he's from one company. Sends from somewhere else entirely.
The ‘I help’ formula
“I help [audience] do [thing] so [outcome].” The LinkedIn bio, pasted into an email.
Bulk-mail footer
An unsubscribe link is a legal admission that this went to a list.
Right-person routing
A qualification question routes you into a sequence. Any answer, including a redirect, confirms a live inbox and feeds the machine.
And one that works in your favor
Real context
Named referrals, a meeting you actually had, an invoice number: things a sequence can't produce. Each one earns points back, and if nothing in the email is selling, the fingerprints barely count at all.
Put the Trish Test in your AI
An MCP server that runs this exact engine inside Claude, Cursor, or any AI that drafts or reads email. Your assistant preflights every cold email before it sends, and scores every pitch that lands. Free. Runs on your machine. Nothing leaves it.
npx trishtest-mcpAdd to Claude Desktop
{"mcpServers":{"trishtest":{"command":"npx","args":["trishtest-mcp"]}}}The judge, the rewrite, the badge
Semantic judgment beyond the pattern tier, drafts rewritten until they pass, and sequence certification with a verifiable badge. API keys are rolling out. Join the list.
Built by people who run B2B outreach for a living. That is why we know the fingerprints.