How Zinta finds a company worth emailing

A company that matches your customer profile perfectly and has no reason to act is a cold lead. You could have bought it from a database for a fraction of what an agent costs. What is worth paying for is the small subset of that list where something changed recently enough that an email is welcome rather than an intrusion.

So the search runs the other way round from most tools. Instead of listing companies that fit and then hunting for news about each, Zinta searches for the events — and the companies arrive attached to them. It is slower per company and it is the whole product: it surfaces businesses you would not have thought to look at, at the one moment they are receptive.

Not every piece of news is a reason to buy. Signals are ranked by how well they predict that someone is actually in the market, and the hunt works down from the top.

And what does not count

This list matters more than the one above it. Every weak lead we have ever seen came from treating one of these as a reason:

• ✕They match the customer profile
• ✕Their website mentions a relevant topic
• ✕An award, or a place on a "top 50" list
• ✕Anything that cannot be dated
• ✕Anything older than twelve months
• ✕Company size, funding stage or industry on their own

Freshness is enforced per signal type rather than as one blanket rule, because signals decay at different speeds. A job posting is interesting for weeks; a funding round for months. Past its window, a signal stops being a reason and becomes trivia.

Every company Zinta surfaces carries a link to the page the claim came from, and that page has to actually say it. This is the rule the whole engine is built around, and it is the reason the reason is written out in full rather than reduced to a score.

• ✓Nothing is asserted that was not read on a page fetched during that run.
• ✓Email addresses are read from a page or left blank. They are never guessed from a pattern, however obvious the pattern looks.
• ✓People are named from a page, or not named at all.
• ✓No numbers, headcounts or funding amounts unless the source states them.
• ✓No date unless the source carries one. "Recently" is not a date.

A guessed address bounces, and bounces are what destroy a sending domain — so the honest blank is worth more than the confident guess. The same logic covers everything else: you are going to click that source link, and a single invented claim would make every other row suspect. An agent that returns two verified companies has done its job better than one that returns five with three embellished.

Each company comes with an email and a LinkedIn note, written from the specific finding rather than from a template with the company name merged in. The standards are not ours; they come from published analyses of cold outreach at scale, and they are unusually consistent about a few things:

• Short. Reply rates peak well under 150 words. A first message from a stranger is read on a phone in seconds.
• Specific or worthless. Genuinely personalised outreach reports roughly double the reply rate of generic templates, and the gap has widened as inboxes filled with machine-written mail. The test we apply: if you could swap in a different company and the message still made sense, it is a template and it gets rewritten.
• One soft ask. Asking whether something is interesting out-performs asking for a slot in a calendar.
• The LinkedIn note is not a short email. Its only job is to get accepted, so it stays close to 150 characters, names one real thing, and never sells. A templated note performs worse than sending no note at all; a specific one lifts replies sharply.

Every batch is then graded mechanically before you ever see it — checked for a missing date, a message with nothing specific in it, a pattern-shaped address, a note that drifted into a pitch. Anything that fails is fixed or dropped.

The engine was not tuned on private data or on anybody's contact list. It was built from public, checkable material, and it is worth saying which kinds:

• Published cold-outreach benchmarks from the teams that run outreach at volume and report on it — reply rates by length, by personalisation depth, and how those have moved as inbox filtering tightened.
• Intent and trigger-event research from the B2B data industry on which company events actually precede a purchase, rather than which merely correlate.
• LinkedIn outreach studies covering acceptance and reply rates with and without a note, at different lengths.
• The mailbox providers' own published rules — Google, Yahoo and Microsoft — on authentication, complaint thresholds and what marks a message as bulk.
• Regulation, read directly: GDPR legitimate interest for business-to-business contact, CAN-SPAM, and CASL.

What is not published here: the prompts, the query patterns, the per-industry source lists, and the scoring weights. That is the part we built, and describing our reasoning is fair game while handing over a recipe is not.

The constraints are as much the product as the capabilities, and all of them cost us something:

• It never sends. You send, from your own mailbox. Your domain reputation stays yours, and nothing we drafted reaches a stranger without a human reading it first.
• It contacts businesses, at business addresses, for a relevant business reason. Personal mailboxes are never used for a business approach, even when they are published.
• It stops when it runs out of good ones. If four companies clear the bar and your plan allows six, you get four.
• It forgets on request. A company you reject is excluded, and the reason becomes a rule the next hunt obeys rather than a preference it weighs.