Keeping an AI lead researcher honest: "word for word" facts and emails that are never guessed

Lead discovery is an easy place for AI to make things up. Ask a model for "the founder of Cedar Creative and their email", and it will happily invent a plausible name and alex@cedarcreative.com. A cold email to an invented address bounces and hurts your sender reputation. A wrong name is worse.

When we built lead discovery for VeloxonHub Reach, we set one rule for everything shown on a lead: every fact has to come from a page we actually read, and the user can see that page. The 30-second reel shows the flow. Here's how the rule works in practice.

A run takes a plain-language description ("independent social media agencies managing content for 5–20 clients") and searches public websites and indexed LinkedIn, Facebook and X pages through Exa. Each run returns up to 10 results, deduplicated by product and normalized URL, so running the same search again doesn't fill your list with copies.

Every lead keeps the text that made it match, shown as "Why this lead appeared". The fit score next to it is a keyword match, and the UI labels it as a heuristic, not a purchase probability. We'd rather undersell the score than have someone treat it as a prediction.

Deep research reads each website lead's contact, about and team pages. An email is added only if it appears in that text. Addresses on the lead's own domain come first, and no-reply and placeholder addresses are dropped. The lead also stores the page the email was found on, so "where did this address come from?" always has a clickable answer.

We pull phone numbers the same way: from tel: links, or from numbers that start with + or follow a phone label. WhatsApp comes only from wa.me or api.whatsapp.com links. We don't use heuristics that "look like a phone number" anywhere else on the page.

To find named people and company facts, the model reads the same page text and returns owners, founders and leaders with their titles, plus team size and founding year. Then we check each one against the text:

each fact is kept only if the page text contains it word for word

If the model returns "Alex Rivera, Founder & Creative Director" and that string isn't on the page, it's dropped. No error and no fallback, just an empty field. An empty field is honest, and a made-up one isn't.

This check also makes failures harmless. If the AI call fails or the workspace is out of credits, the people and company fields stay empty, and the rest of the lead is unaffected.

A run can also write a first message for every new lead. Email goes to leads with an address, and LinkedIn and Facebook leads get a manual DM task with an "Open profile" link. But research never sends anything: every draft waits in review. Private profiles are never accessed.

Try it with a 7-day free trial: https://reach.veloxonhub.com. I'm happy to go into the matching or the Exa setup in the comments.

https://cdn.social.veloxonhub.com/u/eb16653d-76f2-4f5c-b8a4-6521e5547415/2026/10/7127022b-02e5-4bbd-b623-27cd9da6c676.mp4