The claim: “Our AI found thousands of leads”
Software can identify many records quickly. A record's existence does not show that the person wants to buy, knows the business or has agreed to a conversation. AI can assist research and classification; it cannot turn missing evidence of interest into observed intent by choosing a stronger label.
The first question is whether the provider found an address, observed behavior or received a direct response. Those are different facts and support different next actions.
Use an evidence ladder
| Label | Evidence to define locally | What remains unproven |
|---|---|---|
| Contact | An identity or contact route with a documented source | Interest and suitability |
| Prospect | A contact or company matching agreed targeting criteria | A current buying project |
| Inquiry | A person completed an agreed request for contact | Sales qualification |
| Positive reply | A response classified under a stated rule | Meeting attendance or a sale |
| Qualified opportunity | Sales accepted a documented opportunity definition | A won contract |
| Customer | A verified completed commercial relationship | Profitability and retention |
This is a proposed operating vocabulary, not a claim that every platform uses the same definitions. Put your agreed terms in the campaign brief and reports. If a vendor uses another vocabulary, translate it rather than assume equivalence.
An original identity example
Imagine a hypothetical export contains 150 rows. Twenty rows are repeat form submissions, ten are tests, and twenty relate to existing-customer support. Removing those events leaves 100 potentially relevant new inquiry records. If some people appear under different addresses, person-level deduplication may reduce the count further.
Do not publish the smaller number until the exclusion rules are defensible. A repeat submission might contain a new service request, and a support contact might separately ask about a new purchase. The rule should preserve those distinctions rather than mechanically delete every repeated address.
Qualification is a decision that needs evidence
Ask what information changed the label. A job title might support role relevance; it does not establish budget. A company visit can support account-interest investigation; it may not identify a person. A positive answer to a broad email question might indicate curiosity; it does not prove a procurement timeline.
For a qualification rule, identify required evidence, who verifies it, the expiry or review period and the next action. When a model suggests a classification, retain a human review path for ambiguous cases. Evaluate disagreements on actual records rather than treating a model confidence score as a sales fact.
Make a report difficult to inflate
Report counts by defined stage and show transitions on consistent cohorts. If 40 inquiries arrived this month and four customers signed, establish whether those four came from this month's inquiries or an older cohort before implying a ten-percent conversion rate.
Record rejected, duplicate and still-unreviewed entries. Preserve the distinction between a record received and a record accepted by sales. A dashboard that shows only the largest stage makes it hard to know where the work is succeeding or breaking down.
Questions for a lead-generation provider
Ask for a sample record with personal information appropriately protected. Have the provider show why it meets the promised label, how the source is preserved and how repeated records are counted. Then ask what additional evidence would be needed to call the same entry a meeting, an opportunity or a customer.
ZeroDark's role is to make those distinctions easier to inspect. A smaller, traceable number is more useful for decisions than a larger number whose meaning changes during the sales conversation.
Supporting documentation
Google Analytics: recommended eventsGoogle Ads: conversion measurement