Lead Spam Filter Checklist: Protect Genuine Enquiries

LEAD-FILTER EVIDENCE WORKSHEET
Scope ID / form / dates / timezone / snapshot:
Policy version / exact action (allow, quarantine, reject):
Deduplication key and exclusions:
Genuine/spam rubric, unresolved state and reviewer:
How accepted and blocked records were sampled:
Records unavailable before capture / unresolved count:
spamBlocked / spamAllowed / genuineBlocked / genuineAllowed:
Review cost per allowed enquiry / currency / basis:
Expected contribution loss per blocked genuine enquiry (assumption):
Incremental policy and quarantine cost for this comparison:
Cost scope and double-counting checks:
Metrics / null denominators / known sampling limits:
One bounded next investigation / owner / review date:
Security owner / finance reviewer / decision owner:
Stop conditions / appeal route / rollback authority (not granted here):
Evidence references (anonymous aggregate only):

Use this before tightening a lead filter

Use this original editorial synthesis when a marketing, RevOps or sales team wants fewer junk enquiries but cannot tell whether a filter also hides real ones. It fits a fixed policy review with access to evidence from both sides of the decision. It is not a bot detector, a vendor ranking or a reason to disable an abuse control. All examples are synthetic; no client result, personal implementation experience or review by Akshay is claimed. Growthcraft Editorial produced this material with AI assistance.

Do not use the worksheet to infer who is genuine from nationality, a free email domain, failure to answer a call, or whether a deal closed. Define the actual enquiry task and evidence. A genuine enquiry can still be outside the target offer; a malicious submission can look commercially plausible. Keep qualification and security review separate from the two-class measurement problem.

1. Freeze the population and decisions

The analyst records one form population, date range, timezone, export snapshot, deduplication rule and policy version. Decide whether repeated attempts belong to the same enquiry or separate abuse requests; do not alternate units between cells. Keep retries and duplicate technical requests out of a unique-enquiry denominator. Record the policy's actual action: allowing, quarantining and permanently rejecting are not operationally identical.

Security verifies that required protections remain enforced and identifies where records are unavailable. If the system rejects before an enquiry exists, later CRM rows cannot reconstruct those attempts or their intent. Do not invent missing enquiries from pageviews. Choose an evidence-preserving review design approved by security; never create an unprotected public endpoint to collect a comparison sample.

2. Review both accepted and blocked records independently

The sales/rubric owner defines genuine, spam and unresolved before viewing summary results. Use a second reviewer for ambiguous or consequential cases and record disagreements. The filter decision is not the truth label: calling every rejected record spam makes false positives impossible by definition. Keep unresolved records outside the four labelled cells, even if an AI classifier sounds certain.

Reviewing only accepted records can estimate contamination within that selected queue, but it cannot measure genuine enquiries wrongly blocked. If you sample the two queues at different rates, raw combined counts do not describe the full population. Retain stratum sizes and selection rules for an analyst's weighted estimate; the companion calculator intentionally does not perform that inference. Personal records stay in approved internal systems, not in this public worksheet or an AI prompt.

3. Build a four-cell evidence table

For this method, positive means spam and a positive filter prediction means blocked. Count spam blocked, spam allowed, genuine blocked and genuine allowed. Their sum is reviewed N. Keep the remaining unreviewed count separate and display N divided by the total declared cohort as review coverage. Coverage is completeness within the supplied records, not proof of representativeness, truth or statistical precision.

Ask three different questions: what share of blocks were spam, what share of known spam was blocked, and what share of genuine enquiries was blocked? A policy can look excellent on the first and still fail the last commercially. Agree the next investigation's decision criteria with security and the business owner; this framework supplies no universal acceptable error rate.

4. State a comparison before calculating savings

The finance owner supplies review cost per allowed enquiry and a separately stated expected contribution-loss assumption for each wrongly blocked genuine enquiry. For example, three minutes at 30 per hour is 1.50. The second input is not the value of a whole sales opportunity: it should reflect an explicitly assumed probability and net contribution over a named horizon, with a sensitivity range reviewed separately.

The baseline reviews every labelled enquiry. The filtered scenario reviews only allowed enquiries, adds the assumed loss of blocked genuine enquiries and adds incremental policy cost. Quarantine reviews, appeals or extra work must fit a declared cost boundary, without duplication. The comparison applies to the reviewed subset only. Do not allocate a whole month's platform fee against a small sample while presenting the result as full-month economics.

5. Walk through a synthetic review

Of 450 enquiries, 400 have independent labels and 50 remain unreviewed. The four cells are 160 spam blocked, 40 spam allowed, 10 genuine blocked and 190 genuine allowed. Coverage is 88.89%; block precision is 160/170 = 94.12%; spam capture is 80%; genuine block rate is 5%; allowed-queue spam share is 40/230 = 17.39%. None is a universal benchmark.

At 1.50 per review, allow-all costs 600. The filtered queue costs 345; assumed missed-enquiry loss is 10 × 25 = 250; policy cost is 50. Total is 645, a net saving of negative 45. The missed-enquiry loss assumption breaks even at (255 − 50)/10 = 20.50. A precision of 94.12% does not decide whether the policy is worth it. Change the value assumption and inspect how the conclusion changes.

6. Assign one evidence-led next step

Use the cost calculator for the arithmetic, then the review prompt to find gaps. Read the technical row-audit guide before implementing the aggregation. The analyst owns reconciliation, sales owns the rubric, finance owns cost assumptions and security owns control changes.

If blocked records lack independent labels, the next action is an approved shadow review or appeal audit—not a threshold change. If scope mismatches remain, repair them before comparing policies. Keep a review date, owner, evidence reference and stop condition. Small samples, changing attacks, inaccessible attempts, observer disagreement and unequal enquiry values all limit transfer to future traffic. Neither a good score nor a favourable scenario authorizes deployment.

Sources and boundaries

Checked October 25, 2026: Cloudflare validation (updated October 22) separates server verification from a browser widget. HubSpot spam prevention (updated September 22) documents differing filter and reporting effects. Google classification metrics (updated January 12) supplies precision/recall definitions. These sources do not validate this scenario's costs or label accuracy.