Suspicious leads: evidence, duplicates and review costs

A rise in submitted forms is not necessarily a rise in customers. It can reflect real demand, repeated deliveries, accidental submissions, an unsuitable audience or deliberate manipulation. This guide gives an advertiser a defensible review process: preserve the evidence, separate technical duplicates from business outcomes, protect legitimate customers and decide whether a source deserves another test.

The two cases below are fictional training datasets for an ordinary service business. They are not AffiliateCore client results. Amounts are independent educational USD assumptions, not provider prices. The procedure is applicable to a permitted product and country; it does not authorize a restricted vertical or a particular traffic channel. No example requires generating fake registrations, orders or payable conversions.

1. Define what counts before looking for fraud

Write a one-page acceptance brief before launching a partner placement. Specify the service, supported area, qualifying customer request, duplicate window, evidence of an intentional submission, review deadline and dispute route. Separate a valid lead from a sale: a person can genuinely request information and decide not to buy. If payment is for a valid request, the absence of a sale is not automatically a reason to reject it.

Use three independent identifiers: the source or placement reference, the submitted request and the delivery attempt. One request delivered three times is not three customers. Two requests from one household are not necessarily duplicates. Resolve those questions using the agreed business definition, not a blanket rule about matching IP addresses or contact details.

In our AffiliateCore platform, start with the conditions of your own offer and the supported conversion review process. Agree the event and permitted source before connecting a partner. The evidence worksheet described here belongs to the advertiser's operational process; it is not a claim that AffiliateCore contains a case-management screen, a Cloudflare integration or automatic proof of fraud. If the necessary event cannot be represented reliably, resolve that gap before accepting payable traffic.

2. Use signals to select a review, not to issue a verdict

A sudden burst, many similar submissions, an unusual country mix or a very short interval deserves investigation. Each has alternatives: a newsletter can arrive at once, an office can share a network, mobile routing can distort location and an existing customer can complete a form quickly. Country estimates alone do not establish a person's residence or eligibility.

Cloudflare describes its bot score as an estimate of automation. Detailed scores require the relevant Enterprise Bot Management entitlement. A score is not a diagnosis of commercial fraud, and an uncomputed score is missing information. Do not invent scores when the product or data is unavailable. Keep a review category called “insufficient evidence” rather than forcing every record into genuine or fraudulent.

Google's invalid-click guidance includes accidental and automated interactions. That is a useful distinction: invalid clicks, rejected leads and cancelled orders are different populations. A platform's click adjustment does not automatically resolve the acceptance of a lead in your own CRM, and an advertiser's rejection does not prove an ad platform will issue a credit.

3. Prepare a minimal evidence worksheet

Use a restricted worksheet with these columns: case reference; source and placement; request reference; delivery-attempt references; timestamp and timezone; contractual acceptance rule; observed signal; alternative explanation; evidence location; review owner; provisional status; final reason; appeal deadline; payable amount; reversal or refund status. Store a pointer to private evidence rather than copying customer identities into every export.

Keep observations separate from conclusions. “Three delivery attempts share one request reference” is observable. “The publisher fabricated three customers” is a conclusion requiring additional evidence. Record what is unknown and which next check could distinguish the explanations. Do not buy leaked databases or contact unrelated people to enrich a suspicion.

Suggested internal labels are: awaiting review, technical duplicate, accepted, declined under agreed criteria, unresolved and overturned on appeal. These are worksheet labels, not promised interface values. Every decline needs a reason tied to the original brief. Maintain the initial decision as an audit entry when correcting it; do not silently rewrite the history.

Register as an advertiser in AffiliateCore to prepare your own offer conditions and source requirements. Bring this brief to the setup process; an account does not guarantee a source's approval or a supply of ready-made partners.

4. Contain the problem without destroying the evidence

Pause additional spend on the specific questionable placement when exposure exceeds the agreed limit. Preserve the source reference, relevant configuration and a dated read-only export. Do not block every partner because one placement produces a burst. Continue processing already valid requests and honour the agreed review deadlines.

Check the intake system first. Compare request identifiers with delivery attempts and CRM rows. A retry after a timeout can create apparent growth without any new customer activity. Ask the integration owner whether retries are idempotent: processing the same request again should not create another payable result. Compare server receipts with the CRM, rather than inferring missing records from unrelated pageview totals.

Review a bounded sample of suspicious records alongside ordinary records from the same period. Include accepted and declined cases, not only the worst examples. Use the contact process already agreed with the person; repeated unsolicited calls are not a valid investigative method. Have a second reviewer examine borderline rejections when feasible, without showing the first reviewer's conclusion initially.

5. Scenario one: apparent inflation caused by retries

Assume an approved service placement produces 160 incoming event messages over one week. The system contains 120 distinct request references and 40 repeat delivery messages. A reconciliation shows those 40 messages concern already existing requests. They are delivery duplicates, not additional leads, and their presence alone does not establish misconduct by the partner.

Review the 120 unique requests. Under the pre-agreed definition, 80 are accepted, 20 fall outside the service area, 10 are withdrawn before the required confirmation and 10 remain unresolved. The assumed contract makes that confirmation necessary for acceptance; this is not a universal right to reject withdrawn enquiries. The four categories total 120. Keep the unresolved group outside the final cost-per-accepted denominator and reserve its possible liability.

The fee is an assumption, not an AffiliateCore tariff. Cost per accepted lead is USD1,164 / 80 = USD14.55. Dividing by all 160 event messages would give USD7.275 and misrepresent the business outcome. If each unresolved request could add USD8 plus an assumed 10% fee, reserve another USD88, with further review labour budgeted separately.

The pilot's pre-agreed cost ceiling is USD16 per accepted lead. The current cohort passes that arithmetic gate, but scaling waits for the unresolved decisions and a verified duplicate fix. The evidence supports correcting the delivery process and continuing a bounded test, not accusing the publisher of fabricated customers. The first diagnostic expenditure remains part of this pilot's costs even if later runs are cheaper.

6. Scenario two: cheap submissions, expensive accepted results

A different fictional placement records 200 distinct requests. After the agreed review period, 35 qualify, 100 do not meet documented eligibility conditions, 25 are withdrawn before the required confirmation and 40 remain unresolved. A suspicious pattern in timing prompted the review, but rejection reasons come from documented criteria. “Arrived too quickly” is not used as the sole reason.

Costs are USD1,000 for the placement, USD350 for creative and landing work, USD160 for review labour (8 hours × USD20), USD80 for tools, USD70 for dispute handling and USD350 partner payment for 35 accepted leads at USD10. An assumed separate fee adds USD35. Total cost is USD2,045, or USD58.43 per accepted lead after rounding. The assumed maximum is USD30. The 40 pending requests carry up to USD440 additional variable liability, before further review work.

Even the optimistic case in which all 40 pending requests become accepted would produce 75 accepted leads at a cost of at least USD2,485: USD33.13 each, still above the ceiling. This is a bound using the stated assumptions, not a forecast. Stop buying this placement while completing review and paying valid liabilities. Do not manufacture a profitable report by rejecting legitimate requests or excluding creative labour.

The correct conclusion is that the placement fails the economic gate and has an unresolved quality investigation. It is not yet proof that all 165 non-accepted or pending records are fake. If later evidence confirms deliberate manipulation, preserve the finding and the affected records separately from ordinary ineligibility, withdrawals and uncertainty.

7. Validate the protection, including its false rejections

If an external challenge service is used, check its server-side verification and failure handling. Cloudflare's Turnstile validation documentation states that the browser widget alone is insufficient; tokens require server validation and are short-lived and single-use. These properties protect a technical interaction. They do not establish that the person wants the product or meets the offer's commercial conditions.

A safe test uses the provider's documented test mechanism or a staging environment, with records excluded from payable activity. Test a normal permitted submission, an expired validation response, a duplicate delivery, a temporary provider failure and a legitimate retry. Confirm that a rejected technical attempt does not create a payable conversion and that an allowed retry cannot create two. Never send invented customer activity into a live partner campaign to test earnings.

Measure what happens to legitimate users. Record challenge failures, support complaints and decisions overturned after review. Inspect mobile and accessibility paths. A rule that removes both unwanted traffic and most real customers can improve a rejection percentage while damaging the business. Do not silently make a new friction rule universal before measuring that trade-off.

8. Give the partner a bounded right of reply

Send a concise case summary through the agreed business channel: placement reference, cohort dates, counts, applicable criteria, anonymised examples and the response deadline. Ask for evidence of the actual placement and its stated audience. Share only what is necessary for the dispute; do not send full customer lists as routine proof.

Treat a missing response as missing evidence under the agreement, not as proof of identity or intent. Log new evidence and correct mistaken decisions. Where the advertiser's own form or integration caused the issue, say so and correct the affected accounting. Keep any reimbursement request separate from the decision to continue buying traffic.

9. Decide what may restart

Restart only when the source is identified, the event definition is stable, duplicate handling is verified, review capacity is adequate and pending liabilities are funded. Set a small spend cap and a date for a mature-cohort review. Change one major control at a time so that a better result can be interpreted. Compare similar periods and customer groups; a before-and-after difference alone does not prove the new filter caused improvement.

Stop again when the acceptance cost breaches its limit, evidence cannot be reconciled, legitimate-user rejection rises beyond the agreed tolerance or the partner will not substantiate the placement. There is no universal safe rejection percentage. State the business-specific limit before the next pilot and include the cost of investigating it.

Your final pack should contain the acceptance brief, evidence worksheet, duplicate reconciliation, two-sided review sample, full-cost calculation, pending-liability reserve, appeal decisions and restart criteria. Existing advertisers can sign in to AffiliateCore and compare their offer's supported conditions with this pack. New advertisers can prepare an advertiser account. For the broader operating process, see lead review and disputed decisions.

Sources checked on 14 October 2026. The worksheets, scenarios, thresholds and calculations are original educational constructions. They describe recognition, investigation and protection, not methods for producing deceptive traffic.