What 10,000 field submissions taught us about BTL execution fraud in India

1. Sample selection

10,247 submissions drawn from active campaigns running on gOGig across 29 enterprise brands. Stratified across BTL activations, retail audits, OOH installations, promoter check-ins, and field sales visits to ensure cross-format representativeness.

2. Verification dimensions

Each submission evaluated across seven dimensions: GPS coordinate integrity, mock-location detection, EXIF timestamp validity, image fingerprint uniqueness, route consistency, accelerometer cross-check, and clustering pattern analysis.

3. Anomaly classification

Submissions flagged as anomalous were classified into ten anomaly categories. A single submission could trigger multiple flags simultaneously. The total flagged rate exceeds the sum of unique anomalies due to overlap.

4. Stratification analysis

Anomaly rates broken down by format, geography (tier-1, tier-2, tier-3, rural), time of submission, and submitting vendor. Statistical confidence intervals applied where sample size permits.

5. Validation against known cases

A control set of 200 submissions known to be authentic was passed through the same pipeline. False positive rate measured at 1.4%, indicating the anomaly detection is conservative rather than aggressive.

GPS coordinate mismatch with EXIF

The GPS coordinates reported by the device do not match the coordinates embedded in the photograph's EXIF metadata. Indicates a photograph taken at one location and submitted from another. Detected in 12.4% of submissions.

Server timestamp deviation

The timestamp claimed in the submission does not match the server timestamp at the moment of upload. Detected in 11.7% of submissions, often with multi-hour or multi-day gaps.

Mock-location app active

Developer mode and mock-location flags detected on the submitting device. Detected in 7.0% of submissions, concentrated in field sales and promoter check-ins.

Image fingerprint duplicate

The image hash matches another submission in the dataset. Indicates the same photograph submitted multiple times. Detected in 6.3% of submissions.

Accelerometer inconsistency

The device reports stationary positioning during a submission that claims movement, or movement during a submission that claims stationary work. Detected in 5.8% of submissions.

Geo-fence violation

The submission location falls outside the contracted geographic zone for the campaign. Detected in 4.9% of submissions, predominantly in mobile van and field route campaigns.

Clustering anomaly

Multiple submissions originating from a single GPS coordinate within an implausibly short time window. Indicates one device producing multiple 'site visits' without physical movement. Detected in 4.2% of submissions.

Quantity mismatch

Reported execution scale exceeds physically plausible volume given the time and location constraints. Indicates inflated counts. Detected in 3.6% of submissions.

Image quality below threshold

Photograph quality below the verification standard, including blur, glare, dark conditions, or framing that obscures the contracted artefact. Detected in 3.2% of submissions.

Sequence anomaly

Submissions arrive out of the contracted workflow sequence (for example, dismantling photographed before setup). Detected in 2.7% of submissions, mostly in OOH and BTL setup-dismantle workflows.

Late-evening batch uploads are a red flag

If a campaign's submissions concentrate after 9 PM, the underlying execution should be questioned. A working day ends. Submissions should taper through the day, not surge at the end.

Rural campaigns require higher verification investment, not lower

The verification logic that brands apply most heavily in tier-1 is exactly inverted. Rural and tier-3 campaigns carry double the anomaly rate, but historically received the least verification.

Vendor type is a leading indicator

Procurement decisions based on cost alone optimise for the vendor categories with the highest anomaly rates. The 13-point gap between individual contractors and national agencies reflects the structural fraud premium of fragmented vendor models.

Workflow segmentation reduces fraud

Submission types with multiple checkpoints within a single workflow show lower anomaly rates. Designing campaigns with more verification touchpoints, not fewer, structurally reduces leakage.

Detection accuracy is high enough to act on

With 98.6% accuracy and a 1.4% false positive rate, the platform's verification engine produces signals that are reliable enough to trigger payment holds, contract renegotiation, and vendor rotation decisions.

1. Conservative classification thresholds

We tuned the verification engine to err toward accepting submissions, not flagging them. A submission is only classified as anomalous if multiple checks align. This produces a lower headline anomaly rate, but every flagged submission is high-confidence.

2. Cross-format weighting

We did not weight the dataset to over-represent high-fraud submission types. The 22% headline GPS anomaly rate reflects the natural mix of submissions on the platform during the analysis window.

3. Anonymisation of brand and vendor identity

All findings are reported in aggregate. No individual brand, agency, or vendor is identifiable in this report. This is a research publication, not an audit disclosure.

4. Reproducibility commitment

The classification logic used to produce these findings is documented in the full Q1 Report. Future researchers and journalists can validate the methodology against their own samples or platform deployments.