How Costless verifies retail price & receipt data

Online parsers

Each chain parser knows that chain's page structure or API. Parsers run on a fixed twice-weekly cycle, staggered to avoid burst load, respect site protections, and prefer chain-published APIs over scraping. Some chains we don't crawl at all because their terms prohibit it — for those, data comes from receipts, field collection and banners only.

Receipt OCR pipeline

An uploaded receipt is pre-processed with OpenCV (skew, contrast, glare, thermal-paper fade), then text is extracted by Google Cloud Vertex AI vision services as the primary engine, with Tesseract as a fast fallback and an in-house CNN trained on Eastern-European receipt formats for the hard cases. Line items are parsed, products matched against our normalized database via vector embeddings, and totals reconciled against the printed total.

End-to-end accuracy across accepted receipt formats is consistently high. Per-stage benchmarks vary by retailer template, paper type and image quality, so we don't publish a single number that would misrepresent the real distribution.

Fraud detection

Roughly 8% of submitted receipts are flagged as fraudulent before reaching the verified dataset. Signals include duplicate detection, image forensics (photos of screens or printouts, edited images), merchant-consistency checks, structural anomalies, and velocity heuristics. Flagged submissions are quarantined and the user is asked to re-photograph clearly.

Verified field capture

The field-agent app enforces an on-site flow: it reads GPS, reverse-geocodes to an address, and blocks capture unless the agent is at the target store. After an explicit check-in, photos are taken in-app only (never the camera roll), processed on-device by our AI models, and uploaded instantly — so every price ties to a specific store, agent and moment.