Ad Platform API Integrations Built as Production Services

The Google Ads API returns HTTP 200 on a mutate in which 400 of 500 rows failed. The row-level errors sit in partial_failure_error, and an integration that reads only the status code will log a successful upload every night while your offline conversions quietly stop arriving. That failure mode is the reason this is engineering work and not configuration. We build and hand over the services that move data between your commerce backend and the ad platforms: click conversions and store sales into the Google Ads API, order and CRM events into the Meta Conversions API against a dataset, Microsoft Advertising, Amazon Ads and Criteo, and cost, impression and click data pulled the other way into BigQuery so blended reporting stops being assembled by hand every Monday morning. What the code has to survive: Return windows. A sofa ordered in March and collected back in May was never revenue, and a platform that already booked it will keep booking it. We write conversion adjustments off your ERP credit notes — RETRACTION for cancellations, RESTATEMENT for partial refunds — carrying the original conversion timestamp, without which the adjustment is simply rejected. Time and currency. Google Ads reports in the account time zone, your ERP posts in Europe/Berlin, Meta expects Unix seconds in UTC. Every inherited pipeline we have opened had at least one of these wrong, and the symptom is always the same: a day-shaped hole at the month boundary that nobody can explain. Gross or net of VAT gets decided once and written into the spec. Identity. Hashed email and phone go up only for orders whose consent record permits it, normalised the way each platform documents it — lowercased, trimmed, phone in E.164 — and hashed inside your network before anything leaves it. Failure. Every job has a retry policy, a dead-letter table and an alert that names the batch it lost. When a night's upload does not happen you hear it from monitoring, not from an agency asking why ROAS moved. How it is delivered: Python or TypeScript, running on your infrastructure — Cloud Run, Cloud Functions or a container on your own host — in your repository, with service account scopes documented and credentials in your secret manager, never in ours. You also get a reconciliation query that compares what the ERP believes against what the platform accepted, and an agreed tolerance. Drift past that tolerance is a defect and we treat it as one. What we do not do: We do not resell connector seats. If three endpoints are what you actually need, a managed ELT subscription is an expensive way to own less of your own pipeline, and we will say so before you sign for one. We do not build write paths that change bids or budgets. One unattended job with a bad multiplier spends a five-figure day, and none of these platforms offer a rollback. We do not deliver dashboards. You get tables in BigQuery with a documented schema; point whatever BI tool you already pay for at them. We do not start on a pipeline whose numbers nobody inside the company is accountable for. Find that person first, then call us. Who this is not for: Shops where the platform's native shop integration is genuinely enough — under roughly EUR 200,000 monthly media spend it usually is, and the honest answer is to keep it. Teams who want the integration finished before the event definitions are agreed; the order matters and we will not reverse it. Anyone looking for a supplier to sign off that the resulting data flows are lawful. We document precisely what is transmitted, to whom, and on what trigger. Your data protection officer decides. Documentation, code comments and handover in English or German.