A forwarded link reports utm_source=facebook even when the click came from WhatsApp

A published link arrives with its query string intact. The platform hands the destination page the full URL, so whatever label you typed before posting is what the insert sees ^[8]. That fixes the boundary of the method. A link tagged utm_source=facebook, copied by a reader into WhatsApp and clicked there, still reads utm_source=facebook. The post treats this as inherent to the approach and says the code is not where the defect lies ^[9].

The landing table is seven columns: created_at, path, event, source, medium, campaign, content ^[2]. Four of them, source through content, stay empty until you tag links, so three are filled by the request itself and four by whoever built the URL ^[3][1]. The column list includes no visitor identifier. A count over this table therefore counts events ^[2]. Where storage is tight, the post suggests aggregating by day instead of keeping every row ^[4].

Serverless is where a file-based store breaks. The function filesystem is per-invocation, recycled and sometimes read-only, and the post says: "A SQLite file there is silent data loss." ^[7]

Compound labels like utm_source=facebook-group-1-shirts-september-blue-post are called an antipattern, because each question needs its own axis ^[11]. The three worked examples share utm_campaign=camisas-2026-09 and differ on source, medium and content ^[13]. So SELECT source, COUNT(*) ... WHERE campaign = 'camisas-2026-09' GROUP BY source answers which platform worked, while the campaign filter answers whether the campaign worked ^[12].

The tagging itself comes out of config. The offer exists once, id oferta-2026-09-camisas at precio_cup 2300, with a channel list underneath that carries source, medium and content. URLs are generated from that, so a price change is one edit and tagging cannot be forgotten ^[15]. Per-channel copy goes through an adapter interface because text, image, tone and CTA differ by channel ^[16]. What you published is a second table, publicaciones, keyed with a pieza_hash that detects content changes so the same piece is not republished twice, which the post labels deduplication and idempotency, not security ^[17].

Two of the four questions do not automate at this size. Publication to visit measures itself; visit to inquiry to sale does not, and the recommendation is one column or a sheet, manual first, automated when volume justifies it ^[18]. The whole design also rests on a stated assumption: you control the destination page and you generate the links before posting. If a shortener or a third party generates them, the data is not yours ^[14].

Short codes are presented as representation. A parameter like ?c=7f3a resolves to a row holding campaign, channel and content, and the post says it does not make attribution truer ^[19]. Signing is scoped just as narrowly. HMAC with a shared secret verifies one thing: that a value was generated by whoever holds the secret. The post reserves it for parameters with consequences: coupons, discounts, affiliate links, attribution that pays someone ^[20].