Traffic source attribution after the cookieless transition
GA4's attribution model held together for longer than I expected. Here's what's actually trustworthy in 2026 and where the model still hides large amounts of organic traffic.
Three years into the cookieless transition, attribution debates are quieter than they were. Most of the loud predictions were wrong in both directions. GA4 didn't fall apart; it also didn't become reliable. Here's the calibration I use when reading attribution data in 2026.
What you can still trust
- First-touch organic search is reliable for landing pages with a clear referrer. If the referrer header was google.com or duckduckgo.com and the landing page didn't redirect, GA4's attribution is solid.
- Direct traffic is real, but it's not what it used to be. A significant share of what GA4 labels 'direct' is in fact 'we lost the referrer due to HTTPS-to-HTTP downgrades, in-app browsers stripping headers, or paid social link wrappers.' Treat the direct channel as a mixed bag and segment it by landing URL — a long, deep URL receiving 'direct' visits is almost always a lost-referrer case.
- Branded organic search remains the cleanest signal for measuring brand demand. Watch it as a leading indicator for everything else.
What you can't
- Cross-device journeys for non-logged-in users. The signed-in user graph still works inside Google's ecosystem; outside it, multi-device attribution is mostly modeling on top of incomplete data.
- AI search referrals. When a user clicks a citation in an AI Overview, ChatGPT, or Gemini, the referrer is inconsistent and often stripped. A growing percentage of high-intent visits are landing in your 'direct' bucket. We estimate this is now 4–8% of total traffic on most B2B sites we work with.
- Anything that depends on third-party cookies remaining functional in 2026. They mostly don't, in mostly-uneven ways across browsers.
The unit of measurement that actually scales
We've moved most of our reporting away from session-level attribution and toward landing-page-level demand. The questions are: which pages are growing in incoming non-branded sessions, which pages have shifted from referral to direct (a lost-referrer warning), and which pages have a rising share of branded-after-visit follow-up traffic.
That's not a sophisticated model. It's a deliberately simple one — because the underlying data has gotten too uncertain to support the kind of multi-touch attribution that was fashionable five years ago. Simple, defensible, and honest about its limits beats clever and untrustworthy every time.