An attribution model decides how much credit each marketing touchpoint receives when someone converts. If a user clicks a Facebook ad, reads a blog post two days later, then converts via a Google search a week after that, the attribution model determines which channel gets the credit—or how to split it. The most common models are last-click (the final touchpoint before conversion gets 100% credit), first-click (the initial touchpoint gets all credit), and linear (every touchpoint gets equal credit). Google Analytics 4 defaults to data-driven attribution, which uses machine learning to weight touchpoints based on their actual influence. Multi-touch models like time-decay give more credit to recent interactions, while position-based models emphasize the first and last touchpoints. No model is objectively correct. Last-click is simple but ignores earlier research phases. First-click overvalues awareness channels that rarely close deals. Linear treats a quick retargeting click the same as a detailed product comparison, which doesn't match reality. Data-driven sounds smart but becomes a black box—you can't easily explain why budgets shifted. At Ottawa SEO, we typically start clients on last-click or data-driven in GA4 because they're accessible, then layer in multi-touch analysis when they run 4+ channels and need to justify upper-funnel spend. For ecommerce clients spending $15K+ monthly on ads, we'll build custom reports that compare models side-by-side so they see how Facebook looks under first-click versus last-click. The real issue is cross-device and offline conversions. Someone researching on mobile, converting on desktop a week later breaks most models. If you run both online ads and trade shows, standard attribution completely misses half the journey. That's where CRM integration or tools like HubSpot's multi-touch reports become necessary, though they add cost and complexity. Choose the simplest model that reflects how your customers actually behave, then upgrade only when budget decisions depend on more granular data.