Attribution models in Google Ads are rules that assign conversion credit to different ad interactions in a user's path to purchase. When someone clicks three different ads over five days before converting, the attribution model decides which click gets credit and how much. Google Ads provides six standard models. Last click gives 100% credit to the final ad clicked, which is simple but ignores upper-funnel work. First click gives everything to the initial touchpoint, useful for measuring awareness campaigns but blind to closing tactics. Linear splits credit equally across all clicks, which sounds fair but treats a homepage visit the same as a bottom-funnel search. Time decay gives more weight to recent interactions, typically with a 7-day half-life. Position-based gives 40% to first click, 40% to last click, and splits 20% among middle touches. Data-driven attribution uses machine learning to assign credit based on actual conversion patterns in your account. It requires at least 3,000 ad interactions and 300 conversions per model within 30 days, which locks out smaller accounts. When you qualify, it typically outperforms rule-based models by 15-30% in our tests because it learns which touchpoints actually move the needle for your specific funnel. The model you choose directly affects campaign optimization. If you run brand and generic search simultaneously under last-click attribution, brand campaigns look phenomenal because they capture demand that generic search created. Switch to data-driven and suddenly your budget allocation shifts. At Ottawa SEO, we default clients to data-driven when they qualify, then compare it against position-based for 30 days. For lead-gen clients with long sales cycles, we often export raw path data and build custom models in Sheets because someone who fills a form after clicking five ads over two months needs different logic than ecommerce. The critical mistake is changing attribution models mid-quarter and panicking when numbers shift. Your conversions didn't change, just how you're counting them. Lock your model for at least 60-90 days before evaluating performance differences.