A new research report from the University of Toronto examines how generative AI answer engines differ from traditional web search, revealing a fundamental shift in how information is sourced, ranked, and presented.
The study compares Google Search and Google Gemini with leading generative systems including GPT-4o, Claude, and Perplexity across thousands of queries. It finds that AI-generated answers draw from source ecosystems that barely overlap with Google’s top results, often below 15%. This shows that answer engines are not simply re-ranking the web but constructing responses from distinct pools of content.
Generative systems consistently favor earned media and brand-owned sources while underrepresenting social and community content, especially for consideration and transactional queries. Unlike Google’s relatively stable sourcing mix, AI engines adjust their source composition aggressively based on query intent. Across consumer electronics and automotive queries, AI systems also cite significantly fresher content, with median article ages 40% to 70% newer than Google’s.
The report further isolates the impact of large language model pre-training. For popular entities, rankings remain highly stable even when retrieved evidence is shuffled or constrained, indicating that pre-trained knowledge dominates reasoning and citations often act as confirmation rather than discovery. For niche entities, rankings are far more sensitive to retrieved snippets, grounding constraints, and context order, signaling evidence-driven reasoning.
These findings suggest that SEO and Answer Engine Optimization follow different rules. For popular topics, being included in AI retrieval contexts matters more than traditional rank position, while for niche topics, freshness, authority, and source selection become decisive.
For brands, influencing AI answers will require optimizing for inclusion, credibility, and freshness across earned and owned media, rather than relying solely on traditional keyword rankings or page-level SEO tactics.
