How can a company get its website linked in chatbot results? A worthwhile Saturday morning coffee read from researchers at the University of Toronto attempts to clarify that. GEO (generative engine optimization) is an emerging field of marketing that aims to improve product discovery by nudging chatbots to mention specific products, brands, or services. From what I've seen, the techniques employed borrow heavily from the Search Engine Optimization (SEO) playbook: focus content on keywords, provide structured and machine-readable product / services data, include summaries and FAQs, etc. I've always been skeptical of this approach to improving organic discovery because a) it is reactive to model ("algorithm") changes and b) it strips companies of direct control over discoverability. A new paper titled 'What Drives AI Citations? Content, Retrieval and Model Knowledge in Generative Engine Optimization (GEO)' seeks to quantify the role of models' existing knowledge of brands as well as the relevance and structural elements of webpage content in citation in answers from chatbots. What it finds is that content relevance is associated with more citations, while various popular GEO content formatting and structural influence tactics show no reliable citation benefit (note that the scope of the paper is websites). The authors prompted several models through their APIs (OpenAI's GPT-5.4 nano, Google's Gemini 3.1 Flash Lite, and Perplexity's Sonar), with web search enabled, with commercially oriented questions about US home internet providers and niche B2B software once per day over the course of 90 days and recorded the responses. Among other things, they find: Common GEO features show no reliable citation benefit within websites. FAQs, summary boxes, author information, and structured data appear beneficial when comparing *different* websites, but those associations disappear when comparing pages on the same website. They find that these features are concentrated on websites that are more likely to be cited generally, but their presence is not significantly associated with higher citation frequency within the same website. The model's existing knowledge of a brand shapes the searches it conducts in anwsering a question, and retrieval in those searches is strongly predictive of citation. For OpenAI, how often pages appear in results returned by its search tool, and how highly they appear, explains substantially more variation in citation counts within websites than the measured page characteristics: 32% versus 2.8%. So the model is more likely to search for brands that it already recommends without searching, and those search results strongly predict citation. Paper linked below.