LLM Selection Rate Optimization (James Dooley Interviews Charles Floate)

James Dooley: SRO — Selection Rate Optimisation in LLMs. Today I’m joined with Charles Floate. Charles, it’s a pleasure having you on. With regards to Selection Rate Optimisation, for anyone who doesn’t know what it is, can you briefly explain what it is and why it’s important within AI search today? Charles Floate: Yeah. So without going into extreme technical detail about how the system actually works, SRO is essentially the process of the AI selecting which sources it’s going to extract information from and then summarise that information from. Let’s say as an example that ChatGPT performs five grounded searches for a specific query during a conversation. Each of those searches might return around 50 results. That gives you roughly 250 results in total. There will be overlap between those results, so you might end up with about 200 unique sources. Now, from those 200 sources, the AI can only select a limited number. For most queries right now in ChatGPT, it will typically choose somewhere between 14 and 16 sources. Selection Rate Optimisation is essentially the process of getting your content chosen within that final group of selected sources. James Dooley: Right. So if someone does an initial search query, the AI may create additional synthetic queries — part of the query fan-out process — and then collect results from those. It pulls back the top results and then the model needs to decide which documents it’s going to select. After that, it performs chunking to extract specific parts of those documents that help form the final answer. For someone watching this, are there any tips or strategies around trust signals or content optimisation that can help improve Selection Rate Optimisation? Charles Floate: Yes, absolutely. The first and most cost-efficient thing you can do with the highest ROI is content-level optimisation. The AI only has a limited number of tokens it can process from each source. From those hundreds of results it’s evaluating, it can only look at a relatively small portion of your page. So you want to create well-structured chunks of content that are designed to be easily extracted by the AI. These chunks need to be optimised for the query itself, which means the way you structure them will vary depending on the search intent. Because of certain biases within the models, the AI often looks near the top of the article. Typically, the extractable content will be located under an H2 or H3 heading. But the page also needs to be on a strong domain that can already rank in Google or Bing. If you launch a brand-new website and try to rank for something competitive like “best casino websites,” you’re probably not going to be selected by the AI. The page needs to rank first, and then it needs to have an optimised snippet that can be extracted by the AI model. James Dooley: That makes sense. So you mentioned semantic content being placed higher up the page. I remember Dejan talking about using question-based headings. The idea is to structure headings as clear questions, and then directly underneath provide a concise, structured answer that clearly resolves the question. That creates a semantic triple — question, answer, and supporting context — which makes it easier for AI models to extract that chunk of information. But another interesting point he raised was about consensus signals. If the AI sees consistent information across multiple sources — for example across titles, URLs, and meta descriptions — it might not even need to open the page to confirm that information. So this is why I wanted to talk to you about things like parasite SEO, link building, and building consensus. Can you explain why Selection Rate Optimisation isn’t just about optimising your own website, but also about off-site signals and third-party corroboration? Charles Floate: Yeah, absolutely. First, it’s important to understand that a lot of this is model-dependent. Some models are grounded in Bing, others in Google, and some platforms like Perplexity have their own crawlers and caching systems. Each of these systems applies different weighting to sources. For example, OpenAI has partnerships with certain news publishers. Those sites often receive preferential weighting in the model. So authority plays a role not just in traditional search rankings but also within the AI models themselves. Beyond that, there are entity-level signals and knowledge-graph signals that reinforce trust around your brand. These signals help the AI validate and understand your entity. They also influence both training data and grounded retrieval systems. What we’ve seen increasingly is that if a brand appears in an article or list but the AI doesn’t have much background information about that brand, the model may include caveats in the answer. For example, it might not rank the brand first, it might flag it with a warning, or it might include some kind of caution indicator. Those situations are obviously not ideal for your brand. So the goal is to ensure positive sentiment across the web and across all entity signals, so that the model consistently recognises and trusts your brand. James Dooley: Exactly. So what you’re really trying to do is expand the entity attributes connected to your brand. That might include reviews, testimonials, case studies, awards, and other reputation signals. Instead of the AI simply mentioning your brand, it can also pull the reasoning behind why it’s recommending you. For example, it might say that a company has strong five-star reviews or recognised awards. On the other hand, if there isn’t enough supporting information, the AI might say there’s limited evidence or mixed sentiment. As you mentioned, negative or missing sentiment can effectively poison the model’s perception of your brand. Charles, it’s been an absolute pleasure talking about Selection Rate Optimisation. If you want to learn more, check the link in the description — there are several other episodes where Charles and I discuss topics like parasite SEO, link building, and building third-party corroboration. Charles, it’s been a pleasure.