Fatrank Podcast

James Dooley: Selection rate optimisation in LLMs. Today I am joined with Charles Flo. Charles, it is a pleasure having you on. With regards to selection rate optimisation, for anyone that does not know what it is, can you briefly explain what it is and why it is important in AI search?

Charles Flo: Without going into extreme technical detail about how the system works, SRO is the process of the AI selecting which sources it is going to extract information from and then summarise information from a number of sources. For example, ChatGPT may perform five grounded searches for a specific query during a conversation. Those five searches might return fifty different results per query. That gives around 250 different results, although there will be overlap between them.

Charles Flo: From those results there are only a limited number of sources the system can actually pull information from. For most queries right now, that tends to be around fourteen to sixteen sources within ChatGPT. So from roughly two hundred potential sources, the system has to choose sixteen. Selection rate optimisation is the process of getting your content from that larger pool into the selected set of sources used to generate the answer.

James Dooley: So when someone runs a query, the system may create multiple synthetic queries as part of query fan-out. From there it collects the top results and brings those back. Then you have the selection rate optimisation stage for the documents. Within that, the system performs chunking which forms part of the final answer. For anyone watching this, are there any tips or techniques that can improve selection rate optimisation, such as trust signals?

Charles Flo: One hundred percent. The most cost effective and highest return activity is content level optimisation. AI models only have access to a limited number of tokens from the pages they review. That means only certain sections of your page can realistically be selected.

Charles Flo: Because of that, you need to create structured content chunks that are specifically designed to be extracted by AI systems. These sections need to answer queries clearly and concisely. The structure depends on the query itself, but usually the extractable section appears near the top of the article and often within an H2 or H3 heading.

Charles Flo: The content must also exist on a strong domain. AI models often pull sources from search engines such as Bing or Google. If your page cannot rank in those search engines in the first place, it is unlikely to be selected. For example, if you launch a brand new domain and try to rank for something competitive like best casino websites, the AI model will probably not select that page because it lacks authority.

James Dooley: You mentioned that semantic content should appear higher up the page. Dejan once explained that headings written as questions can work well, followed by a concise answer structured clearly in a semantic triple format. That allows the AI system to easily extract that section.

James Dooley: Another interesting point is consensus. If many of the top results reinforce the same statement, the AI may not even need to read the full documents. It could use signals from the title tag, URL and meta description to form a conclusion. Can you explain why third party corroboration across multiple sites is important for selection rate optimisation?

Charles Flo: Yes, absolutely. The process is highly dependent on the model and the underlying search systems it uses. Some rely on Bing results, some rely on Google, and platforms like Perplexity have their own crawling and caching systems.

Charles Flo: All of these systems apply different weighting factors. For example, OpenAI has partnerships with certain news publishers. These sites may receive preferential weighting compared with other websites. This applies both in the search results themselves and within the model’s internal signals.

Charles Flo: There are also entity and knowledge signals involved. These signals reinforce what the model already understands about a brand or entity. Training data, grounding signals and knowledge graph information all play a role in how the model validates a brand.

Charles Flo: We are increasingly seeing that when a brand has little or no existing information across the web, even if it appears in a list of results the AI may include a caveat in the answer. The model might not rank it first, it may flag it, or it might display a caution indicator. These are situations businesses want to avoid. Positive sentiment across the web is important because negative or uncertain signals can influence how the model interprets your brand.

James Dooley: That makes sense. Expanding entity attributes becomes important. That includes reviews, testimonials, case studies and awards connected to the entity. Instead of the AI simply citing a brand, it can also explain why it is recommended.

James Dooley: For example, the system might reference strong customer reviews or industry recognition rather than saying the brand has mixed feedback or limited evidence. If there is insufficient supporting information, that can effectively damage the brand’s reputation inside AI outputs.

James Dooley: Charles, it has been great discussing selection rate optimisation. There are several other episodes where we discuss topics such as parasite SEO, link building and building third party corroboration. Charles, it has been a pleasure.