The Future of AI SEO Explained with Dan Petrovich

**James Dooley:** Hi. So today I’m joined with Dan Petrovic, who has always been someone I’ve admired within the industry. He’s always been two steps ahead of the game. Back in 2013, there’s a famous video now out there about how he said people would be chatting with Google in 10 years’ time and how agentic search would be around. People couldn’t believe it back then. Roll on 12 years and it’s where we are today. So the first topic I want to run through with you is what do you see now for the future of AI SEO? **Dan Petrovic:** Loaded question. The best kind of question because there’s a lot to unpack. Obviously the future of SEO has many additions. But we’ve been through this before with the introduction of various modalities such as voice search, conversational search and mobile first. So we’ve been through these adaptations. I see SEO as evolving and embracing AI as just another modality, another type of interface that sits between search and the end user. There’s going to be a lot more that an SEO needs to do and I would imagine some companies will start specialising. Some consultants will also start specialising. Some will say I’m an AI SEO just like others say I’m an app store optimisation specialist or a mobile SEO or a local SEO. AI SEO is an interest for me because it’s been an interest for the longest time, like you said, and we finally got to the point where the technology is accessible and widely adopted. Technologies that were once only available to very large corporations are now available to everyone. I remember tools like Quill. You could not access it unless you were a multi-million dollar corporation. Now everything is available. I came out of retirement to get back into this space because it is genuinely exciting. Why is it exciting? Because there are opportunities for entirely new methods. One of them, since you are asking for something actionable, is understanding model psychology. That might sound abstract but it is not. I have spent the last two to three years studying mechanistic interpretability. That means understanding what models do internally and why they produce certain outputs. It is similar to human psychology. Humans receive input through sight, sound, touch and smell. That input enters the brain, something happens inside the black box, and then output comes out as a decision. We do not fully understand what happens inside the brain. With models, it is similar. We have input, then internal processing, then output. SEOs now need to act as model psychologists. With open source models like Gemma, Llama and DeepSeek, we can inspect their internal structures. We can examine how semantic concepts cluster in vector space. You create embeddings for one model, embeddings for another, rotate the geometry and you still see similar semantic clusters. Apple as fruit and Apple as technology still separate cleanly. That insight matters because it shows how models organise knowledge. Practically speaking, the future of AI SEO is about understanding how inputs affect outputs. That is where retrieval augmented generation comes in. RAG is essentially search. Models are too expensive to retrain constantly. They are updated occasionally, but for fresh and factual data they rely on search. So AI SEO becomes the optimisation of the interpretative layer that sits on top of search. There is also bias inside models. They have associations toward certain brands and entities. You can probe this by asking what the model associates with a brand. Then reverse the question and ask which brands it associates with a concept. That bi-directional probing maps the model’s internal associations before search grounding even happens. That matters because those pre-trained biases influence the final output. Traditionally we optimised click-through rate. Now we optimise selection rate. Machines do not click. They select. So the goal is to make your content more attractive for machine selection inside RAG pipelines. Surveying humans is messy. Surveying models is clean and repeatable. **James Dooley:** I’ve never heard anyone talk about selection rate optimisation. A lot of people in SEO are obsessed with click-through rate and conversion rate optimisation. But with LLMs there are no clicks. There are only selections. When you spoke about the brain analogy, I started thinking about it in practical terms. If someone goes to a restaurant and gets a plate of food, they judge it by presentation, smell and taste. Different senses contribute to a final decision. With AI SEO, there are multiple signals. Semantic content, entity relationships, third party corroboration, reviews, backlinks. In your opinion, what outweighs what? Is semantic content now the most important thing? Or is third party corroboration stronger? Is it confidence and clarity that matter more than page rank? Some people obsess over on-page. Others obsess over links. What actually moves the needle in AI search? **Dan Petrovic:** Models are actually simpler than traditional search pipelines. They receive plain text or multimodal input. They do not receive page rank, domain rating or authority metrics. Those are part of the search layer, not the model layer. When a RAG system runs, the model does not see your entire page. It receives a trimmed extract. You may write 2,000 words. The model may only see 400 or 500. This is extractive summarisation. It selects verbatim snippets from the highest scoring parts of the page. Those snippets plus the prompt and competing snippets become the input. That means all your sophistication, schema and nuance often get trimmed away before the model even sees them. So your challenge is ensuring that the most meaningful parts of your page are surfaced early and clearly. The model typically receives top results from search. For Google systems it is often around the top five results. The key issue is representation. Will your brand be represented correctly in the small extract that the model receives? That is the real optimisation problem. **James Dooley:** So with query fan out and query augmentation, is the aim to rank for every possible variation? Do you need to dominate across all synthetic variations of a query? **Dan Petrovic:** Fan out queries generate multiple search variations from a single prompt. Each variation retrieves results. Then those results are filtered and trimmed before being passed to the model. You cannot directly control that trimming stage. Your control remains traditional SEO. I use synthetic queries in two ways. First, I ask the model which non-branded queries users would type to find my brand. Then I compare that with Search Console data to identify gaps. Those gaps become content opportunities. However, endlessly generating content for infinite query variations is not sustainable. You must map core entities, cluster related queries and build conceptual graphs. That foundational work matters more than brute force expansion. Prompt tracking as a daily ranking metric is pointless. Instead, probe relationships. For example, ask the model yes or no questions about recommending your brand for a scenario. Repeat it multiple times and calculate probability. That gives you a probabilistic view of brand fit. **James Dooley:** If you receive 68 yes responses and 32 no responses, would you actively optimise to convert those no responses into yes responses? **Dan Petrovic:** Absolutely. There are two paths. The long path is influencing training data. That requires sustained digital PR and brand presence. It can take months or even a year before a new model version reflects that effort. The shorter path is influencing RAG. Improve on-page clarity. Condense content. Insert precise entity signals. Engage in third party citations on domains that are frequently selected by the model. You can create similar content to what is being cited or place your content on those high-performing domains. We tested this with a German sportswear brand called AIO. They were excluded in US recommendations. We strengthened US entity signals and adjusted their on-page localisation. Within a week, grounded recommendations improved. Ungrounded model bias took around six months to shift after a new model release. That demonstrates the difference between influencing search grounding and influencing model memory. **James Dooley:** So realistically, if someone influences a model within a year, they have done well. **Dan Petrovic:** Yes. If you influence a model’s native bias within a year, that is strong performance. But you can influence search-grounded outputs much faster with solid SEO fundamentals. **James Dooley:** Anyone watching this, we planned this to be a short discussion but Dan Petrovic is always two steps ahead of the game and it is hard to keep it short. Check the links in the description for additional videos including how to optimise for LLMs and how AI is impacting link building. Dan Petrovic, it has been an absolute pleasure. **Dan Petrovic:** Thank you.