James Dooley: Hi. So today I'm joined with Dan Petravvic 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 now be chatting with Google in 10 years time and how agentic search would be around. People couldn't believe it back then and roll on 12 years. It's where we are today. So the first topic I want to kind of run through with you is what do you see now for the future of AI SEO?

Dan Petravvic: Loaded question. The best kind of question because there's a lot to unpack. Obviously the future of SEO has many many additions. But we've been through this before with introduction of various little modalities and voice search, conversational search, mobile first and so on and so on. So we've been through all 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. So there's going to be a lot more that an SEO needs to do. I would imagine some companies will start specialising and some consultants will start specialising. Some will say I'm an AI SEO just like some people say I'm an app store optimisation guy or I'm a mobile SEO or local SEO and so on.

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 technology is accessible and widely adopted. All these technologies that were only available to very large corporations are now available to us. I remember Quill, you couldn't get access to it unless you were a multi million pound corporation. Now everything is accessible. So for me it's like letting the dog off the leash. Everything is fun. Everything is useful. I came out of my retirement to get back into this and I'm very much enjoying it.

Why? Because I see fantastic opportunities for new methods and things to develop. One of them is understanding model psychology. That might sound abstract but it's not. For the last two or three years I've been studying mechanistic interpretability. Mechanistic interpretability is understanding what models do and why, and what happens in their neural network when they do that.

It's similar to human psychology. You get input through eyes, ears, touch, smell. That multimodal input goes into the brain, something happens in the black box, then output comes out and we make a decision. SEOs now need to become model psychologists. We have to deal with the black box.

In human psychology you can plug in an EEG and see brain waves, but it's still limited. In SEO we have to probe models and understand what they do and why. With open source models like Gemma, Llama and DeepSeek, we can actually look into their circuitry. I'm not suggesting every SEO needs to go that deep, but I have found that smaller models built on similar data and architecture have the same multidimensional geometry as the larger models. Semantic concepts cluster together. You generate vector embeddings and although they rotate differently, they collapse onto the same entities. Apple the fruit and Apple the technology cluster separately but consistently.

That means we have work to do in model probing and entity associations. On a practical level, the opportunity lies in understanding how different inputs affect model outputs. This is RAG, retrieval augmented generation. Basically search. The future of AI SEO is dealing with the interpretative layer that sits on top of search.

Models are too expensive to train daily. They are trained periodically and then rely on search for fresh information. So we need to analyse that layer. But there is also something deeper, the model’s primary bias towards or against brands and entity associations.

You can probe that. Ask the model what it associates with a brand. Then flip it and ask what brands it associates with a concept. That bidirectional probing maps the model’s internal associations, which reflect its pre training, fine tuning and reinforcement learning. You need to understand that because it influences choices when search grounding is introduced.

From there we move to specifics. We used to have clickthrough rate optimisation. Now we have selection rate optimisation. Machines do not click. They select. So we need to make grounding snippets more attractive for machine selection. Surveying humans is hard. Surveying models is easy.

James Dooley: I’ve never heard anyone talk about selection rate optimisation. Most SEOs obsess over CTR and conversion rate optimisation, but with LLMs there are no clicks, only selections. You mentioned the black box brain. Let me frame it like this. If I order food in a restaurant and it looks great but smells bad, I may not order it again. Humans weigh multiple inputs. With AI SEO, what outweighs what? Is it semantic depth, backlinks, third party corroboration, reviews? Is there a Midas touch right now or is it just do everything well?

Dan Petravvic: Models are simpler than search. They receive plain text or multimodal input. There is no schema awareness or authority signals inside the model itself. That sophistication lives in search. The model gets a trimmed slice of your content, not the whole pie.

The RAG pipeline decomposes the prompt into fan out queries, retrieves search results, trims them and supplies condensed snippets to the model. That snippet, plus competing snippets, plus the prompt, forms the input. No link metrics. No page rank. Just text.

So the challenge is whether your slice represents you well. That is why condensed, substantive content matters. The model performs extractive summarisation, it pulls verbatim snippets. It does not rewrite your content in grounding.

James Dooley: With query fan out, should we aim to rank across all variations? Is there a cumulative score?

Dan Petravvic: It is not cumulative in that way. Fan out queries retrieve many results, but attrition filtering trims them down. Often the top five dominate. You cannot manipulate the internal scoring. You must rank well traditionally.

I use synthetic queries. Ask the model what users would search to find your brand. Compare that with Search Console data. Identify gaps. Close them. But do not endlessly generate content for every possible query. Map primary entities. Build a concept graph.

And do not waste time tracking hundreds of prompts daily. That is nonsense. Instead, probe relevance. If presented in grounding, would the model recommend you? Force a yes or no output. Repeat multiple times. Measure probability.

James Dooley: You mentioned treewalker.ai. What does it do?

Dan Petravvic: It generates a sentence about your brand without search grounding, then with search grounding, then self evaluates. It also explores all probability paths the model could take. We identify high entropy points, moments where the model is uncertain. Those are your optimisation targets. Reinforce them with on page and off page signals.

James Dooley: So link building is more about clarity and association than page rank?

Dan Petravvic: Correct. The model does not see page rank. Authority in its mind comes from training data associations. Influencing that takes time. But influencing RAG is faster. Get into cited domains. Create lookalike content. Piggyback visibility while building long term brand imprinting.

James Dooley: How long does influencing training data take?

Dan Petravvic: I have one case study. A German sportswear brand wanted US visibility. Grounded optimisation worked in a week. Native model association shift took around six months, aligned with a new model release. If you shift model perception within a year, you are doing well.

James Dooley: Anyone watching this, this was meant to be 10 to 15 minutes but I always enjoy speaking with Dan Petravvic. Check the links in the description. Dan, it’s been an absolute pleasure.

Dan Petravvic: Thank you so much. Pleasure.