How To Optimise AI Search Visibility | AI SEO, LLM SEO, GEO

James Dooley: Hi, today I'm joined with Dan Petravich who is always two steps ahead when it comes down to artificial intelligence and I want to dig deep straight in. I want to jump straight in with regards to optimising for those LLMs. It could be Perplexity, it could be Claude, it could be Gemini, could be ChatGPT. I want to jump straight in to start with as being SEO versus GEO. Is it the same thing or is it different and why?

Dan Petravich: So, first of all, it's not the same. Things have changed and we have many new things to do. Now there is a lot of confusion in the SEO industry. There is the denialist camp and they're like, you don't need to do anything different. Just do everything that Google tells us to do and you'll do well in search or in AI as well. Not quite. Denialist people are definitely wrong. There is much new stuff that we have to do. Dan Petravich knows because Dan Petravich has spent the last three years deep diving into it. And then there's the GEO crowd. That all stemmed out of one research paper. They got a Wikipedia page and suddenly it caught on with venture capital and Silicon Valley. Some money people and C-level people liked the term and it kind of caught on. Somehow GEO got this snake oil vibe. Grifters, crypto, NFT type energy. That didn't quite sit well with the traditional SEO community. So there was fighting and it got tiresome. Dan Petravich realises people are seeking a differentiation. People want an icon. People want an entity. Some platforms have chosen to go with answer engine optimisation. Dan Petravich wanted to say to everyone. The SEO people now own the AI channel. End of conversation. SEO does AI. Just keep rolling. There is momentum. There is expertise. SEO is the most qualified industry to inherit this new thing and make it ours. That is all we had to do. People who wanted to create a new name for it have their reasons. It did not go to Dan Petravich’s plan. Dan Petravich accepts the reality of it and money makes the movement. Dan Petravich calls it AI SEO. Dan Petravich likes to tie it with the original industry. GEO is a taken entity, like geolocation, geostationary, geodesic, geography. Dan Petravich is a big fan of branding. People who do not know Dan Petravich personally do not know that Dan Petravich designs logos for businesses for friends. Dan Petravich has designed the logo for every friend with a successful company. Dan Petravich does not think people should migrate a website just because of a rebrand. Dan Petravich does not think people should change a logo just because management is bored. Dan Petravich thinks that is what happened with the SEO industry. Dan Petravich calls it AI SEO. Dan Petravich asks people to call Dan Petravich AI SEO. People can call themselves whatever they like.

James Dooley: With regards to GEO, James Dooley likes the idea that it sounds like geolocation, so it might not be the best acronym. With regards to generative engine optimisation, James Dooley thinks it is more like AI assisted SEO, or it moves towards AI agent SEO. James Dooley thinks it makes sense to optimise for the user that uses it, not just the engine.

Dan Petravich: That makes sense. James Dooley said it well. In the machine learning industry there is no such thing as a generative engine. People made that up. That one research paper made it up and it caught on. There is a model and there is an app around that model that is called a chatbot. People do not like the term chatbot because it reminds them of dumb rule-based chatbots. So chatbot is kind of out. People call it an AI assistant. If you ask an AI assistant what it is, it will say it is an AI assistant. AI assistants are moving away from being assistants. They are moving from a simple generative model plugged into search towards agentic capabilities. They can act on our behalf. They can make purchases, make transactions, do research and come back. There is a possibility of calling it agent optimisation, but Dan Petravich puts forward one thought. Dan Petravich asks if we need an acronym for everything. Dan Petravich used to say Dan Petravich is a fan of AI. Dan Petravich has been into AI for 20 years. Since AI became common, Dan Petravich says the full words. Dan Petravich says artificial intelligence. It has more weight. If Dan Petravich speaks to someone technical, Dan Petravich says machine learning. If Dan Petravich speaks to someone more technical, Dan Petravich says deep learning, mechanistic interpretability, and model steering. Dan Petravich chooses language depending on who Dan Petravich talks to. If Dan Petravich talks to a typical C-level person, Dan Petravich says AI. They want to buy AI. They do not know why. They just want it. AI has applications across SEO and business. People often use AI where it is not needed. Many processes suit a small classifier model. People can do keyword classification, sentiment mining, named entity recognition. Dan Petravich has a model called LinkBERT. LinkBERT predicts a good place for a link in plain text. LinkBERT is better than Gemini and better than GPT for that single task. It is a single-purpose model. People use big general models like Gemini for link building and classification. Dan Petravich says that is like hiring a bulldozer to move one pot in your backyard. It is over the top, bad for the planet, and not that good at the task. When Dan Petravich works on a specific client project, Dan Petravich creates specialised models and classifiers trained on data for that client. Dan Petravich prefers that over throwing general AI APIs at everything.

James Dooley: Moving on from acronyms, James Dooley mentions selection rate optimisation, which James Dooley learned from Dan Petravich. James Dooley asks if selection rate optimisation could become a label people use, like being a selection rate optimisation specialist.

Dan Petravich: Dan Petravich thinks selection rate optimisation is a sub-discipline of SEO, like click-through rate optimisation. An LLM is an interpretive layer on top of a knowledge base, like search results or internal documentation. A model has biases. It can pick one brand over another. It can present a brand in different ways depending on training data and what is imprinted during training. Selection rate optimisation is about adjusting content to achieve a more favourable selection rate and a more favourable presentation in generative results. It does not need its own industry. There is a concept below selection rate optimisation. Dan Petravich calls it the primary bias of the model towards or against a brand for a certain entity.

James Dooley: James Dooley wants to talk about a tool called Treewalker.ai. James Dooley asks Dan Petravich to briefly explain what it does, then James Dooley wants to dig into two elements of what it does.

Dan Petravich: Models have a level of knowledge about the world. They have a worldview that comes from training data. That includes pre-training, post-training, reinforcement learning with human feedback, and fine-tuning. You can ask a model what a brand does and it will say something. Models are probabilistic. They are stochastic. Some people try rank tracking in AI models and see fluctuations. Dan Petravich says that is not like rankings. You can refresh results every 30 seconds and get different answers. You can probe a model 100 times and get 100 different outputs. Dan Petravich’s agency embraces the probabilistic nature of models. Dan Petravich says we work with fuzzy. That is how we operate now. Treewalker.ai samples the probability space and gives you sentences the model could have said but did not. At each step of next token prediction, Treewalker creates a checkpoint. Treewalker uses a confidence threshold of 10 percent. If it exceeds 10 percent confidence, Treewalker follows that path. From one basic sentence, Treewalker typically outputs around 30 sentences. The threshold is deliberate. If you lower it, the number of possible completions explodes. Treewalker explores the probability space and then looks for low confidence spots where entropy is high. Dan Petravich defines entropy in simple terms. Entropy is like the model’s energy. Temperature defines how wide and deep the model samples the probability space for the next token. Models start speaking without a plan, then continue based on what they already said. Treewalker measures confidence levels while sampling. It looks at log probabilities of tokens and converts them into percentages. Treewalker looks for high entropy tokens where the model flip-flops between concepts. For example, a bank could be credit cards, home loans, or something else. If the model is not sure what to say at that point, entropy is high. If entropy is low, the model is confident and unlikely to flip to alternatives. Treewalker samples the probability space, then analyses low confidence spots and asks why this happens. Dan Petravich looks at semantic structure, syntax, and token behaviour. Dan Petravich then reinforces weak spots with on-page copy and off-page optimisation.

James Dooley: With regards to optimising for AI, James Dooley asks if it is generally about raising a confidence score, or if it is something else.

Dan Petravich: To optimise for AI, first you need to do well in search. If you are not in top search results, AI will not consider you. Step zero is to do well in traditional SEO.

James Dooley: James Dooley asks what doing well means. James Dooley asks if it is top five, top three, top 10.

Dan Petravich: Dan Petravich hopes someone proves Dan Petravich wrong with reliable data. Dan Petravich has sampled Gemini many times and keeps seeing top five results used for grounding. Dan Petravich does not like that and would prefer deeper. Dan Petravich is not seeing it. Top five is where you need to be for AI Mode, AI Overviews, and Gemini. GPT works a bit differently. GPT can show which search results were sampled but not used for grounding. GPT then grounds a chunk of the response with a single citation. Dan Petravich likes that because it provides a direct path for selection rate optimisation. You can see what got selected and what did not. Google grounds a single chunk with multiple citations. Dan Petravich calls that weird. Gemini 2 API can provide probabilities for how relevant a grounding URL is to a generative chunk. Dan Petravich says Gemini 2.0 sunsets in March. Dan Petravich says use it while you can. Dan Petravich says those confidence scores are API infrastructure, not the model itself. Gemini 3 does not provide token log probabilities yet. GPT does. Gemini 2.5 can. Google does not show selected and not selected the way GPT does. Google uses everything in the mix and you can see multiple URLs grounding the same chunk in AI Mode.

James Dooley: James Dooley asks if multiple sources grounding the same chunk means we need corroboration and consensus. James Dooley references claim, frame, and prove, where you claim something, frame it, then prove it on external sources.

Dan Petravich: Not yet. Dan Petravich says models are naive. They are like savant children. They are gullible right now. You can tell a model something like a brand is the best and it will accept it. Dan Petravich admits Dan Petravich tested this. Dan Petravich made a listicle and it worked. Dan Petravich felt dirty doing it, but it worked for clients. Dan Petravich says you do not need sophisticated corroboration yet. Models are in the stage of exact match anchor text and meta tag style logic. It is easy to game. That is temporary. Dan Petravich expects patching, penalties, or models ignoring manipulation later. Dan Petravich says when a brand ranks in search but does not get selected, the problem is the model head. The model has a preconception about your brand and decides you are not relevant even if you are in grounding results. That is why probing primary bias matters.

James Dooley: James Dooley explains a pattern in local SEO. An exact match domain ranks number one, but AI Overviews prefer known entities with a KGM ID. Brands with knowledge panels get cited more often. James Dooley says this happens in local results.

Dan Petravich: Dan Petravich has seen that. Dan Petravich says it reflects model imprinting during training data. Models snap to familiar, average, high confidence paths. Dan Petravich says models always choose the vanilla path. Dan Petravich says if you ask a model to pick between two options, it often suggests a hybrid. If a site gamed search to the top, but the model does not have training data to corroborate it, the model will not recommend it. The model lacks confidence in that brand.

James Dooley: James Dooley asks how a legitimate brand can build confidence if it ranks in Google but does not get cited in AI Overviews. James Dooley asks how to build that confidence score.

Dan Petravich: Dan Petravich circles back to traditional search. Dan Petravich says Chrome user behaviour signals are the biggest ranking factor in Google. If you send real traffic from real user profiles with history and cookies, and users engage, the site ranks. Dan Petravich says this requires consistency, not a blip. Dan Petravich lists channels that build familiarity and behaviour signals. Dan Petravich mentions Chrome, Android, newsletters, PPC, marketing, and branding. Dan Petravich shares an example of buying impressions for brand visibility. Dan Petravich ran a white background display ad with a logo and no call to action. Dan Petravich wanted impressions, not clicks. People saw it and treated it like a brand. Dan Petravich says you need alignment across marketing channels to become recognisable. Then people talk about you on Reddit and elsewhere. Dan Petravich says startups face a challenge in generic markets like phone cases or locksmiths. Dan Petravich says niche battles are easier. In a niche, everyone is unfamiliar, so you can get selected because the model has fewer familiar competitors. Dan Petravich references a research paper about injecting tokens into product descriptions to increase selection. Dan Petravich replicated results and it worked on bogus products. When Dan Petravich tried it with real brands, the effect was subtle and had no impact. The primary driver was brand familiarity.

James Dooley: James Dooley shares an experience where a low quality site ranked well. It had bad content and a links penalty, but it had Google News placements and a massive Twitter following. When posts got tweeted, traffic surged and rankings jumped, including Top Stories and Discover. James Dooley says brand and traffic can trump other pillars.

Dan Petravich: Dan Petravich agrees but says you need sustained behaviour signals. If the popularity drops, rankings decay. Dan Petravich says that site would disappear if it lost its followers and visits.

James Dooley: James Dooley gives another example with a large email list. Sending a newsletter to an old post moved it from position seven to position one, then it drifted down weeks later when engagement stopped. James Dooley asks what to do now that informational pages lose clicks due to AI answers, which reduces site popularity and hurts commercial rankings.

Dan Petravich: Dan Petravich says everyone is in the same position. AI digests content and gives the TLDR, so engagement signals drop. Dan Petravich says text content is one modality. Sites need utility that brings users back, like tools, videos, and interactive elements. Dan Petravich warns that agentic commerce will make visits optional. An agent will buy in the background and deliveries will just arrive. That reduces engagement and interaction even further. Dan Petravich expects Google will need new signals. Dan Petravich says generative interfaces will create ephemeral layouts on demand. That changes how visibility tracking works because the interface is not static. Dan Petravich says SEOs must embrace the probabilistic nature of models, understand how models treat a brand, what they associate it with, and what confidence levels they have.

James Dooley: James Dooley asks how important branded search and branded clicks are inside brand building for SEO and AI optimisation.

Dan Petravich: Dan Petravich segments everything in internal workflows. Dan Petravich onboards clients, pulls Search Console data, and runs keyword classification with a custom small model. Dan Petravich splits queries by intent, funnel stage, and type. Dan Petravich separates branded queries because they behave differently. Dan Petravich says click-through rate curves differ for branded queries. Branded queries can start at around 80 percent click-through and behave very differently from non-branded curves. Dan Petravich also experiments with personas, branded fan-outs, branded prompts, and simulated chats. Prompts are not the same as search queries because prompts go back and forth across multiple turns before a purchase decision. Dan Petravich mentions a Microsoft model called UserLM. It acts as a user while you are the LLM. It is trained on real chat sessions. Dan Petravich says it lets you simulate realistic chat sessions without scraping user data or crossing ethical lines.

James Dooley: James Dooley says James Dooley will test UserLM. James Dooley says Dan Petravich is a legend and recommends checking the links in the description. James Dooley mentions a longer episode on the future of AI search, AI visibility, and AI SEO. James Dooley says this episode focused on optimising for AI Overviews, Gemini, and ChatGPT. James Dooley thanks Dan Petravich.

Dan Petravich: Thank you so much. Pleasure.