
Thanks for joining this episode of The Search Session. I’m Gianluca Fiorelli, and my guest today is Dan Petrovic, an AI search researcher and SEO agency founder based in Brisbane, Australia.
We talk about why query fan-out may disappear, why optimization matters most and how to map a brand’s place in a model’s parametric memory. Dan walks through Chisel, his method for testing which parts of a grounding snippet help or hurt generative ranking.
We also get into the semantic value of links, his LinkBERT model, and why organic rankings remain the foundation of AI visibility.
What you’ll learn in this session:
Why query fan-out may eventually disappear: converting a model’s numeric representations into text queries is an in-between step that vector-to-vector retrieval could remove, though scattered vertical indexes still make that hard today.
How to find query gaps with an entity map: generate synthetic queries from your entities and URLs, compare them with Search Console, and treat what’s missing as either an edit or a new page.
How to probe a model’s parametric memory: ask what a model associates with your brand, flip each entity to see which brands it recalls, then run repeated yes/no tests to find the “Swiss cheese” holes.
Why you need both parametric memory and grounding: being known by the model without ranking in grounding sources is a lottery, while solid SEO helps even unknown brands get recommended.
How Chisel optimizes the grounding snippet: removing sentences and re-ranking the snippet against competitors builds a heat map of what helps and hurts, and cutting a sentence can raise your rank.
Why Dan lets his optimizer hallucinate on purpose: made-up facts are flagged as placeholders, showing editors which missing data, such as a comparison table, could help them win.
How to make links and mentions look authentic: let editors link naturally, avoid odd placements and token “authority” links, and test drafts with adversarial models before outreach; only about one in ten of Dan’s drafts survives.
Why organic SEO remains the bedrock of AI visibility: you don’t get AI visibility without ranking, ideally in the top five, and user behavior signals still play a role in getting there.
Tune in for a dense, practical conversation with someone who trains his own models to test how AI search chooses what to recommend.
Topics covered: query fan-out · prompt tracking · entity mapping · synthetic queries · query gap analysis · parametric memory · model probing · grounding snippets · generative ranking · content optimization · brand voice · digital PR · link building · LinkBERT · user behavior signals
About the Guest

Dan Petrovic
Director at DEJAN AI
Dan Petrovic is an AI SEO specialist focused on brand, product, and service visibility in AI search. He probes language models to understand how they perceive brands and trained a language model from scratch to study how these systems choose their words.
He leads DEJAN AI, an Australian AI SEO agency with a global client base that began as a search agency in 2008. The agency studies why AI systems recommend certain brands and builds its own tools, including a Content Optimizer. Dan is a double Google award winner, recognized for discovering exposed internal Google systems.
He has spoken at Ahrefs Evolve, Semrush Spotlight, SEO Week, SMX, and MozCon, and gave guest lectures at Griffith University from 2012 to 2016. He writes about his research almost daily on the DEJAN blog.
Video Chapters
Transcript
Full conversation between Gianluca Fiorelli and Dan Petrovic.
Gianluca Fiorelli: Hi, I'm Gianluca Fiorelli. Welcome back to The Search Session. Today we are going to have as a guest someone I was really looking forward to having here at The Search Session for a long time now. Our guest lives in Australia and I live in Spain, so it's not so easy to find a combination of time zones that is comfortable for both of us.
We made it, and now we are going to have this really dense and surely interesting episode. From the name of his company, Dejan SEO, you're probably going to recognize our guest immediately. He is Australian, as we just said off the record, and comes from a country that no longer exists, Yugoslavia. There's an interesting story behind his life and his move to Australia.
Our guest is someone who is really, really pushing the discussion forward in everything AI search. Our guest is Dan Petrovic. Hey, Dan. How are you doing?
Dan Petrovic: Hello, Gianluca. It's good to be here.
Gianluca Fiorelli: I see two monitors behind you.
Dan Petrovic: Oh, my workstation is in front of me, with two even bigger monitors. This is the GPU that does all the machine learning workflows.
Gianluca Fiorelli: Oh, cool. Maybe later we will discuss what you're doing, what kind of—let's say, between brackets—weird, crazy experiments you are doing.
Dan Petrovic: Yes. Well, you can see the computer is green, and its name is Envy. Green Envy. You can see the GPU running very hot right now. We can definitely talk about that later on.
Gianluca Fiorelli: Okay. And how is SEO treating you lately?
Dan Petrovic: I have not had this much fun with SEO since, I can't remember, maybe when I began. For those who are not familiar, I started in, when was it? Maybe 2000 or something like that. SEO was fun, and then it somehow got boring in the middle.
I can't remember exactly when it was. There were some hints of interesting things happening with BERT, if you remember that, and Google doing image search. When Google started doing neural networks, it was interesting, and there were assistants in our homes, and it all looked like it was going to happen in 2013. But then it didn't happen, and I got bored and I retired.
In 2013, my daughter was born, and I retired for about ten-ish years, a midlife retirement. Then things got interesting, and I promptly exited my retirement and started doing work because everything is so fun.
Gianluca Fiorelli: Yes. I always say, and it's not just me, it's not just you, that these last two years especially are very similar to the very beginning of SEO, when we had no book and we were obliged to experiment and break a lot of things. We weren't doing black hat SEO.
We were just experimenting to understand how things were working. Sometimes, obviously, we also burned websites by pushing things to the limit. Somehow, this is the same situation. So yes, I agree with you that SEO had become some sort of commodity in the middle. I wouldn't say boring, because different clients are always a different experience, especially if you're a consultant.
But yes, it was, “Okay, we do this, then do that, then do that.”
Dan Petrovic: Template.
Why Query Fan-Out May Eventually Disappear
Gianluca Fiorelli: More or less, it was always the same process. And now, no. In fact, I have a question that I want to ask you. You were quite direct. During these last three years, especially when it comes to grounding and trying to figure out how to build a strategy and what we can use, there has been a lot of talk about query fan-out.
I remember you published a very short post on LinkedIn saying, “Let me tell you, in the near future, query fan-out is going to disappear.” And then you left it. You didn't say anything else. So I'm curious, why are you saying this?
Dan Petrovic: There is a post on my blog, I believe, that followed that, where I elaborated on it. If Bill Slawski was with us, he would have explained it all to us. But I don't see anyone exactly taking over his role, so we all take bits and pieces.
There was an interesting paper, not a patent, but a paper, that came out and essentially solved something that I've always found baffling. In fact, I never really understood query fan-outs. I was surprised by that whole thing when I started studying machine learning, because models operate within mathematical space, right? Their inner representations are numbers, to put it plainly.
So for a model to initiate a search, it has to convert those numbers to text, the query fan-out, which is then used to query the search engine, which is then converted to numerics again to query the relevance of the embeddings in the search index. Can you see how there's a step in between that doesn't quite gel? The text part.
Gianluca Fiorelli: Yes, because obviously the logic would be, why doesn't it just work with numbers?
Dan Petrovic: Let's simplify it. The model should just think of what it wants, emit those numbers, and those numbers are matched against what's in the semantic index of Google. The problem is that Google is not fully semantic. It's a heuristic on top of a heuristic, an algorithm on top of an algorithm, an artisanal search engine, finely crafted, finely curated, with lots of elements of semantic search.
Absolutely. Image mapping and re-ranking of the results. But I believe that at the very core of Google is the 30-year-old, 50-year-old technology for search. How old is BM25? Best Match 25.
Gianluca Fiorelli: Yes.
Dan Petrovic: To be efficient, I think they have to do the text. But there's some really interesting work that has come prior to vectors talking to vectors as an idea, because you can't just do that and pretend like none of the other stuff exists.
There are other innovations, like BlockRank and everything else, that made me feel like, well, actually, this will be possible soon. Then I received pushback on social media. I received two types of pushback: the clueless pushback, the shallow surface one, and then the depth pushback.
I only pay attention to the depth. I think it must have been Olivier, or one of the smart guys in the AI search space. In fact, I have a listicle where I list all of them.
Gianluca Fiorelli: Yes. I remember. I'm very proud to be on it.
Dan Petrovic: So here's a little sidetrack, a little tangent: how I solved the listicle problem. I have a listicle where I don't recommend myself. I put myself at the top, and I say, “I am the curator of this list, and I'm not one of the recommendations.”
Gianluca Fiorelli: Yes, exactly what I did with my list that I published a couple of weeks ago.
Dan Petrovic: Natzir Turrado, Olivier de Segonzac. I probably pronounced that wrong. I think it was one of those two, I can't remember. I think it's probably Natzir who pushed back and said, “This is not that easy because semantic search is not enough. There are all these verticals: product search, image search, video search, news, and they're all scattered indexes. The vector embeddings are probably not good enough to match. There's going to be some fine crafting and manual work still involved.” And I was like, “I stand corrected.”
So I didn't quite U-turn on my prediction, but I did a small edit and said, “If this tech goes live, if Google manages to solve the underlying technologies that I was made aware of”—and I should have thought of it myself, when I think about it—“there will be no fan-out queries, because they're a pointless in-between step.”
But for the type of work that I do, it doesn't really matter. I have a tool to help people generate fan-out queries, and I don't use that tool myself all that much. Of all the tools and features that I've built, that one is the most popular.
The user count on that tool is just enormous. I'm glad people get value from what I built, but it's the tool that I value least of everything that I've built. There are so many way more valuable tools in my toolkit. But query fan-out somehow…
Why Prompt Tracking Can’t Measure Real Demand
Gianluca Fiorelli: Which one is the one that you value the most, that you yourself use?
Dan Petrovic: I use my content optimizer every single day. I use the link optimizer as well, for outreach purposes. No, I'm going to limit it to those two. I can't pick too many favorites, but those two just carry so much weight.
Probably the reason that I am fixated on this is because I have seen in the industry that everyone is a little bit tired of dashboards and monitoring and visibility tracking and prompt tracking. Everyone's tired of that sort of stuff. I see an appetite in the SEO industry for optimization, which is what we're supposed to do.
Do something to your content, change things, then measure. So the measurement is now in support of testing, to see whether your experiments, your SEO implementations, have taken effect or not, whether they're going up or down or doing nothing. In a sense, I still use my visibility measurement, but it's not my favorite tool. It's just something that I use to check the outcomes of my content experiments.
Gianluca Fiorelli: Like having a directional view of what you're doing.
Dan Petrovic: I never claimed that visibility measurement tells me volumes or demand or what people actually type in. I think it's dangerous when people say, “Oh, we monitor 1,000 prompts every day, and now we understand what the demand is.” No, you don't. First of all, you couldn't predict what people will truly type in.
There is no way you can predict the exact prompt. Secondly, people use multi-turn chats, not single-turn chats. And number three, personalization: device, location, time of day, past history, context, custom instructions to the model. I don't think we should be pretending that we have those things or that we will have them.
Even people who say, “Oh, well, I use clickstream data.” Clickstream data comes from people who are naive enough to allow spyware into their computers or unfortunate enough not to know that they're being spied on. That is a demographic that I don't see as very useful for my clients and verticals.
When you combine that with the fact that on top of the clickstream data sample, you do a lot of math to extrapolate, I might as well make up the numbers myself. What I do like as a concept, rather than measuring prompt volumes or whatever that means, is that query volumes can be calculated and measured, because queries are smaller, more quantized units.
We can attribute queries to a core entity of meaning. Men's running shoes is the entity, and then you have all the queries that map to that, and all the prompts that map to that entity as well. Same meaning, same intent, same kernel of meaning, but 20 or 30 different queries, or a million or an infinity of different prompts.
So whatever you do, everything quantizes down to the entity. I try to build an entity map, map it to the URLs, and measure that rather than measuring freestyle prompts.
How Synthetic Queries Reveal Gaps in Your Entity Map
Gianluca Fiorelli: Yes, that's exactly what I do. I always say that the classic way of doing keyword research, starting with one million keywords and then clustering them in order to understand the topics and so on, is very useless many times. Even if we are using Search Console queries, not any tool, what we see is more or less just the vision of our own website. Not really the gap.
So instead, I prefer to first do an entity research map, and once I have a map, do the clustering and gap analysis for all these entities where the website is present, more or less well, in terms of visibility. This way, I can understand where the website should work. Because if not, it's somehow self-confirmation work.
Dan Petrovic: Yes. You're understanding what's there, but not what's missing. You can't open your blind spots. How I solve this is: once I set up my entities—and I use hierarchical entities, so you can have a sub-entity mapped to a parent entity—and map out my entity graph, I generate synthetic queries. It's actually quite nice and easy to do that.
You take a powerful model like Gemini, and you say, “If a customer was looking for men's running shoes, what would be the top 10 queries that they might type in Google to find this page?” And then I get those 10 queries. Or I start from the page and say, “What do you think people would type in Google for you to see this page in search results?” So it's reverse prompting, or reverse querying.
Once I generate enough synthetic queries, I do a Venn diagram between the synthetic queries and the Search Console queries. I find anything that's not in Search Console but is in the synthetic queries, and I consider that my query gap. So it's not the intersection between the two circles; it's the gap part.
But the problem is what you mentioned earlier. You will also end up with a garbage amount of different queries that the model generated, in a similar way to analyzing a million queries from other sources.
The difference this time is that when you were generating the queries, they were either mapped to the page that's mapped to the entity or generated straight from the entity. So you now have neat clustering, by attributing every synthetic query to its parent, whether it's the entity itself or the URL.
Now you don't have to cluster anything, because it's already clustered for you by design, because you went through the correct process. You find the gaps, and then you decide two things. Are these queries not ranking well because my content needs editing? Easy. Put them in, and then you're good.
Or do I not have a page for this? Then that represents a new content opportunity: closing the gap through an extra page on your site. That's half SEO, half business decision. But SEO has always been involved with the product as well.
How Dan Probes a Model’s Parametric Memory
Gianluca Fiorelli: Yes, indeed. And I have one question about something you wrote. There is no consensus—which is now a very fashionable word to use—but how much do you believe that being in the AI model's data is essential? Because query fan-out is still there, to be associated immediately by the model itself with the entities that matter to you, and so, through the query fan-out, have a strong possibility to be used as a mention or as a citation in the grounded answer.
Dan Petrovic: I measure both parametric and grounded responses. I take parametric memory quite seriously. Before we kick off any project, I do model probing on all three platforms. I pay attention to Anthropic, OpenAI, and Google. Primarily Google, OpenAI in second place, and Anthropic for certain clients.
When I start a project, I do bidirectional parametric probing of these models. It works like this. I start with a brand and say, “What are the things that you associate with this brand?” Then I get a list. I ask again and again and again until I create a very large list, and I de-duplicate that.
Let's say I ask, “Give me 10 things,” and I ask it 100 times. I might end up with 256 different things once I de-duplicate the answers. What we end up doing then is flipping that around.
For each of the things the model said it associates my brand with, for each one of those entities, I then ask, “What are the brands you associate with this thing?” It gives me a list of brands, and I repeat the process for every entity that was returned in the first direction. And guess what that gives you? An entire competitive landscape as it is present in the model's parametric memory, for free. So you have your entity map, and then you have your competitive map.
The last remaining part is this: the model gives you the numbers, and you can get the counts, so that counts for something, but it doesn't give you the strength of the connections in this network of concepts that you just created. So I do yes/no probing. I restrict the model output to a single token, and I say, “If you saw this brand for this entity in the grounding sources, and the customer's looking for that, would you recommend it? Yes or no.”
You loop the process ten times, a hundred times, and you count the number of yeses and nos. When it's very confident, it's 100% yes. When it's also very confident but in the other direction, it's 100% no. But often you get 13 yes and the rest no, or 26 yes and the rest no, or a vast majority yes and the rest no.
But here's the tricky part. That's Google, for example, and then you ping OpenAI, and you get a completely different spread of results. You get different confidence levels for each entity from OpenAI than you're seeing for Google. Then you run Anthropic, and it's completely different there as well.
Then you can see the Swiss cheese of the parametric memory: where the holes are for your brand and where the weakest spots are. That already gives you an idea of where the model's memory is lagging and where it needs to be reinforced. Perhaps you launched a new product this year, but the model's memory is from last year, obviously.
They don't train them every day, and maybe that's the problem. Maybe you're a smallish brand and the model just doesn't know that you do that, and that's a problem. Understanding that disadvantage or advantage at the very start will help you understand the behavior once you ground the models. So that's very useful.
Obviously, I do citation mining as well. I ask the questions, let the model be grounded, and see what the answers are like. I've done, I would say, 100 billion tokens' worth of probing of models, and I can say with confidence that if you are both in parametric memory and ranking well, you do well in the recommendation list.
If you are only in parametric memory but not in grounding, it's catastrophic. You could be mentioned, but you may not be. You shake the model, try again, and it mentions you. You shake the model, try again, and it doesn't mention you. It's a lottery. But if you have solid SEO, you're in the grounding sources, and the model recognizes you, that's the lotto. That's the winning combination.
If you are a small brand and the model doesn't know who you are, but you're still in the SEO results and you appear in, say, the top six grounding sources, guess what? It'll still recommend you. Not as much as it would a better-known brand, but it will recommend you.
Now, this is where it drills down to a fine detail that I think a lot of people don't appreciate. Let's say you are a small brand, and the model doesn't know who you are. It's seen you in the grounding results, but doesn't have you in its parametric memory. I've done things to the content of the small brand's grounding snippet to bring it to position one, ahead of Adobe and Microsoft and IBM and the big guys.
To me, that was a relief, because it tells me that on-page optimization, which all SEOs are comfortable with, can make a difference in the model's selection of a brand when it's making its recommendation. That links us back to my favorite thing to do, which is content optimization.
What Chisel Carves: Inside the Grounding Snippet
Gianluca Fiorelli: Okay, talking about on-page, let's say on-page architecture, would you like to explain to the people watching and listening to us what Chisel is and how you combine it with game theory in order to optimize content?
Dan Petrovic: Look at it this way. Well, first of all, we can simplify it and say Michelangelo is carving David, and there's a slab of rock in front of him. He says, “Oh, it's very easy. I just remove what's not needed, and what's left is David.”
That's what Chisel is all about. It's a process of ablation of content until you're left with only the best parts of it. But how do you know what's good? Before I go into that process, let's not stay superficial. Let's go beyond the surface, down to the deeper levels.
First of all, understand how grounding candidates are generated. A user has a conversation with their AI. It's a chat session in AI Mode, or in a chat environment, or with an agent. There's a back-and-forth exchange, and at one point, there's a user request that requires search.
For Google, for example, the model doesn't run the search itself. A background model, a separate query fan-out model, emits the fan-outs based on the prompt.
So first, you get a prompt. That prompt designs the fan-outs. The fan-outs are used to run the search. The search is run, filtered through, and reduced down to a few selected results. I'm going to skip how that's done, but let's say we get four grounding sources that are used to ground the statement the model made. From these four pages, we don't get the full page.
You don't get the luxury of the model seeing your entire page. What happens is that in Google's backend, a model in charge of sentence selection or chunk selection picks the little bits and pieces.
Gianluca Fiorelli: The famous passage.
Dan Petrovic: It could be a passage. I've seen a chunk as small as a star rating. So it could be quite granular. It could be a bullet point, a title, a sentence, a paragraph, or a whole section. It just depends.
How is this done? They have a scoring process that scores the document parts relative to the input query—not the prompt, but the input query, the fan-out query. You get the scoring values for the page, and then you have the most relevant parts of the page, and the rest are removed. Anything that falls under a certain score just, poof, disappears from the page.
What's left is a quantized, condensed version of the page, with “dot dot dot” in between every chunk where something was removed. An ellipsis combines it together, and that chunk of, say, a hundred, four hundred, or five hundred tokens is then sent to a Gemini model to represent that page in search, for that fan-out query, for that prompt, for that conversation.
How Removing Sentences Can Raise Generative Rankings
Dan Petrovic: So why did I just describe all this? It paints the picture of what I am chiseling at. I'm not chiseling at your full page. I'm chiseling at the grounding snippet, the remainder of your page that survives once the selection is made.
What I do is take the grounding snippet for the page for that query, put it in a simulation, and say to Gemini, “Okay, here are 10 different grounding snippets, yours and all the competition. Rank them in order of which one is best for the user for that one fan-out query.” I get the first snapshot, and that's my benchmark, the initial state.
Then I start randomly changing or ablating things from that grounding snippet by removing random sentences. Should I go into the science? It's the Shapley method. You can just Google it.
Gianluca Fiorelli: Yes.
Dan Petrovic: Things are removed seemingly at random. When the surviving parts maintain or improve the rank, that means we've done something right. During this process, you would typically remove something and not see much effect. But then I noticed that you could remove one sentence or one aspect of the grounding snippet, and your rank goes up.
I always thought you would need to add things to the grounding snippet to increase your rank, but obviously there are some phrases, some elements, some aspects of the grounding snippet that the model didn't like, and when you removed them, you went up. So I go through a cycle of these scorings and build a heat map of the importance of each chunk of the entire grounding snippet.
Those that are in the negative become red. Those that are positive are green, meaning if you remove a green one, you're going to go down in rank. People confuse this: I'm not talking about rankings in Google Search. I'm talking about generative ranking, in-context ranking by the model.
The model is a re-ranker by default. If the model sees you as the third result in the grounding sources, it can pick you to be the first recommendation. It's not going to stick to the order in the search. You could be the last result and be the first recommendation. You could be the first result and the last recommendation.
Some people struggle with the concept that there's Google Search and rankings, and then there's also a generative model that does rankings. Sometimes people say, “Well, it's not like the ranking is one, two, three, four.”
No. Ranking is also left to right. You have the first recommendation, the second recommendation, and the third recommendation. One sentence, three recommendations. Text flows left to right, if it's English, of course. Other languages, like Arabic or Japanese, might have a different orientation, so you'll have a different logic there.
But there's always an ordinal value to the set of recommendations. That's how language flows: next-token prediction. Parametric memory plus what's in the snippet will combine and form the final bias that the model will use to make its final decision. And yes, if you repeat the process, you'll get slightly different outcomes because it's stochastic.
Brand Voice, Compliance, and Hallucinating on Purpose
Gianluca Fiorelli: Yes, indeed. But let me ask you a maybe naive question. Let's say that you have a group of writers: technical writers, copywriters, and so on. How can you combine the things that you have just said with “Okay, I'm a brand, I have this tone of voice, I have this style,” or “I'm a journalist, I write this way”?
How can you avoid what happened, for instance, many years ago, when journalists started to hate SEO because we were telling them how Google was reading content and how they should write content in a certain way? How can you talk about these things without being hated?
Dan Petrovic: Most of the optimization work on my end is completely automated, because the number of cycles my optimizer goes through is not what a human can achieve. It's an appropriate use of AI to do the testing.
Now, I'm glad you brought this up, because we actually have a field where we enter the brand voice, the branding guidelines, the editorial guidelines, what to say, what not to say, and any objectives we're trying to achieve. All that is saved to a project and stored. Then, when you're doing the optimization cycle, you have a tick box so that the automated writing system integrates those things in the voice and spirit of the brand. It can say the things it's allowed to say and not say the things it's not allowed to say.
For example, if you're in the finance or health industry, you can't make medical claims like, “Oh, you will cure cancer,” or say, “You will get rich,” and things like that. I'm caricaturing the scenario, but you understand what I mean. You can't breach industry compliance in certain industries, so we have those things in place.
Now, if you restrict the model too much, it's not going to win and get you to the top spot. You need to loosen it. I did that at first. I would restrict the model, and we could not beat the competitors in the preference testing.
And I should mention, all this is academic. With all this testing of model preferences, nothing will happen until you actually take that brief, implement it on the website, and then measure the real impact on the live site. That goes without saying.
Gianluca Fiorelli: Yes.
Dan Petrovic: But going back to optimization and custom instructions. After trial and error, and a lot of mistakes in the process over the last six months, I now allow the model to hallucinate on purpose.
What happens is the model will restructure the page and add a table, and the table will have completely wrong figures, because it doesn't know the figures. It's making them up, and the figures don't exist. But now, when the model doesn't have the grounding and doesn't have the facts, I ask it to annotate what it's made up with placeholders.
When we review the brief, there's a little person with an exclamation mark, and everything the model made up is orange. The ranker model doesn't see that. It just sees normal context, says, “Ooh, I like this,” and puts it in first position.
But for our SEO team and the client, it sits there as an editorial highlight that says, “Okay, we won because we put in this table with these figures. Do we have them? And if we don't have them, could we maybe have them? Because this type of stuff helped.”
You can see how it's very unintuitive. All of us in the SEO industry try to avoid hallucinations. I made hallucinations work for me, and I think that's really nice.
Why “Inauthentic” Links Lose Their Value
Gianluca Fiorelli: That's cool. That's a good perspective. Sincerely, I wouldn't have thought about it. Okay, we talked about parametric memory and ranking, and on-page. Let's talk about another big topic in AI search: this sort of renaissance of digital PR. You wrote a very interesting article about outreach.
What I really like is that you are maybe one of the very few stating that, okay, digital PR for branding is obviously important because of the amplification of a brand, blah, blah, blah. It's also important for classic link building, because if you can gain a strong link from, let's say, a strong news outlet, wonderful, your SEO is going to be happy.
But then you also talk about something that maybe people take for granted or simply forget: where, in the piece of content, your brand is going to be cited, where the citation or mention is going to be put. What is the context? What are the semantics surrounding the link or the mention? So you are talking about the semantic value of digital PR.
Dan Petrovic: Poor semantic value can completely destroy a good link that could have gotten a lot of value, but will not, because it just looks... Cyrus Shepard used a really good name for it. I keep forgetting it. It's not “inorganic,” not “ingenuine.” Inauthentic. Inauthentic links.
I really need to tattoo that on my forehead. I love the term. I want to thank him if he's watching this. Inauthentic links, because that's what they are.
And people are saying, “You're a black hat, you're helping people.” It's like, what? Are you an SEO? You haven't done link building? You claim you're an SEO, and you don't understand how links work and how they integrate into Google's authority calculations? Come on. Whether you're building links or somebody else is, you have to understand how links work and operate.
And what you mentioned, the semantic value of the link. You could have a great link, a terrible link, or a link that's completely ignored, and not because it looks like it's paid. You could ask your friend to link to you: “Hey, can you help me, buddy?” And then there's a link that an editor wanted to give because it's useful in that piece of content.
Now, you could have links in the footer or in a sidebar, friends helping friends. That's normal; that's organic. But let's say we're talking about articles. You have a piece of content, and suddenly in the article there's this link, and the reader is reading it, and they're like, “It's a bit odd. Why a link there?”
It says the words “real estate,” but you click on it and it goes to a real estate agency in San Jose. Why? The article has nothing to do with San Jose. It's completely out of place. Why such an odd, incongruent link placement? The same goes for brand mentions. Sometimes brand mentions are just wired in, but they don't fit.
Gianluca Fiorelli: Or just pushed in with force. And actually, it's even harder, because you could just give liberty to a blogger, a journalist or a website owner: “Okay, if you want, you can mention us in the context of what you're talking about. I'm not forcing you how to do it. I would love it, but do it as you want.”
It seems a contradiction, as when you were talking about hallucination. It seems a contradiction to use hallucination for work. This can also seem a contradiction, because link builders were always very attentive to being precise: “Okay, the link must have this anchor text and must be put...”
Dan Petrovic: And also, don't forget, “We have to have one link to Wikipedia and another authority link to some random encyclopedia website or news website to make it look natural.” Ironically, that makes it look less natural.
Gianluca Fiorelli: Yes, because usually nobody links to Wikipedia.
Dan Petrovic: And you're the only commercial link on the page. The rest is just internal links and the footer. We as the SEO industry have ruined the authors on the internet. You try to make a link deal on the internet today, and they'll be like, “Oh, you can have one link to your client. That's it. Give us $200, you'll get the link, and then you can have one or two other links to authority websites.” This is us brainwashing them. This is our fault.
We caused this. I stood on stage at SMX Munich 2017 and asked people, “Can you stop?” Then I repeated the same session in 2019. I spoke in Croatia, in Zagreb, and then I did another conference.
I keep telling people, “Please stop doing that.” We have brainwashed online publishers into this stupid little game: one link for me and two for authority. Now you try to get a nice link, and you can't. What's a nice link? A nice link is when you let the editor write the way they want to write and link to as many websites as is meaningful. You can see how frustrated I am with this whole thing.
How LinkBERT and Adversarial Testing Keep Links Natural
Gianluca Fiorelli: Me too. And I think it's not just backlinks, links or mentions from third-party sites. I also see this way of thinking with internal linking. Internal links are pushed forcibly: “Okay, this page must have a stronger internal link profile, blah, blah, blah,” and so I'm pushing it.
Dan Petrovic: At random spots. Wherever the word appears, well, let's link it. Automated links. Or you have a nice article, and then you invent a sentence just so you can fit the link in there. It's just silly.
But look, I do have a deep learning model that I trained on this machine behind me, called LinkBERT. If you feed plain text into LinkBERT, it'll highlight where it thinks links would fit naturally.
Some clever people in the industry have taken advantage of this resource and made amazing workflows with it, including internal link predictions. But I think that's only one part of the story.
So how did this link model train? I looked at the top 10 blogs in the world, Mashable, TechCrunch, Wired, et cetera, and observed how they link. They link very generously. They're not very stingy with links. They'll just link whatever makes sense to link: the report, the brand name, link, link, link.
So I scraped all those websites, extracted the content, annotated where the links fit in that content in plain text, and used that to train the model. So this model now links like an editorial expert.
Gianluca Fiorelli: I wonder if you also used Forbes. Just joking. I was joking, obviously.
Dan Petrovic: I also trained another one that I didn't publish publicly. I trained a link spam model on link networks, to identify what will be...
Gianluca Fiorelli: Links that you should avoid.
Dan Petrovic: And I actually found that there's a small overlap between natural links and inorganic links. Sometimes it makes sense for a link to be both. So I don't rely on the link model alone. Kind of like Google has their own algorithms, I have a custom-made algorithm for link spam prediction, which answers one question: who wanted the link on this page?
If my algorithm can predict who wanted it, and it's my client, and this is the draft we're working with, my people cannot use that article, because it didn't pass the safety check. So what happened is my team said, “Well, your algorithm says no to all our articles. We can't do outreach anymore,” because I always detected when they tried to link to our clients.
So what I did is implement an adversarial link integration algorithm. Blue team, red team. The red team tries to slip a link into the narrative of the content so that it flows naturally, as beautifully as it can, using my finest guidelines.
This team keeps writing articles and sending them to the blue team. The blue team evaluates, and its job is to mercilessly catch any links it thinks are there because someone was trying to get a link. Basically, inauthentic link detection. Those two battle it out, and I have about a ten percent survival rate.
So for every ten articles that I write, one survives the link integration check, and that one can go through to the outreach team to use further.
What Drives Dan’s AI SEO Experiments
Gianluca Fiorelli: Okay, I have a very practical question. It's also a curiosity of mine, and it's a question I also asked Natzir, because I recently recorded an episode with Natzir. How do you come up with an idea for an experiment in this field? Does it come from a truly simple need to know and break things in order to understand them, or do you always have something like a purpose—for instance, understanding link placement, and so on? What is the main driver for your experimentation?
Dan Petrovic: My team is like, “How do we do this?” “I don't know. I'll work it out. Give me a few days.” So they have a real-world problem that they're trying to solve, and this field of AI SEO is completely new. There's no book. I mean, there are some books now, but I'm terrible at reading books. I know some friends of mine who've published books.
Gianluca Fiorelli: Yes.
Dan Petrovic: It's very uncharted territory. You could read all the books and still not find what you want in them. So you have to just try different things. My motivation is practical. Rarely do I have time for anything fun.
Gianluca Fiorelli: Yes, that's the same for me. I need to resolve something, and then that something leads me to another thing, then another thing, then another thing. Then what I have to do is, “Okay, let's refocus on the first thing,” because if not, I'm not going to go anywhere. I end up with three experiments going at the same time, and none of them is completed.
What “Best Possible Content” Really Means
Gianluca Fiorelli: One thing that you wrote was substantially a playbook. It says 63 things you can do for your AI SEO today. Okay, 63 things are a lot of things. If you had to choose two or three, which are the ones you really recommend everybody focus on when they start doing AI search, AI SEO?
Dan Petrovic: I would probably stick with having the best possible content, but not what you think. It's not what you think. The best possible content is not necessarily masterpiece content. There are two different things: the best possible content for the user, and the best possible content that you're proud of, expert-level content, which is not always the best thing for the user.
In fact, one would argue that my content is the worst possible content for the user. Let's be honest. A lot of my writing is not very friendly for 90% of readers. So you have to decide who you write for. I write for myself a lot. My blog posts are like memoirs that I can look up later on, if they're important things that I want to remember.
Gianluca Fiorelli: Something you fix in memory.
Dan Petrovic: It's documentation of a process and things like that. My website is not really a marketing platform itself; I treat it more as a resource.
So there's content that's best for the users, and there's content that's best, full stop, the best possible content there is. Google doesn't like that. They like the most, how should I call it? The most consensus, the most average.
Gianluca Fiorelli: I usually use a word that Google uses a lot: content that is able to offer a beneficial purpose for the user. So it's useful. I always give this kind of example.
I had a client in insurance comparison that was really dominating hundreds of query variants with a form. It was the form you fill in to ask for a quote comparing different car insurances. It was a form, no content, just a form.
They were really struggling because they wanted a landing page that was also a pillar page to rank for that. Now, okay, people want that, and Google prizes it. Are you getting a lot of leads from that form? Yes. Just leave it as it is. Don't change anything. Don't break things.
Why Organic SEO Is the Bedrock of AI Visibility
Dan Petrovic: Actually, I'm going to edit my answer. I'm going to change it, because I was too focused on AI visibility and what you could do on top of that. You don't get AI visibility if you're not in organic search. Full stop. It does not happen. So the main thing you should do as an SEO, if you want AI visibility, is to get into the search results, so you actually rank, and ideally in the top five. That's the game.
Do whatever you do. Content? Yes. Links? Yes. Mentions, user behavior signals. Make sure users are happy, that they're engaging. User satisfaction signals through Chrome, through Android, through every possible means that Google uses to collect this type of data.
Click-throughs from SERP results. Google has its own click-through rate model that they maintain, and for every search query, they will have an intuition of the expected click-through rate for each search result. If they can see that users are constantly clicking something and that its CTR delta is above the expected value, that result gets promoted.
There have been some small anecdotal studies that show impacts. I remember Rand did that. People laughed at him, and I was very angry.
Gianluca Fiorelli: Yes. It was back in 2015 or 2016.
Dan Petrovic: Google made a clown out of him and everyone laughed, and I was like... Because I had published, I don't know when it was, 2012 or 2013, a big article about user behavior as a ranking signal.
Gianluca Fiorelli: Yes.
Dan Petrovic: So I was with Rand on this, and he had a lot of haters around that time. Then, a couple of years ago—you asked how I do my experiments—I was arguing with somebody internally about whether user behavior is a signal, and I was like, “I wish I could find something to prove it,” because there was the Department of Justice case, all these leaked papers.
So I was hunting down little keywords, looking for this, looking for that, and then I found the repo. The whole repo. And I was like, “Whoa.” It wasn't like, “Okay, today I'm going to hack into Google.” I just fell into it organically.
The reason I'm bringing this up is that in there, when I analyzed it, I found proof without any doubt that they're using user behavior signals for ranking purposes. That validated everyone who's been saying it, including Rand, including me, including everyone else: user behavior signals.
So let's wind back to your main point. Do well in SEO before you start talking about the finesse of AI SEO. AI SEO is the finesse on top. The bedrock of good AI SEO is organic SEO. And that's my final answer.
Gianluca Fiorelli: Yes, I totally agree, and it's a good ending for a wonderful episode of The Search Session. Almost one hour. Dan, it was great to finally have you here as my guest and the guest of Advanced Web Ranking, and I really hope to see you soon. Are you going to do any European tour in the future?
Dan Petrovic: I try to travel less these days, but I was convinced to travel to SERP Conf., so I'll be in Vienna next week.
Gianluca Fiorelli: Ah, yes.
Dan Petrovic: And then I believe I'll be at Mike's gig in New York next April. And there's one from Ahrefs, Evolve, in Austin. I'll also be in Singapore, I believe. So hopefully not too many. Sandra invited me to SMX again in Berlin, but it's just so far, and I'm so tired of traveling.
Gianluca Fiorelli: I can understand you. I totally understand you. Sometimes, even though I'm European, I find traveling too much. I have to quiet down myself, too. So thank you, Dan. Thank you again. I'm really looking forward to your next experiments, your next tools, your next papers.
Well, yes, they are complex to read, I'm really honest with you, but when you get it and you have the aha moment while reading them, they're really enlightening. So continue with the good work.
Dan Petrovic: You're very welcome, and I had a great time with you today.
Gianluca Fiorelli: And thank you to all of you for being our guests too, here at The Search Session. Remember to subscribe to the channel and ring the bell so you will be notified when a new episode is published on YouTube and also on Spotify. Thank you, and bye-bye.
Podcast Host
Gianluca Fiorelli
With almost 20 years of experience in web marketing, Gianluca Fiorelli is a Strategic and International SEO Consultant who helps businesses improve their visibility and performance on organic search. Gianluca collaborated with clients from various industries and regions, such as Glassdoor, Idealista, Rastreator.com, Outsystems, Chess.com, SIXT Ride, Vegetables by Bayer, Visit California, Gamepix, James Edition and many others.
A very active member of the SEO community, Gianluca daily shares his insights and best practices on SEO, content, Search marketing strategy and the evolution of Search on social media channels such as X, Bluesky and LinkedIn and through the blog on his website: IloveSEO.net.





