Nik Ranger Kalicube Summit 2026 session

Nik Ranger on Fan-Out Queries and the Recruited-to-Displayed Cascade — Seven Million Synthetic Queries and the Citability Score

ByBernadeth Brusola

Session identifier

  • Session: Fan-Out Queries and the Recruited-to-Displayed Cascade – Seven Million Synthetic Queries and the Citability Score
  • Day: 1 (June 24, 2026)
  • Speaker: Nik Ranger, Senior Technical SEO Consultant at Dejan Marketing
  • Framework concept: ARGDW — the Recruited-to-Displayed cascade (Gates 6–9 of the AI Engine Pipeline), where the engine selects candidates, grounds content in a live response, and commits to displaying a brand or passes it over
  • SEL article: 5 competitive gates hidden inside ‘rank and display’

Speaker bio

Nik Ranger is a Senior Technical SEO Consultant at Dejan Marketing, an international search industry keynote speaker, and the founder and Chair of SEO Collective Australia. SEO Collective is a practitioner-first community built on the principle that every speaker is a working practitioner and every conversation is real — no sales pitches, no vendor keynotes, no panels engineered for optics. Nik is also a Google Women Techmakers Ambassador, co-host of SEMrush’s Australian Search Marketing Academy, and founder of Women in Search Australia. Her background is in civil engineering from the University of Queensland, and she brings that structural precision to everything she does in technical SEO. In her spare time she plays electric violin, illustrates, and paints.

Session’s place in the Kalicube Framework

Nik’s session covers the Recruited-to-Displayed cascade — Gates 6 through 9 of the AI Engine Pipeline (ARGDW: Annotated, Recruited, Grounded, Displayed, Won), coined by Jason Barnard in 2024 as the activate phase of the Kalicube Framework. The cascade is the sequence of decisions an AI engine makes after a brand’s content has been indexed: whether to recruit it into consideration, whether to ground it in a live response, and whether to display it to the user. A brand can be recruited and never grounded, or grounded and never displayed. Nik’s session is the empirical outside account of that cascade — seven million synthetic queries mapped against what Google, Gemini, and ChatGPT actually do at each gate. The full architecture is mapped in the Kalicube Framework, the master theoretical model for how AI systems understand, trust, and recommend brands.

Episode summary

Nik Ranger opens with the structural change that motivated the study: the travel experiences booking market has shifted from search-and-scroll to AI conversation. A traveller planning a trip to Tokyo no longer types keywords and browses ten blue links — they ask an AI, receive a synthesised answer, and the brands in that answer are the ones that get the booking consideration. Brands outside it do not exist in that moment. For a global platform competing across hundreds of destinations, tens of thousands of activity categories, and multiple languages, the AI recommendation layer is the dominant discovery channel for high-intent travellers — and the client had no systematic picture of where it stood inside that layer.

The fan-out methodology was the session’s most technically specific contribution. Dan Petrovic replicated Google’s fan-out patent and built a process that takes the client’s existing Search Console queries as seeds and uses a language model to generate synthetic expansions representing the full multi-dimensional space of ways a customer might ask something that connects to the client’s content. These expansions are not keyword variants — they are the actual query branches an AI engine opens when a user enters an ambiguous query and the system fans out across multiple plausible intents. Using the client’s own click and impression data to weight the expansions, the team built a custom demand estimation model accounting for event-driven seasonality and destination variation, reducing seven million queries to an actionable priority set skewed toward the consideration and commercial phases of the buyer journey.

The citation gap finding was the session’s most commercially decisive data point. Citation mining across Google and ChatGPT produced three numbers that changed the client’s entire cross-channel strategy: 110 Google citations, 2 ChatGPT citations, and 16 URLs that ChatGPT had retrieved, evaluated, and actively set aside in favour of a competitor. Google was surfacing the platform to travellers at the research stage. ChatGPT was not. In the 16 cases where ChatGPT had retrieved the client’s content, it had decided the content was not the right answer and recommended someone else entirely. That lost consideration is not a rankings problem or an impressions problem — it is a content-to-intent alignment problem with a specific cause the fan-out methodology can identify at the query level.

The weighted citability score was the session’s most operationally actionable contribution. Nik’s team built a brand-to-entity, entity-to-brand scoring model measuring three dimensions simultaneously: presence (how often the brand appears in AI results for a given query set), rank (what position the brand occupies within the result, weighted with first position highest), and volatility variance (how consistently that presence and rank hold across repeated measurement windows). The score is structurally comparable to Laurence O’Toole’s LLM Ladder Strategy from Session 2.2 — both solve the same measurement problem from different research angles, and Nik named the convergence explicitly during the session.

The two-speed ingestion section was the session’s most forward-looking contribution. Search citations are immediately visible and measurable within days through citation mining. LLM primary bias — the pre-trained opinion a model has formed about a brand before any query arrives — shifts on a months-to-year cycle as models retrain. Nik’s case study of a German sports jersey client perceived by US AI models as a foreign vendor, corrected over twelve months of on-page content, PR, and social campaigns, was the empirical proof of that timeline. Two speeds of machine learning require two speeds of practitioner response: fast-cycle citation work for search grounding, long-cycle campaign work for shifting the model’s fundamental perception. The session converges with Jason’s Recruited, Grounded, and Displayed framework from the demand-side: map the query space first, then watch which doors the engine opens and which it walks past.

Key quotes

“Brands outside the AI answer don’t even exist in that moment.” Nik Ranger, on why AI recommendation is now the dominant discovery channel for high-intent travellers — and why the window to get inside it is narrowing”

“We take a client’s existing Search Console queries, use those as seeds, and use a language model to generate synthetic expansions of those queries. These expansions represent the multi-dimensional space of ways a customer might ask something that will hopefully connect to the client’s content.” Nik Ranger, on the fan-out methodology that produced seven million synthetic queries from a single client’s Search Console data”

“One hundred and ten Google citations. Two ChatGPT citations. And sixteen URLs that ChatGPT had pulled into consideration, evaluated, and actively set aside.” Nik Ranger, naming the three numbers that reframed the client’s entire cross-channel strategy”

“The drop-off curve tells you which type of problem you’re dealing with. If the content is largely absent from the consideration pool, that’s a massive presence problem.” Nik Ranger, on how to read the cascade failure and identify whether the issue is Recruited, Grounded, or Displayed”

“You’re measuring the cascade from Recruited through Grounded through Displayed — and that’s exactly right. It’s very interesting that on either side of the globe, we’re coming to the same conclusions.” Nik Ranger, confirming the convergence between her demand-side methodology and Jason Barnard’s AI Engine Pipeline framework”

Further reading

5 competitive gates hidden inside ‘rank and display’ – The article maps the activate phase of the AI Engine Pipeline — from the Annotated gate through to the Won gate — and names the competitive cascade that Nik’s empirical data puts numbers against. Nik’s session is the outside-in measurement account of the same architecture the article describes from the inside.

Related sessions

Link to preview post

This session was previewed here before the Summit.

Link to speaker post

About Nik Ranger

Links to session clips

How Nik Ranger Fixed a Brand’s AI “Nationality” Problem in Under a Year

Why Two Experts on Opposite Sides of the World Agree: Nik Ranger Explains

Transcript for Nik Ranger on Fan-Out Queries and the Recruited-to-Displayed Cascade — Seven Million Synthetic Queries and the Citability Score

Jason Barnard: Welcome to this session with the wonderful Nik Ranger. She’s a senior practitioner in technical SEO, and she did an AI visibility study with seven million synthetic fan-out queries generated from Search Console for a global travel platform. I’m really interested in that fan-out query approach and what’s going on — especially from my perspective within the Kalicube Framework — at the Recruited, Grounded, and Displayed part of the pipeline. So walk me through what the data was and overall what you found.

Nik Ranger: Jason, it’s wonderful to be here. Hello, everyone, from Melbourne, Australia.

The only reason I can’t name the client is that it’s under NDA, but I’m more than happy to talk about the work because it was a really, really interesting project.

To step back: the travel experiences booking market, which is where this client operates, is going through a fundamental shift in how customers find things to do. Their customer base used to search and find competing directories. They would type in their destination, maybe things they want to do, get a little list together, and essentially just click around and collect options into spreadsheets or documents or whatever organizer they use. But now it’s a conversation.

A traveler planning a trip to Tokyo — and I love Japan, I just came back from there — doesn’t just type “things to do in Tokyo” and scroll through ten blue links. They ask an AI. They get a synthesized answer, and the brands in that answer are the ones that get the booking consideration. Brands outside it don’t even exist in that moment.

For a global travel experiences platform, the stakes are absolutely enormous. You’re competing across hundreds of destinations, tens of thousands of activity categories, and multiple languages and local markets. The AI is grounding and selecting pieces of evidence about their brand for consideration — whether it ends up in the results or not — and it’s doing all of that simultaneously using RAG retrieval and generation to formulate the result.

The client reached out to us because they wanted answers to some key questions their stakeholders needed to know, and how they were going to plan around them. They wanted to know: how are Google, Gemini, and ChatGPT deciding what to surface? Do we even surface for what’s important to us? What are we being recommended for, and — more importantly — what aren’t we being recommended for where our competitors are leveraging that ground?

Jason Barnard: And to do that you used query fan-out. How did you figure out what the fan-out queries would be? You had a base query — how did you decide what the fan-out queries to track would be?

Nik Ranger: You open up Search Console nowadays and you get some cool impression data. That’s telling you what’s already working, but it’s not really telling you anything about the query space that the AI is actually traversing when it’s deciding who to recommend.

So Dan Petrovic, who I work with, replicated Google’s query fan-out patent and built a process where we take a client’s existing Search Console queries, use those as seeds, and use a language model to generate synthetic expansions of those queries. These expansions represent the multi-dimensional space of ways a customer might ask something that will hopefully connect to the client’s content. That’s a much more elegant approach because it’s using queries that we already rank for — so we can tie it back to something the team can see and relate to how it connects to content. That’s really helpful in a workshop setting when you’re explaining fan-out to a bunch of content marketers. That’s where we got the seven million synthetic queries.

Jason Barnard: My question there is: what were the types of queries that come into the fan-out? Is it fanning out to try to cover the entire space? Is it fanning out to see what might come next? Or is it fanning out to give the most complete and exact answer?

Nik Ranger: Good question. When you do query fan-out, it typically splits into two different waves — the market share, if you will, of the ways a large language model can conceivably find your brand.

From these seed queries, we already know what we’re currently surfacing for. So we can say: based on everything you could possibly surface for, versus what you already rank for — these are the areas where we’re doing well, and these are the areas where there’s a significant gap.

For this particular client — and I should note, when you have global clients the numbers are absolutely huge — seven million isn’t a number you can take to a content team and just say: great, let’s work with that. We needed to break that number down and create a system of prioritization.

What we did was take their data, train a model on it, and build a custom volume estimation model specific to their Search Console query patterns. This accounted for event-driven seasonality and destination variation, to essentially rank the fan-out data and model the predicted demand using their actual click and impression data. So it’s tied to something the team is already seeing and expecting.

Jason Barnard: So you were able to identify the fan-out queries they could actually aim for to be part of the conversation. Then the question is: how many of the fan-out queries were questions people might genuinely ask? And which ones were purely machine-generated questions the AI used to help it answer the original query?

Nik Ranger: Exactly. And that’s really helpful because now we can go through and map it to the customer journey funnel. We can run query classification on it, map it out to the funnel, and start to see gaps. Here’s where we’re doing well and already surfacing. But these are the specific synthetic queries where we have no grounding whatsoever. And through our prioritization method — a custom demand estimation — we’re starting to reduce the seven million to: let’s go after something more heavily skewed towards the consideration phase. Let’s look at the commercial phase. Maybe even the aftercare — after they’ve made a purchase, do they come back? Do they give feedback? Is there an additional experience they might add?

All those key moments become identifiable. We can then work with their content teams to create really helpful wrap-up pieces that take a large section of this and turn it into content that’s better served for that particular intent.

Jason Barnard: Is there a particular form of content that survives in this world? The machines are picking up content, deciding whether it actually helps, and not citing an awful lot of it. If they’re looking at eighty pieces of content and only citing five or six — do you have data on how much is cited compared to how much is viewed or analyzed? And what kind of content survives?

Nik Ranger: What kind of content survives is really, really hyper-specific to this client. I say that for two reasons.

One, because I think a lot of people ask the same question. We’ve all heard from so many people that listicles work, so everyone makes listicles. And inevitably, anything done with regularity and openly discussed is easily fingerprinted. Come on — we work with algorithms, we analyze them, we figure them out. If we’re making easily replicable content types, they’re very easy to identify algorithmically. So the kinds of content that work are so specific to this particular client.

The way we approach this work is to ask: what is currently working for you? We look at the click-through rate distribution for their current content that ranks well in positions one to ten. What types of content are these? Are these evidence-based pieces where real people are on the ground experiencing what it’s like in Greece, or Athens, or Tokyo — going through Disneyland, riding the Shinkansen? People doing live, real content. And then: what can we replicate based on the resources and business case we can put forward to create more pieces like it?

Content teams work exceptionally hard and have enormous pressure to perform and show the value of heavily resource-intensive investments. So for us it’s saying: here are the choices your team has made that have been doing well. Here’s the new research where we’re seeing significant gaps in the market. Let’s pull this together and create a conversation about how we want to replicate that — using our influencer networks, our on-the-ground journalists in specific locations doing reviews, photos, videos, something that is a point of difference we can see working. Or even just testing different text-based content formats and watching whether the large language model selects them in those specific results. Everything is a minimal test, and that’s where we build the business case to hopefully unlock resources to enrich the content experience for the end user.

Jason Barnard: Thank you for that. The next question is about the differences between Gemini and ChatGPT — because I think that’s a huge question we’re all asking. What are the differences in terms of what they select, what they use, and how they present it?

Nik Ranger: Huge differences. There were a few numbers that really changed the conversation with the client in terms of how they thought about this from a cross-channel alignment perspective. Those numbers were 110, 2, and 60.

One hundred and ten Google citations. When we did citation mining, we wanted to see what the model was actually using as a link pointing out to different service providers, vendors, or platforms. There were 110 of them from Google. There were only two ChatGPT citations. And there were 16 URLs that ChatGPT had pulled into consideration, evaluated, and actively set aside.

So essentially: Google was showing this platform to travelers at the research stage. ChatGPT was not. In those 16 cases where ChatGPT had pulled in their content, it had decided it wasn’t the right answer and recommended someone else entirely instead. That lost consideration is the exact moment where a customer is deciding whether to use them or their competitor.

How that translates to the client — and I keep almost saying the brand name — is a market share story. The AI recommendation layer is becoming the dominant discovery channel for high-intent travel planning. The traveler who asks “what is the best cooking class in Chiang Mai?” is further along in their purchase journey than someone running a generic search — they’re closer to converting. Losing that customer to a competitor because the AI passed on the content is a commercial problem with a specific cause. The drop-off curve tells you which type of problem you’re dealing with. If the content is largely absent from the consideration pool, that’s a massive presence problem.

So that’s where we really look at the selection rate. When the URL is accessible and the page is being processed, a lot of what we found is that only around two thousand characters of any particular page are actually used in its retrieval process to create that result. We’re hoping that within that sample of content the AI pulls from the page, there’s something meaningful it would use for its result.

This is what we use to measure with a brand-to-entity, entity-to-brand score — essentially a weighted citability score that tries to figure out: for your brand, and the entity you most want to be associated with, what’s the frequency with which you appear? And what’s the rank — meaning what position does your brand surface within the result?

So for example: what is the best platform for tickets for Shibuya? The AI might return: this brand, this brand, this brand, this brand. If your brand appears first, we count that as a rank of one. How often you appear — that’s prominence within the result. We use that as a weighted score to measure over time, together with volatility variance. With Laurence O’Toole we were discussing something very comparable: a weighted citability score that is fundamentally the same idea. Are you present? Are you the first one mentioned? Is the sentiment positive for you?

Jason Barnard: Love the idea. And we’re talking about what I call Recruited here — you mentioned “selected,” but the idea is: are you recruited by the algorithm as an option? Are you then discarded, or are you kept and used for Grounding? And then from Grounding are you kept and pushed into Display, where you’re actually shown both as a link and as a brand? So there are three steps. How do you feel about Recruited, Grounded, and Displayed?

Nik Ranger: You’re measuring the cascade from Recruited through Grounded through Displayed — and that’s exactly right. The comparable thing to how we do it is that we approach it from the demand side. We map the query space first, and then watch which doors the engine opens and which it walks past. So it’s entirely comparable. It’s very interesting that on either side of the globe, we’re coming to the same conclusions.

Jason Barnard: I was talking to Kevin Indig, and we were talking about macro measurement. He basically described what I’ve done, and I described what he’s done. There is a lot of meeting of minds with the people who are really thinking this through — looking at it from the perspective that the conversation is happening within the engine, the engine is recommending what it feels is the best solution. The goal is always: get the user to the best solution to that problem as efficiently as possible.

And I find that Recruited, Grounded, Displayed part very interesting. I was talking to Krishna Madhavan about this — he, with Bing, is releasing a lot of data inside the console. Have you had a look at that?

Nik Ranger: Yes, it’s so helpful. We recommend all of our clients, if they haven’t set up Bing, to set it up, because it adds an additional contextual lens. In the same way that when you’re looking at Search Console and seeing engagement metrics like click-through rate, you still want to go to Google Analytics to look at sessions, time on page, bounce rate, where they’re clicking from, that whole flow — Bing provides helpful additional context in the same way. Bing offers way more transparency.

And I think Google has now come to play — that’s why we now have AI visibility impressions in Search Console — but Bing still offers significantly more.

Jason Barnard: I mean, Google is now in a position where suddenly they’ve got competition that didn’t really exist before, to an extent that they need to worry about it. So Bing has been very open. Google is having to be more open. And with ChatGPT in the mix there’s much more focus on that. How much are you seeing Google become more open about data they were previously not pushing quickly to share with us?

Nik Ranger: That feels like more of a political question. We have people like Gary Illyes, Daniel Waisberg, and John Mueller, as well as lots of amazing Googlers who are very willing to have a conversation with us and to triage our questions. I think it’s amazing that they do that at all, because they didn’t have to.

We get visibility when there are core algorithm updates. We don’t get the same kind of visibility for tweaks and changes within the models — for post-training within large language models. But we know when they’re using a whole new model altogether.

At the moment I’m doing a test just for my own curiosity — testing Gemini 2.5 versus 3.1 to see whether post-training quantization has had an impact in terms of how it’s pulling results. It’s interesting, but I feel like more is being obfuscated. And there is still an old guard of people wanting to engage with us.

I think it’s really, really nice that they do that. Always keep that in mind when you’re asking a question — what if they weren’t there at all? What if they completely turned off the page on core algorithm updates and we just saw volatility? That would be really, really tough. But I think they like having a conversation with us. I think they learn a lot from us — what to do and also what not to do. Repeatable patterns get fingerprinted pretty quickly. They are paying attention when things happen. They are the ones tweaking the models and the way they’re responding to us — with post-training, or even running experiments before core algorithm updates happen. That’s why we see, waiting for enough queries, intent shifts. That’s why we have algorithmic suppressions — things that happen without manual actions.

They do still use manual actions. They send out messages. They have content that a lot of people don’t like. But the web is so huge, and the number of people thinking about this every single day in such granularity is massive. They are a team trying to do things at scale, trying to produce the best result — and then they have shareholders and all kinds of other pressures pulling in different directions. I try to keep kindness in mind when I think about the way we talk about this.

Jason Barnard: I’ve found the Bing engineers to be super helpful and super open in the interviews I’ve done with them. That’s been hugely helpful. And as you say, all of the engines — Bing and Google — are engaging now.

The last question is about this recruitment moment. At Kalicube, what we’ve been looking at is the Algorithmic Trinity: LLM for intelligence, search engine for information, and knowledge graph for validation or verification. An LLM to have the conversation, a search engine for fresh information that the LLM doesn’t have in its training data, and then the knowledge graph to make sure it’s getting the facts about entities and topics correct.

The recruitment moment is interesting because for a search engine, I would say it’s immediate — it’s recruiting, deciding, and showing in one motion. But to train the LLM or add information to the knowledge graph, it’s a slower pace — by retraining the model, they select the data, clean it up, and feed it with a particular goal in mind. Same with the knowledge graph. We saw that with Google’s focus on person entities three years ago, when they tripled the number of people in the knowledge graph. They recruited content specifically for that, cleaned it up, and added the person entity.

So there are two speeds. One is search — immediately visible in what we see with citations through the assistive engines. The second is training data for LLMs and training data for the knowledge graph, which is significantly slower — months rather than hours or days. How do you feel about that?

Nik Ranger: I feel there are a lot of levers that control how things are done that are not immediately obvious or recognizable to us when we’re testing this.

Just recently — or it might actually be the culmination of years of campaign work by a team of lawyers — Germany, as a country, has essentially created a precedent that Google as a company is legally responsible for AI-generated results in their search engines. So when Search Generative Experience was first released, one of the first things they did was look at E-E-A-T content — experience, expertise, authoritativeness, trustworthiness — typically the areas that impact your-money-your-life content. I can’t fathom why, out of all types of content, they would tackle that first. But now that this German precedent exists — and no doubt Google’s lawyers will try to overturn it because it sets a huge precedent — if that sticks, if that’s directional, I would imagine we’re going to see a huge shift in the way the results are used.

And even today, Dan just found that Gemini’s native web search grounding is limited to 1,500 daily calls, which is useless for most serious applications. Downgrading to Gemini 2.5 gives you 5,000, which is better but still not great. We see a diminishing return in peak requests per day for the search grounding results.

This is what I was wanting to test — the models alongside post-training quantization, to see whether it’s an efficiency problem rather than a model capability problem. But again, the model is static and will always show the same result until it’s fine-tuned. We can only really speculate about the reasons and mechanisms.

But in the way it’s doing pre-training or post-training on its models, there are a lot of different reasons and levers that govern the way they’re going to steer it.

There are two frames of thought within machine learning that I have done my best to really understand over the last few years. The first is mechanistic interpretability — the machine learning science of the tokenistic probability of one word preceding the next, to actually understand why the model chose that word in the way it surfaces. We’ve done enormous amounts of testing with tree walker analysis, essentially trying to replicate results and understand why they’re happening. And this is from people who don’t have data science degrees. I’m sure Dawn Anderson, who we’re going to be speaking to, has wonderful ideas about this.

The second frame is model steering and model control. Model steering is typically what we’re all interested in — steering the model towards our clients, our sources of information, our sources of truth. This looks at the bias every AI model has already formed about your brand before the query even arrives. A pre-trained model, if it includes you, has been trained with your brand described, rated, compared, and recommended — and sometimes, if there is negative sentiment, criticized. The sum of that training is what we think of as primary bias. Changing that is where citation mining comes in — looking at where it’s sourcing its results — so that we can do digital PR campaigns to put out certain types of content that works on that sentiment and shapes the public’s opinion of the brand.

We’ve even done that with a client that does custom sports jerseys. They were perceived as a German client operating in the US. They were never surfaced because when the model was looking for results, it was looking for American brands and American vendors for sports jerseys — even though this client had decades of legacy, physical locations on-site, US-based representatives and teams, and physical locations. We worked with the client on on-page content, PR campaigns, and short runs of media and social media at different levels to influence the perception that this is actually an American company. Over about a year and a couple of model updates, we got this client to be considered a US-based company for its US operations.

Jason Barnard: Brilliant. And that comes back to the point I was making: the search results will surface immediately, but if you want to change the fundamental opinion of the model, it takes about a year. So you have two speeds of ingestion by the machines — depending on whether you’re talking about search, which is immediately visible in what we see with citations through the assistive engines, or whether you’re talking about the LLM, which is the case you just gave, or the knowledge graph, which is my favorite topic.

Nik Ranger: Absolutely.

Jason Barnard: Brilliant. Thank you so much. That was delightful. Thank you, everybody. See you in the next session. Thank you, Nik.

Nik Ranger: Thank you so much, Jason. It was an absolute pleasure.

This session extends the Search Engine Land article series on AI-era brand recommendation by Jason Barnard at Kalicube.

Canonical metadata footer

Session date: June 24, 2026 Duration: approximately 25 minutes Format: pre-recorded, on-demand from session drop Day: 1 Framework concept: ARGDW — Recruited-to-Displayed cascade (Gates 6–9 of the AI Engine Pipeline) SEL article: 5 competitive gates hidden inside ‘rank and display’ [Share buttons: LinkedIn, X, Facebook, Email]

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *