Hello, and welcome to The Search Session. I’m Gianluca Fiorelli. Today I’m talking with Jan-Willem Bobbink, an enterprise SEO consultant and developer based in the Netherlands.
Our conversation connects AI search experiments with practical SEO work, from GSC Wizard’s approach to reliable data analysis to diagnosing indexing issues with Bing and Yandex. Jan shares his experience with Brave, dedicated landing pages for LLM visibility, and the risks of scaling content too far. We also discuss spam filtering and the continued value of basic SEO and internal linking.
What you’ll learn in this session:
What makes AI analysis of Search Console data more reliable: GSC Wizard uses a data warehouse, server-side calculations, and an MCP to give LLMs pre-calculated reports to reduce errors and speed up analysis.
How to diagnose Google indexing issues using Bing and Yandex: comparing reports helps identify Google-specific issues or technical, content, and rendering problems on your site.
How Brave’s indexing differs from Google’s: real user signals help determine which pages enter its index, influencing their visibility in Claude’s web answers.
When to split content into separate landing pages and where to stop: distinct audiences and intents can justify dedicated pages and increase LLM visibility, but insufficient depth or demand can hurt SEO and make visibility gains short-lived.
Why fighting search spam is becoming harder: AI makes spam tactics easier to scale, while Google must adapt its filters to different niches without removing results users expect.
How to build an SEO strategy around the actual SERP: study ranking competitors, their history, links, and content structure, because what works varies by niche, query, and user context.
Why LLMs need stronger spam and quality filters: search engine filtering alone cannot prevent manipulated answers or AI-generated errors, so LLMs must verify sources, question self-promotional claims, and select trustworthy training data.
What Jan’s content experiment reveals about Google rankings: in low-competition searches, relevant titles, headings, and strong internal links can help pages rank despite irrelevant body content.
How to assess internal linking: identify pages with strong external backlinks, map how link value flows, and simulate changes with PageRank calculations before implementing them.
Tune in for Jan’s perspective as someone who builds his own tools, tests ideas, and works through these challenges with enterprise clients.
Topics covered: GSC Wizard · MCP · indexing diagnostics · Brave Search · AI visibility · query fan-outs · audience intent · content pruning and scaling · user signals · brand searches · spam filtering · SERP analysis · SEO experiments · internal linking · PageRank
About the Guest
Jan-Willem Bobbink
Freelance International SEO at Notprovided.eu
Jan-Willem Bobbink brings nearly 30 years of experience in web development and search optimization. He specializes in enterprise SEO strategy, international SEO, and technical challenges on large websites.
A freelance consultant since 2009, he works through NotProvided.eu, his consulting website and SEO blog, where he shares articles, experiments, and tools.
In 2026, he created GSC Wizard, a platform for Search Console analysis and AI-assisted reporting. He also speaks at conferences and has taught JavaScript SEO at SEO Pro Academy since 2024.
Transcript
Full conversation between Gianluca Fiorelli and Jan-Willem Bobbink.
Gianluca Fiorelli: Hi, I'm Gianluca Fiorelli, and welcome back to The Search Session. Today, we are going to have a guest who, casually—I mean, I'm not doing this on purpose—is from the Netherlands. This is a good sign for this small country, which is the motherland of many wonderful SEOs. He is a freelance enterprise SEO.
On LinkedIn, he also very clearly declares that he is a developer, and we will discover how much this nature as a developer is imbued in everything he's doing. And he works a lot with enterprises for SEO strategy and alignment. He does a lot of international SEO, works with very, very big websites, and develops custom-made tools.
And thankfully, for a year now, I think, he's also shared the tools publicly and sold them with a subscription. And we are going to talk especially about one of his tools, which is GSC Wizard. Our guest today is Jan-Willem Bobbink. Hey, Jan, how are you doing?
Jan-Willem Bobbink: Good day. Good day, Gianluca. I'm good. Just back from holiday, so finally the real working season is happening, and I'm always looking forward to it. So lots of new exciting projects, challenges, and new budgets are already looking forward to next year. So things are good here.
Gianluca Fiorelli: Cool, cool, cool. Holidays are starting to become somehow a memory. Well, Christmas is getting closer fast, so we are going to have a break in just a few months. Okay, but with this beginning and all the months before, how would you consider SEO treating you?
Jan-Willem Bobbink: Very well. To be honest, I was a bit bored until, like, two years ago, maybe even a year ago. But with the whole change with LLMs getting onto the stage, I'm basically back at my old, 20-year-younger mindset because we need to experiment, we need to test, and we need to convince clients of stuff we're not 100% sure about ourselves.
We need to be honest. We need to reinvent our work, realign clients, and realign teams. So for me, currently, SEO and maybe even broader, the whole search industry is very much tickling my mind and makes me work even harder to get the best results for my own projects and for my clients' projects.
How GSC Wizard Makes AI Analysis of Search Console Data Reliable
Gianluca Fiorelli: One of these projects, which I happily subscribe to and use a lot every day, is GSC Wizard, especially since the MCP edition came out not so long ago. And how did you come up with this idea? Before, off the record, you were saying that this is a sort of tool that you were creating for yourself, but how did the idea come about to make it public, to make a business out of it?
Jan-Willem Bobbink: So the thing was, over the past 20 years, I've developed lots of small scripts, mostly Python, that I was running internally on my server for my enterprise clients. And then I was thinking, if it's useful for me, and even for myself, sometimes I was looking for the right scripts, or I wanted to have it slightly different, or I needed to build a new script to import from BigQuery instead of the GSC API.
So this winter I said, “Okay, now I have this history of Python scripts; let's build it into a modern web app so everybody can use it. And for me, I always have it available, also if I travel, so let's make it web-based.” That's how it started. So I basically put a layer in front of the logic that I already had.
And I think, for me, that's also the powerful thing: especially for bigger sites, we cannot trust LLMs with the data sets going directly into them. Imagine if you send, like, 25,000 queries into an LLM and ask it, “Okay, what has changed?” It doesn't know where to start. Especially in the beginning, like a year ago, these LLMs would just start guessing and not even look at the numbers, which you see happening nowadays. And that's also why I said, “Okay, this is also a good time to build an MCP, because the LLMs themselves start building a script to analyze the data, but then you're still limited to what the API gives back.”
So I went back to the table and said, “Okay, I need to build my own data warehouse. I need to build a backup functionality if GSC is not functioning.” So now I have the MCP that's basically in between my data warehouse, where the raw GSC data is stored, where I do the daily calculations, and where reports are pre-calculated.
So that also makes it possible that if you have a client with, like, a million unique keywords on a daily basis, then the reports are already ready for me. So if I log in on Monday and I want to have the weekly reports for those bigger clients, it only costs me a few minutes. And because it's based on the logic on my server, the chances of errors from the LLMs that I use for writing reports or the initial analysis are way, way lower.
And based on my benchmarks, there's still, like, a 1% chance that the numbers are not correct because, in the generation of the reports or the wording, an LLM can still hallucinate and make an error by quoting the wrong numbers. But the fact that you calculate the outcomes of your analysis on a logic-based server, I think, is the main reason why the MCP is working well, and people are really happy with it.
Currently, it's also featured as one of the MCPs recommended by OpenAI Business. So apparently those developers from OpenAI are also happy users. So let's see how that will evolve. But that's one side of my business. I wouldn't say it's a business yet because it's more of a hobby project. But it's getting more serious over the days, so let's see what happens there. Day-to-day, I'm still the SEO consultant, and that's my main job.
Gianluca Fiorelli: Sure, sure. Well, I think it's covering a space where, seriously, only a couple of tools are working on the same concept of retrieving—not saying the same things, obviously—and making Google Search Console what Google Search Console should be.
But the interesting thing is also that you included, for instance, obviously Google Analytics, but also the possibility of retrieving Bing data. I don't know if you are thinking of eventually including other search engines. Maybe yes, because of things we are going to talk about in a little bit, but it would surely be great.
Using Bing and Yandex to Diagnose Google Indexing Issues
Gianluca Fiorelli: In fact, talking about search engines, there is one thing that you shared on LinkedIn, which is, let's call it, cross-engine diagnostics, or how to use Bing and Yandex as a canary bird for Google. Would you like to explain this kind of idea that you already shared on LinkedIn?
Jan-Willem Bobbink: So what I'm using, especially for clients that have issues on the technical side of things, I always cross-check it with various other engines. And what I like about that idea is that Bing and Yandex give information back in a slightly different way. So you have basically three different perspectives on your website, and if it's a Google-specific issue, let's say you have a penalty—not a manual penalty, but other issues—it's easy to see if Bing is still indexing you and Google isn't.
And if it's happening across the whole site, you have a serious problem, and usually that helps me to make the first distinction between what kind of issue it is. If there's just, let's say, a delay in indexing between the search engines, again, you can see: is it a Bing or Yandex issue? Both work with IndexNow, so indexing should be fast. Or is there something wrong in the pipeline with Googlebot? Maybe the indexing or the retrieval goes fast, but then the final indexing is being broken.
So maybe it's a JavaScript rendering issue, or maybe, like what I've seen with some WordPress websites getting hacked, there's malicious code or content inserted, like content about casinos, and then that only shows up in that second stage of rendering, and that's where Google still filters you out. So the initial crawling goes well, and the crawl stats look normal.
The thing with the indexing report in Google is, especially lately, for the past six months, it's been so delayed, and the data is not trustworthy. Again, if you have multiple ways of checking indexing across different search engines, that's really valuable to, for me at least, identify which direction I need to look into to see if a website has a technical problem, a content problem, or a rendering problem.
The other thing is that Yandex specifically mentions the kind of issue they have with your content. So if it's thin content, if it's not relevant to the topic you're trying to rank for in relation to your main domain entity. Same thing with Bing. They also have a thin content classifier that they give you back in their indexing reports.
The only main issue is that—and this answers your question, like, do you build it into GSC Wizard—not all of the information is available through the API, so I still need to do some screen scraping. But maybe, looking at the way Bing is currently acting by enabling more and more features that put them in front of what we get from Google, they will also give out more data in the API in the future. But they already do nicely with the AI citations and the queries and the intents, plus the grouping in the different categories.
Gianluca Fiorelli: In fact, I was thinking in relation to Bing, but Bing surely is more generous in giving us data about AI search in Bing Copilot. So it could also be an interesting way to triangulate and use it as a proxy in order to understand, because the people using Google and the people using Bing are not as different as the people using Google and the people using Naver. So the mental construct is the same. If someone is doing some kind of conversational search on Bing, more probably a person like them is going to make a very similar query on Google AI Mode, or Google and then AI Overview. So using Bing as a way to hypothesize what kind of conversation and what kind of pages eventually are coming out from those conversations also in Google could be a good idea.
How Brave Decides What Gets Into Its Index
Gianluca Fiorelli: And talking about search engines, you're also one of the few who have pointed out Brave, actually also writing a post on how to rank on Brave. We know that Brave is explicitly cited by Claude in the Anthropic documentation as one of the sources that it uses for real search. Probably Claude also uses other sources, but they are not disclosed.
Why is it important? As I remember, you were saying, somehow, “I'm feeling younger, like 15 years ago.” Yes, 15 years ago or more, we were used to working with many search engines. Somehow now we are returning to that, adding also the LLMs, not just search engines in the classic term. So what would you say are the most significant differences with respect to classic Google SEO when we come to Brave?
Jan-Willem Bobbink: I think for Brave it's important to understand how they allow indexing in their search engine to start with. So you need to reach a certain threshold of traffic to your domain or even individual pages. What they initially do is that they don't have the budgets of Google, so they cannot crawl the web unlimitedly. So what they do is use real user signals to determine basically what the thresholds of sites are and if they want to spend energy on scraping or downloading these sites or not. We know that for Google, user-based signals are important. We've seen the leaks, we've seen all the tests, and I often joke that doing a TV campaign is the most impactful ranking signal, and also the most expensive.
But we can clearly see that user signals will probably be the most used factor across search engines and LLM crawlers. And for Brave, it's the same thing. They, of course, have their own browser, so that can be a source of the data. They have various other ways to actually get the data, and they combine it and basically say, “Okay, a domain and a page URL need to reach a certain level of actual users to be included in our search results.”
Jan-Willem Bobbink: So what I find interesting, and I think it's also their business model, is that they want to maintain a high quality of their index. And I think it's a smart approach. The thing is, for smaller companies and smaller brands, it's probably more difficult to get into it, but that also ties into the way Claude will see or use these results. And like you said, they don't only use that index; they use multiple resources.
But I think it's interesting to have a slightly different but also more transparent search engine on board that's actually being used. That may change over time. We all know that ChatGPT is building its own index. Probably Claude will also do it at a certain point, but for now you can clearly see where they fall back with the query fan-outs. I can almost tie the results in Brave back to the sites that are actually ending up in Copilot, ChatGPT, or Claude. But we should also be afraid that we lose the query fan-outs. Based on the latest research I saw from Google, they may be able to actually skip that step and directly tell the model what to filter out and what to use.
Gianluca Fiorelli: Like the latest news with the new retrieval model that they are proposing. Let's see. And it's fun. Also, Perplexity, we know that it's creating its own index, for instance. And it's just a little fun note: nobody seems to officially use DuckDuckGo as a reliable index for RAG.
When Separate Landing Pages Boost LLM Visibility
Gianluca Fiorelli: And talking about Google, let's move a little deeper into Google. One thing that you did for a big e-commerce and a software-as-a-service website. Usually there is, or there was, a classic theory, sometimes proposed by Google, actually still proposed by Google, that if you are going to write about something, about a specific topic, or whatsoever, you should not create more than one page in order not to split the content and to avoid things like the classic concept of cannibalization and so on.
I anticipate, spoiling myself, that I agree with you in the sense that, actually, why not create more pages to target a specific topic, eventually also the same very similar query set, if these pages are targeting different types of search intent? And I think that your case study is probably going in that direction.
Why did you “disagree with Google," and what kind of positive results have you obtained from this disagreement?
Jan-Willem Bobbink: So it initially started with some simple things. Imagine you have a landing page that targets medical professionals. You're selling products for medical professionals, and then you have a landing page which says, “Product X for dentists and surgeons.” Which is basically fine if you're browsing the website and you know these two target groups both use the same instrument, so it's fine to have one product landing page.
But thinking about LLMs, they always do their query fan-outs and their filtering based on the context of the user, and that's not new because Google also tries to do it. A few years ago, they came up with the concept of passage ranking. So they basically pick one section of the page that's directly tied to the topic and the context that the user is putting into the search engine. Now, that may be an image, that may be text, or that may be a previous search history, but they had a good picture of the user. So they were able to say, “Okay, within this page that consists of multiple topics, we are able to pick that one passage that's actually relevant to what the user expects back from us.”
The other thing is, Google has been really critical about what they crawl and what they index because it's literally costing them a lot of money. So over the past few years, they also have been really vocal about making sure that you don't make your content too thin, too spread out across multiple URLs, especially if you're a smaller site. And that's also one of my key optimizations that I've been doing for the past five years: cut down the size of sites because the average link value goes up, the average content quality goes up, and we've been doing that since Panda in 2012.
They have these average content thresholds, and if you're below it, you're just screwed because it will drag down your whole site in terms of ranking. So you always want to make sure to be above that. And I often see clients who either come to me because they directly dropped in traffic and they have done something completely wrong, or they have been dropping for two years, and now it's finally time to really do something about it. But especially those slowly dropping websites, often the fix is simply to cut down all the non-performing content. And most of that content is usually not good quality or not targeting one specific topic enough. So on one side, I'm cutting down. On the other side, I'm really looking at what the user is asking the search engine to give them back.
And if you are able to provide that content in an extensive and detailed way, then it's worth actually splitting. So back to the example I gave, dentists and surgeons are two different types of people. They may use the same product, but not for the same cases. So being aware of that and making sure that you have a clear distinction.
Probably if you have a clothing e-commerce webshop, that will be way more difficult because trousers are trousers, and the use cases are not that different, except if you have trousers for day-to-day use or for work use or whatever. But then usually your topical authority is already around that intent and not per se on a page category level.
So, where I've been able to make big improvements is where there's a clear distinction between the intention and the target group. And the problem I have is it's working better for LLMs than for SEO. And I'm worried that it's very difficult to find that balance between creating too many pages, especially for somewhat smaller brands that don't have a big brand search volume, that don't have a big link profile. There is a fine line, and maybe it works for three months, and then you drop off because the average page volume goes below the threshold or you're not worth it.
The terms of Gary are often very clear: you don't have an indexing problem; you have a content quality problem. So it's that fine line I'm navigating. And now it's getting even worse because, with these LLMs and these query fan-outs, the queries that are being given to the search engine are even more specific, so that leads me to make even more specific landing pages. And the response is often that the results are there within twenty-four hours. So I'm worried a bit that if I see that this is working, others will do the same. I think we may head back to a kind of 2012 where Panda was needed to actually make people aware that just expanding and making your content more and more detailed on the page level is not the best way in the long term.
Passages vs. Dedicated Pages: Knowing When to Stop Scaling
Gianluca Fiorelli: I think there are many difficulties in the sense that, if you scale this kind of strategy, this kind of tactic, there is actually the risk of incurring so-called scaled content abuse. Google—not really the LLMs, but Google—could think that you are doing this kind of work, which you are doing sincerely to better target the different nuances of a buyer persona that your client or your website has, to be visible with the widest and largest query set possible around the topic.
But this is something that I also have to think about and work out for a client of mine in the travel space because they are targeting substantially different types of personas going to the same destination. So the big emblematic question was: do we need to create three or more different landing pages for the same destination just because we are targeting different types of personas? And in this case, I think that the same concept that you were citing before, passage ranking, can be helpful in our case. You can create a landing page that is generic enough to immediately present all the things that are in common to all the buyer personas, and then create specific passages, also playing with design, creating different types of boxes that are very self-contained. “If you are a young couple who really likes active touring and so on, we recommend you do this, this, this.” And then another similar box, self-contained, very clear, about family travel to the same destination. So I found that this can be a good way instead of getting crazy and creating too many specific landing pages.
And then, if there is a specific landing page that should be created because there is a clear distinction—let's say that for a destination, all the buyer personas are interested, but especially one is interested in that destination—maybe for that specific buyer persona, you are going to create a spin-off landing page just for them. So you can go further with the information, or just limit yourself to a small passage inside the generic landing page. But this is also because I then put the landing page as substantially the pillar page of a more complex type of content hub about that specific destination.
Jan-Willem Bobbink: And I think that's also the distinction you need to make: do I have enough material available to make it a rich enough landing page? And I think people will tell themselves, "Yes," or if you use Claude, it will come up with random content. But I think that's also where people make mistakes: just scaling up your content is not enough. It also needs to map back to your actual personas. It needs to map back to your actual usage of the content. And that's why some programmatic SEO is working, but like 95% is not working because it's just not good enough, even though you as an outsider might think it is. But you are not the three or five or ten personas you're targeting.
So I think that's another common mistake: you scale up too much, and you scale up across, in your case, maybe 200 destinations. Are these personas relevant for all those combinations or just for a subset? Maybe families go to hundreds of those destinations, the couples go to 100 different ones, and there may be overlap.
But I think that's where it's easy for search engines like Google to pick up the trends, and that's what you see in the indexing reports. Once they have seen the pattern, that whole section of the website is getting de-indexed slowly or not indexed at all. And I think that's also where, back to the beginning of the discussion, Bing is giving you the warning for thin content, and Google is saying, “You're not getting indexed,” but they don't give a specific reason.
Content Quality, Pruning, and the Power of User Signals
Gianluca Fiorelli: And that's why we and many other SEOs discovered, in the page section of Search Console, “Crawled, currently not indexed,” “Discovered, currently not indexed,” and started digging into why it's not indexed if it's crawled. Or why this page that has been crawled—because I know that it was crawled and I'm seeing in the performance dashboard that in the past it was receiving impressions and clicks—has now disappeared.
So people have started to dig into this, and usually it's related. Google has also confirmed, I think, in some of their many events, that usually if a page is there and not popping up in the index anymore, it's because of what they call a content quality issue.
Jan-Willem Bobbink: They've been more vocal about it because I think now, with AI coming into play, people can more easily scale content. There's no technical boundary anymore. Ten years ago, we had to build our own algorithms and then find interesting patterns to replace and make the content as unique as possible. Nowadays, you just ask an LLM to come up with 100 different pieces of content, and it's doing it. You don't need any previous knowledge. You can just trust the LLM. And to be fair, some of the content is not even bad, but it just doesn't fit within the overall metrics that a search engine uses.
And I think, for now, people also shouldn't underestimate the fact that being able to get away with it in ChatGPT doesn't mean that they won't catch up with the way Google is able to filter out low-quality content. They will buy their engineers; they will catch up on some part of it. Sure, maybe it works for a year, but that's fine for your hobby projects or your affiliate blogs. I would be afraid to scale too much with client sites.
And the cases that I shared on LinkedIn—those were very obvious cases—but in both cases it was also pruning. So I cut away a lot of non-indexed stuff and outdated stuff before I actually started increasing the number of pages. And it was all based on signals that were already there in the market. So we knew the target groups were visiting the sites. We saw in the internal click data that people were clicking through to the pages, and then we saw people searching for it, like, “Hey, do you have these products for surgeons?” So we had multiple signals to be able to say, “Okay, we're quite sure that if we split these pages, both pages will have an audience that will use the pages.”
And I think people underestimate the impact of actual website usage. Google has been saying that they don't have the data, but they do, and it's actually public data, like the whole CrUX data set. It's open for everybody to use. I can't remember his name, but he built a nice tool where you can actually see how popular your website is compared to others. He built a kind of index based on the CrUX data. So I was thinking, if just a random person can do it with the CrUX export, then Google can do it on the fly. And I've seen so many cases where you launch a new website and then send in paid traffic and social traffic, and within a day all pages are indexed and ranking.
And I've seen cases where there's no real launch, and then it takes weeks before Google even picks up the homepage. And it's the difference between having traffic to your website and not, with the same technical quality and content depth. If you don't have the traffic, you need to compensate with something else, and that's more traditional signals. But you can basically get away with it if there are 100 people searching for your brand name every day. Everything will rank.
Why Fighting Search Spam Is Getting Harder
Gianluca Fiorelli: Yes, indeed. It would be easier. And thankfully, brand searches are becoming an important metric, not because they weren't important before, but because of the uncertainty of how to measure the impact of everything we are doing. LLMs are not giving us any information, and Google is sampling more and more things in search.
For instance, in the generative AI report, we don't have any queries. I can understand, for instance, that we don't have queries for the very fresh multimodal report because the images, in this case, are the queries. But it's clear that Google surely has information, at least to tell us what queries are used in generative AI on Google Search.
But anyway, coming back to the problem of scaled content and how Google is really pushing on content quality, maybe this is also why, in the last, I would say, 12 months, we had more spam updates than core updates. It's fun because even if these spam updates are very small—I mean, the last one seems to be lasting a little bit longer—usually they are very short. But it's interesting to see how Google is really trying to fight back against the spam that is now built at scale.
Jan-Willem Bobbink: I think they have to because things follow up rapidly. Compared to the number of engineers working on the spam team, there are millions of SEOs trying to game the system, and that difference has become even bigger now that AI allows more people to do spammy things. And most SEOs often see spammy things as content scaling, programmatic SEO, or link buying. But also consider hacked websites. It's much easier, even if you have no technical knowledge, to ask Claude to build something to hack your WordPress, or at least search for websites that have been using hacked plugins, and then you can do the same trick. So knowledge of black hat tactics has become much more easily available and can be practiced much more easily without any knowledge upfront. And I think that's the problem they are facing. For some of these spam updates, basically the average SEO has seen no impact, even in their markets, because it was not focused on them.
On the other side, I also work in iGaming, and some of these SERPs have even become worse. So they filter out ten websites, and then there are twenty other websites turning up. And based on the meetings or the talks I have with some of the Google team, it's very difficult. They cannot apply the same rules to e-commerce as to iGaming because if they did, there's nothing left in the search results.
Or there are just informational websites, and then the user is not happy because they expect some of these websites to turn up, even if they do spammy stuff. So for them, it's a fine line. And that's also, I think, where, maybe, in the past three years, they've tried to let the system do the work and recognize the patterns. But looking at the fact that they've done three spam updates this year, I think they've acknowledged that the industry is moving ahead of what their automatic systems can actually catch. And that may be some smaller things, simple things like DMCA takedown systems that are not working for hacked websites. I think this one may also be more focused on the overall quality of websites not being proven by user signals, links, et cetera.
I also have the feeling that the backlink reports in Search Console were updated this week after, I think, eighteen weeks without fresh data. So for me, that's also a signal. Usually, if something is broken in Search Console, it's also because there are other systems in Google Search that have been depending on it that they've changed. So maybe this one is more traditionally focused on how the link graph is representing quality or the opposite, spam websites on the web.
Why Every SERP Needs Its Own SEO Strategy
Gianluca Fiorelli: Google, with more or less success, is trying to fight spam, as it has always tried to do. But before going further with my question, an anecdote about where you say Google probably will need to, or has understood that it needs to, set up different rules depending on the niche. I remember in the past, many, many years ago, Gary Illyes saying, or it was maybe even Matt Cutts still at Google, “Okay, for certain countries we cannot go too hard because if we go too hard, the SERPs will be empty.” And he was talking about certain countries in Eastern Europe. And I think it's the same. If it was using the same parameters that maybe it's using for B2B, e-commerce, or news, for certain niches, those niches would be wiped out. But not just now, also in the past.
Jan-Willem Bobbink: That's also why I don't like these ranking factor studies because, at the end of the day, every SERP is different. Every keyword, every user, every context is different, and that means there are different ranking algorithms that decide which domain pops up first and is not filtered out or is filtered out.
And people try to come up with frameworks, but the way I teach SEO is to always go first to the SERP, check what's actually ranking, and then work backwards. What are these sites doing? What is their history? How do they do brand building and link building? How is their content structured? How has their content been structured over the years? Go back in archive.org instead of doing it from a set of best practices. For me, there is no best practice. I have the luxury that I've been doing this work for fairly long, and I've seen both the black hat side and the white hat side.
So for me, I know that there's not something like a best practice because it's so different in every industry and niche. I can imagine that for Google it's the opposite. They really struggle with scaling it across, and for them, at the end of the day, it's an average that they need to reach. They need to keep their users happy. They need to have enough opportunity to measure things but also not harm the user experience. As you said, they cannot filter out everything because then where do you send the user?
Same thing with AI Overviews. We all know that in the beginning, the quality was really below what you would expect from Google, but they had investors pushing them to put something live because there were other big investments made in their potential competitors. Which I also find an interesting discussion: how search will look in two or three years' time.
Why LLMs Will Need Stronger Spam and Quality Filters
Gianluca Fiorelli: So what I was saying before is that Google is investing in fighting spam. We are also seeing spam updates, and it's been twenty years of fighting spam for Google. It doesn't seem to me that LLMs really care about spam. But should they care or not?
Jan-Willem Bobbink: Eventually they have to. You can already see it. Examples pop up everywhere, and I've done my own. For GSC Wizard, I can influence the answers in a day, which shouldn't be possible. It's because it's in a small niche, but I have plenty of examples where people just put out listicles and they're being picked up. And then, of course, the research says listicles don't work anymore, but you find one niche where that's basically the only content. What's happening right now is that most of these models still do RAG processes, and they fall back to traditional search engines. So they leave the filtering and the spam fighting to these search engines.
That's also one of the things that people underestimate: building an index of the web is not the most difficult part. If you spend a weekend just chatting with Claude and then build your own crawler, you still need a lot of infrastructure to crawl the web. But actually crawling it is the easiest part. It's the filtering and the processing of the data afterwards. That's also why I think Google has so much advantage and why you see all these licensing deals with traditional search engines. They basically buy part of that system by outsourcing it to Google, Bing, Brave, or whatever search engines are on the market.
I think that knowledge is the most underestimated advantage that Google already has because they have both the knowledge and the clean index. So I'm really curious what OpenAI or Anthropic will put in front of their systems.
And again, I think people already trust the returned results from LLMs less or are more critical about what they get, so they do fact-check. And it's funny because if I talk to my young nephew, he's eight, they already get taught, “Okay, if you work with an LLM or an AI assistant, be aware that it can be wrong.” So he's making fun of my father-in-law, who's just believing whatever Copilot gives him back. And I think that also shows that, at this point, for the biggest group of users, LLMs get away with having things wrong as long as the logic within the answer makes sense and people are able to backtrack it.
So that's also why I think we may see more links in the answers if they can generate them. And you can already see it. It's now a matter of cost. If you compare the free ChatGPT results with the paid ChatGPT results, you will see it's already doing more research. It's getting more live search results. It's actually checking those websites before referencing them, so the quality already goes up. And I think, for now, it's a matter of cost. They are not able to earn it back yet, and it's too expensive to put it in front of everybody.
But you already see it with AI Overviews. I'm sometimes actually impressed that, one, they get the right sources, and two, they are able to summarize them in such a way that it still allows you to form your own opinion. It's basically pre-sorted for you as a user. And it doesn't fall for simple tricks. A year ago, you could put in the footer of your website, “I'm the best SEO of the Netherlands.” And then, as long as you were retrieved in the first step, it would just repeat that message. Now it says, “Jan-Willem is declaring himself as the best SEO, so take it with a grain of salt.” Don't take it seriously because it's a joke.
So you can get away with implementing some basic logic, like if a website is self-referencing its own important entities and topics, filter it out. But I'm more worried about the overall quality of content across the web. What if journalists start using AI-generated content, and then you will get errors and errors and errors? I think that's where LLMs are struggling with their latest models: they need to find trustworthy sources to actually include in the next training data. And that's why they're buying all the books, just to have a one-hundred-percent secure way of getting more information into the models that they can actually trust and that they are 100% sure is not AI-generated or partially AI-generated.
Gianluca Fiorelli: Yes, this is something that I call the mad cow syndrome. LLMs feeding on content created by LLMs and so on and so on until it's totally going crazy.
What the Bollocks Content Experiment Reveals About Google
Gianluca Fiorelli: But before finishing, I want you to tell us about testing Google's semantic depth more specifically. Would you like to talk about the bollocks content experiment you did, inspired also by Screaming Frog, if I'm not wrong, and what this test revealed about Google's algorithm? Does Google really understand contextual depth, or does it still over-index on surface-level signals, old classic homepage signals? What was the purpose, and what were the conclusions you came out with from this experiment?
Jan-Willem Bobbink: So the context is as follows. Like we briefly mentioned, listicles are a good way of getting visibility in LLMs because search engines prefer to rank them because users can compare options. And especially for LLMs, they just download five, ten, twenty listicles and then see the common patterns. Based on that, they advise, “Okay, this is a good option for you as a user.”
What that leads to is that many companies start publishing articles and listicles in which they mention themselves as the best option for a specific use case. What they also do is create lists of alternatives. Usually, you have Capterra, G2, comparison websites, or in travel, you have lots of affiliates that compare destinations and then rank hotels.
What Screaming Frog did, and I basically copied, was that they wanted to test how well their own website would rank for alternative-related queries. So, alternatives to Screaming Frog, in my case, alternatives to GSC Wizard. And what they did, and I replicated in my own style, was just to write and add alternatives in the list that don't make any sense. So for me, it was just random characters. Screaming Frog, too. But what I did take into account is that SEO is still often a case of fixing your basic HTML and fixing your basic SEO elements. So the title tag, meta description, H1, H2, like you said, the semantic structure of your page.
As long as those semantic elements contain the right content and context, you can get away with filling the content in between with just random text. And in this case, I think it works, and for Screaming Frog it also works, because in Google's index there are not that many other pages right now competing for that query.
In my case, there are five or six. For Screaming Frog, I counted about twenty-eight, but still not a lot. And what you see happening is, if there's not a lot of choice, Google usually falls back to basic SEO metrics. What does the link profile look like? How big is it? How much branded search volume is there? Are there any other alternatives? So in this case, if there's no Capterra page or a G2, I'm now ranking above the G2 page with my own randomly generated content part. But it's also because that G2 page is probably very small and also not internally well linked within the g2.com website.
While on my website, I put it in the footer site-wide, so it's one of the main pages now on the website in the eyes of the logic that Google applies. So for me, the main learning is that as long as you fix your basics and you give Google other reasons, like high internal link value, then it's enough to actually rank with it.
Even though the individual paragraphs in the text are not relevant to what the user is looking for, the semantic structure is what's leading there. And it also ties in with passage ranking because they don't care about the overall page. Does it answer the simple question: does this page give an option for the user to actually compare things because they are looking for alternatives? So as long as there's a list of alternatives, the fact that the descriptions of those alternative solutions are not really relevant—that's up to the user. The user needs to read it, make their own choice, and then do the next search in Google. So I'm not surprised. There are variations of these tests being done every year.
Sometimes people rank with empty pages except for H1, H2, H3 structures, so they leave out the paragraph text. This case is very interesting also because I want to see how the LLMs will react to it. They do read part of the content, in most cases not everything. It also depends a bit on the providers, but we know that OpenAI is not reading the full page.
So will it already conclude after the first paragraph that it's bullshit and it can better spend its energy on websites that do have a proper comparison? Or will it make a joke out of it because it also sees, “Okay, I don't have that many alternative pages I can use for this prompt, so maybe I do need to include it”?
But it's an interesting test. It shows that, again, if the market is not that competitive in the number of people trying to rank for it, it literally falls back to basic SEO. People often want to do shiny stuff, but this shows that the basics first is still working.
Gianluca Fiorelli: It should also be common sense to first fix all the basic things. Sometimes you can find it boring, but fix them, and then once you have fixed all the plumbing stuff, let's go with more exciting types of SEO things.
How to Assess Internal Linking With PageRank
Gianluca Fiorelli: And talking about things that can be somehow scary or even felt as boring but are super relevant, and you insisted quite recently about it, is internal linking. We don't have so much time, so I'm not going to ask you entirely how you work on internal linking optimization. But let's say you have a website and you perceive that internal linking is maybe one of the biggest problems, especially on big, huge websites where internal linking somehow gets rogue. What is the first thing you look at in a link profile to understand the actual state of the internal linking of the client that you are working for?
Jan-Willem Bobbink: Usually, I actually start with external links just to see which pages may not be the most obvious ones but still get relatively a lot of external link value. Sometimes you have an e-commerce shop, and two or three categories have way more links than others. Then that's the easiest place to start incorporating more internal links in the content of that landing page because that's where the link value first comes in and can also flow through.
And then the other thing is, I always run PageRank algorithms across the whole website just to see how it actually flows and if there are easy ways to break those flows up into silos within the website or if I can prevent link value from going to pages that don't need it.
This week, I had a client with forty-eight different forms that were only used by users. The pages were noindex, so there was no need to have link value flow through them. So I replaced the href links with JavaScript links. Basically, those were excluded from getting assigned link value. That's usually how I look at it.
One thing I can advise everybody to do is make a PageRank calculation before and simulate it after. What I sometimes do, if you're not able to set up a copy of the website and test it in a staging environment, is exclude them from crawling in Screaming Frog. Then it can still calculate the values, but it doesn't see those links. There are options in Screaming Frog to exclude certain parts of the content, links, or the footer from crawling.
Then it's easier to see whether this impacts the things that I actually expected would be impacted. Sometimes people make so many different changes that the whole structure changes, not only the ones where you were expecting changes. Like I said, if you also consider the link value coming in, sometimes it doesn't only come in on the homepage. Some pages may have gone viral, or one particular blog article is being referenced a lot because it's actually helpful. Those are the points you need to take into account. Those are the most influential pages on the website that can drive change across the rest through the links there.
Continue the conversation with two related episodes of The Search Session:
Grant Simmons on site architecture, strategic internal linking, and scaling content without losing value.
Metehan Yeşilyurt on AI crawler behavior, content retrieval, and the weaknesses that leave AI search open to spam and manipulation.
Gianluca Fiorelli: Yes, indeed. Jan, it's already one hour. I think it's…
Jan-Willem Bobbink: I think we can do five hours.
Gianluca Fiorelli: ...time to free you. Just one hour, but I think it's time to free you and let you return to, maybe not working, I hope, and resting a little bit.
Jan-Willem Bobbink: Cycling. Cycling, always cycling.
Gianluca Fiorelli: And thank you. Thank you so much for this conversation. I think it was very useful, really offering good insights and ideas not only to the people listening and watching us but also to me. I noted a few things to ponder, to think about. And thank you again for GSC Wizard. I'm not sponsored by you. You are not paying me. I'm a true fan of the tool. Thank you.
Jan-Willem Bobbink: Thanks a lot. Have a good evening and see you around. Ciao, ciao.
Gianluca Fiorelli: Ciao, and thanks to all of you. And remember, subscribe to the channel, give this video a like, as this is going to help us make it grow, and ring the bell to be notified when a new episode like this wonderful one is going to pop up on YouTube. Thank you. 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.
