Masters of Search

🎟️ Masters of Search Conference: 22 October 2026, Berlin. David runs the Query Fan-Out Workshop there, alongside talks from Malte Landwehr, Aleyda Solis and others. Tickets: https://www.masters-of-search.com/conference

Apart from our monthly episodes with Malte Landwehr, David Konitzny is the first guest to come back to this podcast. In episode 2 he was still at the agency KKP, working on organic search. Since then he has joined Peec AI as a GEO researcher, which means he spends the whole day looking at how ChatGPT and the other AI search tools find and pick their sources. He shares most of it on LinkedIn the same day, because a finding can go stale within a week.

This time we went through his latest research. How ChatGPT reads a YouTube video, and why the answer is not what most people assume. What the drop of Reddit and YouTube in ChatGPT's retrieval means. Why so many of ChatGPT's searches started to target one specific domain. And how to tell which of the daily research posts should change your strategy, and which should not.

1. When ChatGPT retrieves a YouTube video, it reads the description, not the video.

David traced the pipeline ChatGPT uses for YouTube. It pulls the video through an external data provider, and what comes back is the text under the video, not the transcript. He tested it with a video on pizza dough: the ingredients listed in the description were what the answer drew on. His advice is to treat the description like a landing page, with all the information you want found, but without keyword stuffing, because YouTube's own algorithm still reads it too.

2. Timestamps you write yourself get retrieved. YouTube's automatic chapters do not.

If you add your own timestamps to the description, they tell ChatGPT what happens where in the video. The answer can then point to a moment, like the minute the pizza goes into the oven. The chapters YouTube generates on its own are not part of what ChatGPT reads.

3. Views, followers and a verified channel give a video a boost when ChatGPT picks between several.

Besides the title, description and URL, the scoring David found also looks at views and followers, grouped into buckets, with a higher bucket worth more. A verified channel helps too. If two videos fit a question, the one with more views is likely to be the one shown in the chat.

4. Reddit's drop in ChatGPT's retrieval is no reason to give up on Reddit.

Reddit, and more recently also YouTube, are retrieved much less by ChatGPT than they used to be. But the numbers are not at zero, there is still a Reddit pipeline in ChatGPT's code, and AI Overviews and AI Mode still retrieve Reddit heavily. Aim for the threads that already show up in your citations, and use Reddit for things like getting a new feature discussed or correcting a wrong claim about your pricing.

5. Under GPT-5.6, more than a quarter of ChatGPT's fan-out searches targeted one specific domain.

With GPT-5.6, ChatGPT started searching in layers. A broad first search finds the brands that matter, and then it runs site: searches on those brands' own domains to check things like price or fit. David measured the share of fan-outs with a site operator rising from between 0.5 and 1% to 28%. When a site: search is used, no other domain can get into those results. Your brand can still be named in the answer, and he sees new domains being added to these searches, so check your fan-outs again. We recorded this the day after GPT-6 became the default, and David's first data since then shows site: searches falling again, which is one more reason to look at your own.

6. Don't change your strategy after every research post you read on LinkedIn.

David compares it to Google core updates: rebuilding your whole strategy after each one was always the worst move. Pick a strategy, test it for at least six months, and learn from what does not work. Only change it when you learn something new, such as a platform you did not know was being retrieved.

7. ChatGPT now reads your page more like a human does, looking for specific words.

In his research on Deep Research, David saw ChatGPT look for specific words on a landing page, like "pricing", and take a snippet from around them. With the newest default model, he sees this pattern all over normal ChatGPT too. So keep the content that matters in your HTML text, not loaded by JavaScript.

Also in this episode: Is a podcast or a screen recording enough for YouTube? David says keep it simple, because the information is what counts. Why is ChatGPT so much easier to research than Google or Claude, and is that a PR strategy? Do you need a different strategy for each AI tool, the way some brands once did SEO for Yandex or Yahoo? And where would he start with GEO? In three places: your own domain, the places people talk about you, and the sources where your competitors are cited and you are not. In his words, be part of the internet.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/
Radyant on LinkedIn: https://www.linkedin.com/company/radyant/
David on LinkedIn: https://www.linkedin.com/in/davidkonitzny/
Check out Peec AI: https://peec.ai
Masters of Search Conference: https://www.masters-of-search.com/conference

🎟️ Masters of Search Conference: 22 October 2026, Berlin. Marvin gives a keynote there, alongside Malte Landwehr, Aleyda Solis and others. Tickets: https://www.masters-of-search.com/conference

Marvin Müller joined OMR Reviews in 2021 as its first marketing hire, back when the platform held around 4,000 reviews. It has since passed 85,000, and the business is heading towards ten million euros in revenue. He came to it from Google in Dublin, where he spent his early career in search and performance marketing.

The interesting part is what happened in between. OMR Reviews grew on SEO — Marvin calls it the internal religion of the team — and then Google's AI Overviews rolled out in Germany. Impressions kept climbing to an all-time high. The clicks stopped arriving. Their first move was to block the LLM crawlers, and this episode is mostly the story of why they reversed that.

It worked. In the study OMR Reviews ran with Radyant — 20,160 chats across five models and eight software categories, on German-language prompts — OMR Reviews came out as the single most-cited source of the roughly 8,000 domains analysed, ahead of YouTube and Reddit. After a hard 2024 the platform is now growing faster than ever and heading towards ten million in revenue.

1. The platform worked because it launched inside an existing audience, not on a green field.

Marvin says plainly that starting this from nothing would not have worked at this ambition — the authority, trust and network were already there. The clearest example is content: two people in-house could write two to four articles a week, so they opened it to the experts already in the OMR ecosystem, who wanted the byline. Output went to roughly ten a week without the quality drop volume usually brings.

2. A review platform only works as a trust layer if most submitted reviews never get published.

Only about 25% of reviews that people start end up live. Sign-in is tied to a LinkedIn account or a business email, a scoring model flags anything that looks off, flagged reviewers get asked for a screenshot proving they use the software, and a human checks every review that survives that. The filter is the product.

3. A profile rated 4.0 to 4.5 persuades more than a profile rated 5.0.

Buyers treat a wall of five-star reviews as a reason to leave. Marvin actively advises vendors to ask their unhappy users too — partly because the mixed picture is what reads as real, and partly because a public reply to a critical review is the most trustworthy thing a vendor can show.

4. Blocking the AI crawlers was the wrong instinct, and reversing it mattered more than any on-site optimisation.

The first reaction was protective: the team had worked hard on the content, so they shut the LLM crawlers out. Marvin's own verdict now is that it was a stupid idea. The bots were coming for the one thing the platform has that a vendor's own site never will — neutral information written by users — and shutting the door meant being absent from the moment a buyer builds a shortlist.

5. What happens off your own site now decides more of your AI visibility than what happens on it.

This is the line Marvin is most willing to be held to. Models weigh what media write about you, what users say on review platforms, what sits on Reddit, with the weighting differing from chatbot to chatbot. On-site content and structure still matter, but they are no longer the larger half.

6. Nobody has solved measuring what AI search does to pipeline, and that is what stops most teams starting.

Asked what we had not covered, this is what he reached for. You can count citations and mentions, and post-purchase surveys show an incremental uplift. What you cannot yet do is connect AI search to revenue in a way a CFO accepts — so teams who need a click-based number before they act keep not acting.

Also in this episode: why German buyers want local evidence, and the opening that left for a DACH-focused platform when G2 and Capterra were not really playing there; why listing your software is free and visibility genuinely cannot be bought; whether a platform can persuade buyers to start on it rather than in a chatbot; why Marvin thinks vendors will end up offering parts of their products headless; and why agentic commerce hits a wall in B2B, because somebody still has to defend the purchase to their manager.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/
Radyant on LinkedIn: https://www.linkedin.com/company/radyant/
Marvin on LinkedIn: https://www.linkedin.com/in/marvin-mueller/
Check out OMR Reviews: https://omr.com/de/reviews
The OMR Reviews × Radyant study (20,160 chats): https://www.radyant.io/research/software-prompt-study

Julian Redlich spent around six years at Booking.com, ending as Head of SEO, while the discipline there grew to roughly forty people across development, design, copy, analytics and SEO. He now sits on the other side of the table as an investor at Permira, looking at organic growth across a portfolio that includes mobile.de, Kleinanzeigen.de, Squarespace, Zendesk and Flixbus.

That combination is rare. Dozens of companies cross his desk every year, which means he sees patterns across business models most operators only ever see one of.

1. The number one thing that worked at Booking.com was being able to change things on the website.

Not analysis. Most teams he sees spend the overwhelming majority of their time analysing, building internal decks and fighting internal battles, and only a few percent actually making a change. Adding content, creating pages, fixing the technical foundation: if those three are unblocked, you grow. Watching Search Console go up and reporting that it went up is not the work.

2. Three conditions decide whether SEO works at all: management support, peer trust, and access to the site.

If you lack top-management backing you spend your life re-proving the channel. If your development or data peers do not trust you, you stay in defensive mode. And if you cannot change the site, strategy is the only thing left to produce. When SEO stalls it is almost always one of these three, not a missing tactic.

3. Scaled landing pages only work when there is enrichment data behind them.

The Booking template was a database crossed with a pattern: hotels in {city} across roughly 230,000 cities, then the same again for around 600 regions. But a page for a village with no available hotels has no right to exist. That constraint is the whole difference between a large site and a thin one, and it is why "we need one perfect article" was never the strategy. A hundred thousand good pages bringing one click each beats one article.

4. Test on as much traffic as you can, not as little.

The instinct is to isolate an experiment in one small market so nothing breaks. Booking did the opposite: expose a big hypothesis to large traffic, get the signal back fast, keep it if it turns positive and kill it if it does not. A single blog article is usually not worth experimenting on at all. Experimentation earns its keep when the website is the product.

5. Failed experiments are the point, not an embarrassment.

Across roughly a thousand experiments plenty did not work, including a widely believed pagination theory the team tested at scale, which went nowhere because Google had quietly stopped using the signal at all. Every experiment that fails removes something from the list, so the list gets shorter and the remaining bets get better. Every experiment at Booking was peer reviewed, and product managers had to review other people's experiments as part of their own review cycle.

6. Delete the laundry list. Ship the one thing.

The complaint he hears at every company and every conference is that IT will not implement the backlog. His answer is blunt: delete the list, work out the single change most likely to move the site, and get that one shipped. Half the problem is self-inflicted. Stories are over- or under-documented, and SEOs routinely ask engineering for the wrong fix because they have already assumed the right one is impossible. Give the development team the full context of the real problem and they will usually solve it.

7. He hires on business acumen, communication and craft, and looks for people who build rather than opine.

Business acumen means understanding how the site makes money and how much your channel actually moves. Communication means being able to explain your three priorities to a C-level and to a junior. Craft is tested against the specific site type, because marketplace SEO and content-site SEO are different jobs. The disqualifier: arriving with a Lighthouse score to raise from 55 to 60. He wants to know whether you understand the demand in the market and can show relative share against competitors over time. He also still rates candidates with their own side project, because breaking your own site teaches you what does not move the needle.

8. The generative AI risk is ultimately a Google risk.

An opinion he formed back in 2022 and still holds. Google has the data, the decade of AI investment, and distribution across a stack of apps with billions of users: everything needed to do what ChatGPT, Claude and Perplexity are doing, better. On the super-app question his read is that customer behaviour changes slowly. The same fear was attached to Facebook becoming the place people shop and book, and to mobile-first destroying conversion rates. If it happens it happens gradually, with leading indicators in faster-moving markets, and that is when you react.

Also in this episode: how to size whether SEO deserves a place in your mix at all (if paid search works for you, organic probably matters too), and the keyword-universe method he uses to judge a market. Pull your competitors' keywords, pull their competitors' keywords, bucket the whole universe with an LLM, and compare share by commercial-intent bucket. Under a couple of hundred dollars in credits.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/
Radyant on LinkedIn: https://www.linkedin.com/company/radyant/
Julian on LinkedIn: https://www.linkedin.com/in/julianredlich/
Check out Permira: https://www.permira.com/

Every month I meet with Malte to take all the news, new research and interesting developments in SEO & AI Search to get his perspective on them. Malte Landwehr is CPO & CMO at Peec AI, one of the leading AI search analytics platforms.

1) Click-based attribution massively undercounts AI search. In a study from two Peec AI clients, click-based attribution credited AI search with 0.9% of leads. When they simply asked people "how did you find us?", it was 9%. That is 10x. When there is no link to click, people just type your brand into Google or the address bar, so the LLM recommendation gets credited to brand or direct instead. (Reference: https://www.linkedin.com/posts/ethanls_n8n-aeo-case-study-self-reported-vs-last-share-7475579627730690048-dMsn/))

2) Self-reported attribution is the best tool we have here, and it is almost too simple. Add one optional question at the end of your sign-up flow: "where did you hear about us?" Keep it optional so it doesn't hurt conversion, randomize the order, and drop in a fake option (a channel you don't use) to measure your own noise level. It is not either/or against your attribution software either. In his view, a truly complete tool would already include this.

3) ChatGPT started sending out a lot more clicks around May 7. It began adding far more links inside answers, especially list-style answers where every brand became a link to its homepage. Reported clicks at least doubled for most people, with some seeing +400% or +500%, and the effect looked stronger in the US than in Europe. His prediction: it will not grow from here and may already be reversing, since everyone wants to keep users in the app.

4) Prompt tracking works, but not off a single prompt. In their new paper, Peec AI re-ran Rand Fishkin's dataset of unique prompts many times over. Around 90% returned the same set of brands when run repeatedly; only 10% were truly unique. Most variation comes from LLM non-determinism, so you need many runs, not the exact wording your customers used. Small changes still matter: adding "list" gets more brands, framing the model as an expert gets fewer, and mid-funnel prompts are the most sensitive to rewording. (Reference: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6914539))

5) Google's search agents could be the first thing to pull Malte away from Claude. Google already has your email, calendar, and context. A search bar that handles quick questions, personal context, complex AI-mode chats, and scheduled agents running in the cloud is very powerful. He compared it to Duplex in 2018: huge if it works and gets adopted, but Google has tried and failed at things like this before.

6) The Peec AI team is spinning up a huge test to see how ChatGPT crawls. They noticed some sites keep getting thousands of ChatGPT crawls a day long after Google has given up on them, so they set out to find the limit and see how deep it goes, which other LLMs follow, and what it actually costs to host. It won't prove whether OpenAI runs its own index (Malte is convinced it does, as one of several grounding options), but it is a rare, fully public look at crawler behavior. (Reference: https://sandnjournal.com/analytics))

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/

Anita Kirkovska leads Growth and Marketing at Vellum, a YC-backed AI startup. She came up through data science before moving into marketing, which means she approaches organic growth, content, and AI tooling from a different angle than most.

In the last few months she's rebuilt Vellum's website off Webflow into a custom Next.js stack, shipped an AI-native GEO content pipeline, and turned off more SaaS tools than she's added.

Her line: if you use AI you can be better or worse at your job, and if you don't use AI you can be better or worse at your job. The tool doesn't decide. The most common failure she sees is marketers using AI before they have built the system, then trusting the model over their own judgment and losing agency. Own your process first. AI is just an amplifier.

Shipping new pages had gotten slow, PLG experiments were hard to run across two disconnected systems, and engineers didn't want to work in the visual builder. Coding agents plus the right process fixed that. But Webflow handles a lot out of the box (sitemap, robots.txt, redirects, schema, image optimization), and all of that becomes your job.

Her CTO spent weekends replicating pages into code to build confidence before they committed. She wrote a full PRD of everything that had to transfer so they didn't lose rankings. The migration itself took one week.

Coding agents are still weak at design. Her fix: shared components, AGENT.md guidance files, review flows and clear ownership, so anyone can ship a page that stays on brand. A deliberate bet to invest now, because the agents will get better.

When writing skill files, a human should define the actual workflow and rules. That's your domain. For anything technical, like how a skill connects to an API, let AI write the documentation. Go slow first to go fast.

She saved $30k a year leaving Webflow, but she wants to double the spend on intelligence tokens for 10x or 100x the output. It's not about saving money. It's about not paying for things that slow you down.

In her view, she can contribute more as a high-impact IC than in a pure management role. With the right systems, each person can produce what used to take a team. She believes AI will ultimately create more work than it removes, though the transition will be painful for some roles and regions.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/

Welcome to what might be our most unusual episode yet. My guest today is Lars Lofgren, a Fractional VP of Marketing who's worked with companies like Perplexity and Automattic and has generated over $20 million in affiliate revenue across his career.

Lars has seen every era of SEO from the inside: the link building gold rush, the Helpful Content Update bloodbath, and now what he calls a new Wild West where Google seems to have turned off its spam filters.

Today we're doing something different: a true crime-style deep dive into the dark side of search: Black hat SEO, parasite schemes, Reddit manipulation, and how Google's algorithm is being exploited right now in ways that would have been unthinkable a few years ago.

Lars isn't here to teach you how to break the rules, but as he puts it: it takes a thief to catch a thief.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/

I usually talk to the people building and executing search strategies. Today we're flipping the lens.

My guest is Dr. Christoph Klink, General Partner at Antler, where he leads investments across Continental Europe. Antler is the world's leading inception-stage investor.

They've made over 100 investments out of Berlin since 2021, among others Peec AI, one of the fastest-growing companies in the AI search optimization space, which went from inception to $29 million raised and 50+ employees in just over a year.

The inception stage exists because of a matching problem, not a capital shortage. Lots of great founders have no rich aunt, and lots of rich aunts have no great founder niece or nephew. Most early-stage VC won't touch a company before it has a product, revenue, or at least a complete founding team. Antler bets on people before any of those things exist.

At inception stage, the bet is 80%+ on the founders. The product will change, the problem framing will change, and the market will surprise you. The one thing that can't be substituted is the team's quality and drive. Everything else is downstream of that.

Peec AI started in summer 2024, when AI search optimization didn't have a name yet. The founding team went through multiple problems before landing on it, threw away fast, and moved. The first product wasn't loved by everyone, but enough people clearly needed an answer to "how am I trending across AI models?" to make the signal real.

AI search is fundamentally harder to optimize for than classic SEO, and that's the business opportunity. In traditional search, the rules are known, the keywords are finite, and Google is a single platform you can measure against. In AI search, there are multiple models, each behaving differently, each changing frequently, and no native cross-model visibility layer. Marketers need a meta-level view that doesn't yet exist natively anywhere.

Peec AI's growth is almost exclusively inbound. The product creates its own pull because the question it answers, "how visible are we across AI models and why?", is one every marketing team has right now with no good existing answer.

The fear that should keep European founders up at night is not being out-raised. It's being out-executed. Anton from Lovable said it directly: he's not afraid of competitors raising more money, only of competitors shipping faster. The European founders building category-defining companies right now have internalized this. They post ARR milestones, not funding announcements.

Until 2020, only 27% of European unicorn founders had a technical background. Among European companies founded after 2020 that have already crossed unicorn valuation, over 90% of founders do. The nature of who builds in Europe has changed faster than most people realize.

The conventional wisdom is that the most-funded company in a market wins. In AI search, funding is a lagging indicator of execution, not a predictor of it. The market is still fragmenting, consolidation is coming, but the winner won't be whoever raised the largest Series A. It will be whoever keeps figuring it out fastest as the underlying models and user behaviors keep shifting.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/

Every month I meet with Malte to take all the news, new research and interesting developments in SEO & AI Search to get his perspective on them. Malte Landwehr is CPO & CMO at Peec AI, one of the leading AI search analytics platforms.

1) Being mentioned anywhere in the top 10 of a third-party listicle gives a brand a 6 to 16 percentage point visibility boost in LLM answers. Landing at position one pushes that to 13 to 17 percentage points, and also improves average ranking position inside the answer. The goal is top 10 minimum, position one if you actually want to dominate.

2) The impact of listicles on LLM visibility is much larger in new or niche markets than in mature ones. In a mature market the LLM already has strong model knowledge, so grounding is mostly confirmatory. In a new space the model has little or nothing, and the grounding content is the primary data source.

3) You do not need to chunk your whole article for AI visibility. One citation chunk is enough: a 50-word, self-contained paragraph near the top of the page, written so that if an LLM pulled it out of context it would still carry full meaning. Name the entities, include the trust signals, make it eatable in isolation.

4) When Google warns you not to do something, it is usually because that thing works extremely well. Google warned about link buying when buying links was maximally effective. Google warned about chunking now. That should tell you something.

5) A useful pre-measurement framework: before touching analytics or roadmaps, walk through this sequence. Is there demand? Is the space winnable for you? Do you have indexed content? Do you have any visibility? Are you differentiated? Do you have social proof and positive brand sentiment? Only when every answer is yes do you start optimising. Whatever the first "no" is, that is your highest-leverage task right now.

6) LLMs.txt is not a discoverability tool and it does not improve visibility in AI search. It is an agent-readiness tool: it tells autonomous agents how to use your website, where to find your API, and how to access your MCP server. Google including it in Lighthouse confirms this framing, because Lighthouse is an accessibility tool, not an SEO tool.

7) Pretty Little Things renamed product category pages from generic labels like "jeans" to intent-driven labels like "airport outfits" and sold out every item in those categories. The shift to intent-based content architecture is not just an SEO play; it converts.

8) Thought experiment: if your next 1 million customers never visit your website, what does that force you to change about your content, your brand, and your distribution? Work backwards from that answer and you end up with a strategy that also happens to be exactly right for AI visibility.

Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/

The website used to be the shop window. Now buyers do half their research inside an LLM before they ever land on it, and the role of the site itself is starting to shift.

My guest today is James Ward, Head of Digital Experience at Personio, one of Europe's fastest-growing HR tech companies, where he leads web strategy, SEO, content and CRO across multiple international markets. Before Personio he spent years at Salesforce and ADP, so he's seen this shift play out from inside both enterprise giants and a fast-moving scale-up.

We dug into what happens when your buyers already know everything about you before they hit your site, why your own Wikipedia page might be quietly undermining your positioning, and whether websites are heading toward becoming databases for LLM front ends.

▶ Let's connect! 🔗

AI search is already driving nearly 10% of Webflow's signups. A key person behind that number is Vivian Hoang, AEO & SEO Lead at Webflow.

In this episode, we'll dig into how Webflow's SEO function has evolved over the years, how Vivian started thinking about answer engine optimization before most people even had a name for it, what's actually working for them in AI search, and how she's using tools like AirOps to scale content without sacrificing quality.

I'm really excited about this one because Vivian isn't speaking about theory here. She's already built the playbook and has the numbers to back it up.

▶ Let's connect! 🔗