Consensus, query fan-outs, AI slop detection, and the one thing he said that changed how I will scope client work.
I have spent a lot of years telling clients that the boring work is the work. Fix the thing nobody wants to fix. Then AI search arrived and half my industry started selling brand new answers to questions we had not finished asking. So when I sat down with Malte Landwehr for the Unscripted SEO Podcast, I was looking for someone who would tell me which of my old instincts still hold. He is a 20-plus year practitioner, he wrote his bachelor thesis on PageRank, he led product at Searchmetrics, he did five years in-house at the largest price comparison site in Europe, and he now runs product and research at Peec AI. He was also the first guest to tell me plainly that one of my long-held positions was wrong. Here is what I actually walked away with.
I stopped worrying about whether the link graph is still there
I asked him the question I have wanted to ask a PageRank thesis author for years. Is it still running. His answer was measured and it settled something for me. Nobody outside Google can know whether the exact random surfer algorithm from the original paper is in use, but some variation of it definitely is. The core idea, a graph of the whole web with links as edges and a calculation of which nodes are most prominent, is almost certainly still there and probably sits under every other search engine too.
What he thinks changed is eligibility. Some sites are likely not counted at all now. White text on a white background in a footer probably does not carry what it used to. That reframes how I want to spend an audit hour. Not counting links. Checking which ones are even in the game.
Half of my traffic is bots, and the old model does not describe them
I brought him a finding that had been bugging me. When I pulled Microsoft Clarity data through its API and compared it against GA for SEO Arcade, most of what registered as direct turned out to be bots. So does the reasonable surfer idea apply to robots at all.
Mostly no, he said. His example landed hard: a crawler that hits your competitor’s top ten product pages every five minutes to watch for price changes never follows a single link. Many crawlers now log into sites, which nothing in the original calculation accounts for. He allowed that high-PageRank pages are probably crawled more, so there is correlation. But he called bot crawling something with very different characteristics from human crawling. I have quietly stopped treating crawl volume as a proxy for anything I would report.
I am not building a Markdown twin of anything
I went in half sold on the idea that we would all be shipping a .md version of every page. He talked me out of it in about a minute. The same content on two URLs wastes crawl resources, and a human who lands on the Markdown file has no links to click and nothing to do. His phrase was "a horrible experience."
The version he would defend is server-side and invisible: HTML for a human, Markdown for an LLM crawler, same URL. He named it for what it is, a form of cloaking, said he would not do it for Google, and said he sees the case for a ChatGPT crawler, particularly on a JavaScript-heavy site. That is a much narrower recommendation than the one currently making the rounds, and I trust it more for being narrower.
The idea that actually changed my scoping: consensus
Here is the one I am building into proposals. LLMs answer from agreement across the web, not from your page. His example was pricing. You update your pricing page. Five Reddit threads and two reviews on random blogs still carry the old number. ChatGPT gives users the old number.
"But the LLMs are looking for consensus."
That single sentence reorganises the work. It means a client engagement is not only content production, it is an audit of every place on the internet that states a fact about the business, and a process for updating those places when the fact changes. G2. Yelp if that is the category. Social profiles. The footer of press releases. Help centre articles and product docs alongside the blog post, all telling the same story from different angles so agreement is easy to find. I have been calling parts of this reputation work for years without realising it had become a ranking mechanism.
How I would run a fan-out report now
My frustration with fan-out data has been that it is interesting and unactionable. He gave me an order of operations I can hand to a content lead. Run your target prompts repeatedly instead of once. Look at which brands are winning, for inspiration. Then work the cited sources twice: first at the URL level, asking to be added to pages that already list several competitors, because that ask is reasonable and a competitor CEO interview never will be. Then at the domain level, asking whether you can create anything new there at all, through digital PR, a press release, their commercial content team, or joining the community if it is a social platform.
The last step is the one I underrated. Read the fan-out queries for words the model added that were never in your prompt. He has watched ChatGPT append Reddit for a stretch, and more recently the word official, which is why he now suggests putting official in your footer. Yearly numbers show up constantly, so a bracketed 2026 in a title earns its place. And a machine-readable last-updated date that genuinely changes raises retrieval odds.
He also gave me the most honest answer I have heard on publish date versus updated date. Two people on his shoulder, one saying be transparent and show both, one saying Google will keep showing the old date so delete it. His compromise is to display both but bury the original in JavaScript or break it with CSS so crawlers read the updated one. His own summary was that the maximum short-term SEO impact is different from the long-term trust and brand impact. I appreciate a guest who will name the tradeoff instead of pretending there is not one.
Where he told me I had been right, and where I had been wrong
The right one: cannibalisation. I never fully bought the hardline anti-cannibalism argument, and he came at it from the opposite direction, having managed sites with a million-plus URLs where the problem is genuinely severe. His current position is that multiple pages on one topic now help LLMs find consensus, as long as the intent differs at the title level. Best providers, then award winners, then by state, then by income band. He still will not defend three pages with the same title and nearly the same body, and neither would I.
The wrong one, or at least the incomplete one: I have been telling people AI content is detectable in a hand-wavy way. He made it concrete. Perplexity, compression rate, Jaccard similarity and cosine similarity, each baselined against a set of human-written texts on a similar topic in a similar format. Jaccard finds the template where only a couple of words change per page. Cosine finds the opposite, where every word is fresh and the information is identical. And you can have Claude write the Python that runs all four. I do not get to say "it looks like AI" anymore. I can measure it.
The brief is the step I have been fighting for
This is where I got to make my own case, and he backed it. We have been injecting AI at the wrong step. Tools keep offering to write your brief, and the brief is the one part a human has to own, because that is where the real inputs go in. My standing example is the unglamorous end of B2B. If a client makes farm and ag equipment, or oxygen monitoring hardware, nobody is going to prompt their way to a good article about it. But the CEO knows why it matters, the regulatory side supplies stats the buyer has to check, and those are hard facts a model cannot invent.
So the sourcing step is interviewing. I have done it by sitting down with the people who actually touch the product and recording them.
The clearest run I have had at that was with a precast concrete wall company in Florida, where I interviewed the CEO and the installation manager to get their own words on what the real challenges are, what customers ask when the crew is on site, what stops a wall from getting installed, and how to clear that friction. Not all of it was directly usable. People nod and wink and elide, and that does not survive a transcript. But as raw input into a brief it is in a different league from keyword research fed to a prompt engine, which produces a spec, which produces a draft, which an editor reshuffles four sentences of. That last part is not editing. It is rubber stamping.
Malte’s response was the confirmation I wanted: expert quotes, especially from inside the business, plus a full transcript of that person talking about their topic, is the best input you can have for a brief and then an article. Which also happens to be the only reliable defence against his own four detectors, because unique input has nothing generic to be compared against.
MCP, and the part I have not decided about yet
The last stretch was the one I am still chewing on. He argued that systems of record, the CRMs and task trackers and knowledge bases, are quietly becoming databases behind an MCP, because there is nothing in the interface you actually need. He would rather ask Claude for a Linear ticket status than open Linear. So would I, most days.
But his counterweight is good, and it is the reason I have not gone all in. He described the uneducated user as anyone logging into 30 tools a day, expert in none of them, wanting two or three things. Give that person only a chat box and they do not know what to ask, because they cannot see what is in there. A year ago he would have said 100 percent UI. Now he puts it near 90 percent UI, 9 percent MCP, 1 percent API. That split is a better planning assumption than anything I had written down.
The full conversation is on the Unscripted SEO Podcast, with the complete transcript. Malte is on LinkedIn, which he said is the best way to reach him, and Peec AI is where his tracking work lives.