A tactical read of the Unscripted SEO Podcast interview with Malte Landwehr of Peec AI.
I book guests for this show on one test: can they give me something I can hand a client on Monday. Malte Landwehr cleared it inside ten minutes. He has been doing SEO for 20-plus years, wrote his bachelor thesis on PageRank, co-founded an agency, led the product team at Searchmetrics, then spent five years in-house at the largest price comparison website in Europe. Now he runs product and research at Peec AI, which measures visibility inside LLM answer engines. So this recap is the tactical cut. What he told me to do, what he told me not to touch, and the numbers he is watching while he says it.
PageRank did not die. It got pickier.
The first thing I wanted from a guy who wrote a thesis on PageRank was a straight answer on whether it still runs. He gave me the honest version.
"We cannot know if the exact algorithm from this random surfer paper that Google published a long time ago is being used. But some variation of it is definitely used still today."
The core principle he is describing is the one worth keeping: build a graph of the whole internet, treat links between sites as the edges, then calculate which nodes are the most prominent. That, he says, is surely still in use at Google and probably at every other web search engine in some form.
What changed is the input filter, not the math. Some websites are probably not counted at all anymore. Links sitting in white text on a white background in a footer are probably not counted as strongly as they once were. Practitioner takeaway: stop arguing about whether links still work and start auditing which of your links are eligible to be counted in the first place.
Your crawl logs are not a reasonable surfer
I brought him a live problem. When I compared GA data against Microsoft Clarity for SEO Arcade, a large share of what looked like direct traffic turned out to be bots crawling, scraping and digesting pages. So I asked whether the reasonable surfer model applies to robots. His answer was mostly no.
His example: a crawler that checks every five minutes whether a competitor changed the price of their main ten products. It refreshes the same URLs forever. It does not care about links and it does not follow any. On top of that, many crawlers now log into websites, which the original PageRank calculation never contemplated. He allows that high-PageRank pages are probably crawled more often, so there is correlation. But he called bot crawling an approximation of user behaviour with very different characteristics.
The practical consequence: do not read crawl frequency as an authority score, and do not build a client report around bot hits. That is exactly the kind of number that looks like performance and is not, which is a habit worth breaking before it reaches a slide. I have written before about measuring the wrong things for precisely this reason.
Serve Markdown by user agent, not at a second URL
Publishers have started serving Markdown for machines, and I asked whether a parallel .md file is where technical SEO is heading. He shut that down fast. Creating example.com/test.md alongside example.com/test is the same content on two URLs.
"It wastes crawl resources. If humans land on the .md version, there are no links to click. There’s nothing for them to do. It’s a horrible experience."
The version he can defend is server-side. When a human comes, serve the HTML. When an LLM crawler comes, serve the Markdown, under the same URL. Not a redirect, not a second address. He is clear-eyed that this is a form of cloaking, and he would not do it for Google. For a ChatGPT crawler, he sees the case, especially on a JavaScript-heavy site with client-side rendering.
One caveat he flagged on the Time magazine version of this story: they also inject ads specifically for the LLM. That is a different animal, because the human never sees the ad. His read is that a product manager at OpenAI or Anthropic eventually says this is not okay, or demands the ads be wrapped in markers so the model can ignore them. For now, he noted it is as far as he knows the only way anyone is monetizing bot traffic at all.
Consensus is why your old pricing keeps winning
This is the section I would read twice. Matt Brooks gave me the line on an earlier episode about treating LLMs like an uneducated support rep who needs constant training material, and Malte liked the metaphor enough to say he would steal it. But he immediately narrowed where it applies, and the narrowing is the useful part.
You do not win by publishing as much as humanly possible, because at some point you are publishing low quality and duplicate work. The mechanism is different.
"But the LLMs are looking for consensus. So if you only talk about your pricing on your pricing website, and then you change your pricing, and then there are five Reddit threads and two reviews on random blogs that still talk about your old pricing, ChatGPT will answer with your old pricing if a user asks about it."
So the deliverable is not another blog post. It is a distribution checklist for every fact you care about. Do you have a G2 profile. Do you have a Yelp profile if that is your category. When something changes, does it get changed there too. Put your core message in the site footer, on every social profile you own, and in the footer of your press releases. For a new feature, ship the blog post and the product page and the help centre article and the product docs, each telling the same story from a different angle, so the models can find agreement quickly. If you have followed Jason Barnard on entity optimization, this will sound familiar, but the pricing example is a sharper way to sell it internally.
Five steps for working a query fan-out
I asked the question every consultant eventually hits. Fine, there are 75 fan-out queries for "podcast producer." What do I write in the action items column. He gave a repeatable order of operations.
- Run the prompt set multiple times. Not once. Pick the prompts you want the brand known for and repeat them.
- Study the brands currently winning, purely for inspiration.
- Mine the cited sources at the URL level. For pages you are not on, ask whether you can be added. His trick: prioritise pages that already list several of your competitors, because asking is reasonable there. A page that is an interview with a competitor’s CEO is a lost cause, so skip it.
- Then work the domain level. Can you create new content on that site at all. Digital PR, press releases, their commercial content team, a paid article, an advertorial, affiliate. If it is a social platform, join the community, or do an AMA.
- Last, read the fan-out queries themselves for common concepts, especially words the model added that were never in your prompt.
That last step is where the fresh tactics come from. If half the fan-outs contain the word review, reviews are now your topic. There was a period when ChatGPT appended Reddit to nearly everything. More recently he has seen it appending the word official, which is why his current advice is to put the word official in your site footer. Yearly numbers show up constantly, so adding "in 2026" or a bracketed 2026 to article titles can work. And because these systems favour fresh content, refreshing a page and making sure there is a machine-readable last-updated date that actually changes raises the odds of retrieval.
Dates and cannibalisation, both loosened
On publish date versus updated date, he described two people on his shoulder. One says be transparent and show both. The other says Google will keep using the old publication date in the snippet, so delete it. What he would actually do is display both, but push the original publishing date into JavaScript, or break the date with a line break and strip it back with CSS, so the updated date is the one every crawler sees. His own framing was that the maximum short-term SEO impact is different from the long-term trust and brand impact.
Cannibalisation moved too. He was firmly in the anti-cannibalism camp, and with good reason after working sites with a million-plus URLs where ten thousand landing pages chase one topic. Now he thinks multiple pages on one topic help LLMs reach that consensus, provided the intent differs. Ten best health insurance providers, then award winners, then best per state, then best by income band. Overlapping, but distinct at the title level. His limit was blunt: he is still not a fan of three pages with the same title and almost the same content.
Four numbers that catch AI slop before you publish
His answer on AI content came in three parts. If you only care about short-term AI visibility and are willing to lose your Google rankings, publish a thousand articles a day, and he admitted he hates that it works, especially on an older established domain. If you are extremely cautious, publish none. Most people live in the middle, and that is where the guardrails matter. The idea that your AI slop has a fingerprint is not a vibe, it is measurable, and he named four measures.
- Perplexity. How predictable the next word is given the previous ones. Baseline it on human-written text first. If your AI text comes back a lot higher, change the prompt and humanise it.
- Compression rate. Borrowed from email spam filtering, because spammers pad messages with filler. Roughly, how many words can be removed without losing information. A generic "write 600 words about this topic" with no data supplied often scores badly.
- Jaccard similarity. Catches the template. Same sentence in every article with a couple of words swapped. His example was a run of city pages where only the capital, population and main industry change.
- Cosine similarity. Catches the opposite failure. Every word is unique but the information is identical across the set.
He was explicit that this is not a research project. You can tell Claude to write the Python script that checks all four, and you do not need to understand the maths underneath to run it in-house on your own content. That is a same-week deliverable for any content team publishing at volume.
The other half of his answer was the workflow. Instead of one prompt that writes 600 words, chain it: a briefing, a second prompt that checks whether the briefing makes sense, a briefing per paragraph, one prompt per paragraph, a fact-checking prompt, then checks for empty paragraphs and repetitive concepts. He put the token cost at five to seven and a half euros per piece, and said the output is usually really good. He also gave the line I keep coming back to on why generated text struggles to earn a ranking:
"Basically, if you can create it with a prompt, why would ChatGPT or OpenAI or Google crawl it, index it, and rank it? They could just use that prompt on their own."
Put the human at the brief, not the edit
I made the case that we have been injecting AI at the wrong step. The tools want to write your brief. The brief is the one part a human should own, because that is where the data and the quotes get in. An editor moving four sentences around at the end is not editing, it is rubber stamping. Malte agreed without qualification, and put it in production terms: expert quotes, especially from inside the business, plus a full transcript of that person talking about their topic, is the best possible input for a brief and then an article.
That is also the honest defence against his own detection stack. Unique input is what the four measures cannot flag, because there is nothing generic to compare it against.
MCP is turning systems of record into databases
He has added an MCP to Peec AI, and some customers now use the product primarily that way. His own habits have shifted the same direction. He is more likely to ask Claude for the status of a Linear ticket than to open Linear, and more likely to have Claude search Notion than to open Notion.
"I think all systems that have this character of being a system of record, like a CRM, task management, knowledge management, I think these are becoming basically databases for an MCP. Because why would I log in? There’s nothing there for me that I need."
He is not calling the death of the interface, though. His counterweight was the uneducated user, by which he means anyone who has to log into 30 tools a day and is therefore an expert in none of them. With only a chat box, you often do not know what to ask, because you do not know what is in there. Navigation and charts still do that job. A year ago he would have said it was 100 percent UI. Today he puts it at roughly 90 percent UI, 9 percent MCP, 1 percent API, with the direction of travel obvious.
The short version
- Audit which of your links are eligible to be counted, not just how many you have.
- Do not treat bot crawl volume as an authority signal or a client KPI.
- If you serve Markdown, serve it from the same URL by user agent. Never build a .md twin.
- Fix consensus, not just your own page. G2, Yelp, socials, press release footers, docs.
- Work fan-outs in order, and harvest the words the model adds on its own.
- Run perplexity, compression rate, Jaccard and cosine against a human baseline before you publish at scale.
- Own the brief. Source real quotes and transcripts into it.
Full interview and the complete transcript are on the Unscripted SEO Podcast. You can follow Malte on LinkedIn, which he named as the best place to reach him, and see what Peec AI tracks at peec.ai.