You Rank Consensus, Not Pages: Malte Landwehr on LLMs

August 10
43 mins

Episode Description

Your pricing page is not what the model is reading. Malte Landwehr of Peec AI wrote his bachelor thesis on PageRank, ran in-house SEO at Europe’s largest price comparison site, and now measures brand visibility inside answer engines for a living. His argument: LLMs resolve to consensus, so five stale Reddit threads and an old G2 profile will happily overrule the price you changed last week.

Arcade chart: up to 3% of the ChatGPT prompts Malte Landwehr tracks fan out to a site: query against a domain that does not exist. Those domains sit parked, and anyone could register one. Key takeaways
  • One canonical page cannot create consensus. The same fact has to be findable in enough places that agreeing with you is the cheapest option.
  • He is against a separate .md URL and for serving markdown by user agent. Same URL, different response. He also calls it cloaking, out loud, and says he would not do it to Google.
  • Fan-out queries are a placement map, not a keyword list. Target the cited pages that already name several of your competitors.
  • Watch the words the model adds that you never prompted. The pattern rotates; the method is to keep watching, not to memorise this month’s word.
  • Two live manipulation surfaces, both already measurable. Hallucinated domains inside fan-outs, and paid advertorials being treated as grounding sources.
On this page Why consensus beats your canonical page Quote card: If you can create it with a prompt, why would ChatGPT or OpenAI or Google crawl index and rank it and use it? Malte Landwehr, Peec AI

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.

— Malte Landwehr

The remedy is distribution of a single fact rather than optimisation of a single page. G2, Yelp, your own footer, social profiles, press release boilerplate, the help centre, the product docs. The goal is to make agreeing with you cheap to verify.

LLMs are looking for fresh content. So refreshing content, updating it, and then also making sure there’s a machine readable last updated date.

— Malte Landwehr
Against a second URL, for serving by user agent

I would not do that. As an SEO, I don’t want the same content on two different 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. What can make sense is that my server makes a decision. When a human comes, I serve them the HTML version. And when a LLM crawler comes, I serve them the markdown file under the same URL.

— Malte Landwehr

He does not pretend this is clean. He names it as a form of cloaking and says he would run it for LLM crawlers and not for Google. He also supplies the fix for the two-URL version if you insist on it: set the canonical in the HTTP header, because a text file has no HTML head to put one in.

They also inject ads specifically for the LLM. And that is again going very much in direction of cloaking.

— Malte Landwehr
Fan-outs as a placement map

I would look at the sources both on the URL and on the domain level and see for these existing URLs, can I get my brand mentioned there if I’m not mentioned yet? And the trick is often to look for the ones that have multiple of your competitors already mentioned because then it’s often reasonable to contact and say, hey, can I also be added?

— Malte Landwehr

The corollary saves you a quarter of wasted outreach: a page that is an interview with a competitor’s chief executive is unwinnable, so do not spend anything on it. Sort the cited sources by how many rivals they already list, and work down.

The terms you did not prompt

Especially terms that the LLMs are adding that were not part of the prompt... A few days ago, ChatGPT started adding the term official to a lot of fan out queries. So right now I recommend everybody to put the word official in the footer of their website.

— Malte Landwehr

Cheap, testable, and explicitly time-bound. He is clear these patterns rotate, which is the actual instruction: watch the added terms, do not enshrine one of them.

Manipulation surfaces and AI-content guardrails

There are some prompts I’m tracking where it’s in the range of 2 or 3 % of the prompts have a hallucinated domain. In these cases, the domain is just parked. But if I was an evil person, I would now register this domain or buy it and put up some completely negative content about the brand that supposedly owns it.

— Malte Landwehr

He raises a second surface alongside it: in an insurance prompt set he monitors, roughly 2% of the sources an LLM leans on are paid advertorials. Humans skip advertorials. Models quote them.

When AI content is worth publishing

If you can create it with a prompt, why would ChatGPT or OpenAI or Google crawl index and rank it and use it? They could just use that prompt on their own.

— Malte Landwehr

His acceptable cases are all data-backed: summarising real reviews on a product page, or writing up structured data nobody else holds. And if you want to know whether your own output is detectable, he names the four measures to run over it.

Perplexity... compression rate... jacquard and the other is cosine [similarity]. And if you use these four measures you will often find that there are many, many very easy to detect footprints in AI written content.

— Malte Landwehr

Benchmark all four on a corpus of your human-written text, run them again on the AI-written batch, and compare. He notes you can have Claude write the Python and that you do not need the maths to read the result.

Chapters
TimeWhat happens
00:00Twenty years of SEO, and a PageRank thesis
01:48Is PageRank still running inside Google?
03:37Bot crawling is not the reasonable surfer
05:57Serving markdown to LLM crawlers, and why a parallel .md URL is not the way
08:57Time magazine, and ads only the model sees
12:55Hallucinated domains in ChatGPT fan-outs
14:23Advertorials used as grounding sources
16:55Consensus: stale threads beat your pricing page
19:43The five-step query fan-out method
24:21Cannibalisation loosens up
29:34Perplexity, compression rate, Jaccard and cosine
38:02MCP turns systems of record into databases
People, ideas and sources mentioned
EntityWhat it is
Peec AISoftware for measuring and improving visibility inside LLM answer engines
Malte LandwehrTwenty-plus years in SEO. Agency co-founder, product lead at Searchmetrics, five years in-house at Europe's largest price comparison site
Query fan-outThe expansion of one prompt into many underlying searches, including terms the user never typed
Time MagazineCited as serving markdown site-wide and injecting ads into the model-facing version
Perplexity, compression rate, Jaccard, cosine similarityThe four measures he recommends for finding AI-content footprints in your own writing
Hallucinated domainsParked domains appearing inside fan-out site: queries, and an open manipulation surface
Questions this episode answers
  • Should I publish a .md version of every page for LLMs?
  • How do I set a canonical on a markdown file that has no HTML head?
  • Why does ChatGPT give an outdated fact about my company?
  • What do I actually do with a list of query fan-outs?
  • How can I tell whether my AI-written content is detectable?
  • Are advertorials being used as grounding sources by LLMs?
  • Is cannibalisation still a problem in the era of AI search?
Go deeper

Every link above goes somewhere different. These are the ones not already mentioned above.

Watch the interview Find Malte Landwehr

peec.ai, and he is most responsive on LinkedIn.

Cite this episode

Landwehr, Malte. Interviewed by Jeremy Rivera. “Consensus, Query Fan-Outs and AI Content Guardrails.” The Unscripted SEO Podcast, 10 August 2026. https://unscriptedseo.com/episode-recap-malte-landwehr-of-peec-ai-on-consensus-query-fan-outs-and-ai-content-guardrails/

See all episodes