|
Almost Timely News: 🗞️ How I Proved Listicles May Harm Your GEO (2026-06-21) :: View in Browser
 |
The Big Plug
👉 My new course, GEO 201 on competitive GEO measurement, is now for sale.
Content Authenticity Statement
75% of this week’s newsletter was made by me, the human. You’re going to see a lot of Claude outputs as part of the testing. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.
Watch This Newsletter On YouTube 📺
Click here for the video 📺 version of this newsletter on YouTube »
Click here for an MP3 audio 🎧 only version »
What’s On My Mind: How I Proved Listicles May Harm Your GEO
This week, let’s talk about some articles that have been going around about the impact of listicles on GEO/SEO. If you are unfamiliar, the term listical means list article. That is a long article that is nothing more than a large list, popularized by places like BuzzFeed and adopted by all kinds of companies as a type of content marketing.
The one that caught my eye first was by SEO expert Lily Ray, who penned an excellent piece that started everything, that her observations empirically were that listicles were helping competitors more than the companies who published them. She found in her research that alistical made competitors more prominent in citations and results. 69% of the time.
I thought that interesting, and wrote up my own LinkedIn post on the topic, which I’ll reproduce here:
Why are your listicles helping your competitors?
Research from folks like Lily Ray and Rand Fishkin are pointing out that listicles might be doing more harm than good for GEO/AEO. Why?
Step back and think about how large language models work. They’re semantic engines, paying attention to not only words, but word order and the relationship of topics to each other. Unlike traditional SEO, LLMs don’t have a concept of things like page rank or validity; one pile of text is equivalent to another. The East Peoria Evening News is just as credible as the New York Times.
Now, suppose you created a listicle that proclaimed you were the best, and all these other competitors were pretty good. In isolation, the concepts in there would be ingested by an LLM and understood. Today’s reasoning models are excellent at understanding context and meaning, and for a while, LLMs would recommend and cite you as the best.
But that’s in isolation. Nothing happens in isolation. When LLMs inside a harness go and grab a bunch of data, they aggregate it together, look at it in context, and draw from the aggregated corpus.
Which means if you’re company A and you create a listicle with companies ABCD in it, and company C creates theirs with CFGD, and company H creates theirs with HJFD, and company K creates theirs with KQMD, what happens?
In AGGREGATE, company D has the greatest prominence in the corpus. It’s mentioned the most, and usually in positive terms. Company F is mentioned second-most in aggregate, also in positive terms. As all that data gets tokenized, processed, and regurgitated, companies D and F - who didn’t even make listicles - benefit from the greatest prominence in the result set, even when companies A, C, H, and K all proclaimed they were the best.
And an additional twist - even in one listicle, the amount of information density about you versus ALL the air time you gave competitors doesn’t work to your advantage.
That’s how LLMs work. They work with text, with semantics, with concepts, and with topical density. So if you’ve been cranking out listicles and you don’t know why your competitors are winning, it’s because of a tragedy of the commons. As more people latched onto “make listicles and win”, it made things worse for all the participants because that’s how AI works. Even if a company is dead last on a listicle, if every listicle mentions them, they have the greatest information density in the result set.
Mitch Joel said it best - it’s not who you know, it’s who knows you. In GEO/AEO, it’s not who you talk about, it’s who talks about you.
Crappy drawing by me.
I published this, and Matt Trifiro commented:
“This seems like a contrived scenario with no evidence...”
Matt was correct in that this post was based on my knowledge of generative AI and how LLMs work, and I didn’t present any evidence. So I commented back:
“Matt Trifiro I do like the idea of taking a corpus and running some diagnostics on it. Maybe I’ll give it a try over the weekend.”
And that’s how we arrived at this week’s issue.
Part 1: Designing the Experiment
One of the great dangers of generative AI is that it’s incredibly easy for us to introduce our biases to it. If I had said to a tool like Claude Code, “prove that I’m right”, it would engineer a sophisticated, compelling, persuasive and intellectually dishonest take. It would design tests and charts and graphs that all looked amazing, and it would use all the symbols of authority to persuade you that I was right.
If we value ego, that’s the approach to take, to be incurious, to tell AI to be our cheerleader, to make us look good.
If we value truth, that’s the approach to avoid.
In this case, I value truth. I genuinely want to know whether listicles help or hurt GEO/SEO.
I started by laying out to Claude that I wanted to build a test setup which would use Google’s Gemma 4 model, mainly for efficiency and no cost, to construct listicles and test my hypothesis, that LLMs would, in aggregate, reinforce the competitor more than us once everyone started making listicles.
We came up with 3 different kinds of tests. The first is BM25, which is a classical machine learning algorithm that search engines used in the old days to compare different texts and see which was the most relevant. The second is a knowledge graph using the old PageRank centrality measure. The third is an LLM’s recommendations.
Part 2: Anticipating Biases
One of the things generative AI has built into it are biases of literally every kind. Every name, every noun, every topic has some pre-existing connotation based on the training data set. In an experiment like this, where we want to measure the impact of the listicle itself, we have to take those biases out.
Lily herself flagged this in her article - LLMs like Google’s Gemini may have biases towards known and trusted brands. If we want to test our hypothesis that the listicle itself, the architecture of information around them, is what’s at play, we have to work with completely fictional, synthetic brands that have no name recognition of any kind.
To do this, I relied on the inherent biases of the English language and had Claude choose company names from non-Western word sources - Japanese, Korean, Indonesian/Malay, Swahili, Sanskrit/Hindi, Tagalog, Arabic, Zulu/Xhosa - and then measured whether Gemma 4 had any pre-existing associations.