The same built-in randomness that makes AI useful for some work makes it exploitable at scale. Best-of-N jailbreaking works across text, image, and audio. The same brute-force attack that cracks a language model cracks visual and audio models (just with different noise).

Spotify only needs 30 seconds to count a stream. LLMs only need one passage to answer a question. A Datawrapper analysis of 10 years of Spotify chart data shows songs and titles are getting shorter. Every element of the format shrank to match what the platform actually rewards. The customer journey

Your AI tool is literally warning you it can be jailbroken. This warning exists because vendors know Best-of-N attacks work. They know malicious prompts can trick models into harmful output. They know copyrighted content can be extracted under the right conditions. They warn you.

Most pages are too long for AI agents to read fully. They don't tell you. They just truncate or skip. Alex Breitenbach built an extension that shows you the token count of any page you're on. Simple fix for a silent problem. Check the thread to download it.

This is what most client calls feel like lately. Jamie Indigo and I are launching a podcast together (we're not but we do have the promo photos for it). Thank you Martin Splitt for this photo session.

TikTok taught artists to engineer their best 20 seconds. Many artists optimize for the fragment that survives the scroll. Brand content has the same problem. LLMs are teaching brands to engineer content with the same mindset.

Marketers, you never had human data. You have human-shaped data. 1. You track clicks, not decisions. 2. You track page views, not attention. 3. You track conversions, not satisfaction. I talk about this at #OMR: the end of the human-only web, choice architecture in the agentic web.

Rankings don't exist in AI search...is true if you're looking at one prompt. Ask the same category question across hundreds of prompts, platforms, and phrasings. Certain brands surface consistently. Others never do. That's a ranking. It just doesn't fit in a spreadsheet

@johnmu.com @stephaniewalter.bsky.social fuck it. I'm trying some zucchinis this year.

Most brands don't have a content gap in AI search. They have a retrieval gap. The ranking can still be there. The citation won't be. Dan Shure tells you about it:

One of the most recognized brands in the world doesn't own its own story in AI search. A complementary business does. A business that benefits from the brand's success. It's writing the "why it matters" layer. Brand prestige does not automatically become citation authority. The brand layer is vital.

3 years of Reddit threads, Bloomberg investigations, grocery TikToks saying "Doritos got greedy" are now LLM training data. ChatGPT learned that Doritos is the brand people switched away from. Good luck cleaning that up.

Nobody likes PDFs. Yet every contract, every compliance document on the planet is one. RAG is heading to the same place in enterprise AI. EU AI Act takes full effect August 2026. RAG allows instant data purging. What survives a compliance audit wins the budget conversation.

The Grand Tour gave us TripAdvisor. AI search is the same inheritance. Anthropic's f***s chart flags male technical rage as the universal frustration signal. GEO for brands that serve anyone else (♀️ 🏳️‍🌈, etc.) means handling a legibility problem inside a system not built to find you.

LLM copywriting best practice #7: Scroll down twice on your homepage so the hero banner and primary H1 disappear. Start reading from there to get a feel for what LLMs deal with. The system cannot identify what you sell without the design context it cannot see so this is a good test.

LLM copywriting best practice #6: lead with your answer. LLMs reliably extract claims near the beginning or end of a text. Open with a dense 40-60 word declarative statement that answers the who, what, why, or how. Name the entity. State the relationship. Include the condition. Then add context.

LLM copywriting best practice #5: vagueness is a disqualifier. Does your copy name a specific subject? Does it quantify the claim? Does it state the condition under which it's true? If not, the retrieval system has nothing to grab. Want to improve? Check how to write for AI search.

LLM copywriting best practice #4: label your sections. H2s and H3s are retrieval signals for machines. A heading above a paragraph improves its relevance score for AI systems by up to 17.54%. No new content required, just correct the structural wrapper (your headings) around what already exists.

LLM copywriting best practice #2: write for fragments not sequence. Most mid-page copy fails because it was written to be read in sequence, not extracted in fragments. The isolation test is the fastest way to find where your content has a utility gap. The test is easy & free.

LLM copywriting best practice #3: state relationships, don't list entities. LLMs need 4 things: a named entity, an explicit relationship, a preserved condition, and a specific detail. "Our platform makes project management easier" fails all four. Read on to find concrete examples.

LLM copywriting best practice #1: don't write more, write tighter. Content density beats length in AI retrieval. You get around 380 words to make an impression. A single self-contained sentence with a named entity, a stated relationship, and a specific condition outperforms 3 paragraphs of fluff.

Aesthetic Study · Dawn of the Eurozone This little blurb: If your local coffee shop has started to look like a Ryanair gate in 2003, you’re not alone. We’re in the era of Barcelonacore with folks cosplaying European summer as defined by Puma Speedcats, low-rise cargos, and sheer tanks.

Absolutely freakin’ loving this: the awesome and insanely smart and funny @myriam.dopamine.builders wrote about my utility-writing framework: searchengineland.com/ai-search-pl...

Parametric memory is the vast repository of information an LLM "absorbs" and stores directly within its internal weights and parameters during training, essentially acting as the model's built-in, long-term intuition. That's it. Now you know more about AI "memory".

Physics says time is a static block, not a river. Humans project "flow," but LLMs don't. To an AI, a 2017 brand disaster is happening RIGHT NOW. Stop waiting for time to heal your reputation; the machine has no clock. There is no 'fading into the past' inside a training set. Read about brand drift:

Most people want a "good grade." Pros want to find the bugs, figure out what doesn't work, get better, build something worthwhile. Assume bugs. Entropy is part of the journey. feeds.feedblitz.com/~/939887294/...

A hallucination cost a company 25 million in sales. The solar contractor Wolf River Electric dealt with Gemini hallucinations. Gemini hallucinated a non-existent lawsuit from the Attorney General. Competitors used that fake news. It snowballed from there. That's another brand drift issue.

AI memory is a recursive trap. It draws on a "cached" version of you from months ago, locking you in a cocoon of familiar, safe, same...same...same. I don't want to be defined by my past biases. Growth requires letting go.

My brain beats an AI because it's lossy. Perfect recall isn't the target. We can only simulate the future because our past isn't fixed in stone. Forgetting is what makes imagination possible. True intelligence isn't defined by what you store, but by how efficiently you discard the noise to hone in.

Infinite memory is a prison. To grow, we must forget. Tech giants want AI to have infinite memory but high-performance intelligence requires curated erasure. We need tools built for revision, not just retention. If AI doesn't know how to let go, it's an anchor dragging us down.

The first "Schema.org" involved a tomato in 1750 Carl Linnaeus gave us binomial nomenclature, the original language for nested hierarchies. Taxonomy isn't just for plants or librarians. I think 2026 is going to be about structuring the web to make it clearer for humans, crawlers and AI agents.

Musing inspired by Carl Linnaeus. He invented index card because he was drowning in global data. Systems aren't for productivity; they’re for survival. Stop storing info. Start arranging it, remixing it. The value isn't in what you know. It's in how expertly you can remix it.

Update: I got the pattern in my head now, so I can just be in the flow and continue working on it. Also apparently I now rewrite patterns on a napkin to adapt it for my niece, who have I become? Please enjoy said niece, looking fab in a dinosaur hat. ROAR.

Executives are professional rationalizers. They are paid to remain calm. If you present a "declining trend" or a "inefficiency," their psychological immune system kicks in to explain it away ("it's seasonal," "market headwinds," "we're pivoting"). That is Region Beta.

When Google’s AI happily invents special deals, that’s AI brand drift: the gap between what you actually offer and what the machines confidently say you do. AI systems will define your brand for you...and not always in ways that are accurate, fair, or even commercially viable. ​

LinkedIn AI content musing: if you don't write it, you don't understand it. If it’s not worth writing yourself, it’s probably not worth reading IMO. Your "messy" original thought is the asset. AI text is a "Frankenstein’s monster" of other people's data. It's a commodity with no authorial intent.

The art of delivering bad news in a nutshell: 1. Don't bury the lead or minimize the loss. 2. Explain the context with data. 3. Numbers come with feelings attached. Validating frustration neutralizes the emotional impact. 4. Go from "we failed" to "we learned X, so now we do Y."

so he doesn't have a fractured hand. the clinic's AI hallucinated it. literally thousands in lost flight, a painful trip to the ER, a cast-like bandaid forrrrrrrrrr nothing. the fractured hand had healed normally.

Stop using "Explain Like I'm 5." It insults your audience. Instead, treat them like a smart peer on their first day: capable of understanding, but lacking context. Your goal is to orient, not dumb down.

Emotion is the new moat. Features are easy to copy. Feelings aren't. Andrew Hogan argues that "resonance" is the only sustainable competitive advantage left. If your product doesn't make users feel something, it is vulnerable to the next clone.

Musing of the day: connecting dots across domains is now more valuable than deep expertise in just one.

Digital PR is not just "SEO with better storytelling." It focuses on audience, narrative, and brand authority. 1. The "Why": Are you fixing a reputation or building one? 2. The Audience: Who are we talking to? 3. The Timeline: 3-month pilots rarely work. You need 6-12 months for impact.

Hey @johnmu.com weird question. I see the "catch-22" for sites when it comes to stuff like the Stanley toolkit: being penalized for something that we can't technically fix on our servers. This has ome "security as UX vibes". Curious to know what you think. www.darkreading.com/remote-workf...

Forget 4K, for AI search optimization, visual anchors are priority #1. Don't block UIs with your head in Loom demos, rotate products slowly for 3D modeling, and keep labels camera-facing. Consistent brand codes help models identify you across different videos. Get more good advice here:

The "semantic gap" between image and text is gone. Google’s Gemini 3 Flash uses reasoning to zoom, inspect, and perform visual math. Read my article on image AI search optimization best practices: searchengineland.com/image-seo-mu...

I'm a #UKContentAwards judge this year! My judging lens for quality content this year: 1. Does the content actually support a clear business goal? 2. No fluff: is the content noise or something that can be scanned by humans and machines? 3. Does it deliver zero-click value and high-intent utility?

"Brand Drift" happens when AI doesn't have enough facts about you, so it assumes you're like your competitors. High-quality video provides the "ground truth" proof to stop AI hallucinations. There are many reasons for that, check out why here and get tips on how to optimize videos in 2026.

Jump cuts are killing your AI search efforts. AI models sample video at ~1 frame per second. If your info flashes too fast, the AI misses it and "hallucinates" your brand details based on competitors. Slow down your roll, quite literally. Want video optimization tips that work in 2026? Read this

I have a world war II question: too many old undetonated things from the war end up in men's rectums and they go to the ER in France. Is this just a French thing. Also why does it keep happening?

Your voice is a highlighter for AI. Use audio bolding: 1. Enunciate: Speak key terms and brand names clearly. 2. A short pause before and after a main point acts like a "comma" or a "period" for the AI. The cadence of speech influences the tokenization. Check out this article on the topic: