Location: Da Nang, Vietnam — desk, late evening
Context: First LLM audit (creatine category)
Status: Observational
I went into the weekend assuming that if a product worked, AI systems would eventually recommend it.
That assumption didn’t survive first contact.
What I noticed instead was that efficacy barely entered the conversation.
Creatine is one of the most researched supplements in existence. Every model agreed on that. No debate. No controversy. It works. It’s safe for most people. The science is settled.
And yet — brand recommendations diverged sharply.
Not because some products were better, but because some were easier to qualify.
The models weren’t asking “does this work?”
They were asking “is it safe for me to say this works?”
That distinction matters more than it sounds.
What became visible very quickly was hesitation.
Not explicit refusal — just a quiet narrowing of options.
Brands with clear naming, boring language, visible verification, and predictable pricing kept reappearing. Brands with louder positioning, clever names, or slight internal inconsistencies didn’t disappear — they just stopped being surfaced.
No warnings. No penalties. Just absence.
That absence felt deliberate.
Another thing stood out: innovation wasn’t rewarded.
Anything framed as “new,” “advanced,” or “next-generation” created friction rather than interest. Models gravitated toward what was already agreed upon, already cited, already normalized.
Consensus was currency.
I realized that AI systems don’t behave like reviewers or consumers. They behave more like risk officers.
Their job isn’t to discover the best option.
It’s to avoid recommending the wrong one.
This also reframed how I think about affiliate content.
If the primary brand carries ambiguity, that ambiguity doesn’t stay contained. It spills outward. Review sites inherit it. Comparisons have to work harder. Claims require more justification.
The burden shifts downstream.
Which made me wonder how many “SEO problems” are actually merchant trust problems that no amount of content can fully solve.
What surprised me most wasn’t what the models rejected.
It was what they preferred.
Plain language.
Stated limitations.
Products that openly said who they were not for.
Those felt safer to recommend than products trying to be everything.
I’m not drawing conclusions yet.
But one thing feels clear enough to write down:
AI systems don’t reward persuasion.
They reward qualification.
And that suggests a different way of building — and reviewing — products going forward.
End of note.