In two previous posts I documented that four AI models build different profiles of the same person. Perplexity defined me as an "SEO specialist." Gemini called me an "independent web developer." Neither could cite where that information came from. The phantom pointer.
I wanted to test the other side. If instead of relying on compressed memories, I gave them directly the file I designed for them to understand me, would the result improve?
I copied the complete content of my llms.txt and pasted it to Claude, ChatGPT, Gemini, and Perplexity with the same prompt. Five questions: what this site is about, who it is for, what its unique value is, what question would trigger a citation, and what the file is missing.
All four correctly identified me as a UX/UI designer with fintech and banking experience. The same AIs that with compressed memories got my profession wrong, with the llms.txt got it right.
The identity correction
This is the strongest finding of the experiment. Two weeks earlier, using only their internal memories, Perplexity said I am an "SEO specialist / digital marketer / web auditor." Gemini said I am an "independent web developer." Both were confident.
With the llms.txt, Perplexity wrote: "a UX/UI designer with banking and fintech experience documents how they integrate AI into design workflows." Gemini wrote: "an experienced UX/UI designer that documents practical workflows within banking and fintech contexts."
Same model. Same person. Different identities depending on whether it reads a structured file or recalls from memory. A 3KB plain text file communicates who I am more clearly than six months of conversations compressed into a profile.
Where all four agreed
All four identified three elements as the unique value of the site.
First: the intersection of practicing designer with GEO implementer. All noted that most GEO content comes from SEO marketers, not from designers who build their own site. Claude was the most precise: "it sits at an intersection almost nobody else occupies: designer + frontend builder + GEO practitioner + multilingual."
Second: free tools with no registration. All four mentioned the llms.txt generator and GEO Tarot cards as concrete differentiators.
Third: the experiential perspective. Not a blog that explains concepts. A blog that documents real implementations in first person with measurable data.
Where they diverged
Claude was the only one that distinguished between being cited as a practitioner anecdote and being cited as an authoritative standard. "I would reach for it as a practitioner anecdote, not as an authoritative standard. For 'what is llms.txt' in the abstract, I'd cite the spec's originators before a personal blog." No other model made that distinction. The other three treated the blog as a potential source without qualifying the category.
Claude also detected a real bug in the file: the Spanish homepage description rendered as "Blog principal en espa彨衞l" instead of "Blog principal en español." UTF-8 corruption that no other model caught. An AI parser reading that text sees broken characters, which undermines the credibility the rest of the file builds.
Gemini was the only one that requested explicit licensing terms and a preferred citation format (exact markdown structure or CC-BY license). A product perspective the other three ignored.
Perplexity asked for separation between what parts of the content are opinion, which are experiment, and which are reproducible methodology. A distinction that connects directly to the critique from the phantom pointer post: do not mix observation with inference.
The gap between the owner and the machines
I answered the same 5 questions without reading the AI responses. The difference was revealing.
I described the site in 6 words: "UX/UI + AI, the future." All four AIs wrote detailed paragraphs about GEO, JSON-LD, banking/fintech, vanilla PHP, no-registration tools, and content structure for citation.
When asked who it is for, I answered: "Anyone who does not want to follow the easy path but wants to experiment." The AIs answered: "UX/UI designers, frontend developers, digital creators working in fintech/banking or multi-language environments."
I see the philosophy. The AIs see the implementation. Neither is wrong. But an llms.txt needs to speak in the language of machines, not in the language of the creator. Machines do not cite philosophies. They cite specific facts from verifiable implementations.
What I am changing
From the four audits combined, five concrete changes.
Adding the author name. Claude flagged it explicitly: "No author name. It says 'a UX/UI designer' but never names the person, so there's no authority to attribute a claim to."
Adding publication dates per article. Without dates, a model cannot evaluate content freshness in a field that changes every week.
Adding a concrete finding per article. Not just the title. One line with the main result: "This post documented that 4 of 5 AIs accessed the real endpoint, one hallucinated everything."
Fixing the UTF-8 corruption in the Spanish text. A silent bug that no human saw because nobody reads the llms.txt in a browser.
Adding a section of verifiable credentials. Not "10+ years in banking/fintech" without evidence. A link to LinkedIn, to freeCodeCamp, to the GitHub repos. Proof that the person behind the content exists and has the track record declared.
The file that defines you better than yourself
The paradox of this experiment is that a 3KB plain text file tells AIs who I am with more precision than six months of conversations. Not because the file is perfect. Because the file is structured and conversations are noisy.
Compressed memories mix facts with inferences, professional data with personal data, and build a different profile in each model. The llms.txt does not mix anything. It says exactly what I want machines to know and nothing more.
That is the lesson. If you want AIs to understand you, do not rely on them remembering you. Give them a file that tells them who you are. They will read it more faithfully than any memory they build on their own.