I have been designing interfaces for banks and fintechs for over 10 years. Onboarding flows, approval screens, transaction dashboards. I have been through usability reviews with product teams, accessibility audits with regulators, and tests with real users who do not understand why a button says "Continue" when there is nowhere to continue to.
I decided to apply the same principle to my own site. But instead of hiring a QA team, I used three AI models with a single prompt: destroy the user experience. Do not compliment the design. Only problems.
ChatGPT, Perplexity, and Gemini audited the same 4 pages with the same prompt, word for word. The scores were brutal. And revealing.
The scores
Four pages. Three models. Twelve scores.
The homepage received 6.2 from ChatGPT, 4 from Perplexity, and 4 from Gemini. The blog post received 7.1, 5, and 5. The llms.txt Generator tool received 5.4, 3, and 3. The search received 4.8, 2, and 2.
The first interesting finding: Perplexity and Gemini coincided exactly on all 4 scores. They did not coordinate. They do not share architecture. But they evaluated the severity of each problem on the same scale. ChatGPT was consistently 1.5 to 2.8 points more generous on every page.
The second finding: the search page, which is the most technically sophisticated feature on the site (a recursive CTE in MariaDB that traverses a graph of 80 connections), received the lowest score from all three. Technical complexity does not compensate for a confusing user experience.
Where all three agreed
Four problems appeared in all three audits without exception.
First: public metadata exposes the database. Internal IDs (P062, P069) and raw view counts (22 views, 43 total) appear in the public interface. All three audits flagged it as database output disguised as interface. It provides no useful information to the reader and communicates "this is a developer prototype" instead of "this is a designed product."
Second: the search displays internal architecture without explaining it. The graph scores (16pts, 14pts), degrees of separation, and connection labels are comprehensible to whoever built the system. Not to someone searching for an article about JSON. Gemini identified a concrete example: a post with "JSON-LD" in the title appears below another that does not have it, because the graph connections accumulate more points. The internal logic is correct. The user experience is not.
Third: the llms.txt Generator tool does not respect the navigation language. If someone arrives from an English ChatGPT citation, they land on a Spanish interface. It is a localization bug that all three detected.
Fourth: the homepage has no orientation for new visitors. All three audits used variations of the same critique: a first-time visitor arriving from an AI citation does not know whether this is a blog, a portfolio, a tool hub, or a consultancy site. There is no visible "start here."
Where they diverged
The divergences are as revealing as the agreements.
ChatGPT was structured but moderate. Its critiques came with context. "The value proposition is vague" followed by exactly which questions it does not answer. Its feedback was the easiest to convert into actionable tasks because it identified the problem, explained why it matters, and suggested the direction of the solution.
Perplexity was philosophical and evaluative. It did not list individual problems. It described the overall impression each page creates and why that impression fails. "The site seems to have strong opinions and lots of content, but not enough guidance for strangers." Useful for understanding macro perception. Less useful for knowing what to change tomorrow.
Gemini was the most technical and the harshest. It cited specific DOM elements. It identified that plain text glyphs (▾, →, ✦) are used as interface decoration instead of scalable SVGs. It flagged that the share section uses raw minimalist characters (𝕏, in, WA) without clear visual CTAs. It found the search bug where "JSON-LD" in the title loses to a post without that word due to graph weight. Of the three audits, it was the most useful for direct implementation.
What I discarded
All three insisted the site should be a consultancy portfolio with case studies, testimonials, client metrics, and a clear conversion path. "After reading three pages I still don't know: Can I hire you? What services do you sell?"
That critique projects an assumption about the site's purpose that is not correct. shinobis.com is a technical content blog and free tools hub. It is not a consulting services sales site. The models evaluated the site against a criterion I never established. That does not invalidate the UX observations. But it invalidates the conclusion that the site "fails" at positioning services it does not offer.
Perplexity and Gemini gave 2 out of 10 to the search, which is excessive. The graph search does something no other personal blog does: it shows why a result is relevant, not just that it is. The problem is not the concept. It is the presentation. Internal scores and degrees should not be visible. But the underlying functionality does not deserve a 2.
Gemini went into accessibility implementation details (aria-live regions, DOM insertion focus management) that are valid as individual bugs but are not UX architecture problems. They are items for an accessibility sprint, not for a redesign.
What I implemented
From the three audits combined, I implemented four changes ordered by impact.
I removed internal IDs and raw metrics from public views. Post IDs (P062, P069) and contextless counts (22 views) no longer appear in the interface. Ten minutes of work in the PHP template. Immediate impact on the perception of professionalism.
I rebalanced the search scoring. The graph is still the engine, but now a direct title match has a guaranteed minimum boost. If the user searches for "json" and a post has "JSON-LD" in the title, that post cannot rank below one that does not have it, regardless of how many second-degree connections the other accumulates. Internal scores are no longer visible in the interface.
I fixed the tool localization. The llms.txt Generator now detects the navigation language and serves the corresponding interface.
I added an orientation block to the homepage for new visitors. A visible section that answers three questions: what this site is, who it is for, and where to start.
What I learned about AI audits
Three models auditing the same site with the same prompt produce three different severity levels. ChatGPT tends to moderate its critiques even when asked for brutality. Perplexity evaluates general impressions. Gemini goes into technical implementation details.
If I had asked only ChatGPT, I would have thought the homepage was at 6.2 and the search was a reasonable 4.8. If only Gemini, I would have thought everything needed rebuilding from scratch. The reality is that the real problems are specific and solvable, not systemic.
The pattern is the same as the JSON-LD audit: you do not ask one AI. You ask three. You implement what overlaps. You evaluate what diverges. You discard what projects incorrect assumptions about your product.
The most revealing thing was not the problems they found. It was the irony. A UX designer with 10 years of experience in banking interfaces had database IDs visible in the public interface of his own blog. Not because he did not know it was wrong. Because when you are the builder and the designer of the same product, blind spots multiply. You see the architecture. You do not see the interface.
The three AIs saw the interface. That was their value.