Three AI auditors reviewed my website: what they found and what only I could see

I had three AI auditors review my own website at the same time, and in under 45 minutes they handed me 45 prioritized issues, each with the file, the line of code and the exact fix. The costliest one was in plain sight: my main call-to-action button, green with white text, had a contrast ratio of 2.28:1. Accessibility guidelines ask for at least 4.5:1.

My site, carlosbetancur.co, “looked fine.” I opened it every day on my laptop and saw nothing serious. That is the trap behind most websites that look good and don’t convert: the owner sees them on one screen, in one browser, always through the front door.

This piece is for business owners and marketing leads in that situation. I explain how I set up the audit, what it found, how I decided what to fix first and what the AI could not see. At the end there is a list of 15 usability problems almost every site has, which you can check this week.

How was the audit set up?

I did not ask an AI to “review the site.” That request returns generic advice that fits any website. I ran three auditors in parallel, each with a different role and the same deliverable format.

Auditor	Focus	Time	Deliverable
1	Usability, navigation, conversion and mobile	23 min	Top 15 with root cause at file:line and exact fix
2	Visual design and premium feel	32 min	Same format
3	Reading experience, blog, content and the English version	44 min	Same format

The shared format is what makes the exercise useful. Every issue had to include severity, page, screen width, the cause in the code and the exact fix. That way the three reports can be cross-checked to see what more than one auditor flagged.

Then four implementers, also AI agents, applied the fixes in Spanish and English. The first change touched 91 files.

What did they find?

These are the findings that come with a number. Some were embarrassing, which is exactly why I am publishing them.

Finding	Measured	Flagged by
Main button, green with white text	2.28:1 contrast (now 5.02:1)	Auditor 1
Grey used for secondary text	4.29:1 (now 7.39:1)	Auditors 1 and 3
AI keynote speaker page on a phone	509 px wide on a 390 px screen	Auditor 1
Mobile menu button	Cut off by 21 px on every page	All three
Desktop menu	Broken between 1,025 and 1,170 px wide	Auditor 1
Hero buttons on mobile	Below the fold	Auditors 1 and 2
Prompts in my 50-prompt guide	49 of 50 cut off on desktop, 50 of 50 on mobile	Auditor 3
First prompt in that guide	11,086 px down the page	Auditor 3
AI consulting service page	18,000 px without a contact button	Auditors 1 and 3
Blog lists	No bullets in 68 of 78 articles	Auditors 2 and 3
Sharing an article on WhatsApp	Counted as a lead in GA4 and Meta	Auditor 1
Emojis used as icons	172 replaced with custom icons	Auditor 2
Links inside English articles	Between 46 and 51 sent readers to Spanish pages	Auditor 3

Two of these deserve a closer look.

The menu breaking between 1,025 and 1,170 px sounds like an edge case. It is not. A 1,440 px laptop with system zoom at 125% renders the web at 1,152 px. At that width, the WhatsApp button in the menu slid off the right edge. I never saw it because I don’t use zoom.

The tracking error was the worst one. Every time someone shared an article on WhatsApp, my analytics logged a lead, and so did Meta. Meta uses that signal to decide who sees the ads. A share button was teaching the platform to find people who share, not people who hire.

How did I decide what to fix first?

With 45 entries on the table, several repeated across auditors, order matters more than speed. I used three rules.

First, whatever several auditors repeat goes to the top. All three independently reported the cut-off menu button, and all three flagged the same unstyled summary block on five pages. When three independent reviews agree, it is no longer an opinion.

Second, objective issues before taste. Contrast, overflow, invisible buttons and broken tracking can be measured before and after with a number. Aesthetics come later.

Third, revenue pages first. Speaking, consulting and training pages before the blog.

The AI did not make the aesthetic calls. The design auditor proposed directions; those became three concepts, A, B and C, and I picked B. AI is good at proposing. The person who knows the brand and the client decides.

What could the AI not see?

An AI can measure contrast to two decimal places. It cannot know whether a photo of me looks the way I want to be seen.

The new design called for blending my photo into the hero, which meant removing its background. It was done twice, the second time with Apple’s edge detection and contour cleanup. The AI reported the job as done. To me the photo looked worse than before: the edge gave the cutout away. We went back to the original photo, background included, framed within the grid, across the whole site.

The book covers went the same way. The 3D mockups had grey boxes and ragged edges. After four attempts to clean those cutouts, we changed approach: the book is now drawn in 3D with CSS, with cover, spine and page edge, on top of the original flat cover. When four attempts at the same approach fail, the approach is the problem.

There is one more thing the AI does not see on its own: how the site looks on a real phone. Auditor 3 discovered that headless Chrome, asked for a 390 px window with --window-size, never goes below 500 px. “Mobile” screenshots taken that way come out cropped and show a site that seems to work. You need real device emulation.

What we learned

“Done” only counts once I verify it. The AI reported the photo and the books as finished before they were. Since then, “done” only counts once I have seen it myself in the browser, on mobile and on desktop.

Caching lies. After the covers changed, the browser kept showing the old ones. One whole commit was just renaming the images to force a fresh download. If you don’t see a fix, don’t assume it isn’t there; if you do see it, don’t assume your client does.

A scheduled job can build from stale code. On 4 October, from 7:02 to 7:35 a.m. Colombia time, carlosbetancur.co was offline. It was not the redesign. A scheduled job, the one that publishes queued articles every morning, was building from a copy of the code dated 24 September. That stale copy clashed with the new server setup and the site ended up redirecting to itself in a loop. An external uptime monitor alerted me. The safeguard I added: every deploy now refreshes the copy that job uses, and the version it builds goes out first with no traffic; it only receives visitors if several pages respond correctly. If not, the site stays on the previous version.

The auditor can be wrong too. I repeated the method on btodigital.com, my agency’s site: 78 pages at 6 screen widths, 468 views, all now at zero overflow and zero cut-off text. On that same site, contrast measured in a real browser had gone from 176 failures to 0 across 41 pages. But the automated contrast checker was blind to gradient backgrounds and passed them. Once that was fixed, 31 real failures surfaced that the zero had been hiding.

Not everything is perfect. The main secondary grey improved, but at least one grey on the site still sits at 4.29:1, just under the minimum. I will not claim the whole site meets the standard.

What business results did it produce?

None I can honestly show yet. The fixes went live on 3 October, and the first useful measurement needs at least three weeks of data. I will look at four things:

• Real WhatsApp leads from the site.
• The share of those leads coming from mobile.
• Book downloads.
• Leads from the AI keynote speaker page, which was broken on mobile.

There is a trap worth naming. The baseline is contaminated: before the fix, every shared article added a fake lead. If I compare the old event with the new one, leads will “drop.” The honest comparison is against stored form submissions and real conversations.

How can you apply this in your business?

The recipe fits in five steps:

1. Split the review into roles: usability and conversion, design, content. A single generalist auditor produces a generic list.
2. Require a shared format: issue, page, width, cause and exact fix.
3. Rank by agreement between auditors, then by what is measurable, then by revenue pages.
4. Leave aesthetics to whoever knows the brand. Ask for two or three concepts and choose.
5. Verify yourself, on a real phone or with device emulation, before calling anything done.

And here are 15 usability problems almost every site has. Go through them with your team, on mobile and desktop:

1. The main button has contrast below 4.5:1 (green with white text rarely passes).
2. Grey secondary text falls short of 4.5:1.
3. The menu button does not fit on a 360 to 390 px phone.
4. The desktop menu breaks at in-between widths, such as a laptop at 125% zoom.
5. The main hero button sits below the fold on mobile.
6. A page is wider than the phone screen and forces sideways scrolling.
7. Long pages have no contact button halfway down.
8. The floating WhatsApp or chat button covers text or form fields.
9. Lists lost their bullets to a CSS reset.
10. Tables or code blocks are cut off with no hint that there is more.
11. The table of contents is longer than the first paragraph.
12. A conversion event counts actions that are not leads, such as shares.
13. The translated version links back to pages in the original language.
14. Emojis are doing the job of icons on buttons and cards.
15. Forms have no labels, don’t validate email, or show errors far from the field.

My site had all 15. If more than five apply to yours, it only “looks fine” on your own screen.

Frequently asked questions

Can an AI audit a website’s user experience? Yes, if you give it a role, a method and a format. In my case three auditors delivered 45 prioritized issues with root causes in the code, and several overlapped. What it cannot do is make aesthetic decisions for you or confirm a fix landed without someone looking.

Why three auditors instead of one? Each role looks at different things, and agreement between them is the best priority signal. All three independently reported the cut-off menu button.

What contrast should a button have? WCAG level AA asks for 4.5:1 for normal text and 3:1 for large text. My green button with white text was at 2.28:1; a darker green brought it to 5.02:1.

Can I let the AI apply the fixes on its own? For objective issues, yes, with verification afterwards: contrast, overflow, tracking. For aesthetics, no. And never accept a “done” without checking it on mobile and desktop.

How long does an audit like this take? The auditors took between 23 and 44 minutes. The main fixes shipped the same day. Verification is what takes the longest, and it is the one step you cannot skip.

Does this increase sales? I don’t know yet and I won’t claim it. What I do know is that my tracking was inflated and my speaking page was broken on mobile. I will publish the first results check once I have enough data.

If you want your site reviewed with this method, that is how I work as a digital marketing consultant. On the conversion side, see raising conversion without buying more traffic, and the page this audit changed the most is my Claude for marketing guide with 50 prompts.