We Need to Talk About AI's Accuracy Problem in UX Analysis
AI tools for UX analysis are everywhere now. ChatGPT prompts for heuristic evaluations. Automated CRO suggestions. AI-powered UX audits. The pitch is compelling: instant insights, faster decisions, democratized expertise.
But here's what too few are talking about: most of these AI tools are wrong 25-50% of the time.
When a single bad UX change can cost millions in lost revenue, those odds aren't just problematic — they're dangerous.
I've spent the past year diving deep into this problem and testing what it takes to build AI tools that are actually safe for commercial use. Read my full article, including our complete accuracy testing methodology and raw data.
The promise vs. the reality
As someone who's spent 15+ years conducting large-scale UX research and UX auditing of Fortune 500 websites, I've been fascinated by AI's potential to accelerate UX analysis. The promise is compelling: instant heuristic evaluations, faster insights, democratized access to UX expertise.
But there's a fundamental question we're glossing over: How accurate are these AI tools, really?
When we first tested ChatGPT 4.0 for heuristic evaluations 1.5 years ago, we documented a 20% accuracy rate. I assumed things would improve quickly. They have improved, but not nearly enough.
In March 2025, two Microsoft UX researchers adopted a similar study approach to see how far generative AI and LLMs have come in 2025 for conducting heuristic evaluations. They found that the AI tools had Accuracy Rates of just 50%, 62%, 67%, and 75%.
Let me be direct: a 50-75% accuracy rate is dangerous for commercial websites.
Why accuracy isn't just a nice-to-have
Here's what keeps me up at night about low-accuracy AI tools:
In UX and CRO, even tiny changes create massive financial impact. I've seen clients gain a 1% conversion lift from switching image gallery dots to thumbnails. I've watched a $10 million annual revenue increase from duplicating a single button. I've witnessed a 90%+ mobile abandonment rate caused by poor form field indicators.
The problem? You can't easily tell good UX from bad UX just by looking at it. Neither can AI.
So when an AI tool (with a 50-70% accuracy rate) gives you 10 suggestions — where 5-7 are genuinely helpful but 3-5 will actively harm your conversion rate — and you can't distinguish between them, what do you do?
You're essentially being asked to implement changes where you're guaranteed to hurt your business, just to capture some improvements. That's not efficiency.
The trade-off no one talks about
Here's what I've learned building UX-Ray — our AI-powered heuristic evaluation tool — over the past year:
With AI tools, there's a brutal trade-off between accuracy and coverage. We could make UX-Ray assess 86 UX heuristics if we allowed a 70% accuracy rate. Instead, we're releasing it with just 39 UX heuristics at 95%+ accuracy.
That decision cost us months of additional development and over $100,000 in accuracy testing. Some may think we're being overly cautious. But I've seen too many businesses lose millions from implementing plausible-sounding but ultimately harmful UX changes.
The maths is simple: The revenue lost from one wrong UX suggestion far exceeds the cost of ensuring your tools are properly validated.
What should you demand?
If you're evaluating any AI tool for UX analysis or CRO — whether it's a sophisticated platform or a "premade prompt" circulating on LinkedIn — ask one question:
"What is your documented accuracy rate?"
Not their marketing claims. Not cherry-picked case studies. Their actual, tested, documented accuracy rate compared to human expert evaluation across a broad range of websites.
If they can't show you that documentation, or if their accuracy is below 95%, you're taking an enormous risk with your website revenue.
We've made our accuracy rate testing public. You can see our raw line-by-line comparison across 45 websites. You can download every screenshot we tested. I believe this level of transparency should be the industry standard.
Why this matters beyond just one tool
I'm not writing this just to talk about UX-Ray. I'm writing this because I'm genuinely concerned about the direction our industry is heading.
I see more AI UX tools launching every week. I see more "AI-powered CRO prompts" being shared. And I see very few people asking about accuracy rates.
We're at risk of creating a generation of UX practitioners who trust AI-generated suggestions without understanding the reliability limitations. We're risking businesses implementing changes that sound good but perform poorly.
Generative AI has incredible potential. But for high-stakes UX decisions on commercial websites, we need to be far more rigorous about validation and accuracy standards.
The path forward
At Baymard, we've decided to only release heuristics into UX-Ray when they achieve 95%+ accuracy. That means we're deliberately limiting what the tool can do today, even though we could technically make it "do more."
Some might call that conservative. I think it's responsible.
Because in the end, the goal isn't to ship fast. The goal isn't to generate the most website changes. The goal is to make UX changes that actually improve website conversion rates and revenue.
That requires accuracy we can document, test, and stand behind.
What's in the full article
I've written a detailed article that covers:
Read the complete analysis: AI Heuristic UX Evaluations with a 95% Accuracy Rate
Whether you're a UX practitioner, a business leader evaluating AI tools, or someone building AI solutions for the UX space, I hope this sparks a broader conversation about accuracy standards in our industry.
We can embrace AI's efficiency while maintaining the rigor that commercial websites require. But only if we're honest about the accuracy question.
What are your thoughts? Are you using any AI tools for UX analysis and CRO suggestions? If so, what standards are you holding these tools to?
Product Designer | UX Designer | UI Designer | UX/UI | Finance | B2C | B2B
2wThanks for sharing this 🙌
UX & Service Design Strategy | Human-Centered Design | 15+ Years Designing for People | Physical & Digital Experiences
3wI found out on my own skin that AI is misleading or simply even not true when providing you with the synthesis of A SPECYFIC research data. Getting close to the dark pattern vibe. Thank you for this article!
Senior UX Researcher | UX Strategist | Local Leader at IxDF Bordeaux | NN/g UXC
3wThat is why UXers are still needed, we are becoming curators and quality assurance experts
Optimising Experimentation: Industry leading Expertise, Coaching and Mentorship
3wOh and one small beef - the link to the "public" document hyperlink needs a login. And a reply to this: "1. That you A/B split-test all the AI-generated website changes that contain critical conversion mistakes, leading to significant revenue lost for that traffic cohort. Meaning that half of your A/B web traffic is now experiencing a serious conversion hit in addition to a significant reduction in the actual time to ship. (In comparison, when doing a well-planned A/B test, a UX/CRO team would never ship a “treatment” design that has known conversion mistakes built into the design.)" Lol - I know how tempting this is, but this assumes that your change actually does shift outcomes or business metrics. It may not. Someone may have broken the implementation, tweaked something that made it worse, introduced a bug, stopped it working on iPhones. Until you A/B test you're not going to get confirmation of the effect size and direction. I know this from many years of implementing 'sure fire' 'usability no-brainers' & I have learned my lesson - context, traffic, customers and business are varied. Seeing an anti-pattern in one place, does not guarantee a change will actually work. To find out, you need to run a test.
Optimising Experimentation: Industry leading Expertise, Coaching and Mentorship
3wAs someone who has bridged UX and Experimentation for a few years now, I wanted to give you some feedback: I've been arguing for a few years now that anti-patterns (verifiable stuff we can see in the lab) are one of the most useful priming agents for test ideation. So it's great to see you fusing all that qual and quant data you have, to provide the right lens for product evaluation. Also wonderful to hear you haven't done the 'let's do everything using an LLM' which is pretty stupid. I like what you've done, how you've gone about it and the high bar you've set. I appreciate you set a high bar but with anti-patterns, I often find impact can vary depending on the audience and product type - that context can affect how severe the impact actually is. For example, users who are highly motivated are more likely to fill out a form for a new beta product they are desperate to get, as opposed to the drag of a long checkout with low motivation to complete. One question I have - would this problem of 'generalisability' of the anti-patterns not be solved by controlling for the context, business type or model? Would you be able to increase many scores by feeding GA4 data and business context in there?