Partly. AI is genuinely good at reading a site quickly, spotting patterns across many pages, and drafting the questions a human should investigate. It is unreliable at judging whether a page will convert, and it cannot confirm accessibility or performance on its own. Treat it as a fast first pass, not a verdict.
Every redesign should start with an audit, and audits are expensive because they are slow. Someone has to read every page, run every test, and hold the whole site in their head at once. That is exactly the kind of work where AI earns its place.
What follows is an honest split of the work. Where we let AI lead, where we let it assist, and where we do not trust it at all. The distinction matters, because the failure mode of an AI audit is a confident document full of plausible statements nobody checked.
To decide what to keep. A redesign that starts from a blank canvas throws away the pages that were working, the URLs that were ranking, and the content that was earning links. The audit exists to tell you which parts of the current site are assets rather than problems.
Most redesigns fail on this point rather than on design quality. We wrote about the pattern in our piece on why website redesigns fail, and the short version is that teams rebuild the visible layer while quietly discarding the invisible value underneath it.
The stakes are easy to underestimate. Ahrefs studied roughly 14 billion pages and reported in December 2023 that 96.55% of them get zero traffic from Google. If some of your pages are in the surviving few percent, they are rare, and a redesign that loses them is a bad trade no matter how the new site looks.
Give it the reading, the sorting, and the summarizing. Keep the measuring and the judging for tools and people. AI is strong on language and pattern recognition across large volumes of text, and weak on anything requiring a real measurement of a real browser session.
So a good division looks like this. AI reads every page and tells you what each one claims to be about, which pages overlap, where the messaging contradicts itself, and which pages have no clear purpose. Lighthouse, PageSpeed Insights, and Chrome DevTools measure performance. WAVE or axe scan accessibility. A person decides what it all means.
The reason for that split is simple. A language model does not load your page in a browser on a mid-range phone, so it has no way of knowing your real Interaction to Next Paint. Google's threshold for good INP is 200 milliseconds or less at the 75th percentile, and that number comes from measurement, not from reading your HTML.
This is where it is genuinely strong. Feed it your page copy and ask what a first-time visitor would think you sell, who you sell it to, and what you want them to do next. If the answers are vague, your positioning is vague, and you have learned something real in about a minute.
The trick is to ask for the reader's experience rather than an opinion on quality. "Rate this page out of ten" produces flattery. "Summarize what this company does in one sentence using only this page" produces a mirror, and the mirror is often unflattering.
It is also good at inventory work. Ask it to group a hundred blog posts by topic and it will show you where you have accidentally written the same article four times, which is a common finding and a useful one before a migration.
Google's own guidance on helpful content asks whether a page provides original information, reporting, research, or analysis, and whether it was written or reviewed by someone who demonstrably knows the topic. Those two questions are a good prompt to run against your own pages, because they are the questions the ranking systems are built around.
As an interpreter, not a measurement tool. Run the real tests first, then hand the output to a model and ask it to explain what the numbers mean and which fix would move the most. Performance reports are dense, and translating them is exactly the kind of work AI does well.
That workflow works because the data is real. A Lighthouse report contains actual timings from an actual page load. Asking a model to prioritize the recommendations in that report is reasonable. Asking a model to estimate your load time from a screenshot is not.
Be skeptical of specific numbers it produces on its own. If a model tells you your Cumulative Layout Shift is 0.18, check it, because Google's good threshold is 0.1 or less and that decision needs a measurement rather than an inference. We keep an average PageSpeed score of 98 across our builds, and none of that came from a model's estimate of anything.
It can help, and it cannot certify. Automated scanning catches mechanical failures reliably and misses everything that requires judgment. WebAIM is explicit about this in its own reporting, stating that all automated tools have limitations and that not all conformance failures can be automatically detected.
The scale of the problem is worth seeing. WebAIM's 2026 Million report analyzed 1,000,000 home pages and found 95.9% had detected WCAG 2 failures, up from 94.8% in 2025, with an average of 56.1 errors per page. That average rose 10.1% from 51 errors the year before.
The failures are also boringly consistent. WebAIM found six issue types accounted for 96% of all detected errors, led by low contrast text on 83.9% of pages, missing alternative text for images on 53.1%, and missing form input labels on 51%. It notes these have been the same issues for seven years, which tells you this is a discipline problem rather than a technology problem.
The critical caveat is WebAIM's own. It states plainly that the absence of detected errors does not indicate a page is accessible or conformant. So use scanning to find the obvious failures, then test with a keyboard and a screen reader, and read our WCAG accessibility guide for what the standard actually requires.
Ask the models directly. Put your key buying questions into ChatGPT, Perplexity, and Gemini and see whether you get named, what they say about you, and who they recommend instead. It is manual, it is unscientific, and it is still the clearest signal available.
This matters more each year. Ahrefs studied 3,000 websites in February 2025 and found 63% received some AI referral traffic, with ChatGPT sending 50% of it, Perplexity 30.7%, and Gemini 17.6%. Ahrefs also cautions that the real figure is likely higher, since platforms such as Copilot and Mistral do not pass referrer data and show up as direct traffic.
Platform vendors are building for this too. Webflow's 2026 plan announcement introduced AEO agents on its new Team plan, which is a reasonable signal about where the industry thinks the work is heading. Do this check before the redesign, so you know whether you are protecting existing visibility or building it from nothing.
It invents specifics. Ask for a competitor analysis and you may get a plausible summary of a positioning that company has never used. Ask for statistics and you may get a real-sounding figure attributed to a real research firm that never published it. Both look completely credible in a document.
It also flatters. Models are trained to be agreeable, and an audit's whole value is disagreement. If you ask whether your homepage is clear, you will usually be told it is mostly clear with room for improvement, which is true of every homepage ever made and useful for nothing.
And it cannot tell you what your customers want. It has no access to your sales calls, your support tickets, or the reason your last three deals stalled. Ahrefs analyzed 600,000 pages in July 2025 and found the correlation between AI content share and ranking position was 0.011, effectively zero, so using AI is not the problem. Relying on it for the substance is.
For reading and drafting, never for deciding. We use it to summarize large content sets, to draft the question list for a client workshop, to translate technical output into plain language, and to check our own writing for jargon. Judgment stays with us.
We also use it as an adversary. Asking a model to argue against a plan surfaces objections faster than a meeting does, and it does not mind being told it is wrong. That is a genuinely useful role that has nothing to do with generating output.
What we do not do is let a model produce a deliverable a client will act on without a person verifying every claim in it. Across the 150-plus projects we have delivered, the trust is the product. An audit with one invented statistic in it is worse than no audit, because the client cannot tell which parts to believe. Our notes on using AI for website QA go into how we structure that checking.
Run the real tools first. Crawl the site, pull Search Console data, run PageSpeed Insights on your top templates, and scan for accessibility errors. Then hand that real output to a model and ask it to find the patterns, rank the problems, and tell you what it cannot determine from the data you gave it.
That last instruction is the one that changes the output. Asking a model to name the limits of its own analysis produces a much more honest document than asking it for conclusions, and it tells you where a human still needs to look.
Then verify every number before it reaches a decision-maker. If you would rather have people who do this daily run the audit and tell you what to keep before you spend money on a rebuild, we are happy to help. Send us your site and we will give you an honest read. Let's talk, over at phoenix.studio.
Tell us where you want to go. We'll tell you how we'd get you there.