The short answer
Use an AI audit to cover ground quickly: every page, every title, every form. Then apply judgement. A good audit shows the evidence for each finding, separates facts (this page has no meta description) from opinions (this headline is vague), says what it could not check, and does not forecast uplifts. Verify findings before acting, change one thing at a time, and measure.
What an AI website audit is
Most AI audits combine three parts:
- A crawler that fetches your pages, follows links and records what it finds: status codes, titles, headings, links, images, structured data and page weight.
- Data sources, if you connect them: Google Search Console for queries and indexing, Google Analytics for behaviour and key events, and field performance data from real users.
- A language model that reads the pages and data and writes findings in plain English: what looks wrong, why it might matter and what to change.
The crawler and the data produce facts. The language model produces interpretation. Knowing which is which is the most important skill in reading an AI audit.
What an AI audit can find well
| Area | Typical findings | How far to trust it |
|---|---|---|
| Technical crawl issues | Broken links, redirect chains, error pages, missing or conflicting canonical tags, pages blocked by robots.txt, orphan pages | High: these are facts the crawler observed. Confirm the important ones by hand. |
| On-page search basics | Missing, duplicate or truncated titles and meta descriptions, missing H1s, thin pages, missing alt text | High for presence and length; lower for advice on wording |
| Copy clarity | Headlines that do not say what you do or who it is for, jargon, buried calls to action, inconsistent messaging between pages | Medium: a useful outside view, but it does not know your buyers |
| Conversion path | Pages with no route to contact, long forms, dead ends, missing proof near calls to action | Medium: good for spotting gaps, not for predicting impact |
| Search opportunities (with Search Console connected) | Pages with many impressions and few clicks, queries ranking just off page one, pages competing for the same query | High for the data; medium for the suggested fix |
| Performance (lab tests) | Heavy images, render-blocking scripts, layout shifts on load | Medium: lab results show likely causes, not what real visitors experience |
| Accessibility (automated checks) | Missing labels, low contrast, missing alt text, heading order | High for what it flags, but it misses many barriers |
For a structured list of search checks to run alongside an AI audit, see our B2B website SEO audit checklist.
What an AI audit cannot find, or cannot know
Who your buyers are and what they care about
A model can tell you a headline is vague. It cannot tell you which of three accurate headlines your buyers will respond to, because it has not sat in your sales calls. The best evidence for B2B messaging is still what prospects say to your sales team, in their own words.
Whether your claims are true
An audit can flag that a case study has no numbers or that a claim is vague. It cannot verify that "trusted by 200 firms" is accurate, that a certification is current, or that a quote is real. Checking claims is a human job, and it matters: a claim a buyer later finds to be wrong costs more trust than a vague one.
Real performance and real search data, unless you connect them
Lab tests load a page "in a controlled environment", while field data comes from "monitoring all users who visit a page"; Google's web.dev guide explains that the two "measure very different things" and that field data is the more accurate view of what users experience1. An audit that only runs lab tests can point to likely causes of slowness but cannot tell you whether your real visitors have a problem. In the same way, a crawler sees what it can fetch, not what Google has indexed: for that you need Search Console's Page indexing report, which "shows how many URLs on your site have been crawled and indexed by Google"2.
Most accessibility barriers
When the UK Government Digital Service tested 10 automated accessibility tools on a page with 143 deliberate failures in 2017, the tools together found 71% of them, and the best single tool found 41% (counting prompts for manual checks). The team concluded that automated tools "cannot be relied on, by themselves, to check the accessibility of a website"3. An AI audit that runs automated checks inherits that limit; keyboard and screen reader testing still needs people.
What will happen if you make the change
No audit can know that rewriting a page will raise conversions by a given amount. Any audit that forecasts a specific uplift is guessing. The honest output is a prioritised list of hypotheses with the evidence behind each.
Pages it could not reach
Content behind logins, pages only reachable through site search or forms, and content that relies heavily on scripts may be missed or misread. A good audit report tells you how many pages it crawled and which it could not.
Where AI audits go wrong
- Confident mistakes. Language models can produce plausible statements that are false. A 2025 paper by researchers including Adam Tauman Kalai describes these "hallucinations" as persisting "even in state-of-the-art systems"4. In an audit, that can look like a finding about a page element that is not there, or a misquoted guideline.
- Generic best practice. "Add a testimonial above the fold" may be right for one site and wrong for yours. Advice that would apply to any website is not a finding about yours.
- Everything is high priority. Two hundred findings with no ranking are not an audit, they are a backlog nobody will work through.
- Rewrites at scale. Some tools offer to rewrite every page at once. Google says it focuses "on the quality of content, rather than how content is produced", but that using automation to generate content mainly to manipulate rankings breaks its spam policies5. Mass rewrites also make it impossible to tell which change did what.
How to run an AI website audit well
- Start with the pages that carry buying intent. Services, solutions, pricing, case studies and contact pages matter more than old blog posts. Our guide to improving a B2B website without a redesign explains how to find them.
- Connect real data. Search Console, analytics key events and field performance data turn guesses into evidence.
- Give it context. Who you sell to, what you sell, your main competitors' positioning in your own words, and the objections sales hear most often.
- Ask for evidence on every finding. The URL, the element, the data point. If a finding has no evidence, treat it as an opinion.
- Verify the important findings by hand. Open the page, check the data, confirm the problem is real before anyone fixes it.
- Prioritise by evidence and effort. High-traffic pages with clear problems and easy fixes first.
- Make changes one at a time, approve each one, and log it. A website change log is what lets you measure what each change did, and an approval step means a person checks every change before it goes live.
- Measure honestly. Some changes will help, some will not, and on low traffic many will be inconclusive.
A checklist for judging an AI audit report
- Does each finding show the evidence: URL, element and data?
- Does it separate facts from opinions?
- Does it say how many pages it crawled, which data sources it used and what it could not check?
- Are findings prioritised, with a reason for the order?
- Is the advice specific to your site and your buyers, rather than generic?
- Does it avoid forecasting specific uplifts?
- Can you check a sample of findings and confirm they are real?
- Does it recommend a human review before anything is published?
Where Artiweave fits
Artiweave treats audit findings the way this page recommends, as hypotheses. Each opportunity it finds in a crawl, Search Console or performance data becomes a specific proposed change with that evidence attached, and nothing is published until a marketer approves it. Afterwards the change is measured, so a finding that made no difference is recorded as such.
Frequently asked questions
What is an AI website audit?
A review of your website by a crawler and a language model, sometimes with Search Console, analytics and performance data connected. It produces findings and suggested changes in plain English.
Can an AI website audit replace a human review?
No. It covers ground quickly and catches technical and on-page issues well, but it cannot verify your claims, know your buyers or test accessibility fully. Use it to find candidates, and have a person check and approve every change.
Can an AI audit tell me how much a change will improve conversions?
No. Nobody can know that in advance. Treat any forecast of a specific uplift as a guess, make the change, and measure what actually happens.
What data should I connect to an AI audit?
Google Search Console for queries and indexing, your analytics key events for leads, and field performance data from real users. Without them, the audit is limited to what a crawler can see.
How often should I run an AI website audit?
A full audit when you start, then regular checks on the pages that carry buying intent, for example monthly, and after any significant release or template change.
Sources
- Philip Walton, web.dev: Why lab and field data can be different (and what to do about it) (accessed November 2026)
- Search Console Help: Page indexing report (accessed November 2026)
- Government Digital Service, Accessibility in government blog: What we found when we tested tools on the world's least-accessible webpage (24 February 2017)
- Kalai, Nachum, Vempala and Zhang: Why Language Models Hallucinate, arXiv:2509.04664 (4 September 2025)
- Google Search Central Blog: Google Search's guidance about AI-generated content (8 February 2023)
Evidence first. Every change approved.
Artiweave proposes website changes with the evidence attached, publishes only what your team approves, with automatic rollback, and reports each outcome honestly. It is in private beta.