The usual AI website audit: someone pastes a homepage URL into a chatbot, asks “what’s wrong with my site?”, and gets a polite list that would fit any website ever built. Add social proof. Make the CTA stand out. Improve load speed. None of it is false, and none of it says what to fix first.
The problem is the input. A model can only judge the evidence in front of it, and a URL alone often gives it very little. An AI website audit gets useful when you hand over a proper evidence pack, ask for findings that point back to that evidence, and rank the fixes yourself by what they cost and how sure you are. Below: what each type of evidence can reveal, a reusable worksheet, a prompt, and an annotated example on an illustrative page.
It is a guide to diagnosing a site that already exists, not a pre-launch checklist. All tool documentation referenced below was checked on 22 October 2026.
What each kind of evidence can and cannot reveal
Screenshots are good for first impressions and terrible for anything under the surface. Analytics tell you where people leave, never why. Measured technical results are facts about a test run; the model interprets them and should never produce them.
| Evidence | What AI can read from it | What it cannot tell you |
|---|---|---|
| Desktop and mobile screenshots (full page) | Visual hierarchy, what sits above the fold, whether offer and CTA are visible, crowding | Load speed, keyboard access, screen reader behaviour, what happens on tap, security |
| Page text (copied or exported) | Clarity of the offer, who the page addresses, jargon, missing information (prices, area, process), how specific the proof is | Whether claims are true, how the text looks on the page |
| Link and navigation list | Menu labels, dead ends, duplicate paths, pages that are hard to reach | Whether a link works, unless you include measured status codes |
| Analytics export (for example GA4) | Pages with low engagement, entry pages, where key events happen and where they never happen | Why visitors left. A low number is a symptom |
| Search Console Performance data | Queries the page appears for, pages with impressions but few clicks | Anything about on-page behaviour after the click |
| Measured results (PageSpeed Insights, Lighthouse, link checker) | Which measured metric fails and which issues the tool flagged | Anything not in the report. It should never invent scores |
Two of those rows need a definition before you trust them. In GA4, an engaged session is one that lasts longer than 10 seconds, has a key event, or has two or more page views, and engagement rate is the share of sessions meeting that bar (Google Analytics Help). In Search Console, click-through rate is simply clicks divided by impressions (Search Console Help). Paste numbers without those definitions and the model will interpret “bounce” however the internet taught it.
What AI cannot certify from a screenshot
This is where AI audits most often overreach. A model looking at an image cannot tell you a site is accessible, fast or secure. It can flag likely problems, and it should be told to stop there.
Accessibility. The W3C states plainly that evaluation tools “can not determine accessibility, they can only assist” and that human judgement is required (W3C WAI). Even Lighthouse, which actually inspects the page code, notes that a perfect accessibility score does not guarantee an accessible page and that manual testing still matters (Chrome for Developers). A screenshot is weaker evidence than either. The model can say “this grey text on a beige background looks low contrast, measure it against the 4.5:1 minimum for body text” (WCAG 2.2, SC 1.4.3). It cannot say the page passes.
Performance. Speed lives in measurements. PageSpeed Insights reports lab data from a simulated load and field data from real Chrome users, and the two can disagree (About PageSpeed Insights). Google’s Core Web Vitals thresholds for “good” are LCP of 2.5 seconds or less, INP of 200 ms or less and CLS of 0.1 or less, judged at the 75th percentile of page views (web.dev). Paste the report and let the model explain it.
Security. Outdated plugins, server configuration and form handling are invisible in screenshots and page text. That is a job for whoever maintains the site and hosting.
Mobile usability sits in between. Full-height phone screenshots show a lot, and Chrome’s device mode helps collect them, but Google calls it “a first-order approximation” and recommends checking on a real device when in doubt (Chrome DevTools docs). Spend two minutes on your own phone. Tap the menu, the phone number and the form.
The audit worksheet
Fill this in for one page at a time, usually the homepage or your top entry page. The evidence column is the point of the exercise: a finding without evidence goes in a separate pile labelled “hunches”.
| Area | Question to answer | Evidence to attach | Finding | Confidence (high, medium, low) |
|---|---|---|---|---|
| Clarity | Within the first screen, can a stranger say what you sell and where? | Desktop and mobile screenshot, hero text | ||
| Audience fit | Does the copy speak to the buyer you actually want, in their words? | Page text, a sentence describing your best customer | ||
| Offer | Is the service, scope, price signal and next step clear? | Page text, pricing or service page text | ||
| Proof | Is there specific, checkable evidence: named work, reviews, credentials? | Page text, screenshot of proof sections | ||
| Navigation | Can a visitor reach the key pages in one or two taps? Any dead ends? | Menu labels, link list with status codes | ||
| Calls to action | Is there one main action, visible early, with a label that says what happens next? | Screenshots, CTA text, analytics key events | ||
| Mobile usability | Does the page work on a phone: readable text, reachable buttons, usable form? | Full-height phone screenshot, notes from a real-device check | ||
| Measured technical results | Which measured metrics fail, and on which device? | PageSpeed Insights or Lighthouse report, link checker output |
The prompt that keeps the audit honest
The constraints matter more than the role-play. This prompt makes every finding quote its evidence and lets the model say it doesn’t know, which removes most of the generic filler.
You are auditing one existing web page for a small business. Business: [one sentence on what they sell, where, to whom]. Main goal of the page: [for example, quote requests].
Evidence attached: [list what you attached: desktop and mobile screenshots, page text, link list with status codes, GA4 export for [date range], PageSpeed Insights report].
Review the page against these areas: clarity, audience fit, offer, proof, navigation, calls to action, mobile usability, measured technical results.
Rules:
1. Every finding must quote or point to specific evidence (a sentence from the page text, a part of a screenshot, a number from a report). If there is no evidence, write “insufficient evidence” for that area.
2. Do not certify accessibility, security or performance. Flag likely issues and name the test a human should run.
3. Do not invent analytics, scores or predicted conversion lifts. Do not interpret analytics beyond the definitions I give you.
4. Separate observations from recommendations.Output a table with columns: area, observation, evidence, likely business consequence, suggested fix, confidence (high, medium, low).
Annotated example: an upholstery workshop homepage (illustrative)
Copperline Upholstery is a fictional business. Every piece of evidence below (page text, screenshot notes, link check, analytics) is sample data created for this article, not taken from a real site or client.
The setup: a two-person workshop reupholsters chairs and sofas and wants more local quote requests from homeowners. The sample pack holds the hero text (“Crafted with passion since 2009. Quality you can feel.”), a three-item menu (Home, Our Story, Gallery), screenshots, a link check and a 90-day GA4 export.
| # | Area | Observation | Evidence (sample) |
|---|---|---|---|
| 1 | Clarity | The first screen never says “upholstery”, “chairs” or a location. Visitors learn the business has passion. That’s all | Hero text quoted above; desktop screenshot shows only the headline and a photo of fabric swatches |
| 2 | Calls to action | There is no quote request above the fold. The only route to the contact form is a footer link labelled “Contact” | Desktop and phone screenshots; menu list has no contact item |
| 3 | Offer | No price signal or process, so visitors can’t tell whether a dining chair costs fifty or five hundred | Page text contains no prices, ranges or “how it works” steps |
| 4 | Proof | The gallery shows finished pieces with no before photos or descriptions, and the testimonial is unattributed | Gallery page text: images only; testimonial reads “Amazing work!” with no name |
| 5 | Navigation | One menu link returns an error | Link check: /our-story/ returns 404, the live page is /about/ |
| 6 | Mobile usability | The phone number in the header is plain text, not a tappable link | Real-device note: tapping the number does nothing |
| 7 | Measured results | The mobile lab LCP is outside Google’s “good” threshold, likely the large hero image. Field data was not available for this URL | Sample PageSpeed Insights mobile report: LCP 4.8 s (lab), LCP element is the hero image |
| 8 | Audience fit | Insufficient evidence. Copy is generic, but without knowing who requests quotes today, it is unclear which buyer to write for | None beyond page text |
The sample GA4 export shows the homepage as the main entry page, with a lower engagement rate than the gallery. That supports findings 1 and 2 without proving them. A visitor who can’t tell what you do and has no button to press is a plausible explanation for a short visit, not the only one, and the audit should say so.
Notice finding 8. If your AI output has a confident finding for every area from a single screenshot, get suspicious.
Ranking the fixes
Eight findings with no order is a to-do list that will be ignored. Rank each fix by the observed problem, the business consequence if it stays, the effort, and your confidence that the diagnosis is right. Confidence is the column people skip, and it is the one that stops you rebuilding a page because of a hunch.
| Fix (sample) | Observed problem | Business consequence | Effort | Confidence | Priority |
|---|---|---|---|---|---|
| Rewrite hero: what, for whom, where | #1, offer not named | Right visitors may leave before seeing the gallery | Low (one hour of copy) | High | Do now |
| Add “Request a quote” button in header and hero | #2, no visible CTA | Interested visitors have no obvious next step | Low | High | Do now |
| Fix the 404 menu link | #5, broken link | Dead end from the main menu | Very low | High (measured) | Do now |
| Make phone number tappable | #6 | Friction for mobile callers | Very low | High (checked on device) | Do now |
| Add price ranges and a four-step process | #3 | Unqualified enquiries, or none | Medium (needs owner decisions) | Medium | Next |
| Before and after photos with short captions | #4 | Weak proof for first-time buyers | Medium (needs photos) | Medium | Next |
| Compress or resize hero image, then retest | #7, lab LCP | Slower first view on phones | Low to medium | Medium (lab only) | Next, then remeasure |
| Define target buyer before rewriting the rest | #8, insufficient evidence | Copy may speak to the wrong customer | Low (a conversation) | Low | Gather evidence first |
There is deliberately no “expected lift” column. Nobody knows how many extra quote requests a better headline produces here, and a model that gives you a percentage is making it up. Count quote requests before and after.
When the audit points somewhere else
Sometimes the website isn’t really the problem. If the copy is vague because nobody agreed what the business offers, fix the brief first; the Brief Decoder walks through turning a vague request into usable requirements. If findings keep circling back to inconsistent tone across pages, a short set of AI brand guidelines will stop the next round of edits drifting. And if the list of “do now” fixes is long enough that a rebuild looks cheaper, this documented prototype workflow with ChatGPT and Claude Code shows how far a fast concept gets, and where it stops.
Run the worksheet each quarter on your main entry pages and keep the evidence packs. You’ll end up with something rare in small-business marketing: a record of what you changed and why. Direktorium keeps publishing practical guides like this for small teams using AI in real work. If you would rather have someone run the audit and make the fixes, BBDirector can help.
