Quality control for AI-built websites: a 20-point checklist before you hand it over
Short answer: An AI-built website is not ready because it looks finished. It is ready when a named person has checked 20 things: facts in the text, links, forms, speed, accessibility, search settings, tracking and ownership, and has signed a record of the result. Run the checks below on a real phone and a real computer, mark each one pass or fail, fix every fail, and only then give the client the keys. The checklist is plain enough for a project manager to run without a developer.
You have a site that an AI tool built in a day, and you need a clear way to decide whether it is safe to put in front of a client or a customer.
I'm Ranjit, and I run Blur Studio: 15+ years building software, 120 apps taken live, and clients who have gone on to raise $10M+. I build AI systems and processes that help businesses do a lot more with a lot less time. I also work differently from most of what else is out there, more on that further down this page. Say hello.
Use this checklist after the site is built and before it goes live or goes to the client. AI website tools, which write the text and lay out the pages from a short brief, are fast, but they do not check their own work. They can state a wrong fact in a confident voice, leave a placeholder in a heading, or build a form that looks right and sends nothing. Google says the same about AI text in general: its page on using generative AI content warns that outputs may contain inaccuracies, which it calls hallucinations, and its helpful-content guidance tells publishers to fact-check and review what they publish. So a human has to do the checking, and this page is the list.
It works for any build method: a no-code builder, a platform such as the AI website builders agencies compare, or a site a developer finished by hand. It is written for an agency or a small business owner, not for a developer. This is information, not legal advice. Where a rule depends on where you and your visitors are, check local law and the regulator's current page.
How to use the checklist
Print it or copy it into a shared sheet. Give every row a status of Pass, Fail or Not applicable, an owner and a date. A site is ready only when every row is Pass or Not applicable and the sign-off block at the bottom is filled in. A Fail is not a bad mark. It is the reason you run the check.
Test on the live address, or on the real preview address, not on screenshots. Use one phone and one computer you do not normally use, a private browser window, and a different network from the office if you can. Cached pages and saved logins hide problems.
Content checks (1 to 5)
| # | Check | Pass looks like | Fail looks like |
|---|---|---|---|
| 1 | Every fact is true | Prices, hours, addresses, licence numbers, service areas, years in business and staff names match a document the client supplied | A number or claim nobody can source, or one the AI made up |
| 2 | No placeholders | No "Lorem ipsum", "Your Company", "[city]", stock names or sample phone numbers anywhere, including page titles | Any template text still visible, or still in the search result title |
| 3 | Claims and reviews are real | Testimonials come from real customers who agreed, and every result or "best" claim has proof on file | Invented quotes, unproven superlatives, "guaranteed" results |
| 4 | Brand and tone | Names, spelling, logo, colours and tone match the brand sheet | Two spellings of the client name, or a voice the client would never use |
| 5 | A named human reviewed every AI-written page | The reviewer's name and date are in the sign-off record | "The AI wrote it, so it should be fine" |
On check 3: the US Federal Trade Commission (FTC), the regulator that polices advertising claims, expects claims to be truthful and backed up, so the rule of thumb is simple. If you could not defend a sentence to a regulator or a customer, cut it or prove it. On check 5, read whether Google penalises AI-written pages: Google's position is that it judges quality, not how a page was made, which is exactly why the human review matters.
Links, forms and tracking (6 to 10)
| # | Check | Pass looks like | Fail looks like |
|---|---|---|---|
| 6 | Every link works | Click every menu, button and footer link. None gives a 404 (page not found) or goes to the wrong page | One dead link on a main page |
| 7 | Phone, email and map links | On a phone, tapping the number starts a call to the right number, and the email link opens a message to the right address | The number is plain text, or points to a sample number |
| 8 | Forms deliver | You submit a test from every form with a real address, and the message reaches the real inbox, not just the spam folder, within five minutes | No email, wrong recipient, or the message lands in spam |
| 9 | Forms are backed up | Submissions are also stored somewhere you can see, such as the platform's inbox or a sheet | Email is the only copy |
| 10 | Tracking fires once | Your analytics tool shows the test form and test call as one event each, in the right account | No event, a double event, or the data goes to the agency's account instead of the client's |
Forms fail quietly more than anything else on a new site. Vendor blogs on form problems name the same few causes: mail sent without sender authentication, a wrong recipient, host restrictions and plugin conflicts. Gmail's published sender rules (effective February 2024) require all senders to authenticate with SPF or DKIM, two settings that prove an email really came from your domain. Senders of more than 5,000 messages a day also need DMARC, a third setting that tells receivers what to do with mail that fails the check. A contact form is not a bulk sender, but authenticated sending is still the fix when form mail goes to spam. Send the test, do not assume.
Speed, mobile and accessibility (11 to 15)
| # | Check | Pass looks like | Fail looks like |
|---|---|---|---|
| 11 | Pages load fast enough | Per web.dev, "good" is Largest Contentful Paint (time until the main content shows) at 2.5 seconds or less, Interaction to Next Paint (how fast the page reacts to a tap) at 200 milliseconds or less, and Cumulative Layout Shift (how much the page jumps) at 0.1 or less, measured for 75 percent of visits | Any of the three outside those numbers on the home page or the main landing page |
| 12 | Mobile layout | No sideways scrolling, text readable without zoom, buttons easy to tap, menu opens and closes | Overlapping text, cut-off buttons or a menu you cannot close |
| 13 | Images | Real photos, sensible file sizes, a short description (alt text) on every image that carries meaning | Giant files, stock photos of the wrong trade, empty alt text on key images |
| 14 | Keyboard and focus | You can reach every link and button with the Tab key, you can always see where you are, and you can submit a form without a mouse | You get stuck, or the focus marker disappears |
| 15 | Automated accessibility scan plus a human look | You ran WAVE and axe, two free accessibility checkers, fixed what they flagged, and checked contrast and tap-target size by eye | "The scanner said zero errors" and nothing else |
On check 11, test with real field data where the tool offers it, and on a mid-range phone, not only your laptop. On check 15, no scanner can pass a site for you. WAVE's own help page says only a human can determine true accessibility, and Deque, the maker of axe, says expert manual testing is still needed (it also claims its tools find up to 80 percent of issues, which is a vendor figure, so do not rely on it). WCAG 2.2, the current version of the web accessibility guidelines, added a minimum target size of 24 by 24 CSS pixels for tap targets (criterion 2.5.8, level AA), plus a rule that the focused item must not be hidden behind other content (2.4.11). Those two are easy to test by hand, and AI-built layouts often miss them. Do not add an accessibility overlay widget and call the site compliant: in April 2025 the FTC approved a final order requiring the maker of one such widget, accessiBe, to pay $1 million over claims that its tool could make any site meet WCAG. Check the regulator's current page for what the accessibility laws in your country cover.
Search, legal and ownership (16 to 20)
| # | Check | Pass looks like | Fail looks like |
|---|---|---|---|
| 16 | Search settings | The staging "do not index" setting is off on the live site, each page has its own title and meta description, one H1 per page, and the sitemap is submitted in Google Search Console, Google's free tool for site owners | The live site still tells search engines to stay away, or every page has the same title |
| 17 | Redirects for a rebuild | If this replaces an old site, every old address that mattered redirects to its new equivalent. See keep SEO rankings when you rebuild a client site | Old addresses give 404 on launch day |
| 18 | Legal pages and consent | Privacy policy and terms exist and match what the site really collects, and any analytics or ad cookies follow the local consent rules. Check local law | A template policy that names the wrong company or tools |
| 19 | Pages are not thin copies | Each page says something specific the others do not, and has a reason to exist for a visitor | Dozens of near-identical pages that only swap a city name. See doorway pages and thin content |
| 20 | The client owns the keys | Domain registered in the client's name, client is owner on hosting, analytics and Search Console, and you hold a backup you have actually restored once | Everything sits in the agency's account, or the backup was never tested |
Check 20 is the one that hurts later. Ownership has four parts: the domain, the content, the code and the accounts, and a platform can give you one and keep another. The full list of questions is in who owns your website and content. On check 19, Google's spam policies target pages created mainly to manipulate rankings, however they were made. A reviewer asking "would this page exist if search engines did not?" catches most of it.
Who signs off
Split the roles so that the person who built the site is not the only person who approves it.
- Builder or producer: runs checks 6 to 15 and fixes fails.
- Editor or account manager: runs checks 1 to 5 and 16 to 19, and names themselves as the human reviewer of the AI-written pages.
- Client owner: confirms the facts in check 1, the ownership in check 20, and signs that they accept the site.
The sign-off record is one short block you keep with the project file: site address, date, the three names, the count of Pass, Fail and Not applicable, and a line that says "all fails fixed and rechecked on [date]". If a client later says "you never told me about that", the record answers it. For how to get the approval itself in writing, see getting client approval on a real website version.
Turn the fails into a fix list
Do not fix as you go. Finish the whole run, then sort the fails into three groups.
- Stop-ship: wrong facts, forms that do not deliver, noindex left on, wrong ownership, broken main-page links. Nothing launches until these pass.
- Fix this week: speed over the limits, mobile layout issues, missing alt text, missing meta descriptions.
- Backlog: polish, extra pages, small copy changes. Write them down with an owner, or they never happen.
Then rerun only the rows that failed, on the same devices, and date the second pass in the record.
A worked example with made-up numbers
Say an agency builds a 12-page site for a plumbing firm in an afternoon with an AI tool. The numbers below are invented to show the method, not a benchmark.
The run finds four fails. The "Free quote" form sends to the agency's test address (check 8, stop-ship). The service-area page says "serving Dallas, Austin and Houston" but the firm only works in Dallas (check 1, stop-ship). Home-page Largest Contentful Paint is 4.1 seconds against the 2.5 second target, because of one 6 MB hero image (check 11, fix this week). Eight photos have no alt text (check 13, fix this week). Both stop-ship items are fixed in 20 minutes. The image is resized and the alt text added the next morning. The rerun shows 20 rows at Pass or Not applicable, and all three people sign. I would plan about one to two hours of checking for a small site like this, though your time will vary. The two stop-ship fails would have reached customers without the run.
Copy-ready one-page version
Paste this into a sheet with columns for Status, Owner and Date.
- Every fact matches a client document
- No placeholder text, including titles
- Claims and testimonials are real and provable
- Brand names and tone are right
- A named human reviewed each AI-written page
- All links work
- Phone and email links go to the right place
- Every form test arrives in the real inbox
- Submissions are stored as a backup
- Analytics counts each test once, in the client's account
- Speed within web.dev "good" limits
- Mobile layout is clean
- Images are real, light and described
- Keyboard and focus work
- WAVE and axe run, plus a human check
- Search settings are live-ready and sitemap submitted
- Old addresses redirect
- Legal pages and consent match reality
- No thin copy pages
- Client owns domain, hosting, analytics and a tested backup
When this checklist is not enough
This list covers a normal marketing site. Add more checks if the site takes payments, holds logins, stores health or financial data, or serves people in more than one country. Those need a security review and a lawyer, not just a checklist. See the AI-built app security checklist for the app side, and finishing a half-built AI website if the build itself is unfinished.
How I'd approach this
I would make the checklist a standing gate, not a favour. No site goes to a client until the sheet is filled in, even for a small job, because the failures that cost money are always the boring ones: a form that goes nowhere, a number that is wrong, a noindex tag nobody removed. If you build many sites a month, put the 20 rows into your project tool and turn the sign-off into a task that cannot be closed without the three names.
First week: copy the list above into a sheet today, run it on your last two launched sites, and see how many rows fail. That tells you how much risk you were carrying. In week one, also assign the three roles by name, and pick the one device pair you will always test on. If you want help turning this into a repeatable gate for AI-built client sites, Blur Studio builds the custom version: AI agents that use your business knowledge, connections to your CRM (customer records tool) and calendar, finishing half-built work, and ongoing care. See what that looks like. The checklist works without any of it.
Related guide: where an AI marketing platform fits in a larger SOP.
I'm not a lawyer, accountant or other licensed adviser, and this page is general information, not advice. Rules and prices differ by country and change often, so treat it as a direction to think along, and please check with a lawyer or the relevant regulator before you act on it.
Questions people ask
Do I need to run all 20 checks on a very small site?
Yes, but a one-page site is quick, often well under an hour. Content, forms, mobile, search settings and ownership matter on every site, however small. Mark rows that do not apply as Not applicable so the record is still complete.
Can an AI tool run the QA for me?
It can help with part of it, such as spotting broken links or listing claims that need proof. It cannot confirm that a fact is true for this client, or that a form reaches the real inbox. A named human still has to run and sign the checklist. Google says to fact-check and review AI-written content before publishing.
Which accessibility tools should I use?
WAVE and axe are both free and run in the browser. Neither can pass a site on its own. WAVE says only a human can determine true accessibility. Use them to find obvious problems, then test with the keyboard, zoom the page to 200 percent, and check colour contrast by eye. Accessibility rules depend on where you and your visitors are, so check local law.
What speed numbers should I aim for?
web.dev, Google's web guidance site, lists good as Largest Contentful Paint of 2.5 seconds or less, Interaction to Next Paint of 200 milliseconds or less, and Cumulative Layout Shift of 0.1 or less, at the 75th percentile of page loads. Check the current page for any change, and test on a phone.
Who should sign off, the agency or the client?
Both, in different roles. The agency signs that the checks were done and the fails fixed. The client signs that the facts are right and that they accept the site. Keep the signed record with the project file. For the approval wording, see getting client approval on a real website version.
How often should I recheck a live site?
Run a short version every month: links, forms, speed and any page that changed. Run the full list after any redesign, platform move or large content change. A monthly routine is covered in the monthly routine for keeping a website fresh.
What if the AI-written text is mostly right but a few facts are wrong?
Fix the facts and keep the page, but treat it as a signal. Review the rest of the AI-written pages more slowly, because errors tend to repeat across pages written from the same brief. Add a source line to your brief template so the AI starts from the client's real facts next time.
Does a clean checklist guarantee good rankings or leads?
No. It lowers the chance of launch mistakes. It does not promise rankings, traffic or leads, which depend on the market, the content and time. See whether Google penalises AI-written pages for what Google says about quality.
More in this topic: Websites: agency delivery, quality control and margins (10 guides) →
Sources
- web.dev: Web Vitals (Core Web Vitals thresholds)
- W3C: What's new in WCAG 2.2
- WebAIM WAVE help
- Deque: axe accessibility tools
- Google Search Central: Using generative AI content
- Google Search Central: Creating helpful, reliable, people-first content
- Google Workspace Admin Help: Email sender guidelines
- Google Search Central: Spam policies
- FTC: final order against accessiBe (April 2025)