Skip to content
All articles
AIWeb DevelopmentWebsite SecurityHiring a DeveloperSmall BusinessPublished · Updated · 14 min read

Built Your Website With AI? A Security, Legal and SEO Checklist

Oblix Digital

Oblix Digital

Web design & SEO

Close-up of a hand holding a phone showing a security checklist, with a blurred laptop displaying a website behind it.

AI can build a complete website from a short description, and the first draft is often better than people expect. It is also often worse than it looks. I build client sites with AI coding tools every day, so this is my account of what happens when you hand the whole job over, followed by a library of prompts you can paste into your own AI tool to check what it missed.

The short version

  • AI is excellent at speed: layouts, forms, repeated components and first drafts.
  • It does not know your customers or your legal duties, and it will not mention a gap you did not think to ask about.
  • The gaps tend to be the same few: security, missing pages, cookies and privacy, search, accessibility, forms and email, and what happens when something breaks.
  • The prompt library further down checks each of those. Working through it takes an afternoon, and it can save you from hearing about a problem from a customer.

The good

First drafts are fast. Something that once took days to reach a first look can be on screen within an hour, which makes it inexpensive to try two or three design directions before committing to one.

It also lowers the barrier for anyone with a small budget, and once the direction is clear, AI is very good at the repetitive work that used to take up much of a build: page templates, form handling, responsive layouts, the same component repeated with small changes.

The bad

AI tools do not know your business. They do not know your customers, what those customers worry about before they buy, or what has worked for similar businesses. What they produce is plausible, which is not the same as effective.

Without someone checking, you can end up with a site that looks fine in a screenshot and has real problems underneath: slow load times, a mobile layout that breaks on certain phones, buttons a screen reader cannot make sense of, or copy that describes your services in words your customers never search for.

The copy deserves a particular warning. AI writes confident sentences about things it cannot know: your prices, how long you have been trading, the awards you have won. Those claims go live unless a person reads every one of them.

The ugly

Then there are the outcomes that cost people time and money.

  • Security nobody thought about. A key left in the code, an admin page anyone can open, a form with no limits, or a package nobody ever updates. AI will not flag a risk you did not ask about, and the first sign is often a hacked site.
  • Pages you did not know you needed. No privacy policy, no cookie choice, no terms, no company details in the footer. Customers, payment providers and regulators notice these before you do.
  • Content you cannot take with you. Some platforms make it easy to build and hard to leave, so moving later means starting again.
  • Weeks going in circles. Without the technical vocabulary to describe what is wrong, you ask the tool to "fix the layout" over and over while it confidently changes the wrong thing.

The last one is especially frustrating, because it feels as if you are nearly there the whole time.

Same technology, very different outcome

None of this means AI-built sites are bad. It means AI building the whole thing with nobody checking is a very different proposition from AI used as a tool by someone who knows what good looks like. The technology is the same. What differs is whether the person driving it can spot a slow page, a broken layout or a security gap before your customers do.

Three ways a site gets built now

A developer working without AI can still produce excellent work. It takes longer, so it tends to cost more. AI on its own, with nobody checking, gets you something fast and inexpensive that no one can properly maintain or vouch for. A developer using AI gets most of the speed and keeps the judgement, and for most small businesses that is the combination I would recommend.

What AI-built sites tend to miss

The same handful of gaps turn up again and again. Each has its own group of prompts in the library below.

  • Security. Keys in the wrong place, admin pages that are not really protected, forms that anyone can abuse, out-of-date packages.
  • Pages and legal. The privacy policy, terms, contact and company details, a proper 404 page, and the placeholder text nobody replaced.
  • Privacy and cookies. What loads before a visitor has agreed, and what personal data you hold without realising.
  • Search and speed. Titles, sitemaps, redirects from your old site, and the images and scripts that slow a phone down.
  • Accessibility and quality. Keyboard use, contrast, layout on real phones, and claims on the page that are wrong or cannot be proved.
  • Forms and email. Enquiries that quietly fail, and email that lands in spam.
  • Running it. Backups you have never restored, accounts you do not own, and nothing to tell you the site is down.

Five quick facts worth knowing

  • Accessibility affects a lot of people. In the UK government's latest Family Resources Survey, for 2024 to 2025, one in four people were disabled. A site that only works for people with good eyesight and a steady hand can turn some of them away.
  • Faint text fails. WCAG 2.2, the main accessibility standard, asks for normal text to have a contrast ratio of at least 4.5 to 1 against its background. Pale grey on white often falls short, and free online checkers will test it.
  • Small buttons fail too. WCAG 2.2 asks for buttons and links of at least 24 by 24 pixels, or enough space around them, because some people cannot tap small targets that sit close together.
  • Search engines may see less than you do. Some AI builders make pages whose content only appears once JavaScript, the code that runs in your browser, has loaded. Google says not all bots can run JavaScript, so some crawlers may see a page with little or nothing on it.
  • Rejecting cookies should be easy. The ICO, the UK's data protection regulator, says refusing consent should be as easy as accepting it. It uses a banner with only "Accept all" and "More options", and no "Reject all", as an example of bad practice.

How to use the prompts

  1. Open your project in an AI coding tool that can read your files, such as Claude Code, Cursor, Codex or GitHub Copilot in agent mode. If your site lives on Wix, Squarespace or a similar builder, you cannot open the code, but the prompts in Pages and legal, Privacy and cookies, and Accessibility and quality still work if you paste in your page text and describe the site.
  2. Back the project up first, or commit it, so there is always a way back.
  3. Copy one prompt, paste it and read the report. Every prompt tells the AI to report only and to leave your files alone.
  4. Fix the Critical and High items first, one at a time, and run the same prompt again to confirm each fix worked.
  5. Tick the prompt off in the library. Your ticks stay in your own browser.

Three cautions. Never paste a real password or key into a chat. The AI can be wrong, so treat what it finds as a list of leads to check, not a verdict. And if you can, run the audit in a different tool, or with a different model, from the one that built the site: two tools rarely share the same blind spots, and one tool checking its own work tends to repeat them. The legal prompts list gaps and are not legal advice.

The prompt library

Pick a group, open a prompt and read what it checks. The bar shows one line of the prompt beside a Copy button, and Show full prompt reveals the whole text that Copy puts on your clipboard.

20 prompts in 7 groups. Run them one at a time in an AI tool that can read your project, and tick each one off as you go. Your ticks are saved in this browser only.

0 of 20 run

Most hacked small-business sites are not hit by anything clever. A key was left where anyone could read it, an admin page was open, a form had no limits, or a package nobody updated had a known hole. These five prompts look for those. Start with secrets and access, because they are the ones that cost the most when they go wrong.

  • AI tools often put a key straight into the code to make something work, then forget it. A key in the code that reaches the browser, or in the project's history, can be found and used by anyone. This prompt looks for keys in files, in the browser bundle, in git history and in logs, and tells you which ones to replace.

    Find passwords, API keys and tokens that are exposed in the code, the browser bundle, git history or logs.

    You are a senior application security reviewer. Audit this project for exposed secrets.
    
    First:
    - Work out what this project is built with (framework, hosting, database, login system, payment and email services) by reading the config and package files, and tell me in two lines before you start.
    
    Ground rules:
    - Report only. Do not change, create, move or delete any file, and do not run anything that changes data, sends email or deploys.
    - If you cannot tell something from the code, write "cannot tell from the code" and say what I would need to check, instead of guessing.
    - Never print a real secret, key or password. Show the file, the line and the first four characters only.
    - Be specific: name the file and line for every finding, and quote the short piece of code involved.
    
    Check each of these:
    1. Search every file, including config, scripts, docs, tests and example files, for API keys, tokens, passwords, private keys, database connection strings and webhook secrets.
    2. Check which environment variables are exposed to the browser. In Next.js that is anything starting NEXT_PUBLIC_, in Vite anything starting VITE_, and so on. Flag any secret, service-role key, admin key or private API key that has been given a browser-visible name.
    3. Check whether .env files, key files or backups are committed, or listed in .gitignore too late. Look through git history for secrets that were committed and later removed, because removing them from the latest version does not remove them from history.
    4. Check the built output and any source maps for embedded keys.
    5. Check logging and error handling for places that print keys, tokens, session cookies, passwords or full request bodies.
    6. Check that example files (.env.example and similar) contain only placeholders.
    
    Report as a table with these columns: Severity (Critical, High, Medium or Low), Where (file and line), What is wrong, Why it matters, How to fix it.
    Add a second table called "Replace these": for each secret that has been exposed, say which service it belongs to and that I should create a new one and revoke the old one, because deleting it from the code does not make it safe again.
    Then list what looks fine, in one line each. Finish with the three fixes I should make first, in order, and what each one would take (minutes, hours or days).
  • A hidden button is not protection. If the admin area, an API route or a database table answers anyone who asks, it is open. This prompt checks that every protected page and action is checked on the server, that one customer cannot read another's data, and that login, sessions and password reset are built properly.

    Check that every admin page, API route and database rule is protected on the server, not just hidden in the interface.

    You are a senior application security reviewer. Audit this project for broken access control and weak login.
    
    First:
    - Work out what this project is built with (framework, hosting, database, login system, payment and email services) by reading the config and package files, and tell me in two lines before you start.
    - List every route, page, API endpoint, server action and database table or collection in the project, and mark which ones are meant to be private.
    
    Ground rules:
    - Report only. Do not change, create, move or delete any file, and do not run anything that changes data, sends email or deploys.
    - If you cannot tell something from the code, write "cannot tell from the code" and say what I would need to check, instead of guessing.
    - Never print a real secret, key or password. Show the file, the line and the first four characters only.
    - Be specific: name the file and line for every finding, and quote the short piece of code involved.
    
    Check each of these:
    1. For each private route or action, confirm the code checks who is asking on the server before doing anything. Hiding a link or button in the interface does not count.
    2. Look for places where a user could change an id in a URL or request (an order, invoice, booking or profile) and see someone else's data. Check that each lookup also checks ownership.
    3. Check roles. Can an ordinary user reach admin functions by calling the endpoint directly?
    4. If the project uses a hosted database with row-level security or security rules (Supabase, Firebase or similar), check that every table that holds private data has rules switched on and that the rules are not simply allowing everyone.
    5. Check login: passwords are hashed by a proven library or the provider, there is a limit on repeated attempts, and error messages do not reveal whether an email exists.
    6. Check password reset and magic links: tokens are random, expire quickly, work once, and are not logged.
    7. Check sessions and cookies: HttpOnly, Secure and SameSite are set, sessions end on sign-out, and sensitive actions (changing email or password) ask for the password or a second factor again.
    8. Check whether two-factor sign-in is available for admin accounts, and say if it is not.
    
    Report as a table with these columns: Severity (Critical, High, Medium or Low), Where (file and line), What is wrong, Why it matters, How to fix it.
    Then list what looks fine, in one line each. Finish with the three fixes I should make first, in order, and what each one would take (minutes, hours or days).
  • Every box a visitor can type into, every file they can upload and every link parameter is a way in. This prompt checks that the site does not trust any of it: that input is validated on the server, nothing is put into a page or a database query unsafely, uploads are restricted, and forms cannot be hammered by bots.

    Check forms, URL parameters and file uploads for injection, script injection, unsafe uploads and missing limits.

    You are a senior application security reviewer. Audit this project for injection, cross-site scripting, unsafe uploads and abusable forms.
    
    First:
    - Work out what this project is built with (framework, hosting, database, login system, payment and email services) by reading the config and package files, and tell me in two lines before you start.
    - List every place the project accepts input from a visitor: forms, URL parameters, headers, cookies, uploads, webhooks and API calls.
    
    Ground rules:
    - Report only. Do not change, create, move or delete any file, and do not run anything that changes data, sends email or deploys.
    - If you cannot tell something from the code, write "cannot tell from the code" and say what I would need to check, instead of guessing.
    - Never print a real secret, key or password. Show the file, the line and the first four characters only.
    - Be specific: name the file and line for every finding, and quote the short piece of code involved.
    
    Check each of these:
    1. Check that every input is validated on the server (type, length, allowed values), not only in the browser.
    2. Look for database queries built by joining strings or using raw SQL with user input. Parameterised queries or the library's own query builder should be used.
    3. Look for any use of dangerouslySetInnerHTML, innerHTML, document.write, eval or template rendering that outputs user-supplied text without escaping, and for user content shown in emails.
    4. Check redirects that take a destination from the URL, and any server code that fetches a web address supplied by the visitor, because both can be abused.
    5. Check uploads: allowed file types are checked by content and not only by name, there is a size limit, files are stored outside the web root or in a private bucket, filenames are not trusted, and uploaded files cannot be run as code.
    6. Check forms that change data (not only the contact form) for protection against forged requests, such as SameSite cookies or tokens.
    7. Check every public form for spam and abuse protection: rate limiting, a honeypot or a bot check, and a cap on message length. Check that a bot cannot use the form to send email to other people.
    8. Check that error messages shown to visitors do not include stack traces, file paths, queries or internal ids.
    
    Report as a table with these columns: Severity (Critical, High, Medium or Low), Where (file and line), What is wrong, Why it matters, How to fix it.
    Then list what looks fine, in one line each. Finish with the three fixes I should make first, in order, and what each one would take (minutes, hours or days).
  • A site is built from hundreds of other people's packages, and old ones get known holes. The way the site is served matters too: security headers tell the browser what it may and may not do, and a few leftover files or debug settings give attackers a map. This prompt checks both.

    Check dependencies for known vulnerabilities, and the site's security headers, HTTPS, CORS, cookies and exposed files.

    You are a senior application security reviewer. Audit this project for vulnerable packages and weak hosting configuration.
    
    First:
    - Work out what this project is built with (framework, hosting, database, login system, payment and email services) by reading the config and package files, and tell me in two lines before you start.
    - If you can run a read-only dependency audit (for example npm audit, pip-audit or the equivalent), do so and include the results. If you cannot run it, tell me the exact command to run.
    
    Ground rules:
    - Report only. Do not change, create, move or delete any file, and do not run anything that changes data, sends email or deploys.
    - If you cannot tell something from the code, write "cannot tell from the code" and say what I would need to check, instead of guessing.
    - Never print a real secret, key or password. Show the file, the line and the first four characters only.
    - Be specific: name the file and line for every finding, and quote the short piece of code involved.
    
    Check each of these:
    1. List dependencies with known vulnerabilities, with severity and the version that fixes each. Say which ones are in production code and which are only used during development.
    2. List packages that are abandoned, very old, or far behind the current major version, and note anything that needs a major upgrade to be safe.
    3. Check that a lockfile is committed and that the install in production uses it.
    4. Check the security headers the site sends (in the framework config, a middleware file, hosting config or a headers file): Content-Security-Policy, Strict-Transport-Security, X-Content-Type-Options, Referrer-Policy, Permissions-Policy and frame protection (frame-ancestors or X-Frame-Options). Say which are missing and suggest values that will not break the site.
    5. Check that everything is served over HTTPS, that http redirects to https, and that cookies are marked Secure.
    6. Check CORS: flag any rule that allows every origin on endpoints that use cookies or return private data.
    7. Check that debug mode, verbose errors, test routes, sample data and admin tools are switched off for production, and that source maps are not publicly served unless intended.
    8. Check for files and paths that should never be public: .env, .git, backups, database dumps, config files, /admin or /phpinfo style pages, and directory listings.
    
    Report as a table with these columns: Severity (Critical, High, Medium or Low), Where (file and line), What is wrong, Why it matters, How to fix it.
    Then list what looks fine, in one line each. Finish with the three fixes I should make first, in order, and what each one would take (minutes, hours or days).
  • The services your site connects to need handling carefully. A payment webhook that does not check who sent it can be faked, a card number should never touch your own server, and every third-party script on the page runs with the same power as your own code. This prompt checks each of those connections.

    Check payment and webhook handling, API key scope and every third-party script loaded on the site.

    You are a senior application security reviewer. Audit this project for unsafe use of third-party services, payments and scripts.
    
    First:
    - Work out what this project is built with (framework, hosting, database, login system, payment and email services) by reading the config and package files, and tell me in two lines before you start.
    - List every external service the project talks to (payments, email, analytics, maps, chat, forms, CMS, storage) and every third-party script or iframe the pages load.
    
    Ground rules:
    - Report only. Do not change, create, move or delete any file, and do not run anything that changes data, sends email or deploys.
    - If you cannot tell something from the code, write "cannot tell from the code" and say what I would need to check, instead of guessing.
    - Never print a real secret, key or password. Show the file, the line and the first four characters only.
    - Be specific: name the file and line for every finding, and quote the short piece of code involved.
    
    Check each of these:
    1. For payments, confirm that card details are collected by the payment provider's hosted fields or checkout and never pass through this project's own server or logs.
    2. For every webhook endpoint, confirm the code verifies the sender's signature before doing anything, rejects repeats of the same event, and does not trust amounts or prices sent by the browser. Prices and totals must be calculated on the server.
    3. For each API key the project uses, say whether it has more permission than it needs, and whether a restricted key or a separate read-only key would do.
    4. List third-party scripts loaded on pages. For each, say what it is for, whether it loads before the visitor has agreed to anything, and whether it is still needed. Flag scripts loaded from a domain that is not clearly the vendor's.
    5. Check embeds (maps, video, chat widgets) for permissions they do not need, and for whether they set cookies before consent.
    6. Check that outgoing emails and notifications cannot be used to send content to arbitrary addresses.
    
    Report as a table with these columns: Severity (Critical, High, Medium or Low), Where (file and line), What is wrong, Why it matters, How to fix it.
    Then list what looks fine, in one line each. Finish with the three fixes I should make first, in order, and what each one would take (minutes, hours or days).

Free tools to check from the outside

The prompts read your code. These tools look at the live site the way a visitor, a search engine or an attacker would, so use them as well.

What to ask whoever builds your site

Most developers will use AI in some form before long, so the useful question is how they check what it produces:

  • Do they test on real phones, not only in a desktop browser?
  • Do they run accessibility and speed checks before launch?
  • Could they explain what is in the code if something breaks a year from now?

Clear answers to those tell you far more than whether the site was built with AI.

When to bring in a person

An AI audit finds a lot, and it still cannot do everything. It cannot give legal advice, it cannot attack your site the way a professional penetration tester would, and it cannot decide which findings matter most to your business.

Get a person who understands code to look at the site if it takes payments, stores customer accounts or personal details, works in a regulated sector, or if the report from any prompt contains something you do not understand. Twenty findings with no priority is a to-do list that never gets done.

What I would do

  • Use AI freely for drafts and for exploring ideas. It is good at both.
  • Before anything goes live, test it on a real phone, run the free tools above, and check that you can export your content.
  • Work through the prompt library, security and pages first, and fix what it finds one item at a time.
  • Have someone who understands code review security and accessibility, because those are the problems you will not see from the front end.

That is how I work: Claude Code does much of the heavy lifting, and I make the decisions and check everything it produces. If you have tried building a site with AI and got stuck, or you would like a second opinion on one that is already live, get in touch and I will take an honest look.

This is my own opinion, based on what I have seen work in practice, and not legal, financial or professional advice. Every business is different, so take advice on your own situation before making a decision. Views are my own and not those of my employer.

Sources

Frequently asked questions

Can AI build a whole website on its own?

Yes, AI tools can produce a complete, working website from a description. Whether it is a good website is a separate question, and that depends on someone checking the performance, mobile layout, accessibility, security and whether the content matches how your customers search.

Is an AI-built website secure?

Not automatically. AI tools do what they are asked, so if nobody considers form protection, admin access, exposed keys or keeping packages updated, those gaps can go unnoticed until something goes wrong. Security needs someone deliberately checking for it, which is what the security prompts in this article are for.

What should I check before launching an AI-built website?

Start with the basics: open it on a real phone, run it through Google PageSpeed Insights, try using it with only a keyboard, check you can export your content, and make sure the privacy policy, cookie choice and contact details are real. Then work through the prompt library above for security, missing pages, cookies, search, accessibility, forms and email.

Can I use these prompts on Wix, Squarespace or Shopify?

The prompts that read code need access to your project files, so they suit sites built with an AI coding tool, a framework or WordPress. On a hosted builder you cannot open the code, but you can still use the pages, privacy, accessibility and content prompts by pasting in your page text, and run the free outside-in tools listed in the article.

Will the AI fix the problems it finds?

Not unless you ask. Every prompt tells the AI to report only, so nothing changes until you decide. Once you have the report, fix the most serious items one at a time, ideally with a backup or a saved copy first, and run the same prompt again to check each fix worked.

Do I need a cookie banner on a UK website?

It depends on what your site stores or reads on a visitor's device. Anything that is not strictly necessary, such as advertising or tracking, needs the visitor's consent first. The Information Commissioner's Office updated its guidance in April 2026 to include an exemption for some statistical analytics, with conditions, so check the ICO's guidance or take advice before deciding. The cookie prompt in this article lists what your site sets.

Does building with AI coding tools reduce the cost of a website?

It is faster, which can bring the cost down, but most of the value in a build is in the decisions and the checking rather than typing out code. My builds start from £500 for a basic brochure site and are priced case by case.

Is a website built with AI bad for SEO?

Not in itself. Search engines care about speed, structure, content quality and technical correctness, not which tool wrote the code. The problems come from output nobody reviewed, such as missing meta descriptions, muddled headings or oversized images, which a developer checking the work should catch before launch.

How do I know if a developer is using AI responsibly?

Ask how they check what it produces. A good answer covers testing on real devices, accessibility and speed checks before launch, and being able to explain what is in the code if something needs changing later. A vague answer is a bigger warning sign than whether they use AI at all.

Was this helpful?

Share this article

Found this useful? Let's talk about yours.

Tell me about your project and I'll come back with an honest recommendation, timeline and cost, with no obligation.

Start a project