Comparison · ChatGPT-generated code vs. custom
The code it wrote you is real. It just isn't a website yet.
ChatGPT, Claude, v0, Bolt, and Lovable generate owned, irrevocable code: clean markup, responsive layouts/forms, current styling. Missing: hosting/domain/certificate, mail, security review/accountability. This comparison examines when that gap matters.
The differences, side by side
Compare a capable non-developer prompting a small-business site and deploying through research/trial-and-error with ArdinGate's hand-coded build including standard launch work. The table credits generation's advantages, decisive in two rows.
| Factor | ChatGPT / AI code generator | ArdinGate hand-coded |
|---|---|---|
| Cash cost to get the code | $0–$20/mo for the AI subscription: unbeatable, and worth saying plainly | $1,200–$2,800+ one-time |
| Your time to a live, working site | An afternoon if you're technical; two to six weekends if you're not | Yours: a kickoff call, content, and a review round |
| Ongoing annual cost | $100–$250/yr self-hosted (domain + shared hosting), plus your own time | $720–$2,160/yr managed, or the same $100–$250 if you'd rather host it yourself |
| Code ownership and portability | Complete: it's your folder of files, no platform can revoke it | Complete: same deal, and delivered documented |
| Hosting, deploys, and updates | Yours. Pick a host, upload, and remember to do it again next time | Set up and maintained, or handed over documented if you self-host |
| Domain, DNS, and TLS certificate | Yours to configure, and the first time is the hard time | Configured, tested, and monitored, including automatic renewal working |
| Contact form and email delivery | Form code written; sender authentication, spam handling, and a tested delivery path are not | Validated, rate-limited, spam-filtered, authenticated, and confirmed arriving before launch |
| Security review | None. The code runs correctly, which tells you nothing about whether it's safe | Input handling, request forgery, rate limits, headers, and error exposure all checked |
| Accessibility | Partial by luck, usually decent markup, rarely focus states, contrast, or keyboard testing | Tested: contrast, keyboard navigation, focus visibility, labels, alt text |
| SEO foundations | Commonly missing entirely: canonicals, sitemap, structured data, per-page titles, social previews | Written by hand for every page, plus Search Console registration |
| Page speed | Often good: generated code is lean, though images usually ship unoptimized | Under 1 second, with images sized and served in modern formats |
| Who fixes it when it breaks | You, plus a new chat that has never seen your site before | The person who wrote it, by email or phone, same day on Priority |
| The site 18 months from now | Code nobody remembers writing, on a host nobody has logged into since launch | Documented, patched, backed up, and someone still on the hook for it |
What this comparison is about
Generated code is fully yours. The question is who verifies correctness and answers for errors? Models confidently produce plausible code without distinguishing subtle mistakes. Supply your own review/operations and generation is useful; otherwise, apparent savings transfer unpriced risk to your business.
When a ChatGPT-built site is the right call
There are situations where hiring anyone would be a waste of money and generating the site yourself is straightforwardly the correct decision. They share one property: either you can supply what the model cannot, or the stakes are low enough that the missing pieces do not matter.
You are technical and the site is simple
Experienced developers, system administrators, and deployers can build a static site Saturday morning: review code, catch escaping gaps, authenticate mail, choose hosting. Hiring buys convenience. Outsource tedious copy/photos/SEO while retaining the build, but commit to maintenance two years later; these projects often stagnate.
One page, no form, no payments, no personal data
A static identity/phone/email page processes, submits, and stores nothing. With free TLS, security largely means protecting the hosting password; generated code suffices without paid risk management. Forms, booking widgets, uploads, or payments turn it into an application requiring security work.
Prototypes, internal tools, and throwaway landing pages
Two-week ad pages, three-colleague dashboards, investor prototypes, and staff references behind existing login suit generation: short lives, friendly audiences, little reputational exposure. Full builds waste effort on fortnight projects. Set a removal reminder; disposable sites often linger.
You are learning, and building it yourself is the point
Read generated code, ask why, deliberately break/fix it: you'll learn faster than tutorials and gain a maintainable site. Hiring defeats that learning purpose. Keep experiments separate from the live business: a site collecting strangers' phone numbers needs knowledgeable review, even if that's all you buy.
Pre-revenue, with more time than money
Pre-revenue, generated files/free static hosting/$15 domain buy portable ownership for sandwich money, better than renting builder files. Spend weekends, but register the domain yourself, keep URLs simple for rebuilding, and verify form delivery before relying on it. Lost early enquiries are expensive.
Where custom development outperforms generated code
None of these are claims that a developer writes better markup than a model. Frequently they do not. Every one of them is about the work that exists outside the file the model produced, and about the fact that correctness is not visible from a browser.
Generated code is plausible, and plausible is not the same as correct
Models confidently produce plausible errors: functions renamed three versions ago, nonexistent options/library versions, leap-year failures, or decade-old practices now known vulnerable. Formatting and certainty don't distinguish wrong answers from correct ones.
Running code is the dangerous case: crashes warn you; perfect rendering with unescaped input doesn't. Clicking tests only normal behavior. Developers contribute knowledge of medium-specific failures, not typing speed.
The contact form is where it goes wrong, and it goes wrong silently
Generated name/email/message forms often pass visitor email directly into server-mail headers. Unstripped newlines permit header injection and mail impersonation. Missing request-forgery tokens, rate limits (ten thousand overnight bot submissions), honeypots, and escaped confirmations compound risk. The form works for attackers too.
Unauthenticated server mail may land in spam or vanish while the page says thanks. Businesses discover months of lost leads only after customers complain. Professional launch work configures sending-domain records and confirms actual inbox delivery.
Everything you did not know to ask for
Unrequested launch work is often absent: canonicals, sitemap/robots, structured data, unique titles/descriptions, social tags/favicons, 404, keyboard focus, image alt text, heading hierarchy, small-text contrast, modern right-sized images, analytics. Models answer prompts without owning outcomes.
These small fixes enable indexing, sharing, keyboard/screen-reader access, and page-to-job attribution. Models can supply each when asked. Knowing the checklist is the expertise; a site can look complete without it.
Operations is the job, and the code was the easy part
After code: register/control the domain, point DNS, issue TLS and verify renewal, test backup restoration, track host-scheduled PHP upgrades. Otherwise, a future deprecation can become a blank rarely tested form.
Monitor uptime before customers report outages; review logs for brute-force attempts and broken addresses before months pass. Managed hosting performs these routine operations on schedule, keeping sites working without demanding your attention.
Nobody is accountable, and you cannot escalate to a model
At eleven Friday night before weekend ads, you're troubleshooting or asking a fresh chat unfamiliar with your site, changes, server, and decisions. It may help confidently, but owes no support and bears no stake in Monday's trading.
An accountable builder remembers decisions, reads logs, and responds. Priority provides same-day critical response and a test environment for verifying fixes before production. For owners, bad-day duration often determines a website's value.
Eighteen months later, when you need to change something
A year and a half later, new services, prices, and browser fixes confront unfamiliar code from nineteen vanished chats. The model remembers nothing; repasting can rewrite three unrelated things and break a fourth unnoticed for a week.
The result is stagnation or regeneration that loses accumulated fixes. Documented, consistently structured custom sites retain a knowledgeable maintainer. Small changes stay manageable; postponing frustrating edits, rather than indifference, ages most sites.
The arguments for shipping the generated version
All four of these are made in good faith and two of them are largely correct. The point of answering them is not to talk anyone out of a reasonable decision: it is to make sure the decision is being made with the actual costs on the table.
"The code it wrote works, though"
It probably works in your browser on your machine for the cases you tried. That has value, but cannot establish safety against unexpected input, successful notification delivery, layouts on unfamiliar screens, keyboard-accessible navigation, or correct indexing. Those failures can look identical to a working site. That invisible gap between working and correct costs money.
"If something's wrong I'll just ask it to fix it"
Models fix described problems, not unnoticed header injection, recipient-specific spam, or missing canonicals splitting rankings across two URLs. Asked for a form, they supply one without raising these omissions. Unread iterative rewrites also drift until neither owner nor model understands the current code.
"AI is twenty dollars a month and you're $1,200"
If cash is binding, generate, host cheaply, and revisit once revenue arrives. Non-developers commonly spend two to six weekends on DNS, failed certificates, opaque form errors, and mail records. Valuing those hours narrows savings; one month of spam-lost enquiries can erase them.
"Don't you use AI to build sites too?"
Yes, AI speeds scaffolding, boilerplate, refactoring, and drafts, continuing thirty years of better tools. Every line is then read, browser/device-tested, security-reviewed, and checked for accessibility/performance. The invoicing developer owns the outcome. You're paying for review and accountability; unchecked launch, not model authorship, is the problem.
The security review nobody runs
Input-handling failures are well-known and largely unchanged for fifteen years. Generated sites look complete when rendered, so nothing prompts the short, necessary review. The missing process creates risk.
Validate all input and escape output, including emails/confirmations. Add form-origin tokens and rate limits: ten thousand overnight bot submissions can trigger host suspension. Strip email-header newlines. Check upload type/size server-side and store outside executable locations. Disable production errors exposing file paths; enable protective browser security headers. Configure sending records so notifications avoid spam.
Models implement named safeguards when asked but don't tell you to ask. Expertise means knowing the list, selecting applicable checks, and completing them before launch rather than after an incident.
Match rigor to function: a form-free single page needs little of this list. Accepting submissions raises risk; storing personal data or payments raises it again. Don't sell fear beyond that scope.
The bottom line
Generated code is a good starting point and a much better one than a rented builder in at least one important respect: you own it, it is portable, and nobody can switch it off. The question is not whether the tool is good. It is whether you can supply the parts it does not.
The verdict
Technical owners, static single pages, and short-lived prototypes can use generated code without paid reassurance. Collecting strangers' information, taking payments, or generating essential leads requires accountable verification. Chat cannot provide that responsibility or warn you reliably. Bring work already started as a useful foundation; the danger is shipping apparently finished code without discovering its gaps.
Pricing
Single-page home/services/contact builds start at $1,200 ($1,200–$2,200). Multi-page home/services/about/contact plus service/location pages start at $2,800 ($2,800–$5,000), including keyword-targeted Google headlines/summaries, business/FAQ labels, duplicate-page signals, sitemap, and Google registration/indexing. Single-page SEO adds $300.
Bring generated code early; it changes the quote. One coherent page needing form/security review, responsive fixes, SEO, accessibility, and deployment often takes a few billable hours. Repeatedly regenerated, inconsistent pages may cost more to untangle than rebuild; you'll be told. Your preferred layout remains the reference either way.
Optional hosting: documented files run on self-managed PHP at roughly $6–$18/month. Core $60/month provides monitoring, nightly backups/30-day retention, SSL, DNS, and server patches, no edits. Care $100 adds application patches/one monthly edit hour. Priority $180 adds three hours, same-day critical response, testing, and monthly analytics. Managed hosting performs scheduled operations for you.
Common questions
Can ChatGPT build a website?
ChatGPT can produce good five-page landscaping-site code: clean markup, responsive design, working navigation/forms, current styling. It doesn't register domains, select hosting, upload files, configure DNS/TLS, authenticate mail, test real phones, register with Google, or catch unsafe form input. These numerous unfamiliar tasks require knowing what to ask. Generation handles the already-fast code work, leaving the slower route to a findable, usable live site.
Is AI-generated website code secure?
Generated code usually exceeds untrained work but falls below careful development. Forms may put visitor email in mail headers without stripping newlines, enabling impersonation. Common omissions: request-forgery tokens, rate limits against ten thousand submissions, escaped output, upload type/size checks, and disabled production errors revealing file paths. Browsers show working behavior, not these vulnerabilities; knowledgeable review is essential.
What is the difference between a ChatGPT-built site and an AI website builder like Wix AI?
Hosted AI builders supply a working site without exportable ownership; cancel and it disappears. ChatGPT supplies portable, irrevocably owned code without infrastructure, operations, or verified correctness. Choose what you can supply: unable to configure DNS or review forms, prefer builder convenience. Technical deployment/review skills make owned generated code the stronger starting point.
How much does it cost to host a website ChatGPT wrote for me?
Static markup/styles can host free with TLS; domains cost $15–$25/year. Server forms need $6–$18/month shared hosting or self-administered $5–$12 cloud servers: about $100–$250/year total versus builders' $300–$800/year or ArdinGate's $60–$180/month. Savings exclude server patches, failed-renewal fixes, tested backups, monitoring, and host incidents. Self-hosting costs attention; decide whether you have it.
Will Google penalize a website built with AI-generated code?
No AI-code penalty or authorship signal: Google ranks content/performance, and clean generated semantic markup can help. Omissions hurt: canonicals, sitemap/robots, structured data, unique titles, social tags, alt text, heading hierarchy, analytics. These aren't violations, but impair indexing, result presentation, and measurement. A knowledgeable afternoon can fix them; finished-looking pages often hide the checklist entirely.
Can I take ChatGPT's code and have a developer finish it?
Yes. Finishing coherent generated pages with security/forms, responsive fixes, SEO, accessibility, and deployment often takes hours. A twelve-page site from three weekends/twenty chats, four styling rewrites, and inconsistent headers may cost more to untangle than rebuild; an honest developer says so. Ask early, around one or two coherent pages. If code already resists you, preserve the layout as a rebuild reference.
What usually breaks first on a ChatGPT-built website?
First: silent form failure from unauthenticated mail or unmonitored inboxes, discovered after months when customers complain. Second: unverified TLS renewal expires, warning visitors. Third: tables/wide images break other screen sizes despite your phone passing. Fourth: host PHP upgrades turn deprecations into fatal errors on rarely visited pages. All routinely escape the owner's browser checks.
Does ArdinGate use AI to build websites?
Yes. Like compilers, linters, and frameworks, AI speeds scaffolding, boilerplate, refactoring, and drafts. It doesn't decide or ship unreviewed work: every line gets human reading, browser/real-device tests, security review, and accessibility/performance checks. The named invoicing developer owns judgment, operations, and results. The problem is unchecked delivery, not generated origins; the same standard applies to your own use.
How long does it take to get from ChatGPT's output to a live, working website?
Experienced deployers: an afternoon. Newcomers: two to six weekends, mostly DNS records, hosting plans without server code, TLS blocked by propagation, opaque 500 form errors, sender authentication, responsive bugs, and nonexistent library versions. Unfamiliar operations without help turn an afternoon into a month. Price that time before calling generation free.
Are v0, Bolt, Lovable and Cursor different from asking ChatGPT?
Project-aware tools preview/deploy within minutes, sometimes provision databases/auth, improving plain chat. You still own security/operations; platform deployment restores subscriptions/lock-in. Review forms, escaping, rate limits, mail authentication, accessibility, structured data, speed, and backups. Faster running isn't completed work.
When is a ChatGPT-built website the right choice?
Generate when you can handle deployment/review, need only a form/payment/personal-data-free single page, have a prototype/internal tool/two-month campaign, or are pre-revenue and learning with more time than money. A technical Saturday doesn't justify $1,200+. Once strangers' information, payments, or essential leads depend on it, someone must verify correctness and own responsibility; chat cannot.
Already have code an AI wrote you?
Send it over. You'll hear what's safe, what isn't, what's missing before it should go live, and whether finishing it costs less than starting again, including if the honest answer is that it's fine as it stands.
Get a straight answer