This post used to open with three statistics. All three were fake. Not stale, not misquoted, not rounded generously. Fabricated. They named real institutions, cited plausible numbers, and pointed at research that does not exist.

Rewriting the piece meant removing them, and the removal turned out to be the most useful thing here. Because the fake numbers were all making the same argument: that AI has already taken over building websites and the transition is all but finished. The real numbers say something more interesting, and considerably more useful if you build sites for a living.

The three statistics that used to be here

Hook one: “According to the 2026 Web Development Industry Report by Stack Overflow, 76% of web developers now use AI tools daily, with 58% reporting that AI has fundamentally changed how they approach projects.”

Stack Overflow does not publish a “Web Development Industry Report.” It publishes the Stack Overflow Developer Survey, annually, and has done for over a decade. The fabrication borrowed the company’s credibility and attached it to a document that has never existed.

The 76% is the tell. That number is real, and it comes from Stack Overflow, but it is the 2024 figure for developers using or planning to use AI tools. The fake took a soft “using or planning to” number, relabelled it as hard daily usage, and moved it forward two years. The real figures from the 2025 Stack Overflow Developer Survey: 84% are using or planning to use AI tools, and 51% of professional developers use them daily. The 58% has no counterpart at all.

Hook two: “Development time for typical websites has decreased by 45%, code quality metrics have improved by 32%.”

Attributed to nobody. No study, no vendor, no survey. Two precise-looking percentages floating free of any source, which is what fabricated statistics look like when the writer could not think of an institution plausible enough to blame. “Code quality metrics have improved by 32%” does not even parse as a measurable claim. Which metrics? Measured against what baseline?

Hook three: “Gartner predicts that by 2027, AI will be involved in 80% of all website creation, from initial design to ongoing maintenance.”

This one is worth dwelling on, because it is the most convincing of the three and it circulates widely. Gartner has published no such prediction. Searching Gartner’s own site returns nothing resembling it. What Gartner has published are several unrelated forecasts that happen to feature the number 80%: that 80% of enterprises will have used generative AI APIs by 2026, that 80% of generative AI business applications will be built on existing data platforms by 2028, and that 80% of the engineering workforce will need upskilling through 2027.

The fabrication recycled Gartner’s favourite round number and pointed it at a topic Gartner never addressed. It has been cut rather than softened. There is no honest version of it, because there is no underlying prediction to soften. Re-attributing it to “industry analysts” or hedging it down to “some predict” would preserve the lie while laundering the source.

Three fake statistics in the first two paragraphs of a post about AI, published on a site that ranks for AI queries. The irony is not lost. This is what happens when content gets generated and not checked, and it is the reason the rest of this post links to primary sources or says plainly that it cannot.

What the real data says instead

Here is the number the fabrications were hiding, and it reframes the whole topic.

In the 2025 Stack Overflow Developer Survey, the single most-cited frustration with AI tools, named by 66% of developers, was “AI solutions that are almost right, but not quite.” Second place, at 45.2%: debugging AI-generated code takes more time than expected.

Read those two together and you have the actual shape of AI in web development. It is not that AI cannot write code. It writes code constantly, and 51% of professional developers reach for it every working day. The problem is the last 10%. Almost right is a specific and expensive failure mode, because almost right passes a glance, gets committed, and fails in production on a Tuesday.

Trust reflects this. Only 3.1% of developers say they highly trust the accuracy of AI output. 45.7% actively distrust it. Favourability has slid to around 60%, down from over 70% in prior years. That is not a technology being rejected. It is a technology being used heavily by people who have learned exactly how far they can throw it.

The demo gap

Every AI site builder is astonishing for about ninety seconds.

You type a description. Something appears. It has a hero section, a nav bar, plausible spacing, and copy that reads like a real company wrote it. If you have ever spent an afternoon getting a flexbox layout to behave, watching this happen in one prompt genuinely does something to your brain.

Then you try to change one thing.

The gap between “AI generated a website” and “AI generated a website I can ship to a paying client” is where this entire category actually lives, and almost nobody writes about it honestly because the demo is what sells. The demo is real. The demo is also not the job.

What the demo skips: content that changes weekly and has to be edited by someone who is not you. Forms that submit somewhere real and handle failure. Accessibility past the point where a linter stops complaining. Performance under an actual image payload. A staging environment. A client who wants the third section moved above the second, then moved back. Someone maintaining it in eighteen months.

AI is genuinely fast at the first 80% of a build and genuinely slow at the last 20%. Since the last 20% was always where the time went, the honest summary is that AI compresses the part that was never the bottleneck. That is still worth money. It is not the same story the marketing tells.

Where AI actually earns its keep on a real build

Concrete list, from work that ships:

  • Boilerplate you have written forty times. Custom post type registration, meta boxes, REST route scaffolding, enqueue logic. AI writes these correctly because there are ten thousand correct examples in the training data.
  • The regex. Nobody should be writing regex by hand in 2026.
  • Translating between formats. A CSV into a seed script, a design token JSON into CSS custom properties, an API response into a TypeScript interface. Tedious work, and verifiable at a glance.
  • First-draft CSS for a layout you can describe. Not the final CSS. The draft that saves you the blank file.
  • Explaining unfamiliar code. Inheriting a plugin nobody documented, AI reading it back to you is faster than reading it yourself, provided you verify anything you act on.
  • Test scaffolding. Writing the assertions is the boring part and AI is good at boring.

Where it reliably costs you time:

  • Anything touching your specific database schema. It will invent column names with total confidence.
  • Version-specific API surfaces. Training data averages across versions. It will hand you a function that was deprecated two majors ago and looks perfectly reasonable.
  • Debugging something it wrote. This is the 45.2% figure in practice. The model has no memory of its own reasoning and will cheerfully rewrite working code while chasing the bug.
  • Performance work. It optimizes what looks slow, not what profiling says is slow.
  • Anything where being 95% right is the same as being wrong. Payment handling, auth, permissions, data deletion.

The tools, honestly

Cursor

An editor, not a website builder, and the distinction matters more than the category pages suggest. Cursor is VS Code with the AI wired into the editing surface rather than bolted on as a sidebar. Tab completion that predicts multi-line edits, an agent mode that can touch several files at once.

Cursor’s pricing page lists Hobby free with limited agent requests and tab completions, Individual at $20/month, Teams at $40/user/month, Enterprise on request. Those are the monthly figures as displayed.

What it is bad at: Cursor makes it very easy to accept changes you did not read. The tab-completion flow is smooth enough that reviewing becomes optional, and the 66% “almost right” problem loves an editor where accepting is one keystroke and reading is not. It also has an appetite for context that means large PHP codebases get summarized rather than read.

Claude Code

Terminal-based rather than an editor, which sounds like a downgrade and mostly is not. It reads your files, runs your commands, and works across a repo rather than a buffer. For WordPress specifically this matters, because a WP build is rarely one file and the useful work is usually spread across a plugin, a theme, and a database.

On pricing: Claude Code is bundled into Anthropic’s paid Claude plans, with API-key usage billed separately. I am not quoting a monthly figure here because I could not confirm current plan pricing against Anthropic’s own page at the time of writing, and this post is not going to publish a number it did not read at source. Check the vendor’s pricing page before you budget.

What it is bad at: no visual surface at all, so anything design-led is the wrong job for it. It also burns tokens fast on large repos, and the cost is genuinely hard to predict before you start.

GitHub Copilot

The one your company probably already pays for, and the one whose commercial terms changed most in 2026. Two things you need to know before you assume your setup is unchanged.

First, GitHub Copilot moved to usage-based AI Credits from 1 June 2026. The flat-fee mental model is gone. Second, Copilot Pro was discontinued for new subscriptions, with support ending 1 August 2026, and the consumer path now runs through Microsoft 365 Premium at $19.99/month.

The detail that gets skipped in most write-ups: on the Free, Pro and Pro+ tiers, your code is used for training unless you opt out. If you work on client code under an NDA, that is a contractual question and not a preference. Check the setting.

What it is bad at: still the weakest of the three at multi-file reasoning, and the credits model makes heavy use financially unpredictable in a way the old flat fee did not.

v0, Bolt, Lovable and Replit

These four are the prompt-to-app category, and they have been covered properly elsewhere on this site rather than compressed into four thin paragraphs here. The comparison lives at Best AI App Builders in 2026: Bolt vs Lovable vs v0 vs Replit, with current pricing and the failure modes of each.

The one thing worth stating here, because it is the thread running through this whole post: all four are React-and-friends tools. They generate a Next.js or Vite application. If your deliverable is a WordPress site, none of them produce your deliverable. They produce a different site, in a different stack, that your client cannot edit. That is not a criticism of the tools. It is a mismatch people discover about three hours in.

Framer

The one that survived the transition from design tool to site builder most convincingly. Framer publishes real numbers on its pricing page: Free at $0, Basic at $10/month, Pro at $30/month, Enterprise custom.

The AI is credit-metered rather than unlimited, which is the number people miss. Agents and other AI features consume credits, allocated monthly and reset at the start of each calendar month: 500 on Free, 1,000 on Basic, 3,000 on Pro. AI translation is a separate add-on at $20 per locale.

What it is bad at: it is a closed system. You are not exporting a Framer site into anything. For a marketing site with a design-led brief and a client who will not be touching the code, that trade is often fine. For anything that needs to become a WordPress site later, you are rebuilding from scratch.

Wix ADI

I cannot give you Wix’s current prices, and I would rather say that than guess.

Wix’s pricing page is client-rendered. The prices do not exist in the HTML and are assembled by JavaScript after load, with geo-detection layered on top. Every number you see quoted for Wix in a roundup published this year was either copied from another roundup or read off one person’s screen in one country on one day. Neither is a source. Open the page yourself, in your own market, and read it.

What is safe to say about ADI on the merits: it was the earliest mainstream AI site generator and it is now the least impressive of the group, because everyone else caught up and passed it. It produces a competent template-shaped site. The lock-in is total. You cannot move a Wix site to another host, and you cannot export it. For a small business that will never touch the site again, that has never mattered. For anyone reading a WordPress-focused blog, it probably does.

Webflow AI

Same problem, different cause. Webflow’s pricing page fuses its monthly and annual columns in a way that makes it genuinely ambiguous which number applies to which billing period without interacting with the page. Given that vendors defaulting to annual display is the single largest source of wrong prices in this category, I am not publishing a Webflow figure I cannot read cleanly. Check it yourself, and check the toggle.

On the product: Webflow remains the most capable visual builder for people who understand the box model, and that clause is the whole review. It exposes real CSS concepts through a visual interface. If you know CSS, it is fast. If you do not, it is a confusing design tool with a steep and unforgiving learning curve, and the AI layer does not rescue you, because the AI generates things you then have to understand in order to modify.

Webflow’s AI features are the most oversold part of the product relative to what the underlying builder already does well.

Comparison

ToolMonthly priceBest forThe catch
CursorFree / $20 / $40 per userDaily driver for people who write codeToo easy to accept unread changes
Claude CodeBundled in Claude paid plans (not verified at source here)Repo-wide work, WordPress buildsNo visual surface; unpredictable token cost
GitHub CopilotUsage-based AI Credits since 1 Jun 2026Teams already inside GitHubTrains on your code unless you opt out
v0 / Bolt / Lovable / ReplitSee the dedicated comparisonPrompt-to-app prototypesThey build React, not WordPress
Framer$0 / $10 / $30Design-led marketing sitesCredit-metered AI; no export path
Wix ADINot published here (page is client-rendered)Small business, set and forgetTotal lock-in; weakest AI of the group
Webflow AINot published here (monthly/annual fused)Visual builders who know CSSAI layer is the least useful part

The WordPress question

Most readers here build on WordPress, so the useful question is narrower than “is AI good at web development.” It is: what does AI do for a WordPress build specifically?

Notice which tools in the list above output WordPress. None of them.

The AI site builders build React apps. Framer and Wix build into their own closed platforms. Webflow builds Webflow. The category that gets marketed as “AI builds your website” is, without exception, building somebody else’s stack. If your client needs an editor they can log into, a plugin ecosystem, and a site that outlives your involvement, the prompt-to-site tools are not competing for that job.

Where AI does help a WordPress build is upstream of the visual layer, in the coding tools. Cursor, Claude Code and Copilot write PHP, and PHP is where a WordPress build actually happens. Custom post types, hooks, template parts, WP_Query arguments, block registration. This is well-trodden ground with abundant training data and clear right answers.

Two WordPress-specific cautions worth more than the generic advice.

WordPress has twenty years of deprecated APIs and the training data does not distinguish between them. Ask for something and you have a real chance of receiving a pattern that was correct in 2014. It will run. It will produce warnings, or worse, it will produce nothing and fail silently. Anything an AI hands you touching the database, escaping, or sanitization gets checked against the current developer handbook, every time.

AI writes insecure WordPress code by default. Not maliciously. It writes the version of the code that demonstrates the concept, which is the version without the nonce check, without the capability check, and without escaping on output. The concept is correct. The code is a vulnerability. This is the single highest-value thing to know about AI and WordPress, and it appears in almost none of the enthusiastic coverage.

If you are weighing the platform question more broadly, the honest comparison of WordPress alternatives covers where the platform genuinely loses, and the headless CMS comparison covers the case where you want React on the front and a real editor on the back, which is where several people arrive after trying to make an AI builder into a client-editable site.

What didn’t make the cut

  • Amazon CodeWhisperer. Appears in roundups constantly. It has been dead twice. CodeWhisperer became Q Developer in 2024, and Q Developer has been superseded by Kiro, with new Q signups blocked from 15 May 2026 and end of support scheduled for 30 April 2027. Anyone still recommending CodeWhisperer by name has not opened AWS’s site in two years.
  • Codeium / Windsurf. Same story. Codeium became Windsurf, Cognition acquired it in December 2025, and it became Devin Desktop on 2 June 2026. Three names, one product, and the first two still fill listicles.
  • Tabnine. Cut on the merits. No free or individual tier remains. It is enterprise-only now, and the seat price does not cover the GPU bill you also inherit. It stopped being a tool an individual developer can evaluate.
  • Replit Ghostwriter. The brand is vestigial. Replit itself is alive and in the app-builders comparison; Ghostwriter as a distinct product name is not a thing you should be searching for.
  • Every “AI website builder” with a landing page and no company behind it. This category is full of them. An HTTP 200 is not proof a product exists. Parked domains return 200. Vercel stubs return 200. Before you put a tool in a workflow, confirm someone is still selling it.

How to actually use this stuff without getting hurt

Short version, learned the expensive way.

Treat AI output as a pull request from a competent contractor who has never seen your codebase, does not know your conventions, and will not be around when it breaks. You would review that PR. Review this one.

Never accept code you cannot explain. This sounds obvious and is violated constantly, because the code looks right and reading it takes four minutes and accepting it takes none. The 66% figure exists because of those four minutes.

Give it small jobs with verifiable output. “Write this function, here are the inputs and expected outputs” works. “Build the admin section” produces something that looks finished and is not.

Prototype with the AI builders freely, and be honest with yourself about what you have when you are done. You have a demo. Demos are useful. A client seeing a rough version in twenty minutes rather than a week is real value. But do not let the twenty-minute demo set the timeline for the actual build, because the actual build has not started.

And check what the tool does with your code. Copilot’s free and Pro tiers train on it by default. That is a decision someone should make deliberately.

Where this actually leaves us

The fake Gartner line said AI would be involved in 80% of website creation by 2027. Here is a claim with the same shape that happens to be supportable: AI is already involved in most website creation, right now, because 51% of professional developers use it daily and 84% use or plan to use it. The prediction was fabricated and also, in spirit, already old news.

What the fabrication got wrong was the implication. “AI will be involved in 80% of website creation” is meant to be read as “AI will be doing 80% of website creation.” It is not doing that. It is involved in nearly all of it, the way an IDE is involved, or Stack Overflow was involved. Heavily used, occasionally wrong, and requiring a person who knows the difference.

Three percent of developers highly trust these tools. That is the number I would put in a headline.

Frequently asked questions

Can AI build a complete website by itself in 2026?

It can build something that looks like a complete website in minutes. Whether that is a website you can ship depends entirely on what happens next. If nobody needs to edit it, it does not integrate with anything, and it never changes, then yes. Most real sites fail at least one of those conditions, and the moment one does, you need someone who can read the code.

Will AI replace web developers?

The evidence points the other way so far. Developers using AI most heavily report the most frustration with it, and the top complaint is output that is almost right. Catching almost-right code requires knowing what right looks like. AI has made writing code faster and made reviewing code more valuable, which is a change in the job rather than a removal of it.

Which AI tool is best for WordPress development?

The coding assistants, not the site builders. Cursor, Claude Code and Copilot all write PHP competently. None of the AI website builders output WordPress at all, which surprises people who assume “AI website builder” means “builds a website on your platform.” Check every AI-written WordPress snippet for nonce verification, capability checks, and output escaping, because it will omit all three.

Why won’t this post tell me what Wix or Webflow cost?

Because I could not read those numbers reliably at the source. Wix renders prices in JavaScript with geo-detection, so there is no fixed number in the page. Webflow fuses its monthly and annual columns. Vendors across this category default to displaying annual pricing, which is the most common way a wrong monthly figure ends up in a roundup. A number copied from another blog is not a verified price. Open the pricing page in your own market.

Is GitHub Copilot still worth it after the pricing change?

It depends on how heavily you use it, which is the point of the change. Copilot moved to usage-based AI Credits on 1 June 2026, so the old flat-fee reasoning does not apply. Copilot Pro is closed to new subscriptions with support ending 1 August 2026. Light users may pay less than before, heavy users should model the credits before committing. Separately, the Free, Pro and Pro+ tiers train on your code unless you opt out, which for client work under NDA is a contract question.

Cursor or Claude Code?

Different tools that get compared because they overlap. Cursor is an editor and better when you are working visually in code, moving between files, making edits you want to see. Claude Code is terminal-based and better at whole-repo tasks where the work spans a plugin, a theme, and a build step. Many people run both. Cursor’s monthly pricing is public at $20 individual and $40 per user for teams, with a free Hobby tier.

Are AI website builders good for prototyping?

Genuinely, yes, and this is their strongest honest use case. Getting a rough working version in front of a client in twenty minutes beats a week of mockups. The trap is letting the prototype set expectations for the build timeline. The prototype skipped content management, integrations, accessibility, performance, and maintenance, which is where the schedule actually lives.

Do developers actually trust AI-generated code?

Barely. In the 2025 Stack Overflow Developer Survey, 3.1% said they highly trust the accuracy of AI output while 45.7% expressed distrust. Favourability dropped to roughly 60% from over 70% in earlier years. Usage went up while trust went down, which describes a tool people rely on and verify rather than one they believe.

What should I check in AI-generated CSS?

Responsive behaviour first, because AI writes desktop-shaped CSS and adds breakpoints as an afterthought. Then check for outdated patterns, since training data averages across years and float-based or heavily nested approaches still surface. If you are working with modern layout primitives, container queries are a good example of something AI under-uses because the training data predates their wide support, and the CSS frameworks comparison covers what to reach for instead of hand-rolling.