Web Development
Last reviewed
Sources
44 citations
Review status
Source-backed
Revision
v3 ยท 4,682 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
44 citations
Review status
Source-backed
Revision
v3 ยท 4,682 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Web Development ChatGPT Plugins
AI in web development is the use of large language models and related generative systems to design, build, deploy, and maintain websites and web applications. The category spans text-to-website generators such as Vercel v0, Bolt.new, and Lovable; traditional website builders that grafted generative AI onto existing editors, including Wix, Squarespace, Webflow, Framer, Hostinger, Durable, and GoDaddy; headless CMS platforms that added AI helpers (Contentful, Storyblok, Sanity, WordPress via Jetpack); developer frameworks for building AI features into web apps, most notably the Vercel AI SDK; component and design-to-code tools such as shadcn/ui and Builder.io Visual Copilot; and AI assistants for accessibility, SEO, and on-site chat. The shift accelerated after the 2022 release of ChatGPT and entered a distinct phase in early 2025 with the popularization of "vibe coding", a style of building software primarily by prompting AI rather than by writing code line by line.
Web development sits on a stack with several layers, and AI tools now exist at each of them. At the top, end-user website builders aim to convert a sentence or two of business description into a finished marketing site. One step down, prototyping and component tools generate React, Vue, or HTML snippets that designers and engineers can paste into existing codebases. At the framework level, libraries like the Vercel AI SDK and LangChain.js help developers wire up chat, retrieval-augmented generation, and tool use inside their own products. Finally, AI assists with cross-cutting concerns such as accessibility testing, search engine optimization, content writing, and customer support widgets.
The split between "AI features inside a builder" and "a builder that is itself an AI agent" is the defining tension in the space. Wix, Squarespace, GoDaddy, Hostinger, and Durable operate the former: their AI layers sit on top of a structured page model that constrains output to known templates and components. Bolt.new, Lovable, v0, and Replit Agent operate closer to the latter: they generate arbitrary code, run it in a WebContainer or sandbox, and let the user iterate by typing follow-up prompts. The first group prioritizes guardrails and predictable output; the second prioritizes flexibility at the cost of higher rates of broken code, hallucinated dependencies, and accessibility failures.
The first widely adopted AI feature in mainstream web tooling was Wix ADI (Artificial Design Intelligence), launched by Wix in 2016. ADI asked users a series of preference questions, then assembled a website from a structured template library that matched the answers. It was not a generative system in the post-2022 sense; the "intelligence" was a rules-based recommender that paired user answers with prefabricated layouts. ADI ran inside Wix for years and was eventually folded into what Wix later called Wix Harmony.
The arrival of ChatGPT in November 2022 reset the field. Within six months, vendors began layering GPT-powered text generation, image generation, and conversational onboarding into their existing builders. WordPress.com launched the Jetpack AI Assistant on June 6, 2023, focused on inline writing assistance inside the Gutenberg editor. Wix announced its AI Site Generator on July 17, 2023, going beyond ADI's question flow to a single conversational prompt that produced a draft site with text, images, and basic business apps. Squarespace introduced Blueprint as a guided design system in 2023 and folded AI text and image generation into it. GoDaddy and Hostinger followed with their own AI builders during 2023 and 2024.
A second wave began in late 2023 with Vercel v0. Announced on October 11, 2023, v0 was a developer-facing generative UI tool. It produced React and Tailwind CSS code rather than rendering inside a proprietary builder, which made it usable across any Next.js or React codebase. Roughly 100,000 users joined the v0 waitlist within three weeks of the beta announcement. The success of v0 prompted a cluster of similar products. Lovable, founded in Sweden by Anton Osika and Fabian Hedin out of the open-source GPT Engineer project, attempted launches under the GPT Engineer App name in spring and summer 2024 before relaunching as Lovable in December 2024. The relaunched product reached $17 million in annual recurring revenue (ARR) in roughly 90 days. StackBlitz launched Bolt.new on October 4, 2024, combining Anthropic Claude 3.5 Sonnet with WebContainers to generate and run full applications inside the browser. Replit launched Replit Agent on September 11, 2024, with a similar idea executed inside its own cloud IDE.
In February 2025, the AI researcher Andrej Karpathy coined the term "vibe coding" to describe a workflow in which a developer prompts an AI assistant such as Cursor Composer with Claude Sonnet, accepts most of the output without reading it carefully, and iterates by talking back to the model. The phrase quickly attached itself to consumer-facing tools like Lovable and Bolt.new and to the broader habit of building entire web apps by prompt. Collins Dictionary named "vibe coding" its word of the year for 2025. By mid-2025, Webflow had shipped a full AI Site Builder that could produce multi-page sites with shared design systems, and Vercel had released the AI SDK 5 and reframed v0 as a general-purpose agent that could deploy apps and websites end to end.
This category covers tools whose primary input is a natural language prompt and whose primary output is either deployable web code or a working hosted site. They differ from traditional website builders in that they generate the underlying code from scratch rather than configuring a structured template.
| Tool | Vendor | Launch | Notes |
|---|---|---|---|
| v0 | Vercel | Oct 11, 2023 (beta); chat UI 2024; v0.app Aug 2025 | Generates React and Tailwind components; v0.app generalizes to full apps and websites |
| Bolt.new | StackBlitz | Oct 4, 2024 | Browser-based full-stack builder using Claude 3.5 Sonnet and WebContainers; reached $40M ARR by March 2025 |
| Lovable | Lovable AB (Sweden) | Dec 2024 (relaunch from GPT Engineer App) | Founded by Anton Osika and Fabian Hedin; reached $206M ARR in 11 months; $200M Series A from Accel at a $1.8B valuation in February 2025 |
| Replit Agent | Replit | Sep 11, 2024 | Agentic app builder inside the Replit cloud IDE; Agent v2 added live preview |
| Create.xyz | Create | 2023 | Generates landing pages and small apps from a prompt; targets non-developers |
| Magic Patterns | Magic Patterns (Y Combinator W23) | 2023 | AI prototyping aimed at product teams; outputs editable designs and code |
| Subframe | Subframe | 2024 | Visual editor that emits production React plus Tailwind code; integrates with existing design systems |
| Wireframer | Framer | 2024 | Generates responsive multi-page wireframes inside Framer from a prompt |
What distinguishes these products is the loop. Most of them keep the conversation, the code, and a live preview on the same screen, so a user can describe a change in English, see the rendered result, and accept or reject it. Bolt.new and Replit Agent additionally run the generated code inside an isolated runtime so that errors and console output feed back into the next prompt automatically.
Lovable's growth curve in 2025 is among the most extreme in software history. The company reported reaching roughly $17 million ARR within 90 days of its December 2024 relaunch and $206 million ARR within 11 months. Bolt.new went from zero to $20 million ARR in roughly two months and reached $40 million ARR by March 2025. Both products' core differentiator over generic chatbot wrappers is the integration of a runtime: Bolt.new uses StackBlitz WebContainers, and Lovable runs generated code on its own hosting layer with Supabase for data.
This group includes platforms whose business model predates the generative AI era. They retain template-driven editors and have layered AI on top, usually for onboarding (turning a business description into a starter site), content (writing copy, generating images), and assistance (chat helpers for editors).
| Platform | AI feature | Launch |
|---|---|---|
| Wix | ADI question-driven builder | 2016 |
| Wix | AI Site Generator (conversational) | July 17, 2023 |
| Wix | Wix Studio AI tools for designers | 2024 |
| Squarespace | Blueprint guided design system with AI text and images | 2023 |
| Squarespace | Design Intelligence suite | September 2024 |
| Webflow | Webflow AI (AI Assistant, AI in pages) | 2024 |
| Webflow | AI Site Builder (multi-page generation) | February 2025 |
| Framer | AI features in editor, Workshop AI components | 2023 |
| Framer | Wireframer AI multi-page generator | 2024 |
| Hostinger | AI Website Builder (Zyro-derived) | 2023 |
| Hostinger | Kodee AI assistant for hosting and site building | May 2023; LLM relaunch September 2023 |
| Durable | 30-second AI site generator with CRM and invoicing | 2022 |
| GoDaddy | Airo AI for domains, sites, logos, social posts | February 2024 |
| GoDaddy | Airo for WordPress | 2025 |
| 10Web | AI Website Builder on WordPress | 2022 |
| Hocoos | AI builder generating three template variants | 2022 |
The Wix AI Site Generator, announced in July 2023, asks for a business description, then chains together calls to a language model for copy, an image model for hero art and gallery imagery, and a layout engine that picks structured sections. The output is a fully editable Wix site rather than free-form HTML, which means it inherits Wix's existing hosting, SEO, e-commerce, and bookings stack. Wix Studio, aimed at agencies and freelance designers, layered separate AI tools for code, copy, and motion in 2024.
Squarespace Blueprint launched in 2023 as a guided design system. The full AI-driven build experience, branded as Squarespace Design Intelligence, expanded substantially in September 2024 to include AI content generation, automatic image selection, and design recommendation. Blueprint is free to use but requires a paid plan to publish.
Webflow took a longer path. It first shipped narrower AI features inside its existing visual editor in 2024 (page-level AI assistance, content generation, an AI Assistant for editing layouts). In February 2025, Webflow announced its AI Site Builder, which can generate up to five pages from a prompt, complete with shared styles, animations, and a CMS skeleton. Because Webflow output is already production HTML and CSS, the AI Site Builder is closer in spirit to the developer-facing tools than to Wix or Squarespace.
Framer combines two AI surfaces. Wireframer generates a responsive site from a prompt and is aimed at the early sketching phase. Workshop generates individual interactive components (cookie banners, tabs, visual effects) inside an existing Framer project. Framer also offers an AI plugin framework so users can chain calls to OpenAI, Anthropic, or Google Gemini for tasks like alt text or translation.
Hostinger ships an AI Website Builder that asks for a brand name, a description, and a type, and then produces a draft site. Kodee, Hostinger's AI assistant, launched in May 2023 on a hybrid stack of Rasa plus GPT-3.5, then was rebuilt on a pure LLM backend in September 2023; by 2026 Hostinger reports Kodee handles 83% of support interactions and can take over 350 admin actions including migrations and backups.
Durable markets itself as a 30-second AI website builder. Users answer three short questions about their business, and Durable produces a complete site with copy, imagery, and a contact form. Durable bundles a CRM, invoicing, blog builder, and marketing AI in the same dashboard. The company reports more than three million businesses have signed up since its 2022 launch.
GoDaddy launched Airo in February 2024 as a suite that starts at the domain step. A user types a description of their business, Airo proposes available domains, and after purchase produces a draft website, logo, email setup, product descriptions, social media posts, and an email calendar. GoDaddy reported in late 2024 that small businesses using Airo generated 28% more sales than a control group during an August-September 2023 test. Airo for WordPress launched in 2025 with similar capability targeted at the WordPress ecosystem.
10Web's AI Website Builder is unusual in this group because it generates standard WordPress sites rather than a proprietary format, which means the output can be self-hosted and edited by any WordPress developer. Hocoos focuses on speed: it asks eight questions and produces three layout options.
Content management systems sit between the visual builders and developer frameworks. They store structured content (entries, fields, relations) and expose it via APIs.
WordPress remains the most widely used CMS by share of websites. AI in the WordPress ecosystem is fragmented across plugins. Jetpack AI Assistant, launched by Automattic on June 6, 2023, runs inside the block editor and handles writing, tone shifting, translation, summarization, and image generation. The first version used GPT-3.5 with the OpenAI API. Third-party WordPress AI plugins include Bertha AI, Rank Math's AI features for SEO, and the Yoast AI features for title and meta description generation.
Among headless CMS vendors, Contentful, Storyblok, and Sanity all shipped AI features during 2024 and 2025. Contentful introduced an AI Content Generator app in its marketplace and a framework called AI Actions for translation, localization, SEO field extraction, and reusable content automations. Storyblok added AI-generated alt text inside its image editor and AI-assisted translations, and in January 2025 introduced Storyblok Labs as a sandbox for new AI features. Sanity ships Content Agent for conversational bulk edits and audits, Agent API for schema-aware generation and translation, and an MCP server that lets external AI agents read and modify Sanity content through the Model Context Protocol.
Ghost, Drupal, and other CMS platforms have followed similar paths, usually adding writing assistance, alt-text generation, and translation as first-party features. The recurring pattern is that AI sits between the editor and the structured content store, suggesting copy and metadata that humans approve.
A separate slice of the market is libraries that help developers add AI to their own web applications. The leading example is the Vercel AI SDK.
The Vercel AI SDK was introduced on June 15, 2023 as an open-source TypeScript library for building chat and streaming AI interfaces with React or Svelte. It abstracts streaming, tool calling, structured outputs, and provider switching across OpenAI, Anthropic, Google, and others. Version 4 shipped in November 2024 with PDF support, computer use integration, and a new xAI Grok provider. Version 5 followed on July 31, 2025 with type-safe chat hooks, agentic loop control, a redesigned specification, tool enhancements, and speech generation. Version 6 was released later, adding agents, tool execution approval, DevTools, a full MCP implementation, reranking, and image editing.
Vercel maintains the Next.js AI Chatbot template, an open-source starter that pairs the AI SDK with Next.js 15, React 19, and Auth.js for authentication. The template uses the Vercel AI Gateway, which lets developers swap among OpenAI, Mistral, Moonshot, DeepSeek, and xAI providers without changing application code. The latest revisions ship with a side-by-side layout that keeps chat and generated output visible at once.
Other frameworks in the JavaScript ecosystem include LangChain.js and LlamaIndex.ts, which port the popular Python orchestration libraries to TypeScript. Both run inside Node.js, Bun, Deno, and edge runtimes, and both integrate cleanly with Next.js, SvelteKit, and Remix (now React Router v7). The Vercel AI SDK is the more popular choice for chat interfaces because of its tight integration with React Server Components and streaming. LangChain.js and LlamaIndex.ts are preferred for retrieval-augmented generation, agent orchestration, and complex multi-step pipelines.
At the framework level, Next.js, SvelteKit, Nuxt, and Remix all ship sample apps demonstrating AI integration. Next.js 15 added improved support for Partial Prerendering and React Server Components, both of which suit streaming AI output. Next.js 16 continued the trend with caching and turbopack improvements. SvelteKit's AI sample applications demonstrate the AI SDK's Svelte adapter. The choice of framework rarely matters for the AI layer itself; it matters more for hosting and edge runtime support.
Automated accessibility testing has been part of the web stack since the 2000s. The de facto open-source rules engine is axe-core from Deque Systems, which checks pages against WCAG and other accessibility standards. Deque markets a commercial product family called axe DevTools that wraps axe-core with browser extensions, CI integrations, and reporting.
Deque added AI features to axe DevTools in stages during 2024 and 2025. Automated Intelligent Guided Tests use AI to walk through traditionally manual checks (such as evaluating whether an image's alt text accurately describes the image), surfacing candidate issues for human approval. A separate AI ruleset uses machine vision and screenshots to catch issues that pure DOM analysis misses, such as visual contrast in rendered content. Deque also ships axe Assistant, a chatbot trained on accessibility documentation, and supports the Model Context Protocol so coding agents like Cursor and Claude Code can call axe Assistant during development.
The WebAIM Million report, published yearly, is the standard reference for the state of web accessibility. In its 2024 edition, WebAIM found that 95.9% of the top one million home pages had detected WCAG 2 failures, down marginally from 96.3% in 2023, with low-contrast text the most common failure (81% of pages). Multiple 2024 and 2025 studies have shown that AI-generated UIs amplify these failure rates. A 2024 ACM study examined six websites built entirely by AI models and found 308 distinct accessibility errors, more than half of them cognitive. AudioEye's 2024 analysis reported that 73% of AI-generated alt text was wrong or meaningless. The most consistent failure modes are missing visible focus styles, color contrast below the WCAG AA threshold, missing or inaccurate alt text, and broken keyboard focus traps during dynamic updates.
Generative UI tools that target developers, rather than end users, focus on producing reusable components. v0 by Vercel is the highest-profile example. It is trained on best practices for React, Tailwind CSS, and shadcn/ui, which means generated output drops cleanly into a standard Next.js project. Every component on the shadcn/ui documentation site has an "Open in v0" button that loads the component into v0 for customization.
shadcn/ui itself is not an AI tool; it is a set of unstyled React components built on Radix UI primitives and Tailwind, distributed via a CLI that copies source into your repository. Its popularity in the AI generation space comes from two properties. The components are local source files in the consumer's repo, so an AI assistant can read and modify them. The styling layer is Tailwind, which AI models handle well because Tailwind's utility classes appear verbatim in training data.
Builder.io's Visual Copilot focuses on a different angle: converting Figma designs into framework code. Visual Copilot uses a custom AI model trained on more than two million data points to flatten Figma layouts into a code hierarchy, then a fine-tuned LLM emits React, Vue, Svelte, Angular, Qwik, Solid, or HTML targeting the user's preferred styling library (plain CSS, Tailwind, Emotion, or styled-components). Visual Copilot 2.0 adds natural language interactivity, so designers can describe behaviors and have the AI translate them into working code.
GitHub Copilot, Cursor, Windsurf, Cline, and Claude Code all serve as general-purpose AI coding assistants that work inside web framework codebases. Their effectiveness on web work depends heavily on the framework. Next.js, SvelteKit, and Remix are well represented in training data, and AI assistants produce reasonable scaffolding for routes, server actions, and data fetching. Less common frameworks (Astro, Qwik, SolidJS) receive thinner support, though all have improved through 2024 and 2025 as the model providers expanded coverage.
Directly inside CSS, Tailwind CSS dominates the AI-generated category. Tailwind's class-based syntax means model output for styling appears as literal text rather than as a separate CSS file, which suits the way LLMs work. UI libraries that complement Tailwind, including shadcn/ui, Radix UI, Headless UI, and Origin UI, have all become standard targets for AI generation tools.
Search engine optimization has its own ecosystem of AI tools, treated in detail at SEO. The relevant items for web development are the in-editor AI features that propose meta descriptions, page titles, alt text, and internal linking. Yoast SEO, Rank Math, and All in One SEO (AIOSEO) all added AI-generated meta description and title tools during 2023 and 2024. Webflow's AI Assistant produces page metadata as part of site generation. Wix and Squarespace generate metadata automatically as part of the site build flow.
Separately, the rise of AI search engines such as Perplexity, ChatGPT Search, Google AI Overviews, and Claude web search has introduced a related practice called Answer Engine Optimization (AEO) or Generative Engine Optimization (GEO). It focuses on making content quotable by AI systems rather than (or in addition to) ranking in classic search. Webflow's 2025 AI Site Builder advertises native AEO features. The full landscape of AEO and GEO tools is covered at SEO.
At the user-facing layer, AI now lives in widgets embedded on otherwise traditional pages. The most common categories are chatbots, forms, and search.
Intercom Fin, launched in 2023, is a customer support chatbot that lives in a website's chat widget. It is positioned as a tier-one support agent that handles a portion of incoming questions automatically using a customer's own help content. Tally and Typeform both added AI features that generate form structures from a prompt and rewrite questions for clarity. Site search products including Algolia AI Search, Mintlify AI search, and Kapa.ai integrate retrieval-augmented chat for documentation and product sites.
The rise of AI in web development comes with concrete failure modes that have been documented in peer-reviewed studies, security reports, and industry surveys.
Maintainability of generated code. AI-generated codebases tend to accrete technical debt quickly. Output is consistent in style within a single session but inconsistent across sessions because each prompt produces a fresh interpretation of the project. This makes refactoring harder and code review slower. The problem is most visible in the vibe-coding workflows popularized by Bolt.new and Lovable, where users may not read the generated code at all before deploying it. Karpathy himself has noted that vibe-coded output can still be "awkward" and "gross".
Slopsquatting. When AI assistants suggest code that imports a package, they sometimes invent package names that do not exist. Attackers register these hallucinated names on npm and PyPI with malicious payloads, betting that the next developer to receive the same hallucination will run npm install or pip install and execute the malicious code. The term "slopsquatting" was coined by Python Software Foundation developer-in-residence Seth Larson. A 2024 study found open-source code generation models invent package names roughly 21.7% of the time and commercial models around 5.2% of the time. One real-world example, the Python package huggingface-cli, was registered by a security researcher after AI tools repeatedly recommended it and was downloaded over 30,000 times in three months.
Accessibility deficits. Multiple studies during 2024 and 2025 found that AI-generated UIs perform worse on accessibility than human-authored UIs. Common failures include missing visible focus indicators, low-contrast color choices, missing or inaccurate alt text, and inaccessible custom controls. WebAIM's 2024 Million report found WCAG 2 failures on 95.9% of the top one million home pages, with low-contrast text on 81% of pages. AudioEye's 2024 analysis found that 73% of AI-generated alt text was wrong or meaningless. Researchers attribute the problem in part to the training corpus: because most existing web content is inaccessible, models trained on that content reproduce its mistakes.
Hidden costs and lock-in. AI website builders that host the resulting site (Wix, Squarespace, GoDaddy, Durable, Lovable) typically lock the output to their platform. Migration off these systems can require rebuilding the site by hand. Tools that output standard code (Bolt.new, v0, Webflow, 10Web on WordPress) avoid this problem but shift the cost onto running and securing the resulting code.
Hallucinated APIs and frameworks. Beyond package names, AI assistants regularly invent functions, hooks, and CSS classes that do not exist in the libraries they claim to use. The Tailwind classes text-xs, flex-center, and gap-2 look indistinguishable to an inattentive reader, but only two of those are real Tailwind classes. The same pattern affects Next.js APIs, React hooks, and the AI SDK itself, especially across version boundaries (AI SDK 4 to AI SDK 5 was a substantial breaking change).
Quality drift in long sessions. Tools that keep an entire session in context (Bolt.new, Lovable, Replit Agent) degrade in quality as the session grows. Once a project exceeds the model's effective context window, the assistant begins to forget earlier decisions and reintroduces bugs. Users learn to start fresh sessions periodically, but this fights against the value proposition of an AI that holds the whole project in its head.