AI Website Builders Compared — Lovable, Bolt.new, v0 (Stack, Prompting, Marketplaces)

Source: ai-research/lovable-prompting-docs-2026-07-02.md, ai-research/lovable-monetization-guide-2026-07-02.md, ai-research/bolt-prompting-best-practices-2026-07-02.md, ai-research/bolt-new-github-readme-2026-07-02.md, ai-research/v0-templates-docs-2026-07-02.md, ai-research/vercel-how-to-prompt-v0-2026-07-02.md, ai-research/framer-creator-program-2026-07-02.md, ai-research/webflow-template-creator-commission-2026-07-02.md; business-metrics update from raw/Former_Intel_CEO_on_What_Went_Wrong_What_s_Next_+_Lovable_CEO_on_the_Real_Promise_of_Vibe_Coding.md (All-In Podcast, 2026-07-16 — Lovable CEO Anton Osika interview, recorded in Paris)

A comparative reference for the three AI website/app builders that prompt libraries like MotionSites target: Lovable, Bolt.new (StackBlitz), and v0 (Vercel). Covers default stack, prompting conventions, official template marketplaces (and how their economics compare to dedicated template marketplaces Framer and Webflow), pricing, and documented rendering quirks. Written to close a standing gap flagged in the topic’s research agenda — the wiki previously had zero standalone coverage of any of the three.

Key Takeaways

• Stack flexibility is the clearest differentiator. Lovable and v0 are both fixed-stack (React/Vite/Tailwind/shadcn + Supabase for Lovable; React/Next.js App Router/Tailwind/shadcn for v0). Bolt.new is the only one of the three with no fixed default — it runs on StackBlitz’s in-browser WebContainer and supports React, Vue, Svelte, Angular, Next.js, Astro, Remix, SvelteKit, and Expo/React Native.
• All three converge on the same prompting advice: build incrementally, not in one shot. Lovable: “A full-page prompt gets you noise. A section-based prompt gets you signal.” Bolt: add components “incrementally rather than all at once.” v0: plan → build core → add features one request at a time. This is directionally consistent with (but doesn’t validate the magnitude of) the spec-prompt framing in Claude Design Prompt Examples.
• None of the three has a monetized first-party template marketplace — a real gap relative to dedicated template marketplaces. Lovable’s own official guidance tells creators to sell elsewhere (Gumroad, 10%+$0.50/transaction). Bolt has no official marketplace at all. v0’s gallery is publish-for-discovery only — no pricing, purchase flow, or revenue-share exists in the docs.
• Framer and Webflow, by contrast, pay creators 95-100% because template sales are their actual business model. Framer: 100% revenue retention on paid templates + an undisclosed-% referral commission on subscription conversions. Webflow: 95% creator / 5% platform (effective 2025-10-21), plus “fulfillment links” for 0%-cut off-Marketplace sales.
• This answers the standing “does a hero-prompt library compete with a template marketplace?” question: they’re not real substitutes. A prompt (MotionSites) requires an AI builder plus a generation step plus iteration risk; a finished template (Framer/Webflow) is install-ready with no generation uncertainty, but locked to one builder’s ecosystem. The economics gap tracks the product gap — AI-builder galleries aren’t monetized yet because the “product” is still a starting point, not a finished deliverable.
• Bolt.new’s defining constraint is WebContainer, not prompting style. It can’t execute native binaries or compiled C/C++; packages needing native bindings (sharp, bcrypt, most native DB drivers) fail unless WASM-compiled. Its own README frames this as a tradeoff for the opposite gain — Bolt is the only one of the three that gives the AI “complete control over the entire environment including the filesystem, node server, package manager, terminal, and browser console,” unlike “Claude, v0, and similar tools that can’t install packages, run backends or edit code.”
• No comparative benchmark of animation/video-background rendering reliability exists across these builders. See Open Questions — this is a genuine, confirmed gap, not an oversight.

Builder Profiles

Lovable (lovable.dev)

Default stack	React + TypeScript + Vite, Tailwind CSS, shadcn/ui, React Router (frontend) — Supabase (Postgres, Auth, Storage, Deno Edge Functions) as the entire backend. Fixed, not user-selectable.
Prompting convention	Section-based over full-page. A 4-question planning framework (what is it / who’s it for / why will they use it / what’s the one key action) precedes prompting. “Atomic language” for UI elements (name buttons, cards, modals, badges specifically). Aesthetic buzzwords (“minimal,” “cinematic,” “premium”) steer style. Structured prompts of 400-1,200 words work for a v1 build; unstructured prompts can drift after ~200 words.
Template marketplace	lovable.dev/templates is an official gallery, not a paid creator marketplace. Lovable’s own monetization guide directs creators to sell on Gumroad (10%+$0.50/transaction)ortheirownsite,suggesting$99-299 price points. Lovable takes no cut because it isn’t the seller of record. Independent third-party sites (loadable.dev, zeroqode.com, tem.place) exist outside Lovable’s control.
Pricing (2026)	Free: 5 credits/day (~30/mo). Pro: $25/mo,100credits/mo.Business:$50/mo (adds SSO + training opt-out). Enterprise: custom. ~1 credit ≈ 1 message.
Rendering quirks	Not documented specifically for video/animation. Public troubleshooting docs cover generic blank-preview causes (bad vite.config, missing env vars, JS console errors).

Bolt.new (StackBlitz)

Default stack	None fixed — runs on StackBlitz’s in-browser WebContainer. Supports React, Vue, Svelte, Angular, Next.js, Astro, Remix, SvelteKit, plus Expo/React Native. Broadest framework range of the three. Hard constraint: WebContainer can’t execute native binaries or compiled C/C++; packages needing native bindings fail unless WASM-compiled — docs steer toward pure-JS databases (libsql/sqlite) over native drivers.
Prompting convention	”Plan before you build” — treat the prompt as a blueprint. Add components incrementally, one at a time. Be explicit about what should/shouldn’t change (targeted file selection, file locking, code-section highlighting). “Don’t expect the LLM to have common sense” — spell out specifications. Context window is finite; request a summary or duplicate the project to reset history.
Template marketplace	No official first-party marketplace from StackBlitz/Bolt. Third-party sites (BoltStarter, boltmarketplace.com) sell or give away Bolt-compatible starters, unaffiliated with Bolt. The Teams plan supports private internal team templates only — not a public sales channel.
Pricing (2026)	Free: 300K tokens/day, 1M/mo. Pro: $25/mo,10Mtokens/mo(rollsoverupto2months).Teams:$30/member/mo, same 10M allotment per member (not pooled). Enterprise: custom.
Rendering quirks	The well-documented failure mode is native-binary/npm incompatibility inside WebContainer — not CSS or animation specifically.

v0 (Vercel)

Default stack	Fixed — React, Next.js (App Router), Tailwind, shadcn/ui. Per a third-party technical comparison, only React components actually execute live in v0’s preview pane when fetching live data; other-framework output lacks that live-preview capability.
Prompting convention	Vercel’s official “How to prompt v0” guide names three required inputs: product surface (concrete components/data/actions, not “a dashboard”), context of use, and constraints (“tell v0 what not to invent”). Explicitly iterative — plan, build the core, add features one request at a time; up to 10 prompts can be queued while one generates. Vercel’s own A/B tests (three paired comparisons) measured detailed prompts running 19-40% faster, 15-25% fewer lines of code, and slightly lower credit cost than vague prompts — the headline claim is “30-40% faster… fewer credits,” not a large multiple.
Template marketplace	v0.app/templates is a genuine two-sided public gallery — anyone can “Publish as Template” from a chat (needs a 1920×1080 preview image, name, description, category), appearing under Community/Your Templates and the creator’s public Showcase profile. No purchase flow, pricing, or creator revenue-share is documented anywhere — it’s a free discovery/portfolio gallery, not a monetized marketplace.
Pricing (2026)	Free: $5/mocredits.Premium:$20/mo (legacy, sunsetting for new signups). Team: $30/user/mo($30 credits + $2dailylogincreditperuser).Business:$100/user/mo (same credit base plus admin/security). Enterprise: custom.
Rendering quirks	No official documentation of breakage specific to video or complex CSS. A separate Vercel blog post (“How we built the v0 iOS app”) documents general animation-fragility lessons (fade-ins re-triggering on remount) on a different Vercel product — informative about the team’s general practices, not a claim about v0.app’s own web-generation output.

Lovable at Scale — Revenue, Usage, and Model Strategy (2026-07-16 update)

Business-metrics and strategy detail from a 2026-07-16 CEO interview (Anton Osika, All-In Podcast), absent from the stack/prompting sources above:

• Revenue and usage, as stated by Osika: reached **$500M\ast \ast inannualrevenueinMay2026,sincegrownfurther(heconfirms"another100million"beyondthat,consistentwiththehost{s}^{\prime }runningjokethatLovableisdeclared"dead"roughlyevery6monthsandthenaddsanother$100M). 50M+ apps built to date, ~1M new projects built every week, 700M+ monthly visits to applications built on the platform. Lovable has been in market 20 months as of this interview.
• Customer mix: ~20% technical/engineering background, ~80% non-technical. Osika frames Lovable as “a great bridge” for the non-technical majority to figure out what to build, while the technical minority appreciates the opinionated architecture (one-prompt payment setup, automatic background security scanning after every change).
• Usage economics: roughly 60% of customers hit their subscription’s usage cap and pay overages, according to Osika — framed as evidence customers see enough value (“this is a $500,000 piece of software” for 0.1-1% of the equivalent cost) to keep paying past the included quota rather than churning.
• Model strategy: multi-model routing, not single-vendor lock-in. Every prompt is routed to whichever model (commercial frontier or increasingly Lovable’s own open-weight fine-tunes) is best-suited to the task. A Stockholm-based research team runs post-training / reinforcement learning specifically targeted at the mistakes frontier models make in Lovable’s own agent harness (wrong tool call, wrong integration choice) — trained on Lovable’s own enormous token-distribution signal from a million weekly builds. Osika frames this explicitly as European-ecosystem strategy, not just a cost play.
• “Co-opetition” as an internal workflow pattern. Osika (ex-CERN, where isolated teams built the same particle-accelerator component independently and compared results only at publication) actively encourages parallel independent builds of the same feature by different people inside a client organization, then manually cherry-picking the best elements from each into one project rather than picking a single team’s version upfront — treating this as a rapid-experimentation advantage unique to how cheap iteration has become, not a coordination failure to eliminate.
• On Fable 5 specifically: “It can in the first attempt create very sophisticated things that look really good” — but Osika is explicit that raw model capability is not the current bottleneck for Lovable’s customers. The bottleneck remains human judgment on what to build and what strategic direction to take — planning with the agent, not model intelligence, is “not changing as fast” as the models themselves.
• AI-as-co-founder positioning (in pre-release testing). Beyond generation, Lovable is testing giving customers an AI “co-founder” with access to all of a project’s operational data (not just the code) that proactively surfaces strategic suggestions — an evolution from build-the-software toward operate-and-grow-the-business.

Template Marketplace Economics — AI Builders vs. Dedicated Marketplaces

The research agenda asked: “at what point does a hero-prompt library compete with or complement a finished-template marketplace?” Comparing the actual creator economics answers it:

Marketplace	Creator take	Notes
Framer Templates	100% of direct sale revenue	Plus an undisclosed-% referral commission (12mo, subscription conversions via remix links, templates-only, excludes enterprise). No listing fee documented.
Webflow Templates	95% (Webflow keeps 5%)	Effective 2025-10-21; the 5% is framed as covering payment processing. “Fulfillment links” let creators sell entirely off-Marketplace at their own price with 0% Webflow cut.
Lovable Templates (gallery)	N/A — not a paid marketplace	Lovable’s own guide routes paid sales to Gumroad (10%+$0.50/transaction) or a creator’s own site.
Bolt.new	N/A — no official marketplace	Only third-party, unaffiliated sites exist.
v0 Templates (gallery)	N/A — not a paid marketplace	Publish-for-discovery only; no pricing or revenue-share mechanism in the docs.

Why the gap exists: Framer and Webflow sell finished, install-ready files — the transaction is the entire business, so paying creators 95-100% is sustainable. The AI builders’ galleries currently distribute a starting point that still requires a generation step and iteration — a fundamentally less finished product, which is likely why none of them has built out paid-marketplace infrastructure yet. A hero-prompt library like MotionSites sits in between: it sells the instructions to generate a page (builder-agnostic, portable across Lovable/Bolt/v0/Claude Code/Claude Design) rather than either a finished file or a free discovery template — complementary to, not competing with, either category. See MotionSites for the fuller treatment of that middle layer, including its competitive landscape and reseller-rights tier.

Try It

1. Match the builder to the framework constraint, not just taste. If a project needs Vue, Svelte, or a non-React stack, only Bolt.new supports it natively among these three — Lovable and v0 will push everything into React regardless of prompt wording.
2. Apply the shared incremental-prompting lesson regardless of builder. All three vendors independently converged on “build one section/component at a time, don’t ask for the whole page at once” — treat this as the one cross-builder rule worth following even before reading builder-specific docs.
3. Don’t expect to monetize a template built in Lovable, Bolt, or v0 through the builder itself. If the goal is selling finished templates, route through Framer or Webflow’s Marketplace (95-100% creator take) or an external storefront (Gumroad) — not the AI builder’s own gallery.
4. If evaluating MotionSites or a similar prompt library, test the same prompt across at least two of these three builders before buying a paid tier — no independent benchmark exists yet for how consistently a given spec-prompt renders across Lovable/Bolt/v0 (see Open Questions), so this remains something you have to verify yourself.

Open Questions

• No comparative animation/video-background rendering benchmark exists. A dedicated 2026-07-02 research pass (targeted searches plus direct fetch of two informal single-prompt Claude/v0/Lovable comparison blogs — andrew.ooo and p0stman.com) found no systematic test of “cinematic” hero-section reliability across these builders. Both informal comparisons found are qualitative single-prompt reviews with zero claims about video or animation rendering specifically. One search-engine-synthesized claim that “Claude Design handles video backgrounds better than v0” could not be traced to an actual source page on inspection and is not reported as fact here — flagged as a likely synthesis artifact. Closing this gap requires original testing (the same video-background prompt run through each builder, screenshotted), not further literature search. (researched 2026-07-02: still open — no benchmark exists; would require original hands-on testing to close.)
• Framer’s referral commission percentage is undisclosed in the public Creator Program docs — only the 100% direct-sale figure is stated.
• Rendering-quirk documentation is thin for all three builders on anything beyond WebContainer’s native-binary constraint (Bolt) and generic blank-preview causes (Lovable). Whether video backgrounds, complex CSS, or WebGL degrade differently per builder is not documented by any vendor and would need direct testing.
• v0’s live-preview-only-for-React claim comes from a single third-party comparison (betterstack.com), not confirmed in Vercel’s own docs — worth re-verifying if this becomes decision-relevant.
• **The 2026-07-16 revenue/usage figures (50M+ apps, 700M+ monthly visits, $500M+revenue,60$100M every 6 months” framing) but not verified.

Related

• MotionSites — Premium Hero Prompts — the builder-agnostic prompt library this comparison exists to contextualize
• Claude Design Prompt Examples — Spec-Style Prompting — the spec-prompt pattern these builders’ own docs independently converge toward
• Refero Styles — brand-DNA DESIGN.md prompts, complementary to page-level prompting on any of these builders
• Neuform — 400+ free DESIGN.md design systems, same spec-prompt-catalog cluster
• AI Web Design — topic index
• Claude Design — Anthropic’s own hosted design surface, the fourth (non-open-web) target these prompt libraries write for
• awesome-claude-design — aesthetic-family prompt system usable alongside any of these three builders
• AI Website Builders vs. the Claude Code Design Stack — A Decision Framework