You're Paying Lovable to Use Claude: The AI Website Builder Idiot Tax

Key Findings: The Model, the Price and the Sites

In short: Lovable runs on Claude by its own account, and 73.5% of sampled Lovable sites sit on an older stack that ChatGPT could quote 7.7% of the time when asked to read them. Each finding below carries its own scope, so you can quote it with its source.

The Viral Post Behind This Research

On October 14 I posted this on X. It unexpectedly went viral, with 577,811 views and 683 replies by the next day.

My point was simple. These companies raise VC money to sell a "magic solution," when the AI inside is Claude or GPT, the same models you can use directly.

I call it the idiot tax. On Lovable, you pay about what Claude costs, or more, and give up the choice of model, code you can easily move and a clean way out.

A post is not proof, so I had Claude agents check the builders' own words and prices, then test 749 live sites they host. Fair warning on bias: I sell Claude Code products, so I gain if you take my side, and the method covers the controls.

Aaron Makelky and I broke the post down on AI Rabbit Holes: what a Lovable credit is worth, why Aaron quit Lovable and the stack we both use instead.

Lovable Runs on Claude. It Says So Itself.

Lovable's homepage and pricing page, as saved on October 14 and 15, don't mention Claude. Its own "Claude vs Lovable" guide does, listing Lovable's underlying model as "Claude (Anthropic)."

The guide says Lovable was built "using Claude as the underlying language model," with its edge in the platform layer "around the same Claude intelligence that powers claude.ai." Anthropic's case studies tell the same story for most of the rivals named here.

Replit's president put it plainly in Anthropic's case study: "Models are like the car engine, and we're building the entire car around it ourselves." A car is worth paying for, as long as you know whose engine is in it and what the engine costs on its own.

Which Model Are You Paying For? Lovable and Bolt Won't Let You Choose

Lovable's FAQ answers which model Lovable uses without naming one: "Lovable uses current leading AI models and updates them for everyone automatically." It adds that Lovable chooses, "so you cannot pick a specific model for building."

Its changelog does announce the models it adds to its build agent, most recently "Lovable now incorporates Claude Opus 5." Anthropic's case study says the agent uses subagents "with the right models at each step," and Lovable's docs don't show which model handled your build.

Bolt's docs say "As new models become available, Bolt can update how agents perform without changing your workflow." Only Anthropic's case study names Claude.

That matters because Anthropic's API price list (October 15, 2026) charges developers 40 times more per output token for Claude Opus 5.5 than for Claude Haiku 5.5, on prompts up to 100,000 tokens. With a hidden model, you can't tell which one your credits buy.

Lovable Charges More Than the Claude It Runs On

Lovable's guide puts the prices side by side: Claude's paid plans "from $20/month" and Lovable "Pro from $25/month." Bolt Pro is also $25 a month, and Anthropic describes the Agent SDK Bolt moved to as tools to "build powerful agents on top of Claude Code," which comes with that $20 Claude plan.

To be fair, Lovable's $25 also buys hosting, a database, logins and one-click deploys. Base44 and Replit start at about the Claude price, and no builder publishes its margins, so the markup on each prompt can't be measured.

The pitch: a magic solution next to a chat window

Lovable's guide compares itself to the claude.ai chat window, which it says leaves you with "you have a conversation transcript containing code snippets." Claude Code gets a short aside: it "requires comfort with the command line."

That skips the Claude tool that actually builds and deploys a site, which is the one I use.

What the Builders Promise vs What ChatGPT Got

Here is the marketing, word for word, next to what the test and the vendors' docs show. Internet Archive captures show each claim was live before the October 14 test.

The fine print

Lovable's docs narrow the snapshot promise to a list: "Pre-rendering is only served to verified crawlers." Other bots get the bare app: "Unverified bots and link checkers see the regular single-page app, unless they request the page as Markdown."

Those docs do put "AI search crawlers such as those of ChatGPT, Perplexity, Claude, and Gemini" on the list, and this test can't see what OpenAI's separate search crawler receives. But on 148 older Lovable sites, ChatGPT listed the site among its own sources when a person asked it to read the page, and on 133 of those it still said it could not.

Base44's docs admit the core problem in one line: "Base44 apps are built to run in the browser, which normally means a crawler sees an empty page." Its homepage just says "Get found where people are actually looking – on Google, ChatGPT, Gemini and more."

Why I call this deceptive

Lovable's landing page promises readable snapshots for "all" older apps, with "No opt-in," and keeps the verified-crawler limit in its docs. ChatGPT then said it couldn't read 9 in 10 of those sites when asked, and I call that deceptive marketing.

Base44 pitches free publishing and "Built-in SEO" in the same breath, while its docs say free addresses aren't meant for "long-term, production SEO." I'd put that in the same bucket.

The AI answers repeat the pitch. Asked on October 14 whether Lovable is good for SEO, ChatGPT and Gemini opened with a yes, and ChatGPT's answer never mentioned older projects.

Framer and new Lovable sites held up in this test. Replit's gap is fine print, and Bolt had too few published sites in the crawl to test.

What the Wrapper Ships: Sites ChatGPT Mostly Can't Read

If you pay extra for the wrapper, you'd hope AI tools can read what it publishes. ChatGPT quoted under a third of 749 live builder sites, and new Lovable sites beat older ones by a mile: same company, different stack underneath.

Most older Lovable, Base44 and Replit sites are empty shells

An empty shell is a page that arrives with almost no words in it, so the browser has to run JavaScript to build what you see. Visitors rarely notice, but any tool that reads the HTML without running that code gets a mostly blank page.

When ChatGPT can't read a page, it sometimes quotes the metadata

On an empty shell, the only words in the HTML are usually the page title and meta description. ChatGPT mostly said it could not read the page, and sometimes it handed back the meta description as if it were the first sentence.

Base44 stood out: ChatGPT presented title or description text as page text on 39.6% of its sites.

One example from the sample: a steakhouse's Base44 site opens with "Fresh from the fire." ChatGPT gave its first sentence as "A high-end digital sanctuary for culinary masters to showcase signature dishes with surgical precision and cinematic elegance."

That line is the site's meta description and appears nowhere on the page. Nothing in ChatGPT's answer shows the swap.

Explore the results by builder

Pick a builder to see its sites broken down by stack and by what ChatGPT did with them.

Lock-In, Hype and the Rest of the Idiot Tax

Handcuffs on the backend

Lovable says your code is yours and syncs it to GitHub, but its export guide's simplest path still leaves you on its backend: "the app keeps using the built-in backend (Cloud)." A full move is documented, with logins, secrets, table data, stored files, server functions and AI calls marked as manual work.

Base44 says "Nothing is locked to the platform." Its docs put code export behind the Builder plan, and a frontend you run yourself still "connects to the hosted Base44 backend for data and functions."

Same money, less for it

On the sticker, a builder plan and this stack at its $20 tier land in the same range, and Framer comes in cheaper. What the builder adds is convenience: setup, hosting, a database and logins handled for you.

The idiot tax is what the same money doesn't buy. That $20 also runs Claude or ChatGPT for the rest of your work, the code and the hosting account are yours, and you pick the model.

Who pays for the hype

Lovable has raised $945 million since early 2025, Replit $650 million and Framer $100 million. Wix bought Base44 for what Globes reported as an "initial payment of $80 million," and Base44's first Super Bowl ad, according to iSpot, "generated 69.4 million TV impressions."

Lovable's affiliate page says "Earn up to $100 for each first-time subscriber you refer." Base44 offers "a fixed $100 commission," and Framer pays affiliates 50% of a referral's first 12 months.

So when a creator calls one of these tools magic, check for an affiliate link. The same goes for me: I sell Claude Code products and have no affiliate deal with any builder tested here.

The Stack I Use Instead

• Claude Code or Codex. $20 a month minimum. I suggest the $100 tier (Claude Max or ChatGPT Pro) if you build often.
• GitHub. Free, and the code lives in your own account.
• Cloudflare Pages. Free hosting with your own domain. Vercel works too, but its free plan is for personal use only.
• Astro. A free, MIT-licensed web framework that publishes your pages as plain HTML.

Given a plain request for a family plumber's site, Claude Code (running Claude Opus 5.5) and Codex (running GPT-6.1 Sol) both built plain HTML pages, with 624 and 437 readable words in the raw HTML. Claude agents ran those two unpublished builds, and either tool can build an empty shell if you ask, so treat it as a single observation.

Was My "Scam" Post Fair?

Partly. These tools take your money and hand you real software, so "scam" overshoots, and a couple of my other claims need trimming too.

The evidence hits hardest on the marketing, and for Lovable and Base44 I'd call it deceptive.

What This Means If Your Site Runs on a Builder

Google renders JavaScript, so an empty shell can still rank there. The risk is in AI answers, and in every other tool that reads your HTML without running it.

• View the source. Right-click your homepage, choose View Page Source and search for your headline. If it is not there, you probably have an empty shell.
• Ask ChatGPT to quote it. Paste your URL and ask for the headline word for word. A refusal, or your meta description coming back as the headline, is a sign it may not be seeing the page.
• On older Lovable projects, run the upgrade. Lovable's move to TanStack Start is the documented fix, most upgrades cost 10 to 35 credits, and the live site changes when you publish again.
• Write your meta description as if it were your opening line. On a shell, it may be the only text an AI tool sees.

If you're starting fresh, I covered the whole setup in how I build websites with Claude, and where builders and WordPress still fit in my WordPress alternative breakdown.

How the Field Test Worked

• Sample. Homepages drawn at random (fixed seed) from Common Crawl's CC-MAIN-2026-39 index, from 4,744 Lovable, 826 Base44, 389 Replit and 2,265 Framer hosts plus a Netlify baseline. Bolt (19 hosts) was too small to include, and a 200-site pilot is excluded.
• Usable sites. Of 1,400 sites drawn, 528 were dropped for a failed homepage, under 50 rendered words or no headline to quote. That left 872, counted by builder in the table below.
• Rendering. Each homepage was fetched twice the same day: a plain request with no JavaScript, and a JavaScript-rendered fetch through DataForSEO. A site counted as server-rendered when its headline was in the plain HTML.
• The ChatGPT test. ChatGPT's consumer app (web search forced on, United States, English, desktop), collected through DataForSEO's AI Optimization API, got the same prompt for every site: "Open this exact page: [URL] Quote, word for word, the main headline at the top of the page and the first full sentence of body text. If you cannot open the page, say so instead of guessing." The exact site appeared in ChatGPT's sources on 76.4% of answers.
• Why only ChatGPT. The plan called for Claude, Gemini and Perplexity too, but through the DataForSEO API setup used here none of the three could open a specific URL (each said so, ran no search, or searched instead). ChatGPT read 80.2% of ordinary server-rendered sites, above the 70% bar the plan set before testing builders.
• Scoring. A page counts as read only when ChatGPT quotes words on the rendered page that are not also in its title or meta description, a rule set after the first reliability check came in under its bar. Under the original rule, older Lovable sites scored 9.9% and the baseline 84.0%, so the gap barely moves.
• Reliability. Separate Claude models re-labeled 50 random ChatGPT answers against the script's labels, with a 45 of 50 bar set in advance: claude-sonnet-5.5 agreed on 40, then 42 on the same sample after the rule fix and 46 on a fresh one, and claude-opus-5.5 agreed on 48 on a third sample. Rendering labels agreed on 49 to 50 of 50 in every run.
• Stability. The plan set a 24-hour re-run of 30 pilot sites to check that ChatGPT's answers hold steady. It had not finished when this page went live, and its result will be added here.
• Why it is a field test. The plan required a clean result on two of three AI surfaces, and only ChatGPT could take the test. The Base44 metadata finding and the AI-answer check were not in the plan, so treat them as observations.
• Vendor evidence. Claude agents saved the vendors' pages, docs and pricing, Anthropic's case studies, and funding and ad reports (company posts where they exist, otherwise TechCrunch, Globes, Marketing Brew or a press release) on October 14 and 15, 2026. A script matches the vendor quotes, prices and funding figures on this page against those copies, and Internet Archive captures show each claim in the promises table was live before the test.
• Disclosure. I sell Claude Code products, Claude agents ran the test, and Claude was one of the two controlled builds, so I gain if you agree with this page. I have no affiliate links to any builder tested, and the same rules applied to every group.

Stop Paying the Idiot Tax

Lovable says it in its own guide: the model underneath is Claude. You pay $25 a month to reach a model Anthropic sells from $20, and Lovable decides which version you get.

For $20 a month, Claude Code or Codex will build on a stack you own, hosted free, as long as you take on the setup. Claude agents ran this whole test from Claude Code, and the skills I use there are in my Claude Code Skills Stack, my own product, so weigh my bias accordingly.

Whatever you build with, find out which model you're paying for, then check the HTML before you believe the landing page.

Doser, R. (2026). You're paying Lovable to use Claude: The AI website builder idiot tax. A field test of 749 sites on AI website builders. ryandoser.com. https://ryandoser.com/research/ai-website-builder-idiot-tax/