Your AI-Built Contact Form Says "Message Sent!" 27 of 105 Sent Nothing. Test the Inbox
A study of 105 Lovable and Bolt contact forms found 27 that showed "Message sent!" and sent nothing. Even a form that does send can fail after the click, so the only test that proves a form works is a real message arriving in the inbox you read.
In October 2026, a test filled in the contact form on 105 public sites built with Lovable and Bolt, pressed send, and recorded every request the page tried to make. A request is what a web page sends to a server, such as the text of a form going out to be stored or emailed.
27 of the 105 forms showed a success message such as "Message sent!" and sent nothing at all.
A success message on an AI-built contact form proves only that the page ran its own code.
Only one test proves the form works: you send a message from a signed-out browser on the live site, and it arrives in the inbox you actually read.
Everything before that inbox can report success while nobody is told.
You have already checked the success message. If you built a site with Lovable, Bolt or a similar tool, filled in its form once and watched the green toast appear, you tested the one piece the AI was sure to build.
A fake form is a timer and a toast
The forms that faked success had nothing broken in them.
The study's authors read the submit code of 21 of the 27. They found no request at all. The handler waits about a second, shows the message, and clears the fields.
The other six were compressed too tightly to read. The study counts them only because the test flagged them.
The study's write-up prints the pattern it found "again and again", made readable again. Look at what is missing between the click and the toast:
onSubmit={(e) => {
e.preventDefault();
setSending(true);
setTimeout(() => {
toast({ title: "Message sent!", description: "We'll get back to you soon." });
form.reset();
setSending(false);
}, 1000);
}}
No line in it sends anything.
The button changes to "Sending…". The toast appears. The form empties. The visitor leaves believing they reached you.
You never hear from them. Nothing on the page warns you.
The study does not say why the timer-and-toast handler is there. Its authors call their own reading of it "our reading, not a measurement". They say they did not test what Lovable or Bolt generate by default.
The study does show that a form can look and behave finished with no sending code behind it.
How many AI-built contact forms send nothing?
The first doubt most readers have is that these are demos nobody depends on.
The study planned for that doubt. Before counting, it fixed a rule that put every site into one group. A site was a demo (a hackathon entry, or placeholder contacts such as a 555 phone number), a portfolio, or a site with no demo signs.
Table: Forms that showed success and sent nothing, by kind of site
| Kind of site | Fake forms |
|---|---|
| Demos (hackathon or placeholder contacts) | 9 of 15 |
| Portfolios | 2 of 17 |
| No demo signs | 16 of 73 |
| No demo signs, on the site's own domain | 5 of 30 |
Demos fake their forms most often, and the problem stays on sites that look like real businesses, including sites on their own domain name instead of a lovable.app or bolt.host address. Source: RESULTS.md, final run 8 October 2026.
The authors call "no demo signs" a cautious label, not proof that a business depends on the form.
Still, 16 of 73 is roughly one in five sites that show no sign of being a demo.
By builder, 12 of 67 Lovable forms and 15 of 38 Bolt forms were fake.
The study comes from Formgong, whose own method post opens with "We build a form backend", and its write-up ends by recommending form backends including its own. The number holds up despite that, because the authors wrote the rules down before testing and logged every change with its reason. The test script and its 23 known-answer pages are public under the MIT licence. Those pages are small test forms whose correct result is already known.
Anyone can rerun the test on their own sites. The site list and per-site results stay private, so the 105 forms themselves cannot be rechecked.
The study saw requests leave, and never saw an inbox
To test other people's forms without bothering them, the study stopped every message inside the browser.
Once the test found the form, it recorded and aborted every outgoing request. No site owner received anything. A form counted as sending if a request carried the typed test text.
By that measure, 54 of the 105 forms sent.
Of the rest, twelve did nothing visible, seven opened the visitor's own mail app, and five left or reloaded the page without the message.
The authors are plain that, because they blocked every request, they "saw whether a form tried to send, not whether the email arrived". They add that "forms counted as 'sends' can still fail later".
Their own summary of the lesson goes one step further than the number: "A green message proves nothing. A request proves the form tried, and only your inbox proves it delivered."
Why a form can send and you still hear nothing
A reasonable objection is that the outgoing request is the hard part.
If the browser sends what the visitor typed, the form works. Anything after that is the email service's job.
The documentation of the services these forms use says otherwise. Each of the steps after the request can fail without the page ever knowing.
Start with the code that sends.
In JavaScript, a page's code usually sends a request with fetch(). That command does not treat an error reply from the server as a failure. Every reply carries a status code. The 200s mean accepted, and 4xx or 5xx codes such as 404 mean refused. Your code must check the status.
Supabase is the hosted database many Lovable and Bolt projects store their data in. In the study, 27 of the 54 sending forms sent to it. Its JavaScript library returns an error object for a failed request. The code carries on unless told to stop. A handler that waits for the reply and then shows the toast, without checking for that error, thanks the visitor even when the server has refused the message.
The destination can also be quieter than you expect.
On Lovable's documented path, a contact form saves each submission as a row in a table. Email is a separate step. Lovable wires the email only after your sender domain is verified, which means you have added the records your email provider gives you to your domain's DNS settings to prove you own it. Until then, Lovable keeps saving rows.
So the form sends. The messages can wait in a table you may never open.
Bolt's guide sends email through a server function. When you don't name a provider, Bolt typically picks Resend, an email-sending service, and that function then needs your own Resend API key, the password-like code Resend issues. It also needs a published project with a custom domain attached. Bolt lists both as prerequisites. A form wired up before they are in place has no way to email you.
Email adds three more stops.
Resend's shared test sender delivers only to the Resend account owner's address. A notification to any other inbox is refused until you verify a domain.
EmailJS, another email service, was used by 13 of the sending forms. It allows 200 requests a month on its free plan. Past the limit, requests "will stop being processed and will be ignored". EmailJS does warn first. It sends the account owner several notifications as the quota runs out, so those warnings are worth looking for.
And Lovable warns that mail from a newly set-up domain "may initially land in spam".
Chrome's Network panel, part of the browser's developer tools, lists every request a page makes, with its status code. The test below shows how to open it.
Table: Where a contact form message can stop, and what you would see
| Where it stops | Success message | Request in the Network panel | Message in your inbox |
|---|---|---|---|
| Fake handler (timer and toast) | Yes | None | No |
| Server refuses, code ignores the error | Yes | Yes, with an error status | No |
| Saved to a table, no email wired | Yes | Yes | No |
| Email service refuses or ignores it | Yes | Yes | No |
| Delivered to spam | Yes | Yes | Only in spam |
| Works end to end | Yes | Yes | Yes |
All six rows show the same success message. The inbox column tells you whether the form works, and the Network panel column tells you where a failed one stopped. Sources: the study's write-up, MDN, Supabase, Lovable, Resend and EmailJS documentation linked above.
The first four steps can each report success while the message stops; only the last one proves the form works.A check for a request answers the objection only part of the way. It catches the first row, where nothing leaves. An error status flags the second. The next three failing rows show a request that went out with no error, exactly as a working form does.
The test that counts runs from outside
So the test has to start where a visitor starts and end where you read.
That rules out the builder's preview. One site-testing company, QAlaunch, explains why. The preview runs signed in as you. If the database's access rules allow writes only from the signed-in owner, a submission succeeds for you and fails for an anonymous visitor on the live site.
The same requirement rules out reading the screen.
The Network panel earns its place after a failed test, because it tells you which step to look at.
How do you test a contact form end to end?
The test takes about ten minutes. Most of it is waiting.
Use Google Chrome. Other browsers have the same panel under similar names.
- Open the live site in a private window. Use the published address, not the builder preview, and do not sign in. You now arrive as a stranger would.
- Open the Network panel before you touch the form. Press F12 (Cmd+Option+I on a Mac) and click Network. DevTools records only while it is open, so open it first, and tick Preserve log so a page change does not clear the list.
- Write a message you can search for. Put a unique marker in it, such as
form-test-1118, and use an email address you can check as the sender. - Press send and watch the list. Click the Fetch/XHR filter to hide images and scripts, then look for a new row that appears after the click.
- Wait five minutes, then search your inbox and spam for the marker. If you set up a sender domain today, its DNS changes can take up to 72 hours, so test again once it shows as verified. If the marker is there, the form works for strangers. If the form also stores submissions in a table or a dashboard, check that it arrived there too.
If the marker never arrives, the Network panel tells you where to look next:
- No new request after the click. The page sent nothing, which is the timer-and-toast pattern. Ask your builder to connect the form to a real destination and to show the success message only after the server confirms.
- A request with an error status (4xx or 5xx in the Status column). The server refused it. Click the row and open the Payload tab to confirm your text was in it, then read the Response tab for the reason.
- A request that succeeded, and still no email. The message reached your backend and stopped there. Check the table where submissions are saved, whether your sender domain is verified with your email provider, your provider's monthly quota, and the spam folder.
Run the same test again after every fix. Repeat it after any later change to the form or its email setup.
Other gaps between a working prototype and a site people depend on show up after launch. Build to Launch's 8 Vibe Coding Mistakes That Break Apps After You Ship covers them.
Can you run the study's test on your own site?
You can, and it is safe. Once the study's test finds the form, it aborts every request. Nobody receives anything.
The study's test needs Node, Google Chrome and a terminal.
The README asks you to run its 23 known-answer pages first. All of them must match before you point it at a site.
The test's own history shows why a check of this kind needs checking itself. Before the final run, the authors found twelve faults in their own test, from analytics pings counted as sending to toasts that vanished after two seconds. A hand check or a known-answer page caught each one.
For a builder checking one form, the study's test answers less than the inbox test does.
The script blocks the message. It can still tell you a request left, which the Network panel already shows. It can never tell you the message arrived.
The setup is worth it if you look after many sites and want to sweep them for the timer-and-toast pattern.
FAQ
How do I know if my contact form actually sends?
Send a test message from a private window on the live site. Then look for it in the inbox the form should deliver to, including spam.
The study's quick check, a request carrying your text in the Network panel, shows only that the form tried to send.
Why does my Lovable form save submissions but never email me?
On Lovable's documented path a contact form saves each submission as a row, and you read them under More → Database.
Lovable adds email only when you ask for it, and wires it only after your sender domain is verified. The DNS changes for that domain can take up to 72 hours.
Why did my form work in the preview but not on the live site?
The builder preview runs signed in as you, so it can succeed where an anonymous visitor fails, according to QAlaunch's account.
QAlaunch adds that an environment variable, such as an API key, that never made it to the published site has the same effect. Test the published address from a private window.
Sources
- Formgong, dead-forms-study repository, accessed 2026-11-18
- Formgong, RESULTS.md, final run 8 October 2026, accessed 2026-11-18
- Formgong, PREREG.md, accessed 2026-11-18
- Formgong, We tested 793 Lovable and Bolt sites. 27 of 105 contact forms said "sent" and sent nothing, accessed 2026-11-18
- Formgong, How we tested AI-built contact forms, and 12 bugs we hit, accessed 2026-11-18
- Lovable Documentation, Database, accessed 2026-11-18
- Lovable Documentation, Send branded emails from your own domain, accessed 2026-11-18
- Bolt Help Center, Send emails from your Bolt app, accessed 2026-11-18
- Resend, 403 Error Using resend.dev Domain, accessed 2026-11-18
- MDN, Using the Fetch API, accessed 2026-11-18
- Supabase, JavaScript API reference, accessed 2026-11-18
- EmailJS, Pricing, accessed 2026-11-18
- Chrome for Developers, Inspect network activity, accessed 2026-11-18
- Chrome for Developers, Network features reference, accessed 2026-11-18
- QAlaunch, Contact Form Not Working? Test It in 5 Minutes, accessed 2026-11-18
- Build to Launch, 8 Vibe Coding Mistakes That Break Apps After You Ship
Links
- Formgong, dead-forms-study repository (documentation)
- Formgong, dead-forms-study RESULTS.md (documentation)
- Formgong, dead-forms-study PREREG.md (documentation)
- Formgong, We tested 793 Lovable and Bolt sites. 27 of 105 contact forms said "sent" and sent nothing (documentation)
- Formgong, How we tested AI-built contact forms, and 12 bugs we hit (documentation)
- Lovable Documentation, Database (documentation)
- Lovable Documentation, Send branded emails from your own domain (documentation)
- Bolt Help Center, Send emails from your Bolt app (documentation)
- Resend, 403 Error Using resend.dev Domain (documentation)
- MDN, Using the Fetch API (documentation)
- Supabase, JavaScript API reference (documentation)
- EmailJS, Pricing (documentation)
- Chrome for Developers, Inspect network activity (documentation)
- Chrome for Developers, Network features reference (documentation)
- QAlaunch, Contact Form Not Working? Test It in 5 Minutes (article)
- Build to Launch, 8 Vibe Coding Mistakes That Break Apps After You Ship (article)
New guides on building with AI come in the Build to Launch newsletter.
More guides
Fable 5.2 vs Opus 5.5: Raise Effort Before You Switch Models
Anthropic says to start on Opus 5.5 and switch models last. Give Claude a check, raise effort one level, then try Fable 5.2, which costs about the same as Opus at max in independent tests, not half.
Vibe Coding Development Services: A €3,000 and a €12,000 Quote Can Buy the Same Screens
The screens in an agency quote are what your AI tool already builds. Compare vibe coding quotes on the five lines a prototype leaves out: data locked to each user, a security review, deployment, code ownership and fixes after handoff.
How to Build Claude Interactive Apps That Still Work When Claude Withholds the New Features
Claude interactive apps are MCP Apps: a tool on your remote MCP server opens a sandboxed web view in the conversation. The inline view reaches every install; the side panel, app tools, entry points and embedded sites added on 2026-11-10 depend on each user's organization, interface and listing, so build them as upgrades.