Launching today
Verol
Stop AI hallucinations
75 followers
Stop AI hallucinations
75 followers
Stop asking LLMs "is this true?" when they hallucinate, they will just lie again. Verol fixes this by adding an independent verification layer to ChatGPT, Claude, and Gemini. It parses answers, executes real-time web lookups, and validates sources through a dedicated backend pipeline. You get instant verdicts, confidence metrics, and clickable sources. No data tracking; history stays local. 5 free runs to test it. Plans from $4.99/mo.
Free Options
Launch Team / Built With
Verol
Hey Product Hunt 👋
I'm Mark, creator of Verol.
I built this tool because I got tired of manually copy pasting AI answers into Google one sentence at a time to check if they were real. It was fine for a couple of days, but it quickly became exhausting and I still couldn't trust half of what came back.
Verol fixes this by sitting right on top of the AI chats you already use. Instead of asking the model "are you sure?" (which usually just triggers another confident, hallucinated paragraph), Verol extracts individual claims and passes them through an independent backend pipeline with live web lookups.
It currently works on ChatGPT, Claude, and Gemini. And soon this list will increase. You get a clear trust score, per claim verdicts, and actual, clickable source links. Plus, your verification history never leaves your browser it stays completely local in your browser.
Out of curiosity, I also ran a small benchmark (100 identical prompts across 3 models) to see how they stack up. It’s not a formal scientific study, but the charts are in the gallery if you want to see how often confident-sounding answers fall apart under a live source check.
There's a free tier with 5 verifications so you can test it on your own workflows. I’d love your honest feedback, especially if you rely on LLMs for research, coding, or anything where a subtle hallucination completely ruins your day.
I'll be here to answer questions all day. Thanks for checking it out! 🙏
The plausible-sounding hallucinations are the real danger, and they're the hardest to catch because nothing about the answer looks wrong. I work on the data side of this, feeding models pre-verified facts so they don't have to guess, but that only helps when you control the source. For everything else, a verification layer like this is the right shape. The honest "couldn't find a source, here's my confidence" behavior is the part that earns trust. Asking the model to check itself just burns credits and produces another confident paragraph, like Luke said. One question on the real-time mode: verifying every response live sounds great but also expensive in latency and lookups. Do you let people scope it to certain chats or claim types so it isn't running on every throwaway message?
Verol
@hunter_upscale For now real time is a toggle, which user can turn on when he needs to verify data as he chats with AI, it has an eco mode where it finds all claims then by priority picks up only most necessary ones to verify based on whole context picture. Thats why we introduced as well manual mode where user can define by itself which claims to verify manually and which messages are necessary to verify and which ones are just small talk messages.
This is great! @mark_prod I would love to try this out soon. AI hallucinations has been the bane of my existence lately, just 2 days ago, I was working on a project that required a lot of research, I caught it hallucinating, pointed it out and it immediately changed it's direction. This got me worried if all of my previous research had been wrong and I had to start validating them all. So great job! This is really a problem I'm sure a lot of people face.
I'm curious to know though, what triggers Verol? Does a kind of kinds of question trigger it? Or does it work on the background, validating very response the AI chat gives?
Also, what does it validate against?
Verol
@alochukwu Hey, the work pipeline of this tool is simple. We have manual verification where the user chooses their latest message to be verified manually, and a real-time one which picks up every message from the AI chat and verifies it as soon as the response is written. We did both to give the user the freedom to choose when they need to validate and when it’s just a small message not worth verifying.
About validation, it works with multiple search engines and different LLMs based on the nature of the claim extracted. Different search engines have different sources of truth, so we combine them to get the best result out there.
Honestly the part that worries me as a non-coder isn't the obvious hallucinations, it's the plausible-sounding ones I'd never think to question. When I lean on AI to research a post or sanity-check an idea, I'm just... taking its word, because I can't always check the source myself. What does Verol do when a claim isn't really web-checkable, or when the sources it pulls disagree with each other?
Verol
@luca_capone Yeah i agree with that, and our tool is honest with you if it couldn't find any web resources to validate with, it won't frabricate some random resources instead we have a confidence score which says how confident the pipeline is.
Independent verification layer is the right framing. Asking an LLM to grade itself is a dead end, so source-backed verdicts and confidence metrics are valuable. How do you handle claims that require paid/logged-in sources or rapidly changing data?
Verol
@sarveshsea I am using different search engines, when a claim needs actual data it uses an according search engine which searches through most actual data. Unfortunately paid resources can not be validated as the DOM elements fetched by the pipeline are hidden by a paywall or something else.