Useful answers require more than fluent text
An answer is only as good as the sources behind it. What retrieval, citations and abstention should look like, and how to check a dark web answer for yourself.
By Dark Web Search editorial · Updated
How can I search the dark web?
You submit a question to a service that retrieves permitted public sources and returns an answer together with the passages it relied on. That needs no VPN subscription and no Tor install, because you are using an ordinary website. What comes back is a reading of sources rather than a live view of the network, and opening an original onion site remains a separate action that requires Tor-capable software. The rest of this guide covers the harder half of the task: deciding whether the answer you received is worth anything.
What is a dark web search?
It is a query run against a collection that already holds material from onion sources, not a scan of the whole network performed on demand. That distinction decides what a result can mean. A hit tells you a particular collection recorded a particular source at a particular time. An absence tells you that collection holds nothing matching, which is a statement about the collection rather than about the world. Any answer built on top of a search inherits both limits, and a fluent summary can hide them.
Retrieve before composing
A research agent should work from permitted evidence tied to the question. It has bounded attempts, source counts, time and upstream budget. Broader autonomy is not a substitute for correct sources or resource control.
Treat sources as data
A retrieved page can contain instructions intended to manipulate a model. Those instructions cannot change the user's task, invoke tools, spend money, reveal secrets or save memory. The answer process receives source data in a separate evidence field.
Check the output
A citation identifier must correspond to a supplied source version and supporting passage. If support is missing, the system should return sources or abstain. Passing a schema check alone does not establish semantic truth; judged evaluation is still needed.
Measure actual usefulness
Report answerability, essential-intent coverage, exact-literal preservation, supported-claim coverage and false abstention. Do not replace those with model confidence or an untested claim to search everything.
What an AI answer should show
Look for a direct answer followed by the evidence that supports it. Important statements should be traceable to a relevant passage, with qualifications preserved. The answer should distinguish what a source says from the system’s own inference. If only normal web sources contributed, an onion-only result should not be implied. As a reader, try removing the citations mentally. If the remaining answer makes a broad claim, open the source and check whether the claim survives with its original scope and date. This is a practical review technique, not a guarantee that every error will be found.
A useful follow-up question
Instead of asking “tell me everything,” ask “which primary source supports this point, and what does it leave unresolved?” A narrower follow-up gives the system a testable task. For a comparison, name the two statements and the difference you want explained. For a current-status question, specify the time period that matters. Do not add private identifiers merely to make a question seem precise. A public research question can often be bounded by document title, publisher and date. Keep unnecessary personal context out of the source query.
When an answer should stop
Stop when the available evidence cannot support the next conclusion within the allowed scope. More generated text is not additional evidence. A useful response can return the best sources, explain the missing fact and identify the next observation needed. For example, an archived page might explain a service’s former access method but cannot establish that the service still operates today. The correct outcome may be an explicit current-status gap. This product’s methodology page explains the distinction between output checks and actual claim support; neither fluent prose nor a valid citation identifier proves truth by itself.
Primary references
Tor Project: onion servicesThe Tor Project · read · primary source
Related help
- Dark web search, with evidence you can inspect
- From your question to evidence you can inspect
- Evidence before confidence
- Compare evidence, coverage and limits
- A citation starts the checking process
- Connect with scoped access
Content revision 2026-09-16. AI-assisted editorial content; check the primary references and their dates.