Duck.ai Voice Chat Can Search the Web: Audit Sources, Location, and Answer Drift
A test protocol for comparing Duck.ai voice and text web-search answers across transcription, location, sources, citations, timestamps, and repeated runs.
Protocol status: This page publishes a test protocol and no result. DuckDuckGo’s Q2 2026 product update says Duck.ai can search the web during voice chats and names a time-sensitive local-weather question as an example. The announcement does not disclose retrieval providers, citation coverage, selection rules, or parity between voice and text.
A voice run adds speech recognition, microphone conditions, transcript handling, location state, and interface behavior to the retrieval and answer stack. Any difference must be attributed to the complete recorded path, not casually labeled a “voice ranking” effect.
Preregister time-sensitive prompts
Choose prompts that genuinely require current web information: local weather, a current event, a product status, an opening time, a live score, or a recently changed policy. Include prompts with and without local intent and define which answers should refuse or request clarification.
Freeze language, country, app and browser version, device, selected model when visible, signed-in state, location permission, date, time window, and network. Keep personal data out of the published fixture and obtain consent before recording another person’s voice.
| Layer | Save | Why it can change the result |
|---|---|---|
| Spoken input | Audio condition and exact words | Recognition errors change the query |
| Transcript | Visible transcript and corrections | System may search different wording |
| Context | Location, time, device, model | Local and current answers vary |
| Evidence | Sources, citations, answer, timestamp | Retrieval and generation can drift |
Pair voice and text runs
For each item, run a typed prompt and a spoken prompt with the same intended wording. Save the actual transcript before editing, any user correction, visible or inferred query text when legitimately available, answer, sources, citations, latency, and errors. Randomize order and use a clean chat where context could carry over.
Repeat runs on a declared schedule because current search results and generated wording can vary. Do not replace a failed transcription with a perfect manual edit and still call the result “voice.” Preserve both the original and corrected path as different conditions.
Audit sources and answer drift
Open every source and verify whether it supports the nearby answer claim. Record source owner, publication or update date, location relevance, citation placement, broken links, unsupported statements, and important sources omitted from the answer. Separate cited, retrieved when visible, and merely mentioned.
Compare transcript fidelity, answer correctness, completeness, uncertainty, source support, source overlap, location behavior, and repeated-run stability. Do not use source count alone as quality: one primary source may be stronger than several copies.
Report product behavior within the sample
Publish the prompt set, voice conditions, environment, location state, model, timestamps, transcripts, source ledger, rubric, failures, exclusions, and stopping rule. Bound every conclusion to the tested platform version and dates.
An observed difference can motivate a follow-up, but it does not reveal a private retrieval provider or universal voice-search rule. If DuckDuckGo later documents source selection or parity, update the claim ledger and rerun rather than silently rewriting the old observation.
Use the DuckDuckGo preview audit for another bounded product check and the AI visibility protocol for citation terminology.
Ask a question or join the discussion