Duck.ai vs Perplexity: Cited Answers or No Stored History
Perplexity cites live-web answers and trains on your chats by default. Duck.ai keeps no chats on a server and strips your IP, but has no confirmed live search.
Published October 6, 2026· Updated October 6, 2026· By Juan Martinez
Duck.ai and Perplexity end up in the same place in people’s routines: the box where a question goes when a search engine would have been the old answer. Both let you type that question without signing up for anything, so the usual first test of a privacy comparison, whether an account is required, does not separate them. What separates them is what happens next. Perplexity sends the question to the live web and returns a written answer with numbered citations to the pages it used. Duck.ai forwards the question to a model built by someone else, with your IP address removed, and keeps nothing on a server, but has no confirmed live search, so the answer is only as current as the model’s training data. The scores are 7.5 for Duck.ai and 5.4 for Perplexity, and nearly all of that gap sits in one category.
| Duck.ai | Perplexity | |
|---|---|---|
| Rating | 7.5 | 5.4 |
| Account Required | none | optional |
| IP Anonymization | proxy-stripped | unknown |
| Storage Model | client-local-only | server-encrypted |
| Jurisdiction | United States | United States |
| Free Tier | full | limited |
| Platforms | Windows, macOS, Android, iOS | Windows, macOS, Android, iOS |
| Est. | 2024 | 2022 |
Want to compare every field side by side? Open the full comparison table.
What each one is for: answer engine versus model front end
Perplexity is built around search. Every answer is assembled from live web results, and each claim in the reply links back to the page it came from. That is the product’s core, not a feature added to a chatbot, and for fact-finding it is a more useful shape than an unsourced reply. Perplexity AI, Inc. launched the search product in December 2022.
Duck.ai is a different kind of thing: DuckDuckGo’s chat interface to models from Anthropic, OpenAI, Mistral, Azure and Tinfoil, launched in autumn 2024. DuckDuckGo does not build the models. It sits in the middle, removes the identifying parts of the request, and passes the prompt on. DuckDuckGo’s AI-assisted answers on its search results page are a separate feature, and nothing in Duck.ai’s documentation says the chat pulls in fresh search results.
So the choice is not between two chatbots with different settings. One is a search tool that writes its results up for you; the other is a privacy layer in front of several models that know only what they were trained on.
The account and the IP address: a shared floor, then a split
On the first visit the two look alike. Neither asks for an email, a phone number or a card before you type. For a one-off lookup, both keep your name and email out of the picture.
The floor ends there on Perplexity. Signing in is what turns on saved history, and that history is stored on Perplexity’s servers, tied to the account, for as long as the account exists. Staying signed out avoids the account, but not the network identifier: Perplexity’s Privacy Notice says it may collect your general location from the IP address you connect from, and nothing in its documentation says the IP address is separated from the question before it is processed. Cookies, pixels, device information and server logs are collected automatically either way.
Duck.ai has no account to sign in to, so there is no version of the product where history accumulates against a profile. DuckDuckGo’s documentation states that metadata containing personal information, with the IP address as the example, is completely removed before the model provider is prompted. Its privacy policy adds that IP addresses are never logged to disk alongside Duck.ai chats in a form that could be tied back to a user. The stated position is that the link between you and the question is not kept at either end.
Where your question goes and what is kept
Storage is the clearest difference. Perplexity keeps signed-in history server-side, encrypted in transit with SSL/TLS, but readable by Perplexity: no zero-access encryption, and no consumer-specific statement about encryption at rest. Duck.ai keeps recent chats in local storage on your device, not on DuckDuckGo’s servers; you can delete them one by one or all at once with the Fire Button, and pin up to five.
Retention follows from that. Perplexity says personal information is kept as long as the account is active and removed within 30 days of account deletion, which you request through a form. Its incognito sessions expire within 24 hours and cannot be recovered, the one short-lived mode it offers. On Duck.ai, model providers commit to deleting prompts and responses immediately after a reply, with one disclosed exception: prompts may be held in short-term memory for up to an hour for caching, never written to disk. That hour is a real window, and DuckDuckGo states it rather than rounding it down to “nothing is stored.”
Training is where the defaults point in opposite directions. On Perplexity’s Free, Pro and Max plans, AI data retention for training is on by default. The opt-out is a single toggle in account settings, which is better than a support request, but by Perplexity’s own statement it applies only to data collected after you switch it off, and data already collected for training cannot be deleted or removed. Someone who finds the toggle after a month has already contributed that month. Duck.ai’s position is contractual instead: Anthropic, OpenAI, Mistral, Azure and Tinfoil are all barred from training on Duck.ai conversations, and there is no setting to find or forget.
Duck.ai’s protection is also weaker than the headline suggests on one point. The hardware-enforced part, a Trusted Execution Environment where the host cannot see the prompt or the response, covers only the two Tinfoil-hosted models, gpt-oss-120b and Gemma 4 31B. On the other eight models the protection is the contract plus the stripped metadata. Perplexity makes no enclave or confidential-compute claim for its consumer product at all.
Trust and track record: a gated audit against no audit
Both companies are in the United States, which sits inside the intelligence-sharing group commonly called the Fourteen Eyes. The difference is what a legal demand could reach. Perplexity holds readable history for signed-in users. Duck.ai, by its own description, holds no account, no server-side chat store and no IP-to-question mapping on disk. Both products are closed source, so in both cases the claims rest on documentation.
Perplexity’s trust picture is better than its reputation in some places. It lists a SOC 2 Type 2 report and an ISO/IEC 27001:2022 certificate on its Trust Center, plus annual third-party penetration tests and a bug bounty. The help articles on training, retention and incognito are short, dated, and state their own limits, including the uncomfortable ones. The Privacy Notice says Perplexity does not sell personal data and does not send queries, prompts or conversation content to advertisers. The limits: the SOC 2 report sits behind a request-access form aimed at business buyers, with no named auditor and no public report period, and the same notice says cookies, pixels and server-to-server integrations measure Perplexity’s own advertising, with consent assumed from continued use.
That last point sits next to the most serious question raised about the company. A class action filed on 31 March 2026 in federal court in San Francisco, first as Doe and later as Noel v. Perplexity AI, Meta and Google, alleged that tracking code in Perplexity sent users’ chat data, email addresses, device information and IP addresses to Google and Meta without consent, including in incognito mode. The plaintiff dismissed the case voluntarily and without prejudice on 1 May 2026, before any defendant had answered. No court ruled on the claims, and the plaintiff can refile. The allegations remain allegations. A separate disclosure concerns Comet, Perplexity’s agentic browser rather than the assistant: a prompt-injection flaw whose first fix was incomplete. It is a vulnerability in a different product, not a breach of user data.
Duck.ai has no breach on record, detailed public documentation of its data flow, no usage telemetry, and an owner that has been independent since 2008 and is majority-owned by its founder and team. What it does not have is any verification. No independent party has audited the IP stripping, the no-disk-logging claim or the enclave setup, and closed source means nobody outside can check them. Perplexity has an attestation most readers cannot open; Duck.ai has only the operator’s word. Trust & Transparency is 6.2 against 4.0 because Duck.ai’s word describes a design that leaves little to misuse, not because it has been proven.
Answers: live citations versus training-data limits
This is Perplexity’s category, and the Features score says so: 8.5 against Duck.ai’s 7.0. Cited answers from the live web handle the questions a search box usually gets, such as this week’s news, a current price, or a recent change to some rule, and Duck.ai cannot. Around that core, Perplexity offers Pro Search, Deep Research for longer multi-source reports, file upload (limited on Free, ten times the allowance on Pro, up to 50 files per project), image and video generation, document and app creation, memory, and Computer, an agent that carries out tasks. Its lineup includes GPT-6.1 Sol, Gemini 3.8 Flash, Claude Sonnet 5.5, Kimi K3, GLM 5.3, Grok 4.7 and NVIDIA Nemotron; on Free the model is picked for you, and Pro and Max add a picker.
Duck.ai’s case is narrower. It offers ten models from four families, six of them free, with switching per conversation, so comparing how two families answer the same question takes one dropdown. Image and PDF uploads work on some models but not all, PDFs are capped at 15 pages free, 35 on Plus and 50 on Pro, and spreadsheets are not supported. Image generation has an unpublished daily cap. With no confirmed live search, the model’s training data is the limit, which works for explanations, drafting and reasoning but not for anything that changed recently.
Price and platforms
Both free tiers are real. Perplexity’s offers practically unlimited basic searches, search history once you sign in, very limited Pro Searches and limited file uploads; advanced models and image generation are not included. Duck.ai’s gives six models, including both enclave-hosted ones, with no account and no card, under a daily cap DuckDuckGo does not publish.
Paid plans are not comparable like for like. Perplexity Pro is $20 a month ($17 a month billed annually) and Max is $200 a month ($167 billed annually). Paying buys capability and nothing else: Free, Pro and Max share the same training default, storage model and tracking. The protective terms, with no training on queries and 7-day file retention, are Enterprise-only. Duck.ai’s advanced models are not sold on their own. The $9.99 a month Plus plan ($99.99 a year) is DuckDuckGo’s whole four-in-one subscription covering a VPN, advanced Duck.ai models, personal information removal and identity theft restoration. Pro at $19.99 a month ($199.99 a year) adds Claude Opus 4.8, GPT-5.6 Sol, higher reasoning effort and double the Plus usage limits. If you want only the models, that bundle is poor value, which is why Duck.ai’s 8.0 for Price & Value rests mostly on the free tier. Perplexity scores 6.5 because its paid tiers buy features but no privacy.
Platforms are close. Perplexity has native apps for macOS, Windows, iPhone and iPad, and Android. Duck.ai has no standalone app, but it is built into the DuckDuckGo browser on Windows, macOS, iOS and Android, with a Search/Duck.ai toggle in the address bar and a Chats menu. Neither has a Linux client. Duck.ai’s local storage means no sync: a chat started on a laptop is not on the phone, and clearing browser data erases it. Perplexity’s Usability edge, 8.0 against 7.5, comes with its own cost in a product line that has sprawled across the app, Comet, Computer and a credit system.
What we didn’t compare, and why
Perplexity’s Enterprise plans. These have a different data policy: no training on Enterprise data, 7-day file retention, and zero-retention agreements with model providers. They are sold to organisations at per-seat pricing, not to the individual reader, and the Perplexity review scores the consumer plans for that reason.
Comet and the Computer agent. Both extend Perplexity beyond answering questions, into browsing and acting on sites. Duck.ai has no equivalent, so there is nothing to put them against, and the Comet flaw belongs to that product rather than to the assistant.
The enclave as a single yes-or-no row. Duck.ai’s Trusted Execution Environment applies to two of ten models. As a table row it would read as a property of the whole product when it is a property of a model choice, so the scope is stated in prose.
Model quality. Neither review scores whose answers are better, and several models on both sides come from the same families, so the comparison would mostly reflect which model was picked.
Bottom line
Read the Perplexity review if your questions need today’s web with sources you can click: news, prices, recent changes, anything where an answer limited to training data is the wrong answer. Use it signed out where you can, turn the training toggle off early (it cannot reach back), and keep anything sensitive out of a history Perplexity can read.
Read the Duck.ai review if the question itself is what you want protected: no account, no IP address passed to the model provider, nothing stored on a server, and no-training commitments in contract rather than in a setting. You give up live search, citations, sync and much of the feature set, and you accept that nobody independent has checked any of it.