Perplexity Focus Modes: Which to Use When

TL;DR: Perplexity no longer documents Focus modes as a picker. The control today is a mode selector offering Search, Deep Research, Learn Mode, Model Council, Comet and Computer, plus a Choose sources picker inside organisation accounts. This guide maps each mode to the job it is documented for, flags the plan gates, and gives 27 prompts.

If you came here looking for Perplexity Focus modes, the picker with Academic, Writing, YouTube, Reddit and Wolfram|Alpha in it, you are looking for something that Perplexity's current documentation does not describe. That is not a small editorial quibble. Nearly every guide ranking for this phrase tells you to open a control that no current help article explains how to find.

So this post does two jobs: establish what the selector actually offers today, from Perplexity's own pages with dates attached, then answer the question you came with. Which mode fits which job, and what to paste into each one.

Do Perplexity's Focus modes still exist?

Not as the picker guides describe. On August 29, 2026, Perplexity's English help centre contains 207 articles. Exactly one of them names Focus mode, and it exists to explain why you cannot use it.

That article is titled "Why can't I see focus mode on my search bar?" and its summary line reads: "For Enterprise Pro users, Focus has been replaced with Choose sources, allowing you to use Internal Knowledge Search by selecting between Web, Org Files, Web + Org Files, or None." Inside, it is blunter: "When you sign into an organization, it's not possible to use focus mode."

That is the entire published record. No article documents a consumer Focus picker, its options, or where to click for it.

The same quiet rename hit Spaces. The old URL for the Spaces article still returns 200, and serves a page titled "What are Projects?".

What is the mode selector actually called today?

Perplexity uses three different names for it across three live articles, so there is no single right answer to give a colleague.

• The Research article says you access the mode by "selecting" it "from the mode selector in the search box in the Mobile, web browser and Mac App version of the app".
• The Learn Mode article tells verified students they will "see a Learn icon in the search mode toggles", and elsewhere calls it "the toggle panel in your Search bar".
• The Model Council article, the newest of the three at August 20, 2026, gives the clearest click path: "Click Search below the search bar." Its figure caption reads "The Search menu is the entry point for Model Council on web."

Mode selector is the phrase I would use, because it is the one Perplexity's own Research documentation uses. But if a guide tells you to open the "Focus" dropdown, it is describing an interface Perplexity stopped writing about.

Which modes does Perplexity document in 2026?

The canonical list is not in a modes article. It is in the model-availability article, under a heading called "Modes in Perplexity", last updated July 29, 2026.

Two clarifications from that same page, both of which guides get wrong. First: "Best is not a separate mode. It is the automatic model option inside Search: Best picks a suitable model for each query so you do not have to." Second, on Deep Research, the Research article says the mode "automatically selects the optimal models for your research needs, so you cannot manually choose a specific model when using Research". Picking Opus and then picking Research does not give you Opus.

Why does a Perplexity prompt want to be a research brief?

Because the model is not the first thing that reads your words. A retriever is.

In a chat model, your prompt influences one process: what gets generated. In Perplexity, the same text steers what gets searched, what gets read, and only then what gets written. That makes a good prompt structurally closer to a research brief than a conversation opener. A brief names the question, the class of source that would settle it, the recency that matters, and the shape the answer should arrive in.

Perplexity's own prompting guide says the same thing in plainer language. Its golden rule: "State your intent clearly and provide detail." Its explanation: "a vague request gets a vague response, but a detailed prompt with clear intent delivers exactly what you need." And one tip most people skip, "Say how you want the answer", which is free precision.

What that demotes is the persona preamble. "You are a world-class analyst with twenty years of experience" gives a retriever nothing to search on, and displaces the terms that would. If you want the mechanics of that argument with the API evidence behind it, our search-shaped prompting guide has it. This post assumes it and moves on to the mode question.

Which mode should you use for which job?

Here is the decision procedure, in the order I actually run it.

1. Do you need an answer in under a minute? Search. Leave the model on Best unless you have a reason.
2. Would you accept a five-minute wait for something you could send to a colleague? Deep Research.
3. Do you need to understand the thing, not just have it? Learn Mode, if you have access.
4. Is being confidently wrong expensive here? Model Council, if your plan includes it.
5. Is the output a file, an email, or a recurring job rather than an answer? Computer.
6. Does the task happen inside pages you are already browsing? Comet.

If you have used Grok's modes, the shape will feel familiar: a fast default, a slow research mode, and a set of surfaces that are really separate products wearing the same search bar.

Search: prompts for when you want an answer in seconds

Search is the default and the one you will use ninety percent of the time. The prompt that works here is one question with one constraint on it. Naming a date window and a source class does more for the answer than any amount of role-play.

1. What is the current [metric] for [entity], and as of what date is that figure? Cite the primary source, not a summary of it.

2. Compare [option A] and [option B] on [criterion 1], [criterion 2] and [criterion 3]. Return a table. Mark any cell you could not source as "not published".

3. What changed about [product or policy] in the last 30 days? Only include changes announced by [organisation] itself.

4. Give me the three most recent [regulator or standards body] publications on [topic], with dates, and one sentence each on what they require.

5. Is [claim] accurate? Answer yes, no, or unclear, then show the strongest source for and against.

6. What does [organisation] officially publish about [topic]? Quote the exact wording and link the page. Do not include third-party interpretation.

7. I need [number] for a slide. Give me the figure, the publishing body, the publication date, and the URL. If two credible sources disagree, show both.

Deep Research: prompts for when you want a report

Research is the mode Perplexity describes as doing the legwork: "It performs dozens of searches, reads hundreds of sources, and reasons through the material to autonomously deliver a comprehensive report". Perplexity says it "completes most research tasks in under 3 minutes", then adds in the same sentence that this "means it can take around 4 to 5 minutes to generate a response". That sentence contradicts itself; plan for the longer number.

The August 2026 update added something that changes how you prompt: "For broad queries, Research now asks clarifying questions before starting". So a deliberately broad opener is no longer wasted, because the mode will narrow it with you. But a brief still beats an interrogation.

8. Research [topic] for a decision about [decision]. Cover [dimension 1], [dimension 2] and [dimension 3]. Prioritise primary sources and filings over commentary. Flag anything where the evidence is thin.

9. Build a market landscape for [category] as of [month, year]. For each named player: what they sell, who they sell to, and what they publish about pricing. Say explicitly which of those three you could not find.

10. I am evaluating [vendor]. Compile everything the company itself publishes about [capability], everything independent testing shows, and every documented limitation. Keep the three groups separate.

11. Summarise the state of the evidence on [research question]. Separate peer-reviewed findings from preprints and from press coverage. Note sample sizes where reported. Include the strongest published counter-argument.

12. Trace how [policy or standard] changed between [year] and [year]. Give me a dated timeline, then a short section on what is currently in force and what is only proposed.

13. Attached is [document]. Cross-reference it against current public sources on [topic]. Tell me which of its claims are still accurate, which are outdated, and which you cannot check.

14. Prepare a due-diligence brief on [company] covering ownership, funding, leadership, litigation and public commitments. Cite a primary source for every claim, and list every question you could not answer.

For research templates you can reuse rather than retype, our Perplexity research pack goes deeper on structure. And if losing citations between the search and the write-up is your actual problem, that workflow has its own post.

Learn Mode: prompts for when you want to understand it

Learn Mode is the one whose own documentation is at war with itself. The article states plainly: "This search mode is available for all users, including free and logged-out users." Four lines later, under "How do I access Learn Mode?", it says: "At the moment, learn mode is only available via the toggle to verified students." Both sentences were live on August 29, 2026. I cannot resolve which is true from the outside, so check your own toggle panel.

What it does is documented clearly enough: "Instead of just giving you the answers, Learn Mode will ask thoughtful questions and provide gentle hints to help you discover solutions on your own, turning the learning process into an experience both engaging and memorable." It generates flashcards and multiple-choice quizzes inline, and it takes uploads.

The prompt shift here is real: you are asking to be taught, not answered. Say what you already know, or it will pitch at the wrong level.

15. I understand [prerequisite] but not [target concept]. Teach me the gap, step by step, and check my understanding before moving on.

16. Walk me through this problem without giving me the answer. Ask me what I would try first. [Attach screenshot.]

17. I have an exam on [topic] in [number] days. Diagnose what I do not know with a short quiz, then build a study plan around the gaps.

18. Turn these lecture notes into flashcards covering the terms I am most likely to be tested on. [Attach notes.]

19. Explain [concept] three times: once for someone who has never heard of it, once at undergraduate level, and once the way a practitioner would describe it to a peer.

Model Council: prompts for when being wrong is expensive

Model Council runs your question across several models at once and synthesises one answer. Its opening argument is the honest one: "Every AI model has blind spots. It may overlook context, lean toward certain perspectives, or fill gaps with confident guesses."

The mechanics, from Perplexity's article: the default lineup is "one frontier model each from Anthropic, OpenAI, and Google", a synthesiser model reads their outputs and resolves them, and on web "A council on web always runs three models, so deselect one to make room for another." In Computer you can run between two and eight.

The plan gate is strict and worth knowing before you plan around it. Perplexity states Model Council "is not available on the Free, Pro, Education Pro, or Enterprise Pro plans" on web and iOS, and notes it "is not yet available as a standalone search mode in the Android app or in the desktop apps".

The prompt shape that earns the extra compute is a decision with stated criteria, because disagreement between models is the output you are paying for.

20. I am deciding between [option A] and [option B] for [purpose]. My constraints are [constraint 1] and [constraint 2]. Give me your recommendation, then say specifically where the models disagreed and why.

21. Verify this claim before I publish it: "[claim]". Show where the models converge and where any single model found something the others missed.

22. Assess the risks in [plan]. Rank them by likelihood times impact. Highlight any risk that only one model raised, and say which one.

23. Here is my reasoning: [argument]. Find the weakest step. If the models disagree about which step is weakest, show me all of their answers rather than picking one.

Computer and Comet: prompts for when the output is work

These two are not really search modes, which is why treating them as Focus-mode successors goes wrong. Computer is "an independent digital worker that completes tasks and workflows for you", with connectors, persistent memory and scheduled runs. Comet is "a Chromium-based browser" with Perplexity as its search engine.

Prompt them like you would brief a contractor: deliverable, format, deadline, done criteria.

24. Research [topic] and produce a 10-slide deck with a one-page executive summary. Cite every figure. Export as PPTX.

25. Every weekday at 08:00, check [sources] for anything new about [topic]. If nothing material changed, send nothing. Otherwise send me three bullets and the links.

26. Build a spreadsheet comparing [entities] across [columns]. Leave a cell blank rather than estimating, and add a Source URL column.

27. Read the open tabs and tell me which of these [items] meet [criteria]. Give me a shortlist with one line of reasoning each.

What if you need a real source filter?

Here the app and the API genuinely part company, and it matters for anyone who came looking for an academic focus.

In the app, the published controls are soft. Organisation accounts get Choose sources, with Web, Org Files, both, or None. Everyone gets Projects, whose Settings tab lets you "Add and manage web links and domains to be prioritized" and set instructions up to 8,000 characters that apply to every query inside. Perplexity's word is prioritized. That is a weighting, not an allowlist.

In the API, the filters are hard and documented. search_domain_filter accepts "a maximum of 20 domains or URLs" in allowlist or denylist mode. search_recency_filter takes a fixed vocabulary of hour, day, week, month, year. And there is a genuine academic mode: search_mode set to "academic", which prioritises scholarly and peer-reviewed sources.

One deadline to diary: every Sonar documentation page carries the notice that "Sonar will be supported until September 27, 2026", with Chat Completions moving to the Agent API. If you have scripts using the filters above, that date is close.

Does a citation mean the answer was checked?

No, and this is the part worth internalising more than any mode chart.

Retrieval means the answer is only as good as what the retriever found. A citation records that a page was fetched and used. It does not certify that the summary read it correctly, that the page is current, or that the page even says what the sentence above it claims. All four failure modes leave the citation looking perfectly healthy.

Perplexity says as much in its own source-labels article, published August 7, 2026: "A label describes the website as a whole, not any single article or claim on it." And its closing line is the one I would put on a poster: "Perplexity aims for accuracy, but reviewing the original sources yourself remains the best way to build confidence in an answer."

There is external evidence too, and it needs scoping carefully rather than quoting as a headline. The Tow Center for Digital Journalism, writing in Columbia Journalism Review on March 6, 2025, ran 1,600 queries across eight AI search tools on one specific task: given a verbatim excerpt from a news article, identify its headline, publisher, date and URL. Collectively the tools "provided incorrect answers to more than 60 percent of queries", with "Perplexity answering 37 percent of the queries incorrectly". The excerpts were chosen so that pasting them into Google returned the original source in the first three results.

The saving grace is that Perplexity is the one tool where checking is cheap. The links are sitting right under the sentence. You do not need a research protocol, just a habit: open the citations behind any claim you would be embarrassed to be wrong about, and skim for the sentence the summary is standing on. Two clicks, and most hallucination-shaped errors fall out immediately. If you want a genuinely source-locked alternative for some jobs, NotebookLM only answers from what you upload.

What will date this post first?

In likely order:

1. The mode list. Model Council arrived recently enough that Perplexity's own article is still dated August 2026. Expect additions.
2. The model names. Perplexity already contradicts itself across three articles, and its stated tiebreaker is the selector in your own account.
3. Sonar's retirement, dated by Perplexity as September 27, 2026.
4. Learn Mode's access rule, which is currently documented two incompatible ways.

The one thing I would bet stays true is the shape of the advice: name the question, name the source class, name the window, name the output.

Where does a prompt tool fit in this?

Honestly: not in the retrieval. Prompt Architects does not search the web, rank sources, or check a citation. Perplexity does the retrieval, and no browser extension changes what its retriever sees.

What it does is the boring half. Prompts like the ones above are worth writing once and reusing, with the entity, date window and output format as variables rather than a fresh retype each time. Our extension works on perplexity.ai and is listed on our integrations page as available. In fairness, our own FAQ still lists Perplexity as an upcoming integration, which is stale and contradicts the integrations page. The integrations page is the accurate one, and we should fix the FAQ.

The free plan includes 5 prompt enhancements per day, forever, per our FAQ page. That is enough to test whether a saved, variable-driven research brief beats retyping one, which is the only claim being made here.