How to Prompt Perplexity (Search-Shaped Prompts)

TL;DR: How to prompt Perplexity comes down to one inversion: it retrieves before it writes, so your prompt steers the search first and the prose second. Constraints beat personas, vocabulary beats instructions, and the citation panel matters more than the answer text. Eighteen copy-paste search-shaped prompts are below.

What makes a Perplexity prompt different from a ChatGPT prompt?

In a chat model you are steering generation. In Perplexity you are steering retrieval first, and generation only afterwards, using the same block of text for both jobs.

Perplexity states this itself. Its Agent API prompt guide opens with the warning that "prompts that work well with single-shot LLMs often underperform here, because the same text shapes tool selection, search query generation, and final response together" (docs.perplexity.ai, accessed August 27, 2026).

Once you internalise that, a lot of imported habits stop making sense. A three-sentence role preamble is not free in a search product. It is text the query builder has to read past. A worked example is worse: Perplexity's Sonar prompt guide says outright that "pasting a written-out example answer can cause the search step to latch onto the example topic," and that few-shotting structure is fine while few-shotting content is not.

The consumer help centre says the same thing in gentler language: "Perplexity understands you like a search engine. Use for in-depth context or keep it simple with 2 or 3 keywords."

So the useful vocabulary in a Perplexity prompt is not "act as a senior analyst". It is the words that would actually appear on the page you are hoping to land on: the statute number, the product SKU, the drug's generic name, the fiscal quarter, the regulator's own phrasing.

What does Perplexity actually look like in August 2026?

It is no longer one search box with a focus picker. Perplexity's help centre article on model availability lists these as the current modes, alongside a note that model options depend on your plan:

Two things in that table are worth stating plainly, because guides get them wrong. Best is not a mode; Perplexity's article says so explicitly, calling it "the automatic model option inside Search". And Deep Research does not let you pick a model at all. The research-mode article says the mode "automatically selects the optimal models for your research needs, so you cannot manually choose a specific model."

Do Perplexity's Focus modes still exist?

Not in the form nearly every ranking guide describes. If you are looking for the picker with Academic, Writing, YouTube, Reddit and Wolfram|Alpha in it, that presentation is gone.

Here is what is actually published. On August 27, 2026 the entire Perplexity help centre contained exactly one article naming Focus mode, and it is scoped to organization accounts. It says: "When you sign into an organization, it's not possible to use focus mode." The control it documents instead is a Choose sources picker offering Web, Org Files, Web + Org Files, or None.

The same quiet renaming happened to Spaces. The help-centre URL .../articles/10352961-what-are-spaces still resolves, but the page it serves is titled "What are Projects?" and describes Projects throughout.

So source scoping in the app today is these four things, and none of them is a hard filter:

1. Choose sources, where an organization account picks between web and internal files.
2. Connectors and sources under the + button, plus @ mentions of premium data sources. Perplexity documents @statista, @wiley and @pitchbook as inline mentions that pull those corpora into a thread.
3. Project instructions, capped at 8,000 characters, applied to every query inside the Project.
4. Project Settings, Context, where you "add and manage web links and domains to be prioritized".

Which source controls does Perplexity actually publish?

This is where the app and the API genuinely diverge, and it is the strongest argument for moving a recurring research job onto the API even if you never build a product with it.

Two footnotes matter more than the table. First, the recency filter is a fixed vocabulary, not free text: hour, day, week, month, year. Second, and this one bites people:

Why do personas make a search prompt worse?

Because a persona is generation guidance, and in a retrieval product it competes for space in the step that decides what gets read.

Persona prompting earns its place in a chat model, where the only thing your words influence is the next token. In Perplexity, the same words are also the seed for a web query. "You are a world-class pharmacologist with 20 years of experience" contributes nothing a search engine can use, and it displaces the terms that would.

Sonar has a sharper version of this problem, and it is documented. Its prompt guide states that Sonar "runs a web search before generating its answer, and only the user message is used to drive that search. The system prompt is not visible to search." The docs then add the warning most people need: "Do not put search instructions in the system prompt. Phrases like 'search only on Wikipedia' or 'look for the latest results' have no effect."

So a beautifully engineered system prompt cannot rescue a vague user prompt here. Perplexity's own summary: "If retrieval quality matters, invest there first."

What shape does a search-shaped prompt take?

Six slots, in roughly this order. Not all six every time, but the first four earn their place on almost any research question.

[ENTITY]      Name the thing exactly as its own documents name it.
[CLAIM]       State what you want established, not what you want discussed.
[SOURCE CLASS] Say which kind of source wins a disagreement.
[WINDOW]      Give a date range or a recency bound.
[SHAPE]       Table, list of N, one paragraph per item. Say the number.
[UNCERTAINTY] Require an explicit list of what could not be confirmed.

That last slot is the one almost nobody writes, and it is the highest-value line in the whole prompt. Perplexity's own hallucination guidance for both APIs is to hand the model an explicit exit: "If searches do not return relevant results after trying alternative phrasings, say so explicitly rather than providing speculative information," plus a second rule requiring disclosure of near-misses such as "a different year, a parent company, or a subsidiary."

You can paste those two sentences into the app as ordinary prose, or into a Project's instructions so they apply to every thread. Here is the generic template:

Find [CLAIM] about [ENTITY, named precisely].

Prefer [SOURCE CLASS: primary regulator filings / peer-reviewed papers /
the vendor's own documentation] over secondary coverage. Where a secondary
source is the only option, label it as such.

Restrict to material published between [DATE] and [DATE].

Return a table with one row per finding: claim, source domain, publication
date, and a one-line quote supporting it. No more than [N] rows.

Before the table, list anything you could not confirm from a source, and
anything you found that is related but does not match — a different year,
a parent company, a similar product. Do not fill gaps from memory.

The same question, asked badly and asked well

The bad versions below are not strawmen. They are the shapes that work fine in ChatGPT and quietly degrade in a search product.

BAD   Act as an expert energy analyst. Tell me about heat pumps.
GOOD  Compare energy efficiency ratings of residential air-source heat pumps
      against gas furnaces, using manufacturer specification sheets and
      government efficiency programmes published since January 2025. Table:
      system, rated efficiency, source domain, publication date.

BAD   Here's an example of the answer I want: "Company X raised $40M in
      Series B led by Sequoia in March 2024…" Now do the same for fintech.
GOOD  List the five largest disclosed Series B rounds in European fintech
      announced between 2026-01-01 and 2026-06-30. One row each: company,
      amount, lead investor, announcement date, source domain.

BAD   What's the latest on the EU AI Act?
GOOD  What obligations under the EU AI Act take effect in 2026, and on which
      dates? Prefer the official EU text and national implementing guidance
      over law-firm summaries. Flag any date where sources disagree.

BAD   Summarise the research on creatine and cognition.
GOOD  Summarise peer-reviewed human trials on creatine supplementation and
      cognitive performance. One paragraph per trial: sample size, dose,
      duration, primary outcome, journal. List trials you could not verify.

BAD   Is [Product] any good? Give me pros and cons.
GOOD  What do the manufacturer's own specification page and the last two
      firmware release notes say about [Product]'s battery life and charging
      time? Quote each figure with its source. Separate manufacturer claims
      from third-party measurements.

The pattern across all five: the good version contains words that would plausibly appear on the target page, a bound on time, a named source class, and a stated output shape. The bad version contains instructions about the assistant.

Do Perplexity's citations mean the answer is right?

No, and Perplexity does not claim they do. Citations tell you what was read. They do not tell you that what was read supports the sentence they are attached to.

Perplexity now labels some cited domains with a shield icon reading Government, Academic or Trusted. Its help centre article on those labels is unusually candid about their limits: "A label describes the website as a whole, not any single article or claim on it." And: "Perplexity aims for accuracy, but reviewing the original sources yourself remains the best way to build confidence in an answer." Most domains carry no label at all, which the same article says is not a negative judgment.

The independent evidence points the same way. The Tow Center for Digital Journalism at Columbia tested eight generative search tools in March 2025, running 1,600 queries that asked each tool to identify the headline, publisher, date and URL of an article from a pasted excerpt. Collectively the tools answered more than 60 percent of queries incorrectly. Perplexity was the better of the two per-tool figures the article names and still got 37 percent wrong. The study also found that Perplexity Pro cited syndicated republications of Texas Tribune articles for three of ten queries despite Perplexity having a partnership with that publisher.

That study is now around eighteen months old and every tool in it has shipped new models since, so treat the figure as a calibration exercise rather than a current benchmark. The behavioural finding has aged better than the numbers: premium tiers produced more confidently wrong answers than free ones, because they declined less often. Confidence is not evidence, which is the same trap that makes ChatGPT invent things look plausible.

The practical consequence for prompting is small and concrete. If you are building anything on the API, read source URLs from the structured search_results items in the response payload, not from the prose. Perplexity's docs are blunt about why: the model "is prone to mistyping or paraphrasing" URLs it writes into the answer text. In the app, the equivalent move is to open the sources panel rather than trusting the bracketed markers, and to click through on anything you intend to quote. That discipline is the same one that makes summarising papers without losing citations survive review.

Where is Perplexity the wrong tool?

Four cases, stated plainly, because a guide that says "use it for everything" is not a guide.

Anything that needs deterministic output. Perplexity's help centre states the product "doesn't offer manual control over modeling parameters like temperature or top-p" and directs developers to the API. If you need reproducibility, you need the API or a different tool.

Long-form writing in your own voice. Retrieval keeps pulling the register towards whatever it just read. A chat model with a good style brief will beat it, and so will prompting Grok when what you want is live commentary rather than sourced research.

Closed-corpus work over documents you already have. If the answer is inside your own PDFs and the open web would only add noise, a source-bounded tool is the better fit; NotebookLM's source-based prompting exists for exactly that job.

Anything where the web is the wrong corpus. Debugging your own codebase, reasoning over a contract you pasted in, or working through a problem that has no published answer. Search does not help, and the retrieval step will find something regardless.

18 search-shaped prompts you can paste today

Grouped by job. Every one of them is constraint-first; none contains a persona.

Competitive and market research

1. List the pricing tiers [COMPANY] publishes on its own pricing page as of
   today. One row per tier: name, monthly price, annual price, the three
   headline limits. Quote the page. If a figure is not on the page, write
   "not published" rather than estimating.

2. What has [COMPANY] announced in the last 90 days? Prefer the company's
   own newsroom, blog and filings over press aggregation. One line per
   announcement with date and source domain. Nothing older than 90 days.

3. Which companies does [ANALYST FIRM] name in its most recent [CATEGORY]
   report, and what is the report's publication date? If you cannot reach
   the report itself, say so and name what you did read instead.

4. Find every public statement by [COMPANY] about [SPECIFIC FEATURE].
   Separate documentation from marketing pages from conference talks.
   Flag any place where two of its own pages disagree.

Regulatory and compliance

5. What does [REGULATION] require of [ROLE], and from which date? Prefer the
   official text and the regulator's own guidance over commentary. Give the
   article or section number for each requirement.

6. Has [REGULATOR] issued any guidance on [TOPIC] in the past 12 months?
   Cite the document title, date and URL for each. If nothing was issued,
   say nothing was found rather than describing older material.

7. Compare how [JURISDICTION A] and [JURISDICTION B] define [TERM] in their
   respective statutes. Quote both definitions verbatim. Note explicitly
   where the two texts are not directly comparable.

Academic and technical

8. Summarise peer-reviewed studies on [INTERVENTION] and [OUTCOME] published
   since [YEAR]. One paragraph per study: sample, method, effect size,
   journal. Exclude preprints, or label them clearly if included.
   (Note: in the Sonar API, use search_mode="academic" and put the year
   range in this text, because date filters are ignored in academic mode.)

9. What does [VENDOR]'s own documentation say about [PARAMETER]? Quote the
   sentence and give the doc URL. If two of its pages contradict each other,
   quote both rather than resolving it.

10. Which papers cite [PAPER] and disagree with its conclusion? For each,
    name the point of disagreement in one sentence.

11. What is the current version of [SOFTWARE], and what changed in the last
    two releases? Use the project's own changelog or release notes. Give the
    release dates.

News and time-bounded

12. What happened with [EVENT] between [DATE] and [DATE]? Chronological,
    one line per development, each with a source domain and a date. Mark
    anything reported by only one outlet.

13. What is the current status of [ONGOING SITUATION] as of today? Lead with
    the single most recent confirmed fact and its date. Then list what is
    reported but unconfirmed, clearly separated.

14. Which outlets have reported [CLAIM], and does any primary source support
    it? If the claim traces back to a single original report, say so and
    name it.

Product and buying decisions

15. Compare [PRODUCT A] and [PRODUCT B] on [THREE SPECIFIC ATTRIBUTES],
    using each manufacturer's own specification pages. Table. Where a
    manufacturer does not publish a figure, write "not published".

16. What do the last six months of user reports say about [KNOWN ISSUE] with
    [PRODUCT]? Separate the vendor's acknowledgement, if any, from user
    reports. Do not treat forum consensus as a confirmed fact.

17. What is [PRODUCT]'s published warranty and return policy in [COUNTRY]?
    Quote the policy page. Note the retrieval date, since these change.

18. Which retailers currently list [PRODUCT] in [COUNTRY], and at what
    price? Give the retailer domain and the date you saw the price. Flag
    listings that appear to be resellers rather than authorised sellers.

What changes when you prompt Perplexity through the API?

Enough that it is worth 200 words even if you never write code, because the API is where Perplexity documents what it actually believes about prompting.

The first thing to know is that the API you have read about has moved. Perplexity's docs now carry a notice on every Sonar page: "Sonar Chat Completions is now Agent API. Sonar will be supported until September 27, 2026." The Agent API is the recommendation for both new projects and existing Sonar integrations.

The Agent API splits your prompt in two. instructions carries role, tone, formatting and grounding rules, and is re-read on every turn of the agent loop. input carries the question, and it seeds the first search query. The docs warn that instructions bloat "compounds across tool calls", and that setting instructions alongside a preset replaces the preset's system prompt rather than appending to it.

Then the rule that generalises back to the app: use parameters, not prose, for hard constraints. Perplexity's own words are that parameters "are applied by the search backend on every call, while prose-based filters are interpreted by the model and may not carry through every turn of the loop." Its documented anti-pattern is writing "Using only Wikipedia as a source" into the prompt instead of setting search_domain_filter.

Honest note on where a prompt tool fits here

We build a prompt generator. We are not a search engine, we do not index the web, and we are not competing with Perplexity for anything.

What a prompt tool can do on this page is narrow and real. Search-shaped prompts are structural: entity, source class, window, shape, uncertainty clause. That structure is the same on every question and different only in its variables, which makes it a good fit for a saved template with fields you swap. Our extension supports Perplexity natively and biases its output towards search-shaped structure and source awareness rather than the persona-first shape it uses for a chat model.

What it cannot do is more important. It cannot make a citation accurate, it cannot reach a page Perplexity's crawler never saw, and it cannot substitute for opening the source. No prompt tool can. If you take one habit from this post, take the uncertainty clause: ask what could not be confirmed, every time, and read the answer to that question before you read anything else.