Perplexity is an answer engine: you ask a question, it searches the live web, and it writes an answer with numbered citations attached to the claims. The citations are the product. The paragraph above them is a summary written by a language model that can misread a source as easily as any other model - so the skill worth learning is not prompting, it's clicking through. This guide covers what each mode actually does, what the plans cost as of August 2026, prompts that keep answers close to their sources, and the specific ways Perplexity gets things wrong.
What is Perplexity AI, and how is it different from Google or ChatGPT?
Google hands you ten links and leaves the reading to you. A plain chatbot hands you a paragraph and leaves the sourcing to you. Perplexity does both halves at once: it runs live web searches, then writes a direct answer with numbered citations inline, so every claim points at the page it came from.
That is a genuine improvement over an unsourced chatbot, and it is also the trap. A confident three-sentence paragraph with four links under it feels more verified than it is. Nobody checks four links. The links are there so you can check - the value only exists if you do.
What do Perplexity's modes actually do?
The mode selector sits in the search box, and the names change more often than the behaviour. As of August 2026, Perplexity's help centre lists four search modes plus a build mode:
- Best - the default. It picks a model for you based on the query and runs without quota limits. Use it for anything you'd otherwise type into Google.
- Pro Search - multi-step: it breaks the question apart, runs several searches, and returns more citations per answer. Pro subscribers get 10x the citations per answer that free users do.
- Reasoning Search - for questions that need working-out rather than looking-up, on models built for it (Sonnet 4.6 Thinking, Grok 4; Max adds o3-Pro and Claude 4.5 Opus).
- Research - runs dozens of searches across many sources and returns a structured report with a long citation list. Minutes, not seconds. This is what you use when you need a paper trail, not an answer.
- Create files and apps - takes the research and produces the artefact: a report, a spreadsheet, a dashboard, a small web app.
Separately, Focus in the search box restricts where it looks - Web, Academic, Social, Finance, or your uploaded Files. Academic pushes it toward journals and preprints instead of blogs summarising journals, which is the single highest-leverage setting for research work. On Enterprise, Focus is replaced by Choose sources, which lets you point at internal org files instead of, or alongside, the web.
Free users can also select a preferred model per query; Pro and Max widen the list. Model names churn every few months, so treat any specific one you read here as a snapshot, not a spec.
Before you open Perplexity, Aiden turns a fuzzy topic into the precise question and source criteria that get a citable answer on the first try. Install it in Claude, or paste SKILL.md into ChatGPT's custom instructions.
View Aiden - Research Assistant AI Skill →What does Perplexity cost in 2026?
Checked on Perplexity's own pricing page in August 2026 - confirm current pricing with the vendor before you buy, because these tiers move:
- Free - $0/month. Citations in every answer, basic models, limited daily use.
- Pro - $20/month. 10x citations per answer, extended Research, model selection, larger file uploads, app connectors, image generation, limited video.
- Max - $200/month. The top of the consumer range: 10,000 Computer credits a month, frontier models, the highest usage and memory limits, early feature access.
Enterprise is priced per seat and quoted separately. For ordinary research - a few dozen questions a week, occasional deep reports - the free tier does more than most people expect, and Pro is the sensible upgrade only once you hit its daily ceilings often enough to notice. Max is priced for people running it as infrastructure, not for reading.
How do you ask so the answer stays close to the sources?
Perplexity's failure mode is drift: the summary says something slightly stronger, newer or tidier than any of the pages under it. These prompts pull it back. Copy them and replace the bracketed parts.
1. Bind claims to citations
[YOUR QUESTION]. Use only sources published after [DATE]. Put the citation immediately after each factual claim, not at the end of the paragraph. If sources disagree, show both figures and say who reported which. If you cannot find a source for something, write "not found" rather than estimating.
2. Force the primary source
Find the original [REPORT / FILING / DATASET / PAPER] behind this claim: [CLAIM]. Give me the publisher, the publication date, and a link to the document itself - not to an article about it. Quote the sentence or table row that contains the number.
3. Ask for the other side
What is the strongest published argument against [CLAIM]? Cite sources that disagree with it. Do not balance the answer - I want the opposing case and who is making it.
4. Brief for Research mode
Topic: [TOPIC]. Audience: [WHO WILL READ THIS]. Decision it feeds: [DECISION]. Cover: [SUB-QUESTION 1], [SUB-QUESTION 2], [SUB-QUESTION 3]. Prefer sources from [DATE RANGE] and from [SOURCE TYPES]. Flag every figure you could find in only one place. End with a list of what you could not find.
5. Verification pass on an answer you already have
List every numeric or dated claim in your answer above as a table: claim | source | source publication date | does the number appear verbatim in the source, or was it inferred? Mark anything inferred.
That last one is the one people skip and shouldn't. It routinely surfaces claims that no single cited page actually makes.
When Research mode hands back twelve pages and eighty citations, paste it into this pack and get the method, the findings and a quality assessment pulled out - instead of rereading the whole thing.
View AI Research Summary Generator →How do you check citations without rereading everything?
You don't check all of them. You check the ones that carry weight. Three rules:
- Click through on anything you will repeat. If a number is going into a deck, an email, a pitch or a published post, open the source and find the number with Ctrl+F. If it isn't there in that form, it isn't your number.
- Read the date on the source page, not in the summary. Old pages rank well. An answer to "the latest figures for X" is often built on a page from three years ago that happens to have good links.
- Ask whether the source originated the claim or repeated it. A news article about a study is not the study. A vendor blog citing "industry research" is usually citing its own marketing. Follow the chain one more hop; it often ends nowhere.
Where does Perplexity fail?
- Citations that don't support the claim. The most common defect: a link attached to a page that discusses the topic but never states the specific figure. It looks sourced and isn't.
- Numbers stitched together. A sentence combining a figure from one source with a timeframe from another produces a claim neither source makes.
- SEO surface over primary depth. It reads what it can crawl. Paywalled journals, PDFs behind logins, filings, and paid databases are often missed while a listicle summarising them ranks and gets cited.
- Consensus mistaken for truth. It reflects what is written most often. On contested or fast-moving topics, the majority write-up and the correct answer are frequently different things.
- Thin coverage on private and local reality. Anything not published - internal benchmarks, small markets, non-English regional data, prices that live behind a sales call - comes back vague or wrong.
- Confident tone regardless of evidence quality. The writing is equally assured whether it read three peer-reviewed papers or three affiliate blogs. The prose gives you no signal about the strength of the underlying sources.
None of that makes it a bad tool. It makes it a tool whose output is a starting draft with receipts attached, not a finished fact.
What do you do with the research once you have it?
Finding sources is the cheap half. The value is in the decision it feeds. For content and SEO work, Walter, the AI Keyword Research Agent, takes a topic and returns a ranked keyword plan, so a folder of Perplexity threads turns into a publishing order. Working across a whole site? Pair it with the SEO agents, or browse every role in the agents library. For a wider comparison of assistants, see our guide to the best AI agents in 2026. Want to test output quality before spending anything? The free AI generators are a no-risk start.
In summary:
Perplexity's citations are worth more than its summaries - prompt so the two stay close, and click through on anything you'll repeat. To sharpen questions before you search, Aiden - Research Assistant AI Skill is $14.99; to compress long Research reports into a brief, the AI Research Summary Generator is $9. Both run in Claude, ChatGPT and any AI chat, no coding, 30-day money-back guarantee.