Last updated
Perplexity is usually introduced as a search engine with an AI on top, which undersells the only thing about it that actually matters. A general assistant gives you an answer you have to take on faith. Perplexity gives you an answer with the sources attached, which means you can do something no chatbot output permits: check it. Everything worth saying about the product follows from that one property, including the ways it fails.
An unsourced answer from a language model is, epistemically, a rumour. It may well be right, and you have no route from reading it to knowing it. That is fine for drafting an email and disqualifying for anything you will assert to someone else.
Perplexity retrieves live sources, synthesises an answer from them, and links each claim back to where it came from. The practical effect is that the tool stops being an oracle and becomes a research assistant: it finds and reads faster than you do, and you retain the job of deciding what is true. For anyone who has to defend a number in a meeting, that is a categorical difference rather than a feature.
It also changes the shape of the work. The right way to use it is to read the answer as a map of where to look, then open the two or three sources that carry the weight of your conclusion. People who use it well spend meaningful time in the citation panel. People who treat the summary as the deliverable have bought a chatbot with extra steps.
This is the section every other review of this tool leaves out, and it is the one that determines whether you can rely on it.
The first and most common failure is that a citation is present but does not support the sentence it is attached to. The source is real, the link works, the page is roughly on-topic, and the specific claim is either not in it or is a distorted version of something adjacent. Because a footnote reads as verification, this is far more dangerous than an obviously unsourced assertion — the apparatus of rigour is doing the opposite of its job. Any claim you intend to rely on needs the source opened, not counted.
The second is source quality. Synthesis inherits the reliability of what was retrieved, and on thin, contested, or commercially crowded topics what gets retrieved is content marketing, aggregator pages, and forum posts. The answer will be written with exactly the same confidence as one built from primary research. Notice what is in the citation list before you notice how well the paragraph reads.
The third is recency and drift. A live index means the answer to the same question can change between Tuesday and Thursday, which is a feature when the world changed and a problem when you are trying to reproduce a piece of analysis. If a finding matters, save the sources rather than the answer.
And the fourth is that summarising several sources into one paragraph can manufacture a consensus that does not exist. Disagreement between sources is information, and a synthesis that smooths it into a single confident statement has destroyed the most useful thing on the page.
Against a conventional search engine, the trade is speed for control. Perplexity is faster when your question has an answer that exists across several pages and you would otherwise assemble it yourself. A search engine is better when you know what you are looking for, when you need the primary source rather than a description of it, or when the ranking itself is the information — sometimes what you want to know is which pages exist, not what they collectively say. Most people who adopt Perplexity do not stop searching; they stop searching for the class of question that was really a research task.
Against a general assistant, the boundary is cleaner than it looks. ChatGPT and Claude can browse when asked, and that is not the same as being built around retrieval — the citation is an option in one architecture and the foundation of the other. Conversely, Perplexity is not the tool for extended creative writing, for working through a long document you supply, or for anything inside a codebase. ChatGPT vs Perplexity and Claude vs Perplexity both come down to whether the output needs provenance, and our longer write-up goes through the cases.
Do not use it as your only assistant. It is a research instrument, and asking it to be a general-purpose companion means accepting a worse version of a product that costs about the same.
Do not use it for anything where the answer must be right rather than probably right, without opening the sources. Medical, legal, financial, and regulatory questions are exactly where the citation-that-does-not-support-the-claim failure does the most damage, because the reader is least equipped to notice.
Do not use it as a citation manager. The links in an answer are evidence that something was read, not a bibliography you can paste into a paper. Academic and professional citation requires you to have read and cited the source yourself, and reviewers are increasingly good at spotting work where that did not happen.
And do not use it for questions about your own material — your documents, your data, your internal decisions. It is pointed at the open web. Something workspace-grounded like Notion AI is aimed at that problem, and a general assistant with your file uploaded is aimed at it too.
Any question where the next step is telling someone else the answer. The citation trail is what turns a plausible paragraph into something you can stand behind, provided you actually open the two or three sources doing the real work.
The deeper research mode runs a longer, multi-source investigation and returns something structured. It is at its best as a well-organised starting point for reading rather than as a finished report, which is roughly the standard you would hold a capable intern to.
Because retrieval is live, it handles the last few months better than a model relying on training data. This is the clearest everyday advantage over a general assistant and the reason many people keep it open alongside one.
Perplexity offers Free ($0, 5 Deep Research + 3 Pro Searches per day), Pro ($20/mo or $200/yr, full Sonar family + selectable GPT-5.6/Claude Fable 5/Gemini 3.5 Flash, Spaces, Pages, Labs), Max ($200/mo, adds Perplexity Computer orchestrating 19 sub-agent models), Education Pro ($10/mo for students), Enterprise Pro ($40/seat/mo), and Enterprise Max ($325/seat/mo). The Comet browser is free for everyone, with Comet Plus ($5/mo, or included with Pro/Max) unlocking premium publisher content. One caveat if you are subscribing for a specific model: the Claude release named in that picker, Fable 5, is no longer Anthropic's current one. Anthropic lists it among the legacy models that remain available, with Fable 5.1 and Opus 5 shipping above it. Third-party model pickers lag the labs by design, so check Perplexity's own page for what is selectable today rather than assuming the newest Claude is in there. The pricing note: the free tier is unusually generous for casual research, and the leap to Max is only worth it for power users who need the multi-agent Computer feature.
Provenance. A general assistant produces a fluent answer you have to trust; Perplexity produces one with sources you can open. Both can be wrong, but only one of them lets you find out from the output itself. For anything you will repeat to another person, that difference is the entire product.
No, and this is the single most important thing to understand about the tool. A citation can be real, live, and roughly on-topic while failing to support the specific sentence attached to it. Footnotes read as rigour, which makes this failure harder to catch than a plainly unsourced claim. Open the sources for anything you intend to rely on.
For research-shaped questions, largely. For navigation, for finding a specific primary source, and for cases where the set of results is itself the information, no. Most people end up using both and routing by question type without thinking about it much.
For occasional research, generally yes — the free plan includes a daily allowance of the more thorough search and research modes, which is enough to learn whether the workflow suits you. Daily researchers hit the limits. Because those allowances are adjusted periodically, check the current plan page rather than any figure quoted in an article.
On the paid tiers you can generally select between Perplexity's own models and frontier models from other providers. It matters less than you would expect, because the quality of a retrieval-based answer is dominated by what got retrieved rather than by which model wrote the paragraph. Improving your question beats changing the model.
Cite the sources, never the tool. Perplexity is a way of finding material, equivalent to a database search, and the scholarly requirement that you read what you cite is unchanged by how you found it. Passing along a citation you have not opened is a failure mode that predates AI and is now much easier to commit.
It is competent and it is not what the product is for. Extended creative writing, work on a long document you supply, and anything inside a codebase are all better served elsewhere. Buying a research tool and using it as a general assistant gets you a worse assistant at a similar price.
That it is a narrow product and its core promise degrades quietly. The narrowness is manageable — you keep a general assistant too. The quiet degradation is the real risk: weak sources and unsupported citations produce output that looks more rigorous than an unsourced answer while being no more reliable, and the only defence is the discipline of actually reading what it links to.
Full review coming soon.