Perplexity review: the AI search that answers with links to sources

The mechanics are straightforward. You write a question in plain language. The service turns it into a set of search queries, visits the pages it finds, pulls out relevant fragments, and assembles an answer from them. Source cards sit next to the answer, numbered footnotes sit inside the text. From there you can ask follow-ups in the same thread: the context carries over, and each new answer is backed by links again.

That binding is the whole difference. A language model in a chat produces a claim from a blend of training data and whatever it pulled off the web, and the reader cannot separate the two. In Perplexity every claim leads to a page you open and read with your own eyes. The answer works as a table of contents to the primary sources, and it does not spare you from reading them.

Neighbouring tools solve adjacent jobs, and mixing them up costs time:

Perplexity and NotebookLM together cover most factual work: the first gathers material from outside, the second answers from your own documents. I reviewed the second one separately – the NotebookLM review.

Several modes of different depth live inside one search box. The choice affects response time and how many pages get read.

Fast search

The default mode: a few seconds, a short answer, a handful of sources. Good for factual questions like "what is the current central bank rate" or "when do the new escrow rules take effect". There are no limits on it even on the free plan.

Pro Search

The service splits your question into sub-queries, visits noticeably more pages, and sometimes asks a clarifying question first. The answer comes out longer with a wider source list. This is the working mode for questions like "how has demand for business-class apartments shifted over the past year".

Research

The deep mode, sometimes called Deep Research. The query runs for several minutes: the service builds a plan, walks through dozens of sources, fills the gaps it finds, and returns a structured report with sections and a full link list. The result exports to a document or turns into a deck or a spreadsheet. In my experience it is a draft of an analytical memo that still needs editing by hand, but starting from it beats starting from a blank page.

Labs

A separate mode for finished artefacts: spreadsheets, dashboards, small web pages, and simple apps built on the collected data. Available on paid plans. In marketing it earns its keep when research output has to reach the team as a comparison table rather than a wall of text.

Spaces (previously Collections) are workspaces for a specific project. Inside a Space you set a persistent instruction ("you are a residential real estate analyst, always state the publication date of the source"), upload your own files, and optionally restrict the search to a chosen set of sites. Every query inside the Space then carries that context, so you stop re-explaining the task each time. Spaces can be shared with colleagues.

The model picker is available on paid plans: Perplexity's own Sonar runs under the hood, and the switcher offers models from OpenAI, Anthropic, and Google. Perplexity refreshes the list as vendors ship releases, so naming exact versions in an article is pointless – check the switcher in settings. In practice the gap between frontier models shows up on long analytical queries and stays invisible on factual ones.

Comet is Perplexity's Chromium-based browser for macOS, Windows, iOS, and Android. It launched in 2025 as an expensive add-on for top-tier subscribers and went free across every platform during 2026. Inside it is familiar Chromium with tab and extension import, plus a sidebar assistant that sees the open page, answers questions about it, and compares the contents of several tabs. Agentic browsing, where the assistant walks through sites and completes a sequence of actions on its own, is unavailable on the free plan and opens up on Pro and Max with limits.

There is also Comet Plus, a paid add-on at roughly 5 dollars a month that unlocks material from partner publishers and shares revenue with them. Worth it if you regularly hit business-media paywalls.

Mobile apps for iOS and Android give you the same sourced search, a voice mode, and camera input. My phone scenario is narrow: check a fact, or pull together background on a company on the way to a meeting.

The Sonar API is for teams wiring search into their own processes. The model family covers different depths: a lightweight Sonar for quick factual lookups, Sonar Pro for complex queries, Sonar Reasoning Pro with step-by-step reasoning, and Sonar Deep Research for full reports. Responses come back with the list of source URLs, which is what makes the API interesting: mention monitoring, automated market news collection, enriching CRM records. Direct chat completions calls are now legacy; the current path is the Agent API, which combines the model, web grounding, and tools in a single request. On corporate plans API access is billed separately from seats.

What follows are structures that map onto ordinary marketing work. No borrowed case studies with numbers, just setups you can repeat tomorrow.

Market and competitor analysis

A Research query: who the main players in the segment are, how prices and volumes shifted, what restrictions the regulator introduced over the past year. Out comes a report broken into sections with a source list. Then you open 10-15 links from that list, discard the rewrites and advertorials, and read the rest properly. The saving is not in the reading, it is in the finding – which normally eats half a day.

Meeting prep

Half an hour before a call with a developer or contractor: what the company does, which projects it delivered, what trade media wrote about it recently, whether there were lawsuits or missed deadlines. All of it linked, so anything contentious can be verified before you ask an awkward question out loud. At Matveo this pre-read is a standard step before a first call with a new partner: it takes 20 minutes and removes half the awkward moments.

Checking a trend or a number

A colleague arrives with a claim: "the share of mortgage deals has fallen to X percent". A Perplexity query with the period and region spelled out shows where the number came from, which month it refers to, and whether other sources contradict it. Often the number turns out to be correct but for a different segment, or six months older than assumed.

Gathering sources for an article

Before writing a piece I build a link base: research, industry reports, commentary from subject-matter experts. Perplexity produces them in bulk within a couple of queries, and a Space for the article topic holds the context while the material is in progress. Writing the text off the answer itself is a bad idea – treat it as navigation to the primary sources.

There is a flip side. Perplexity displays sources prominently and clickably, so a share of users leaves the answer for the site, and those visits are usually higher quality than average because the person already knows why they are opening the page. The site is crawled by PerplexityBot, and getting into the source list depends on things long familiar from SEO: the page answers a specific question directly, its author and date are visible, and it backs its own claims with links.

At the same time an on-the-spot answer removes part of your visits: the user got what they needed in the results and went nowhere. This channel measures poorly, and it is worth treating as a source of awareness rather than a replacement for organic search. What to do with pages specifically so they get cited by AI search, I covered separately – GEO optimization for AI search.

The free plan is enough to work out whether this is your tool. Key numbers at the time of publication (August 2026):

Sonar API access on corporate plans is bought separately and is not bundled into the seat price. The exact daily Pro Search allowance on the free plan has changed several times over the past year, and reviews quote anywhere from three to five queries a day – trust the official pricing page over articles, including this one.

Now the honest part. Here is what you will run into:

• The answer is never better than the search results. If a topic is covered online only by rewritten press releases, you get a tidy retelling of rewritten press releases. On narrow niches and non-English markets this shows up fast.
• Deep analysis still requires reading. A Research report is a map of the terrain plus a link list. Conclusions you will put budget behind should come after reading the primary sources, not after reading a summary.
• Hallucinations are rare but real. The links are genuine; the retelling sometimes shifts emphasis, and a number lands in the answer with the wrong period or without an important caveat. Verify anything headed for a deck.
• Source freshness needs manual checking. The service happily picks up three-year-old articles when they rank well. I ask for publication dates directly in the Space instruction.
• Team use needs a decision on data. Before uploading client files into Spaces, check what your plan says about retention and access – on Enterprise tiers these are configurable, on consumer plans they are not.