Perplexity

Where does your data go?

Perplexity's cloud — with the most tangled privacy posture of the tools reviewed so far:

• Consumer plans train on your data by default. Free, Pro, and Max all have "AI data retention" enabled unless you opt out in Account → Preferences. Opting out is forward-looking only: already-collected data can't be recalled from training.
• The ads story is the real problem. A pending class-action lawsuit alleges Perplexity embedded Meta and Google ad trackers that shared full conversation transcripts — including tax, family-finance, and investment questions — alongside email addresses, and that "Incognito Mode" did nothing to stop it. Perplexity denies the claims; the case is live.
• Enterprise plans are genuinely different: never trained on, 7-day file retention, Zero Data Retention contracts with OpenAI and Anthropic. But when an enterprise subscription lapses, the account reverts to consumer rules — training back on by default.
• No ads inside answers yet, and the privacy notice states Perplexity doesn't sell personal data or send queries to advertisers. The lawsuit alleges the tracker layer works around that promise.

What happens when it's wrong?

Perplexity's entire pitch is "answers with citations" — so the failure that matters is whether the citation holds:

• The signature finding: a September 2026 audit fetched 1,826 citations Perplexity attached to sentences containing figures. 34.7% pointed at a page that either wouldn't open or didn't contain a single number from the sentence they supported. Per-claim, 14.4% had no working support at all.
• Dead links are the small part — just 1.3%. The real failures: pages behind paywalls (1 in 6 citations, uncheckable by an ordinary reader) and readable pages that simply don't say the thing being cited. The model reached for the nearest plausible source and attached it.
• Best-in-class, still wrong. Columbia's Tow Center tested eight AI search engines: Perplexity's free tier was the best performer — with a 37% error rate. The paid Pro tier scored worse. A medical-literature study found Perplexity hallucinated references at the highest rate of three leading models.
• A quarter of its sources have no archive copy. The citation layer leans heavily on directory and lead-gen pages built to rank, not to last — 25% have never been captured by the Wayback Machine even once.

What's it actually bad at?

• The citation layer itself. The core selling point fails verification about a third of the time on numeric claims. No other mainstream tool sells provenance this hard — or misses it this often.
• Six tiers of naming soup. Free, Education Pro, Pro, Max, Enterprise Pro, Enterprise Max — different models, limits, and privacy rules at each level, blurred on the pricing pages.
• Model gating. The strongest reasoning models (Claude Opus 4.6, o3-pro) are reserved for the $200 Max tier — not available on Pro at any price.
• The publisher war. The NYT, Dow Jones, and others are suing over content scraping; a court already ordered its Comet browser to stop scraping Amazon. The content supply that makes Perplexity work is contested.
• Legal clouds on two fronts — the privacy class action and the publisher suits — both live as of this review.

What does it cost really?

The traps: six tiers total (Free, Education Pro, Pro, Max, Enterprise Pro, Enterprise Max) with different privacy rules at each — Enterprise Pro at $40/seat is where the no-training guarantee lives. Free-tier limits reset on a rolling 24-hour window, so quota math is unpredictable. And Education Pro ($10/mo, verified students/faculty) is genuinely half price for the same Pro features — rarely mentioned on the pricing page.

What does it replace — and what do you still do?

It replaces the first hour of web research: market scans, competitor lookups, supplier checks, "what's the current rate for X." The citation format forces a verification habit pure chatbots don't. You still own: reading the actual sources, judging source quality — directories and SEO pages dominate the citation layer — and every conclusion you draw. It finds and summarizes; it does not verify.

What the ecosystem offers

• Projects (Jul 2026) — hubs for ongoing work with a shared file system and memory. The assistant ("Computer") reads, edits, and saves files across sessions, and Brain — the memory system — reviews your project between tasks so the next one starts with full context. Instructions set guardrails; templates handle repeatable work. Connects to 400+ tools and binds Slack or Teams channels. Available on every plan.
• Privacy inside Projects is genuinely careful: teammates see shared files, but your personal connectors and local files stay yours; every tool connection runs under your own credentials; forking a thread doesn't carry over anyone's private items.
• Comet Enterprise (Aug 2026) — Perplexity's AI browser as an IT-managed product: install it across a whole company silently, set what the assistant is allowed to do (full control / read-only / nothing) with per-site exceptions, and get audit logs. Includes an optional cybersecurity layer from CrowdStrike. $40/seat (Pro), $325/seat (Max). Early customers: Fortune, AWS, Bessemer.

If I stop paying, what do I lose?

• Export: available — Spaces (collections of threads and sources) can be exported; individual answers copy or share as links.
• The downgrade cliff is the sharpest here: lapse from Enterprise to consumer and your data protection disappears — training defaults back on, and you must manually opt out again.
• Enterprise files purge at 7 days; consumer history persists under the standard policy.
• Cancellation: self-serve, clean — refund window is 24 hours (monthly) or 48 hours (annual) from purchase; after that, you keep access through the billing period.
• The lock-in reality: Spaces and Threads accumulate research context that doesn't migrate — but the real asset, the sources themselves, lives on the open web.

Who is this NOT for?

Anyone researching sensitive subjects on a consumer plan — the tracker allegations, the training defaults, and the enterprise-only privacy wall make it the wrong tool for confidential work. Also not for anyone who wants "right more often than wrong" on facts: the best-in-class citation error rate is still 37%. And not for heavy-model users on a budget — the frontier models sit behind $200/month.