Perplexity's output teardown

What works

• Long answers use clear hierarchy , summary, “Current favorites” list, “Best single guess” , without leaving the thread.
• Answer / Links / Images tabs let users switch modality without re-querying.
• Completed N steps label signals research happened before the prose appears.

What we would push on

• Tabs compete with the answer for attention on first load. Default tab must carry the full story for skimmers.

Takeaway

Treat search output as a small report: structured prose plus parallel Links and Images views.

What works

• Steps collapse by default so the answer stays primary; expand reveals concrete actions taken.
• Plain-language step names (Searching the web) beat opaque tool names for trust.
• Step count in the header sets expectation for latency on complex queries.

What we would push on

• Collapsed steps are easy to miss. High-stakes queries may need a stronger “we searched N sources” cue inline.

Takeaway

Show what the system did before answering, collapsed, expandable, in human terms.

What works

• Follow-ups are contextual , squad leaders, France form, betting odds , not generic “tell me more.”
• Each chip is one tap into the composer; users iterate without rephrasing the original question.
• Sources row (favicons + count) sits beside share/copy/rewrite actions on the same answer block.

What we would push on

• Long follow-up lists push the composer down. Cap visible chips or collapse after three.

Takeaway

Pair a complete first answer with specific next questions, refinement without blank follow-up boxes.

What works

• Images tab reuses the same query , brackets, experts, trophy photos , without a separate image search.
• Domain label under each thumbnail ties visuals back to provenance.
• Grid layout supports scan-first research; click-through opens the source image.

What we would push on

• Image quality varies by query. No inline caption explains why each image was selected.

Takeaway

When answers are visual (sports, products, places), give a dedicated Images view beside prose.

What works

• Export sits in the per-answer action row , share, download, copy, rewrite , not buried in settings.
• Three formats cover slide decks (PDF), docs (DOCX), and dev workflows (Markdown).
• Download targets the current answer, not the whole thread.

What we would push on

• Export may strip citation links depending on format, verify sources survive in PDF/DOCX.

Takeaway

Let users ship the answer out of chat in the format their workflow expects.

What works

• Rewrite regenerates the same question with a different depth or model, not a new thread.
• Modes are labeled by outcome (Deep research, Learn step by step) with lock icons for Pro tiers.
• Current mode (Search) shows a checkmark so users know what produced this answer.

What we would push on

• Model names (Sonar, GPT-5.4) mean little to casual users. Pair with one-line capability hints.

Takeaway

Non-destructive rewrite with explicit mode/model choice beats “try again” with no knobs.

What works

• Thumbs up opens structured praise, Good sources is explicit for a search product.
• All chips optional; users can submit without categorizing.
• Thank-you toast confirms receipt without blocking the thread.

What we would push on

• Modal interrupts flow for simple satisfaction. One-tap up with optional expand might feel faster.

Takeaway

On search answers, include source-quality in positive feedback taxonomy.

What works

• Wrong sources is first-class, critical for a product whose value is citation quality.
• Length chips (Too long / Too short) separate formatting issues from factual ones.
• Optional modal keeps downvote fast while still capturing structured signal.

What we would push on

• No way to flag a specific citation inline from the feedback modal. Users must remember which source failed.

Takeaway

Negative feedback on search products needs a source-specific issue type.

What works

• Users can select prose inside the answer, not just react to the whole block.
• Add to follow-up scopes the next query; Check sources jumps to evidence for that claim.
• Split actions match intent: iterate vs verify.

What we would push on

• Selection menu is discoverable only after highlighting. No persistent hint for new users.

Takeaway

Pair refinement (follow-up) with verification (check sources) on the same selection.

Perplexity’s output model is report-first: tabbed Answer / Links / Images, visible research steps, follow-up chips, export, rewrite, feedback, and selection actions.

Inline citation chips, the sources summary row, Links tab audit, and Check sources on selection live in the citations teardown under Trust: the evidence layer this output sits on.

Compared with ChatGPT’s regenerate and writing blocks, Perplexity optimizes for search iteration: new focus, rewrite mode, or a follow-up chip rather than editing one bubble in place.