Perplexity's citations teardown
Updated June 15, 2026
Perplexity’s trust model is evidence in the reading flow. Sourcing stays tied to claims as you scan, verification depth matches how skeptical you are, and you can challenge bad citations without leaving the answer.
Feedback even includes Wrong sources: citation quality is a first-class failure mode, not an afterthought.
Domain chips in the prose

What works
- Chips show publisher domain first, users recognize ESPN vs FanDuel before clicking.
- +N communicates multiple sources for one claim without cluttering the sentence.
- Chips sit inline so skeptical readers spot sourcing while scanning bullets.
What we would push on
- Domain abbreviations (sportsbook.fanduel) may confuse non-US users. Tooltip with full title helps.
Takeaway
Use publisher-first inline chips with a count, not footnotes at the bottom of the page.
Pattern: Citations
Steps show what was searched

What works
- Research steps make retrieval visible, users see the system looked before asserting favorites.
- Expand/collapse keeps auditors happy without wall-of-log noise for casual readers.
- Step wording mirrors user mental model (Checking predictions) not API names.
What we would push on
- Steps do not list domains fetched, auditors still jump to Links or chips for specifics.
Takeaway
Trust starts with showing retrieval happened. Then show which sources matter.
Pattern: Progress Steps
Pattern: Citations
Sources row on the answer block

What works
- One glance shows how many sources backed the answer and which domains dominated.
- Sources row lives in the action bar, same elevation as copy and share, not footer debris.
- Count sets expectation before users open the full Links tab.
What we would push on
- Favicon stack truncates; users cannot see all ten without clicking through.
Takeaway
Summarize source breadth on the answer chrome, inline chips for local, row for global.
Pattern: Source Browser
Pattern: Citations
Links tab for full audit

What works
- Links tab is a dedicated audit view, same query, every reference as a scannable card.
- Each card shows title, URL, snippet, and preview image before opening a new tab.
- Tab sits beside Answer and Images so audit does not require leaving the result.
What we would push on
- Users must discover the Links tab, no inline jump from a chip number to card N in list.
Takeaway
Offer a full source browser as a peer tab, not only hover cards on chips.
Pattern: Source Browser
Pattern: Citations
Feedback includes Wrong sources

What works
- Wrong sources separates citation failures from generic Inaccurate, routes triage correctly.
- Sits beside Out of date for stale odds/news, common on time-sensitive search.
- Optional chips keep the modal lightweight.
What we would push on
- No follow-up to mark which source was wrong, product learns less than per-citation flagging.
Takeaway
Search products need source-specific negative feedback, not only “bad answer.”
Check sources on a selection

What works
- Users can challenge one paragraph , “Why this is still uncertain” , without re-running the whole query.
- Check sources pairs with Add to follow-up: verify vs iterate on the same selection.
- Selection stays highlighted until the user acts.
What we would push on
- Discoverability is select-first only. Skeptical users may never find Check sources.
Takeaway
Let users audit a claim span directly, trust UX is local, not page-level only.
Pattern: Citations
Pattern: Response Refinement
Chip opens a source popover

What works
- Chip click surfaces source detail in place, favicon, domain, headline, and excerpt before opening a new tab.
- 1/2 navigation when a claim cites multiple sources (+1), users browse stacked evidence without the full sidebar.
- Popover anchors to the chip so the connection between claim and source stays obvious.
What we would push on
- Popover competes with sidebar and Links tab, three paths to sources may confuse first-time users.
Takeaway
Let inline chips expand into quick-verify popovers with pagination when +N > 0.
Pattern: Citations
Pattern: Citation Tooltips
How citations fit with output
Perplexity layers trust: domain chips on claims → research steps → sources summary row → chip popovers → sources sidebar → Links tab audit → Check sources on selections → Wrong sources in feedback.
Output teardown covers tabs, follow-ups, export, and rewrite. This guide is the evidence stack those answers sit on.
Steal the layered model if your product makes factual claims: inline chips, summary row, popover verify, sidebar or tab for depth.
Steal this
- Publisher domain +N chips inline on claims
- Visible research steps before the answer
- Favicon stack and source count on the answer bar
- Links tab as a full audit view beside Answer
- Sources sidebar opened from the 10 sources row
- Citation chip popover with 1/N source navigation
- Wrong sources in negative feedback taxonomy
- Check sources on text selections
Skip this
- Links only at the bottom of long answers
- Citation markers with no domain name visible
- Feedback that cannot distinguish bad sources from bad writing
- Forcing users to open ten tabs to verify one claim
How others trust, privacy & settings
Same job, different product bets, and what each tradeoff reveals.
Original gallery pages: Citations & Trust
