What model powers Perplexity?

The answer engine that cites its sources on almost every query.

As of July 2026, Perplexity’s free default answers with Sonar — its in-house model, built on Llama and tuned for cited, retrieval-first search. Perplexity routes the default search experience to Sonar automatically; subscribers can switch to other frontier models. Discoverably reads the model from Perplexity’s own answers, so this tracks the default your customers get, not a paid arm they’d have to choose.

Current model
Sonar
Evidence
Observed
As of
Feb 2025
What we capture
perplexity.ai, the free default

Discoverably reads this model from the engine’s own answers, so it reflects what the free experience actually returns — not only what the vendor last announced.

How Perplexity picks its model

Perplexity is the most retrieval-first engine we track: it runs a live web search on essentially every query and composes the answer from what it retrieves, citing sources inline. The default search experience is served by Sonar, Perplexity’s in-house model family (publicly documented as built on Llama). Which exact Sonar variant serves a given query is not user-visible.

Paid subscribers can select other models — including frontier models from other labs — whose answers and citation behavior differ from the default. Discoverably tracks the default that free and unswitched users get, and follows it as Perplexity changes it.

Model history

Every family-level change to the model behind Perplexity, newest first, each linked to its primary source.

Sonar
Model changeObserved

Perplexity ships the new in-house Sonar model (built on Llama 3.3 70B) as the default for its search experience.

Perplexity ↗ (opens in a new tab)
  1. Sonar

    Perplexity ships the new in-house Sonar model (built on Llama 3.3 70B) as the default for its search experience.

    Perplexity ↗ (opens in a new tab)

What decides whether Perplexity names your brand

The purest retrieval-first engine — it searches the live web on essentially every query and cites what it used, so there is no meaningful parametric regime.

How answers form & what drives brand mentions

Retrieval-first. VENDOR-DOCUMENTED Perplexity searches the web in real time, identifies sources, synthesizes, and attaches numbered citations. All leverage concentrates into one question: does your content get retrieved and selected for citation?

The index & crawlers. VENDOR-DOCUMENTED PerplexityBot indexes content; Perplexity-User fetches on a live user’s behalf; both are independently robots-controllable (~24h propagation) and Perplexity advises WAF whitelisting. Sites disallowing PerplexityBot won’t have full text indexed (domain/headline/summary may remain); crawled content is not used to train foundation models; Perplexity also partners with third-party search providers.

Selection. Partly documented, partly inference Perplexity describes ranking for helpfulness/trust/freshness rather than click likelihood, but not in a single canonical ranking doc. OUR INFERENCE (RAG two-step): a page must both rank in Perplexity’s index for the query and contain an extractable passage that supports a sentence of the answer.

Citation data. INDEPENDENT STUDY Perplexity is the most community-weighted engine measured — Semrush’s Reddit study (248k cited Reddit URLs across 217k queries) found Reddit the top-cited domain on Perplexity, Q&A threads supplying over half of cited Reddit content. Perplexity’s own materials emphasize freshness, but we found no rigorous public quantification of its magnitude, so we treat freshness as directional only.

Factor weights

Weight and confidence are our editorial estimate of each factor’s influence, and of the evidence for its direction — not a vendor’s ranking weight. No engine publishes a brand-selection formula.

FactorWeightConfidenceBasis
Crawlability by PerplexityBot + WAF allowance (index-inclusion gate)High (binary gate)HighVENDOR-DOCUMENTED
Ranking in Perplexity’s own index (helpfulness/trust/relevance)HighMedPartly vendor-described; specifics undisclosed; weighting OUR INFERENCE
Passage extractability (answer-first, one claim per sentence, quotable facts)HighMedOUR INFERENCE; KDD-2024 GEO measured its largest gains on Perplexity
Presence in community/UGC, especially Reddit Q&AHighHighINDEPENDENT STUDY (Semrush 2025)
Content freshnessHigh (time-sensitive)MedVendor-described emphasis; magnitude not rigorously quantified
Domain-level trust/authority (internal scoring)MedMedVendor-described concept; scoring undisclosed
Google/Bing organic rankingsLow-Med (indirect)MedINDEPENDENT STUDY (low top-10 overlap); third-party partners make some coupling plausible
Generic schema markupLowLow-MedOUR INFERENCE; no vendor doc of schema in ranking

What you can do (ranked)

  1. Verify PerplexityBot + Perplexity-User access end to endengine-specificVENDOR-DOCUMENTED

    robots.txt, WAF whitelisting per Perplexity’s guidance, and server-log confirmation; misconfigured bot mitigation silently zeroes citation potential.

  2. Create the strongest evidence page per high-value questioncross-engineOUR INFERENCE

    Retrieval-first architecture makes a retrievable answer-bearing page a prerequisite for dependable citation.

  3. Earn authentic Reddit / category Q&A presenceengine-specific in intensityINDEPENDENT STUDY

    The highest measured community-citation concentration of the six; substantive, well-received answers are the pattern its retrieval selects. Astroturfing violates platform rules and is filtered.

  4. Lead every page with the answercross-engine · strongest hereINDEPENDENT STUDY

    The KDD-2024 GEO benchmark found adding quotations, statistics, and citations improved generative-engine visibility ~28–41% (quotations +41%, statistics +32%, citations +30%), with the strongest tactics validated on Perplexity in the real world.

  5. Update time-sensitive pages without URL churncross-engineOUR INFERENCE

    Stable canonical URL, effective date and changed facts in crawlable text.

  6. Monitor by citation, not rankengine-specificOUR INFERENCE

    Perplexity has no stable SERP; measurement is mention/citation rate over repeated question samples.

Grounding, citation & locality

VENDOR-DOCUMENTED Every answer includes numbered citations linking sources — the densest, most auditable of the six. OUR INFERENCE: visible citations are a selected subset, not the full candidate set or retrieval rank. Locality: approximate location from IP; API city/region controls should not be assumed to describe the consumer UI.

Sources

Three steps to get started.

A couple of hours a month is enough to make real progress.

Enter your website.

We look up your business and the questions your customers actually ask AI.

Review what we found.

Confirm your business, your questions and your competitors. Then we ask the live apps your customers actually use: ChatGPT, Gemini, Google AI Overviews, AI Mode, Copilot and Perplexity. Not the developer API, and not the model's training data. Both give a different answer from the one your customer sees.

Get your free report.

Where you appear, who wins instead, and how AI describes you, in plain English with clear next steps. A one-time snapshot to start from.

See what AI says about you.

Get a free report: no sign-up, no credit card, results in minutes.  ·  See a sample report ›