Articles ·
Which Perplexity model your prompt is running on — and why it changes your citation odds
The one-line version: Perplexity runs at least five distinct model tiers — Sonar, Sonar Pro, Sonar Reasoning, Sonar Reasoning Pro, and Sonar Deep Research — and Sonar Pro returns 2-3x more source citations per response than the equivalent competing grounded models (Bae Seok Jae’s Sonar API guide). The tier your prompt gets routed to changes both how many sources are cited and which sources are eligible. If your Perplexity monitoring only tests one tier, you’re measuring one-fifth of the surface.
Why this matters for GEO
Most GEO monitoring treats “Perplexity” as one thing. It isn’t. When a user selects the free web interface’s default mode, they get Sonar. When they use Pro, they get Sonar Pro. When they use Reasoning, they get Sonar Reasoning Pro. When they use Deep Research, they get Sonar Deep Research. And when developers integrate via the API, they can call any tier directly with different pricing, context windows, and citation-density behavior.
Every tier has a different citation footprint. Your page can be the top source on Sonar and never appear on Sonar Deep Research — or the reverse. The Perplexity source-selection article covered the reranker mechanics within one Sonar run. This article covers how those mechanics vary across tiers.
The five tiers, side by side
| Tier | Context window | Input price | Output price | Reasoning | Best-fit query type |
|---|---|---|---|---|---|
| Sonar | 32K | $1/M tokens | $1/M tokens | Search-first, no CoT | Conversational search, Q&A, fact lookup, summarization |
| Sonar Pro | 200K | $3/M tokens | $15/M tokens | Search-first, denser citations | Document-length research, multi-turn queries, source-heavy work |
| Sonar Reasoning | 32K | $1/M tokens | $5/M tokens | Multi-step reasoning + search | Reasoning + live search on lighter workloads |
| Sonar Reasoning Pro | 200K | $2/M tokens | $8/M tokens | Chain-of-thought over Pro search | Complex research where intermediate steps matter |
| Sonar Deep Research | not published | $2/M tokens | $8/M tokens + citation tokens | Exhaustive, agentic search | Long comprehensive reports, literature reviews, market analyses |
All figures from Bae Seok Jae’s 2026 Sonar API guide and Perplexity’s own model docs. Note that citation tokens were removed from billing in 2026 for Sonar and Sonar Pro — Sonar Deep Research is now the only tier that still bills for citations.
There’s also Sonar Pro Search (sonar-pro-search), an agentic mode with autonomous multi-step orchestration that adds $14-$22 per 1,000 requests. It’s explicitly described as overkill for standard chat interfaces and only appropriate for research-agent workloads.
The citation-density gap
The one hard number that matters for GEO: Sonar Pro returns 2-3x more source citations per response than equivalent Gemini models. This is not a soft “generally more” — it’s a documented multiplier from production testing.
Practical implications:
- On Sonar Pro, more pages get cited. A page that ranks 4th-8th in relevance may still get a citation slot on Sonar Pro that it would not get on plain Sonar. The floor for eligibility is lower.
- On Sonar, only the top-cited sources survive. Standard Sonar cites more selectively. Being the passage-density winner matters more, not less, when the total citation count is smaller.
- On Sonar Deep Research, source count expands dramatically. Deep Research is designed to conduct exhaustive searches. The trade-off is that citation tokens are still billable — which means Perplexity itself is optimizing for source volume differently on this tier.
Separately, 2026 production benchmarking found Sonar models delivered inline citations on 94% of test queries with latency under 2 seconds — and that using Sonar reduced hallucination rates by approximately 60% versus non-grounded LLMs. This 94% floor is what makes Perplexity’s citation share measurable in the first place. On other engines, the “did the answer even try to cite?” question is a real variable.
How user-facing modes map to Sonar tiers
The web interface obscures this, but based on public documentation and observed behavior:
| Web interface mode | Underlying tier | What changes |
|---|---|---|
| Auto (default) | Sonar | Fast, cheap, single-pass retrieval. Most consumer queries land here. |
| Pro | Sonar Pro | 200K context window; markedly denser citations. Multi-turn conversation-friendly. |
| Reasoning | Sonar Reasoning Pro | Chain-of-thought over Pro search. Longer latency, deeper analysis. |
| Deep Research | Sonar Deep Research | Multi-minute agentic runs. Produces long structured reports. |
If a query is complex enough, Perplexity’s Auto mode may promote to a higher tier automatically. But the default routing means the majority of queries — especially casual mobile queries — hit standard Sonar, not the Pro tiers.
The Search Arena result
Independent evaluation via Search Arena — a blind head-to-head evaluation of grounded LLMs — scored Sonar Reasoning Pro at 1,136 and Gemini-2.5-Pro-Grounding at 1,142 (Sonar API guide). The margin is small enough to be a statistical tie. For GEO, the interpretation is that Sonar Reasoning Pro is competitive with Google’s best grounded output — not dominant, but on par. Your content targeting Sonar Reasoning Pro will be judged against Google-caliber sourcing standards.
What this means for your monitoring
If your Perplexity citation-share dashboard tests one tier, it captures roughly 20% of the surface area. To measure your true Perplexity presence, run your fixed 30-100 category prompts through each tier separately on a monthly cadence:
- Web interface, Auto mode — the majority query experience (Sonar).
- Web interface, Pro mode — the power-user query experience (Sonar Pro).
- API,
sonar-reasoning-pro— the reasoning-heavy path. - API,
sonar-deep-research— the long-form-report path.
Log which sources appear at which tier. Track over time. The pattern that usually emerges: your pages perform well on 1-2 tiers and weakly on the others, and the fix for each is different.
What moves the needle at each tier
Given the tier characteristics, here’s where to focus for each:
Sonar (Auto mode) — the passage-density race
- Direct-answer opening paragraphs. Sonar’s smaller context window and lower citation count mean only the passage-density winners survive.
- Short, self-contained sections that map to sub-questions.
- Front-loaded numbers, dates, and named entities.
- Server-rendered HTML. The 32K context window means every extraction pass is precious — JS-hidden content is much more likely to be dropped than on Sonar Pro.
Sonar Pro (Pro mode) — the source-breadth race
- Depth pages that reward the 200K context. Long, well-structured articles with multiple sub-topics benefit disproportionately here.
- Multiple citable sub-passages per page — Sonar Pro’s denser citation output rewards pages that offer more than one answer.
- Robust internal cross-linking so retrieval can chain from one relevant page on your domain to another.
Sonar Reasoning Pro (Reasoning mode) — the chain-of-thought path
- Content that supports intermediate reasoning steps, not just final claims. Compare-and-contrast articles, “why does X work” explainers, and methodological guides win here.
- Explicit statements of assumptions, edge cases, and trade-offs. Reasoning Pro’s CoT engine will consume this and cite it.
- Named studies, statistics, and authoritative citations of your own — these get re-cited when Reasoning Pro synthesizes multiple sources.
Sonar Deep Research (Deep Research mode) — the source-volume path
- Comprehensive coverage. Deep Research conducts exhaustive searches; being one of many good sources on a topic beats being the single best source on one narrow slice.
- Structured, section-heavy articles that a report-generation model can quote from.
- Long-tail sub-topic coverage — Deep Research pulls from a much wider funnel of sources than Sonar.
The strategic point
Perplexity is not one engine — it’s five (or more) engines wearing the same interface. The citation-earning content strategy that works for the median Sonar query is not the same one that wins on Sonar Deep Research, and the passage-selection tactics that dominate Sonar Pro may be less decisive on standard Sonar where the reranker cuts more aggressively.
For GEO measurement, this means splitting your Perplexity monitoring by tier. For GEO content, this means recognizing that a “Perplexity strategy” is really 3-4 overlapping strategies — and the pages that win across all of them are the ones that combine answer-first opening paragraphs, long-form structural depth, entity-grounded schema, and visible presence on the sources Perplexity’s reranker manually boosts.
The Perplexity pillar page puts this into the wider engine playbook. The Perplexity source-selection article explains what happens inside one Sonar run, once your page reaches the reranker.
Related reading: How Perplexity actually chooses its sources — the reranker mechanics within a single tier. Answer-first writing — the passage-level tactics that especially matter on standard Sonar. Measuring GEO in 2026 — the four-metric framework, now split across Perplexity tiers.