How Does Perplexity Choose Sources?

Perplexity's source choice splits into stages, and they are not equally visible.

• Access: whether PerplexityBot may fetch and store your page. Documented, and in your control.
• Routing: which model writes the answer. Partly reported, not in your control.
• Retrieval: which indexed pages enter the answer for a given question. Mechanism unpublished.
• Generation: which retrieved pages get cited. Exposed in the API output, but the selection logic is unpublished.

The gap that matters most is between retrieval and citation. Perplexity tells developers what it returns. It does not tell anyone how it ranks. For a fuller tour of the index and citation format, see our guide to Perplexity SEO. For the same stages across every engine, read how AI search picks sources.

Perplexity runs its own crawler. Its crawler documentation describes PerplexityBot as the search indexing bot and says it is not used to crawl content for AI foundation models. That is why the question of whether Perplexity has its own search index has a clear answer at the crawler level.

A second agent, Perplexity-User, fetches pages when a user's request triggers it. Perplexity's own docs say this fetcher "generally ignores robots.txt rules." So blocking it in robots.txt is not a reliable control.

Three things can keep your page out of the index:

1. robots.txt. A disallow for PerplexityBot removes eligibility. A blanket AI block often does this by accident.
2. The network layer. A firewall, bot-management rule or CDN can return a 403 or a challenge page to a crawler that robots.txt allows. Nothing in your dashboard flags it.
3. JavaScript rendering. Vercel's crawler study found ChatGPT and Claude crawlers fetch JavaScript files but don't execute them. An independent retest in June 2026 reported the same for PerplexityBot. Content that only appears after hydration may not exist for it.

Passing all three gives you eligibility. It does not give you citations.

The Perplexity app is not one model. As of a September 2026 snapshot, secondary reports describe free users getting "Best", an auto-router, while Pro and Max users pick from a menu of eight models, including Sonar 2, GPT-5.6, Gemini 3.7 Flash and Claude Sonnet 5. Those reports are not from Perplexity's own pages. The practical point holds anyway: model routing means the same question can get different answers and different sources.

Retrieval is not guaranteed either. On Perplexity's Agent API, web search is opt-in, per its migration guide. The older Sonar Chat Completions endpoint retired on 27 September 2026, and code that was never migrated returns ungrounded answers. An ungrounded answer comes from the model's training data alone. Nothing you publish this quarter reaches it.

Perplexity also does not publish how it turns a question into search queries. Other engines document rewriting prompts or fanning out into several searches. Assume the query Perplexity runs may differ from the words your buyer typed.

On the Agent API, each search result carries an id, url, title, snippet, date, last_updated and source. The answer then cites whole sources by integer index. Claude's API works differently, citing short text spans. Our page on Perplexity citations walks through the format.

Two cautions apply.

• Exposed is not weighted. Perplexity returns date fields. It does not say those dates influence which source wins.
• Cited is not the same as used. Mechanistic research on how language models cite (arXiv:2606.28358) found citation output can be repaired without changing the sources at all. A citation count is a proxy for influence, and a noisy one.

No study has measured Perplexity's production ranking. The closest evidence comes from simulated answer engines, and it points in one direction.

• C-SEO Bench (NeurIPS 2025) found most conversational SEO methods "largely ineffective" and often harmful to a document's ranking. Moving a document higher in the model's context was far more effective. That arm placed the document at position 1 directly, so it shows position matters, not that you can earn it.
• A 252,000-trial factorial study (arXiv:2605.25517) found "topical relevance and list position are the biggest drivers of being cited first." It used a synthetic corpus, two candidates per trial, and no live search engine. Its authors also ship a GEO checklist.
• A critical survey of 45 GEO studies (arXiv:2607.14035, July 2026) concluded that no reviewed technique shows a stable, cross-platform causal effect on discoverability. It also warned that citation-oriented rewrites can impair retrieval.

Read together, topical relevance is the one content factor these sources agree on. It means fit to the question, not keyword coverage. The original GEO paper found keyword stuffing lowered its own visibility metric. Our answer on whether staying on topic helps AI choose your page covers the limits.

Brand size sits underneath all of this. A 2025 study of AI search citations (arXiv:2507.05301) found heavy concentration among a few news sources, with a Gini of 0.77 for Perplexity, on data from March to May 2025.

The controllable rows are all about removing blocks. That is where the evidence is strongest. Everything below them belongs to Perplexity, and no vendor can see inside it.

Work top-down. Each step only matters if the one before it passes.

1. Check robots.txt. Confirm PerplexityBot is not disallowed, either by name or by a wildcard rule. A minimal explicit allow looks like this:

User-agent: PerplexityBot
Allow: /

2. Test the network layer. Request a real content page with the PerplexityBot user agent and look for a 200, not a 403, a 429 or a challenge page. The AI Crawler Checker does this per crawler. If you use Cloudflare, review your AI crawler settings: its defaults changed on 15 September 2026, blocking Training and Agent crawlers on ad-displaying pages for new domains and free-tier accounts while leaving Search allowed.
3. Diff raw HTML against the rendered page. Fetch with JavaScript off and on. Move prices, product names and key claims into the server-rendered HTML.
4. State specific facts plainly. A price, a number or a one-line definition on a fetchable page gives the engine a sentence it can quote. It is the only positive experimental signal for on-page content, and it promises nothing.
5. Work the third-party pages. When Perplexity cites other sites for your buyer questions, being named on those sites is associated with appearing in the answer. Use the playbook on finding the pages AI engines cite, then fix the sources Perplexity cites where they get you wrong.
6. Measure Perplexity on its own. Blended scores hide engine differences. The free checker on this site shows whether AI engines name you for three buyer questions, as a snapshot rather than ongoing tracking.

• Treating robots.txt as the whole gate. The firewall can still block PerplexityBot, and Perplexity-User generally ignores robots.txt.
• Rewriting pages in an "AI-friendly" style. The best-designed studies found citation rewrites ineffective or harmful. Schema and llms.txt are not levers either: no engine documents parsing llms.txt, and Ahrefs' matched test found schema produced no major lift.
• Bumping dates to look fresh. Perplexity exposes date fields, but no documentation says it weights them. Editing dates mostly changes your own measurement.
• Trusting one test answer. Routing, retrieval and query rewriting all vary. One screenshot is one sample.
• Injecting hidden instructions. Prompt injection and cloaked text aimed at AI crawlers exist. Research on LLM search agents (arXiv:2606.16821) found attack success varies widely by model. It is illegitimate, it can get content flagged, and the same weakness means third-party pages can misrepresent your brand. Watch for it; never use it.