Does Perplexity use its own model or others like GPT and Claude?

Perplexity is not one model. It runs its own Sonar family alongside models licensed from other labs, and the plan decides which one writes the answer.

• Free plan: the default is "Best", an auto-router. Perplexity picks a model per question, and the user does not see the choice.
• Pro and Max plans: subscribers pick from a menu of eight models. Secondary reports list Sonar 2, GPT-5.6, Gemini 3.7 Flash and Claude Sonnet 5 among them.

What this means: "what does Perplexity say about my brand" has no single answer. It depends on the plan, the model the router chose, and whether the user switched models. The retrieval side is Perplexity's own. PerplexityBot crawls pages for Perplexity's search index, which Perplexity's crawler documentation says is not used to train foundation models. The model on top of that index varies.

The next move is to stop treating Perplexity as a fixed system you can test once. Treat it as a family of answers that share an index. For the index side, see whether Perplexity has its own search index.

The plan-to-model details come from a 99Visibility snapshot of what the consumer AI apps run, dated 21 September 2026. For Perplexity, every line in that snapshot is marked as a secondary report. It was not confirmed on a Perplexity page.

Here is what that means in practice:

• The existence of the free auto-router and the paid model menu is consistent across secondary sources.
• The exact list of eight models is a point-in-time reading. Vendors add and retire models often, so expect the menu to have changed since.
• Perplexity does not publish how "Best" routes a question, so nobody outside Perplexity can say which model answered a given free query.

Sonar itself is Perplexity's own model, announced on its blog as a fast model built for Perplexity search. That makes it the closest single model to pin a measurement to. It is still not the app, because the app can hand the question to someone else's model.

Be plain about the limit. Any claim that "Perplexity uses model X" is true for some users, on some plans, on some days.

An AI answer passes through stages. The access and retrieval stages decide which pages reach the model. The generation stage decides which of those pages the model actually uses and cites. Model choice sits in that last stage.

So the same question, on the same index, can produce different answers when a different model writes it. One model may name your brand. Another may cite a competitor's review page instead. Nobody publishes how each model weighs the sources it is handed.

Three more things add noise:

• Retrieval is per query. Perplexity decides what to fetch for each question, so the source set shifts between runs.
• Citations are whole sources. Perplexity cites sources by an integer index, not by quoted sentence, which Perplexity's citation format covers in detail.
• Users switch models. A paid user can rerun the same question on a different model and get a different brand list.

This is the same effect covered in whether switching AI models changes what it says about your brand. On Perplexity it happens inside one product.

Tracking tools query Perplexity through its API, not the consumer app. That gap matters more than most dashboards admit.

• No API model equals "Best". The router is a consumer feature. An API call pins one model, so it measures that model, not the mix free users see.
• Sonar Chat Completions retired on 27 September 2026. Callers had to move to the Agent API.
• Web search is opt-in on the Agent API. Perplexity's migration docs make it a setting you turn on. Code that was not migrated correctly can return answers with no web access at all, and nothing flags it.

That last point is the trap. An ungrounded answer reflects what the model learned in training, not what Perplexity's index holds today. If a tracker reports that Perplexity "stopped mentioning" you after late September 2026, check whether it is still searching before you change anything on your site. The Sonar to Agent API migration notes explain what changed, and prompt tracking with the Agent API covers how to set it up.

You cannot choose which model Perplexity uses for a given user. You can control whether Perplexity can reach your pages, and you can measure it honestly.

1. Check access first. Make sure your robots.txt allows PerplexityBot and your firewall does not turn it away. Blocking it removes you from the index every model draws on.
2. State key facts in plain HTML. Put prices, product names and one-line definitions on a page that loads without JavaScript, so any model gets a sentence it can quote.
3. Repeat each question. Run every buyer question several times and read the spread, not one answer.
4. Score Perplexity on its own. A blended score across platforms hides how differently each engine behaves. Compare which AI platforms get brand facts right separately, and note that Google has the same split between its surfaces, covered in which models power AI Mode and AI Overviews.
5. Confirm retrieval is on. If you use the API, verify web search is enabled before trusting a result.

For a quick snapshot, the free checker on this site shows whether AI engines name you for up to three buyer questions. Treat it as one reading, not a trend.

What is Perplexity Sonar? Sonar is Perplexity's own model family, built for Perplexity's search product. Secondary reports list Sonar 2 in the paid model menu as of September 2026. Developers used to call it through Sonar Chat Completions, which retired on 27 September 2026 in favour of the Agent API. Sonar is the closest single model to the app, but the free router and the paid menu mean many answers come from other labs' models.

Does Perplexity's Best mode always use the same model? Perplexity does not publish how Best routes questions, so you should assume it can vary by query. The user does not see which model answered. That makes a free-plan answer hard to reproduce exactly, even with the same wording. If you test Perplexity, record the date and plan, and repeat the question rather than trusting one run.

Does the model choice change which websites Perplexity cites? It can. Retrieval pulls candidate pages from Perplexity's index, then the model decides which of them to use and cite. Different models can make different choices from the same set. No vendor publishes how each model weighs sources, so nobody can predict the effect for a given brand. What you control is whether your pages are reachable and state your facts clearly.

Can I stop Perplexity from using my content? You can disallow PerplexityBot in robots.txt, which keeps pages out of Perplexity's search index. Perplexity's documentation says PerplexityBot is not used to train foundation models. Its user-triggered fetcher, Perplexity-User, generally ignores robots.txt rules, so blocking is not complete. Blocking PerplexityBot also removes you from the pages every Perplexity model draws on.

Why did my Perplexity tracking change in late September 2026? Check the tool before your site. Sonar Chat Completions retired on 27 September 2026, and web search is opt-in on the Agent API that replaced it. A tracker that migrated without turning search on would get answers from the model's training data alone. Those answers can drop recent facts and brands, and nothing in the output says retrieval was off.