Did Perplexity's Sonar retirement affect AI visibility tracking?

The retirement only matters if your tool handled the switch badly. The risk is real, but it is silent.

Here is the chain:

• Perplexity announced that Sonar Chat Completions would retire on September 27, 2026.
• Its replacement is the Agent API. As of October 2026, Perplexity's Sonar to Agent API migration docs describe web search there as something the caller opts into.
• A request without web search still returns fluent text. It just isn't grounded in any page Perplexity fetched.

An ungrounded answer reflects the model's training data, frozen at a cutoff. It can't show whether Perplexity's index holds your pages, whether your new pricing page is visible, or whether a competitor's review site is being cited instead. Those are the questions most people buy a tracker to answer.

So the retirement didn't change how Perplexity treats your brand. It may have changed what your tool measures, without telling you.

Three situations are worth separating.

Still calling Sonar Chat Completions after September 27

Code that never migrated was calling an endpoint scheduled for retirement. The evidence we work from flags that unmigrated code can return ungrounded answers silently, rather than failing with an obvious error. A tool in this state may have kept producing Perplexity scores that look plausible but rest on no retrieval.

Migrated, but web search left off

This is the quieter failure. The code works, the answers arrive, and nothing breaks. But because web search is opt-in, the tool is testing the model with no web access. If you want the background on this default, see whether AI model APIs search the web by default.

Migrated with web search on

This is the setup that measures retrieval. Each answer should come back with search results and citation references attached. The tool may still differ from the consumer app (see below), but it is at least testing a grounded answer.

Some of this is documented. Some of it isn't.

• Documented: the Agent API exists, it is the migration target for Sonar, and web search on it is opt-in, per Perplexity's migration documentation.
• Documented: grounded Agent API responses include search_results with fields such as id, url, title, snippet, date and last_updated. Citations point to whole sources by integer index. Our page on Perplexity citations and citation IDs covers the format.
• Less firm: the September 27, 2026 date traces to a Perplexity forum announcement. It did not appear in the migration documentation itself when we last checked.
• Not measured by anyone: how many tracking tools missed the migration or left search off. No published study counts this, so treat it as a thing to check, not a known failure rate.

One more caveat applies even to a correct setup. Perplexity's consumer app doesn't map to one API model. As of a September 2026 snapshot based on secondary reports, free users get an auto-router called "Best", and paid users pick from a menu of models. An API call is the closest pin, not a copy of the app. The gap is covered in API answers vs consumer app answers.

You can often tell from the output alone.

1. Look for missing sources. A grounded answer carries search results and citation indices. If your tool shows Perplexity answers with no sources at all, assume search was off. Our guide to telling whether an AI answer used live web search has more signals.
2. Look for a step change in late September 2026. A sudden jump or drop in Perplexity scores around September 27 that matches nothing you shipped points to a measurement change, not a visibility change. Work through why an AI visibility score drops before reacting.
3. Test recent facts. Ask about something you published after the model's training cutoff. A grounded answer can find it. An ungrounded one can't.
4. Compare Perplexity with the other engines. If Perplexity suddenly behaves like a model with no web access while the others don't, that's a clue.

None of these is proof on its own. Together they tell you whether to ask your vendor hard questions.

Send these questions to whoever runs your tracking. Good vendors answer them in one reply.

1. Which Perplexity endpoint do you call today, and on what date did you move off Sonar Chat Completions?
2. Is web search enabled on every Perplexity request, and can you show me the setting?
3. Do you store the search results and citation indices for each answer, and can I see them?
4. Did any Perplexity answers between late September and your fix run without web access? Were those scores flagged or excluded?
5. Does your score blend grounded and ungrounded answers? If so, how is each labelled?
6. Which model do you pin, and how does it relate to what the Perplexity app serves?

The fourth question matters most for history. If a tool can't say which answers were grounded, its Perplexity trend line across this period isn't comparable from one side to the other. Whether a tool should test with search on, off, or both is a separate design choice, covered in testing with web search on or off. Whichever choice it makes, it should disclose it per answer.