← All issues

AEO/GEO RADAR · 26 September 2026

Perplexity's old API shuts off tomorrow, and a new measurement paper shows why asking an assistant once is not enough to know who it recommends.

Confirmed at source

Perplexity Sonar retires tomorrow

Perplexity's migration page still says, verbatim, "Sonar will be supported until September 27, 2026." It does not say what happens to requests after that. A third-party router, LLM Gateway, told its own customers on 25 September that "Perplexity retires its Sonar chat completions API on September 27" and that it has moved its `perplexity/sonar` traffic onto the Agent API, while `sonar-pro` and `sonar-reasoning-pro` stop routing on the 27th. That is LLM Gateway's routing decision, not a Perplexity statement, but it confirms the date from outside Perplexity. The engine's local `src/adapters.ts` still has `perplexity: "sonar"` on line 14 and `https://api.perplexity.ai/chat/completions` on line 91. The remote-branch check that the last four briefs ran was not run today, so whether a migration branch exists on GitHub is unverified this morning.

Perplexity docs, read 26 September 2026
Reported, not confirmed

Google's September spam update: day three, volatility spiking

The Search Status Dashboard still shows the update as open, with no incident posted after 24 September. Search Engine Roundtable's follow-up on 25 September says "the Google volatility is indeed spiking." Google has described it only as "a normal spam update" and has not said which spam policies it targets.

Search Engine Roundtable (Barry Schwartz), 24 September 2026, updated 25 September
Measured, with numbers

One ask is not a full answer: a repeated-query audit of 50 buying questions

The author asked 50 English buying questions (software, consulting, fintech, e-commerce, health tech) to six API configurations 15 times each on 3 and 4 September, 4,500 responses in all. Only one configuration, Perplexity Sonar, had web search switched on; the other five (GPT-5.6 luna, Claude Sonnet 5, Gemini 3.7 Flash, Grok 4.5, Mistral Large) were called without a search tool. Ten asks recovered a median 92.6% to 95.2% of the brands that 15 asks surfaced. A single ask recovered a median 80.0% to 92.5% of a five-ask set in February and 90.0% to 100.0% in September, depending on the question. The paper declines to name a safe number of asks and says results depend on how names are extracted and which questions are chosen. The author is CEO of Rankfor.AI, which sells repeated-query brand audits, and discloses it.

arXiv 2609.05059 v2 (Dmitrij Żatuchin), revised 23 September 2026, first posted 4 September

The Radar is free to read and always will be. If you want to know whether AI names your business when a buyer asks, the free check asks ChatGPT 25 real buyer questions and emails you the answer within the hour.

Get the Radar by email →