18 Days After the Perplexity Sonar Shutdown: What Actually Happened

Starting September 19, we published three pieces warning that Perplexity's Sonar API retirement on September 27 would likely cause widespread AEO monitoring tools to return zero citation data -- silently, with no error to indicate anything was wrong. Developers would keep running audits. Dashboards would show numbers. The numbers would be wrong.

That was the scenario we were watching for. We have now spent 18 days tracking whether it materialized.

Here is what actually happened.

The Predicted Failure Mode

The concern, documented in Scout session 143 on September 19 and updated through session 162 on October 8 in our edge case vault (`edge-cases/perplexity-sonar-api-retirement-2026-09-19.md`), was specific.

Perplexity's Sonar Chat Completions API -- the endpoint most AEO monitoring platforms used to run Perplexity citation checks -- returned citations in `choices[0].message.content`. The new Agent API returns them inside `output[]`. A tool that migrated its endpoint but left its citation parser on the old response structure would execute successfully, return HTTP 200, and quietly discard every citation it received. No error. Zero citations. No signal that anything had changed.

We called this the "silent failure" vector. It was not theoretical: a GitHub issue filed the same day as the retirement (garrytan/gbrain#5534) confirmed that the old endpoint was returning HTTP 403 with the error code `agent_api_migration_required`. But a tool that swapped endpoints without updating its parser would be on the new endpoint -- and still parsing nothing.

There was a second failure mode documented in session 157 (October 3). Sonar accepted location as natural language in the prompt: "You are checking local citations in Toronto for a plumber." The Agent API requires location as a structured field -- `user_location: {city, region, country}`. A tool that migrated the endpoint but kept passing location via prompt text would return results without geographic grounding. Local business citation rates would drop, silently, for no visible reason.

What 12 Consecutive Checks Found

Between September 28 and October 8 -- 12 daily checks across developer forums, LinkedIn, Reddit, GitHub issues, and practitioner communities -- we found no public reports of AEO monitoring tools returning zero Perplexity citations. No forum posts. No LinkedIn complaints. No agency blog posts describing tools returning anomalous results.

The first practitioner-level signal we found was a Perplexity API Platform community forum post, filed September 29, requesting sonar-pro support on the Agent API. One developer hit the hard model-not-found error and posted a bug report. That is the totality of the publicly visible breakage signal from the observation window.

The developer ecosystem migrated without drama. Major gateway tools (LLM Gateway migrated September 25, two days before the deadline), application repositories (Canonry, Raycast, Vercel AI SDK all confirmed Agent API migration PRs in the first week of October), and general developer tooling handled the transition cleanly.

Our October 8 assessment, logged in session 162: "Widespread, obvious breakage did not happen." The watch was closed.

Why It Didn't Break

The cleaner-than-expected outcome has a few likely explanations.

Most production integrations were routed through API gateways (LLM Gateway, similar), which handled the migration transparently for their users. A developer whose Perplexity calls ran through LLM Gateway got the migration for free.

The hard 403 error -- `agent_api_migration_required` -- was noisier than anticipated. Rather than silent zeros, tools on the old endpoint got an explicit error code. Operators with standard 4xx alerting would have seen it. This pushed some teams to migrate when they would otherwise have drifted.

The exception: tools using plain `sonar` (base model) faced a more ambiguous path. Their old endpoint returned 403, not a silent zero. The silent failure risk shifted to tools that correctly migrated the endpoint but left the citation parser on the old response format. That second-order failure is harder to detect and impossible to rule out from no-reports alone.

The One Open Flag: HubSpot AEO Grader

We have not confirmed that HubSpot AEO Grader migrated.

HubSpot AEO Grader uses Perplexity Sonar as one of three scoring models, alongside GPT-5.2 and Gemini 2.0 Flash, based on September 2026 documentation. HubSpot has 205,000+ AEO platform customers (documented in Scout session 148). Between September 27 and October 8, we searched HubSpot's AEO Grader product page, blog, and changelog 12 times. No migration announcement appeared.

This means HubSpot AEO Grader Perplexity scores carry LOW confidence status from us for any report generated after September 27. Not because we have confirmed they are wrong -- we have not. But because HubSpot has had 18 days to publish a migration notice and has not done so, and because the technical failure modes documented above (silent zero or de-localized results) would not surface in normal product review.

If you have run a HubSpot AI Search Grader report since September 27, treat the Perplexity component with skepticism until HubSpot confirms migration.

Why Perplexity Data Accuracy Is Worth Caring About

This might feel like a technical footnote. It is not.

In our June 9, 2026 platform divergence analysis (`knowledge/platform-divergence-data.md`, filed 2026-04-22, updated session 42), we documented what three independent studies -- Averi (680 million citations), Whitehat SEO (118,000 responses), and Passionfruit (cross-platform analysis) -- all found with essentially identical results: only 11% of domains cited by ChatGPT are also cited by Perplexity.

Perplexity and ChatGPT are nearly independent citation systems. The reason is architectural. Our June 2026 update to the same file added the mechanism: ChatGPT generates 91% unique sub-queries from a single user prompt (fan-out architecture). Perplexity retains 88% of original prompt vocabulary and runs close to literal retrieval. They are effectively searching different versions of the web for the same question.

The citation rate data is even starker. From the Averi B2B SaaS Citation Benchmarks Report 2026: Perplexity generates 8.79 citations per response. ChatGPT generates 2.78. For brand citations specifically, the gap is 46x: 13.05% brand citation rate on Perplexity versus 0.59% on ChatGPT.

What this means: a business that looks invisible on ChatGPT may be well-cited on Perplexity -- or vice versa. These are not interchangeable scores. A tool that silently zeros out Perplexity data is not giving you a slightly degraded view of your AI visibility. It is hiding the most citation-rich platform you are likely tracking.

How to Verify Your Perplexity Score Is Accurate

The most reliable check is manual. Open Perplexity.ai directly in a browser -- the consumer web app is unaffected by the API retirement -- and run the queries your monitoring tool is supposed to be tracking. If your business appears in the browser at rates that match your monitoring tool, the tool's data is probably accurate. If your monitoring tool shows zero or near-zero while the browser returns citations, the tool has a problem.

For tools where you can ask the vendor directly: ask whether they run Perplexity checks via the Agent API or via browser-based/manual data collection. API-based tools that have not confirmed Agent API migration are the ones to verify. Manual and browser-based tools were unaffected by the September 27 retirement.

Sourcepull's Signal Check uses a methodology we verify against manual controls. If you want to see where your Perplexity citations actually stand right now, that is the fastest way to get a current, accurate read.