Perplexity retires its Sonar API in 11 days, and Cloudflare's "Block" now shuts out Google too.

Perplexity's own wording: "Sonar will be supported until September 27, 2026", covering the sonar, sonar-pro and sonar-reasoning-pro chat completions models. The replacement is the Agent API at a new endpoint, which takes an input instead of a messages array and returns sources and the answer as separate typed items. Perplexity's docs map plain Sonar to the "fast" preset, and its August changelog says that preset "now uses openai/gpt-5.6-luna". Sonar's announcement was in August; this desk read Agent API changelog entries on 22 and 27 August and never flagged the retirement.

Cloudflare added "Disallow AI Training", which writes Google-Extended and Applebot-Extended disallow lines into robots.txt and keeps Googlebot, Bingbot and Applebot crawling for search. Existing Training selections of Block migrated to it on 15 September. The "Block" setting itself changed meaning: it "will stop Applebot, Bingbot, and Googlebot from reaching your site, search included." Sites that had the legacy "Block AI Bots" switch on now show Search allowed, Training disallowed and Agent blocked, and Agent covers "chat fetch bots and browser-use agents". This reverses Cloudflare's 1 July plan, which said Training blocks would catch those crawlers. Bing will not read the robots.txt no-training line until "early 2027". Cloudflare says 17% of its sites block training in some way.

Method: 520 prompts in three categories (AI search monitoring, public sector contracting, retail), run once a day for 40 to 91 days on ChatGPT, Perplexity, Gemini, Google AI Mode, Google AI Overviews, Copilot and Claude, US-based, 252,407 answers. Only 12% to 19% of brands appeared on a single day in a 30-day window; for cited sources that reached 67%. ChatGPT and AI Mode reused about 26% of the previous day's sources; Perplexity about 75%. 48% of 45,919 URLs were cited once. Perplexity was excluded from brand comparisons because "82% of its answers did not contain a brand mention." Moving from 10 to 100 prompts cut variation in brand coverage by about 73%: with 10 prompts, 90% of results fell between 39% and 71%; with 50, between 49% and 62%.

Limits: vendor study on its own tool, one run a day, none of the categories are local services, and Otterly says the figures "should not be treated as fixed stability rates for each engine."

Between 21 May and 21 July 2026, a result_source field in ChatGPT's server-side events carried four values: Labrador (OpenAI's own index), Bright, Oxylabs and SERP. The author lists Labrador verticals including web, news, local and shopping, and says Bright "very likely Bright Data" scrapes Google results and Google Maps for local listings. It also cites 2024 antitrust testimony that OpenAI began building its own index in 2023, and a Peec test where asking ChatGPT about a zero-traffic site showed up as a spike in that site's Google Search Console data on the same days.

Limits: reverse-engineered from a field visible only from 21 May to 21 July, from a vendor selling ChatGPT tracking. Nothing here is confirmed by OpenAI, and it describes the consumer app, not necessarily the API.