ChatGPT stopped citing several major newsrooms almost overnight in July, and Google rankings do not explain it. Anna Crowe, writing as a contributor for Search Engine Land on November 16, traced a pattern of citation losses that tracks OpenAI model launches more closely than any search update.

Crowe discloses her employer, Muck Rack, which is the company behind both the Generative Pulse research she cites and the summit where it was presented. She ran the Semrush comparison herself. Because Generative Pulse pulls from model APIs rather than the consumer apps, she cautions that the numbers show direction, not gospel.

The headline finding concerns July 9, when OpenAI began rolling out its GPT-5.6 family. Within a week, citations in ChatGPT to The Atlantic, the LA Times, TechCrunch, TIME, and AP dropped by 50 to 90 percent. AP has since recovered about half of its previous level, a rebound visible from Aug. 13.

Earlier drops follow a similar rhythm. Reuters lost ground in January. The Financial Times dropped in March, after GPT-5.4 Thinking launched on March 5. In May, seven outlets slid (Politico, Barron’s, People, Investopedia, WSJ, Business Insider, and The Verge), around the week GPT-5.5 Instant became the default. CNBC followed in June. The January and June waves match no launch Crowe could find, which weakens any claim that releases alone are the cause.

Across the models Generative Pulse tracks, news outlets supplied about 27 percent of AI citations in May and 22 percent in October, and ChatGPT drove most of that decline. Inside ChatGPT alone, the news share went from roughly 33 percent in January to 22 percent by September, falling in two stages, one in April and one in July. Gemini and Claude barely moved. A fixed panel of 102 questions, run once in April and again in September, shows where the citations went. Government and NGO sources climbed from 10 percent to 35 percent, and brand sites from 14 percent to 20 percent. News fell from 28 percent to 22 percent. The later run used a newer model, so the comparison mixes a model change with any source preference. Earned media overall stayed near 84 percent.

Crowe then checked Google. Comparing top-3 keyword counts between June 15 and July 15 in Semrush showed no matching decline, and AP actually gained. The reverse also held: when the May core update hit, top-3 counts for four of the same outlets fell 17 to 22 percent, yet their ChatGPT citations stayed flat. Ahrefs data points the same direction: just 12 percent of links cited by AI answers also sit in Google’s top 10 results for that prompt. Her argument is that citation changes in ChatGPT behave like unannounced core updates, with no changelog and no explanation. OpenAI did not comment.

Her possible explanation is about retrieval behavior. Seer Interactive found that GPT-5.6 Luna leaned on site: searches for 76 percent of its fan-out queries, and it retrieved 43 percent fewer sources than GPT-5.5 did. Ahrefs found that 75.9 percent of those site: searches targeted brand sites. Licensing does not settle it either. Many of the outlets that dropped have deals with OpenAI, while four outlets without one (the LA Times, TechCrunch, CNBC, and Reuters) lost citations as well.

Volatility is the wider backdrop. Of more than 20,000 cited sites, only about 200 to 250 showed up across all four models, and asking the same question twice on the same day produced 41 percent overlap in cited domains. A single week of movement is therefore weak evidence by itself, and the timing correlation with launches is not proven causation.

The practical advice is an audit. Crowe recommends pulling each model’s AI citations for the window from June 25 to July 25, marking July 9 and any later launch dates, and comparing the result against Search Console and rank data. She also suggests checking server logs for the OAI-SearchBot and ChatGPT-User crawlers.

For publishers and the teams reporting on their visibility, a ChatGPT citation loss should not be filed under a Google ranking problem by default. Track citations per model against launch dates now, so the next drop is diagnosed in days, not months.

Search Engine Land, contributor analysis by Anna Crowe, published November 16, 2026.