A study of 4,320 AI answers found that 77% of brands were named by only one of the three models tested, and just 11% were named by all three. That is the strongest public number yet for why one engine is not a measurement.

Google posted on its data anomalies page: "A logging error caused a decrease in impressions on the Generative AI performance report in Search for data starting on August 13, 2026." John Mueller added on Bluesky that "this is just a logging issue and not representative of visibility changes in Search." The issue was still open when Search Engine Land published, and Google said it planned to add an annotation inside Search Console. Separately, Barry Schwartz reported practitioners seeing drops in the main Search performance report from around 12 August, but that part is practitioner observation and is not what Google confirmed. Only the generative AI report is in Google's statement.

Fractl put the same 96 industry-specific prompts to three models, 15 runs per prompt per model, for 4,320 responses and more than 8,500 unique brand references across eight industries. Only 900 brands, 11% of the set, were named by all three models. About 77% were named by only one. More than 9 in 10 brands behaved the way an SEO would expect, with stronger traditional search authority tracking stronger model visibility, but roughly 5% were underexposed relative to their search authority and roughly 4% overperformed it. By industry, insurance had the largest underrepresented share at 8%, with legacy carriers displaced by digital-native brands, and education had the highest overperforming rate at 7%. Travel was the tightest match to Google, and three brands (Booking.com 285 mentions, Airbnb 227, Expedia 215) took roughly 20% of that category's mentions.

The method caveat is large and has to travel with the number. Fractl called the models directly through their APIs with no web search, using GPT-4o, Gemini 2.5 Flash and Claude Sonnet 4.6. That measures what the models remember, not what the products answer after retrieval, and two of those three models are well behind what a buyer talks to today. It also did not test coaches, did not test ChatGPT-the-product, and did not test AI Overviews at all.

Older material, new to us, and worth the space. Across 4 million responses from 13 models on a benchmark built from Wikipedia profiles, GPT-5 and Gemini-3 encoded 95 to 98% of the tested facts but failed to recall 26 to 34% of them when asked directly, falling to 11 to 12% when the model was allowed to think. The same facts came back when offered as multiple choice, which is how the authors show the information was there all along. Recall failures accounted for more than 70% of GPT-5.2's errors. The specific finding Search Engine Journal led with: subject and object entities are learned in a particular order, and a question that reverses that order is harder for the model to answer.

iPullRank analysed 13.1 billion search events from 9.1 million opted-in users between October 2024 and December 2025. ChatGPT was the sixth most clicked destination from Google Search and drew a greater share of paid clicks than any other leading destination, and only 11.1% of searches heading toward ChatGPT were intercepted by a Google zero-click feature. Logged for the sample size rather than the finding: the window closes in December 2025, which is eight months stale in a market that Demandbase measured inflecting in May 2026, and the subject is people navigating to ChatGPT rather than anything about being recommended inside it. No consequence for us today.