ChatGPT Ads hits $1B, Google opens AI Search controls, and a new test for visual hallucinations (2026)

The coverage window runs from Aug. 31 through the morning of Sep. 1, 2026 (Asia/Dhaka). OpenAI's advertising business crossed a meaningful revenue threshold, Google finished rolling out controls for generative search, and a new benchmark showed why multi-image vision still fails in ways single-image tests miss.

Quick scan

Development	What changed	Scale or status	Why it matters
OpenAI ChatGPT Ads	OpenAI says ChatGPT Ads reached a $1 billion annualized revenue run rate in under 200 days and is expanding self-service buying. 1	Tens of thousands of advertisers; self-service buying is expanding across India, Europe, the Middle East, and North Africa. 1	Ads are becoming a central part of the free ChatGPT business model, with conversation context shaping relevance.
Google AI Search controls	Google says a Search Console toggle and new AI-search visibility insights are now available worldwide after an Aug. 31 rollout update. 2	The toggle controls inclusion in AI Overviews, AI Mode, and AI Overviews in Discover; opting out gives up traffic and impressions from those features. 2	Publishers get a direct distribution choice, paired with a measurable cost in AI-search reach.
MIOH vision benchmark	A new arXiv paper tests object hallucination across multiple images rather than treating multi-image reasoning as one broad score. 3	The authors evaluate 29 models across four tasks, three reasoning patterns, and three controlled pressures. 3	Model evaluation can separate perception errors from failures that appear while combining information across images.

OpenAI turns ChatGPT Ads into a billion-dollar business

OpenAI says ChatGPT Ads reached a $1 billion annualized revenue run rate in less than 200 days after launch. The platform now serves tens of thousands of advertisers, and OpenAI is opening direct Ads Manager purchases across India, Europe, the Middle East, and North Africa. 1

OpenAI positions advertising alongside subscriptions, enterprise products, and usage-based APIs. The company says the mix supports an advertising-funded free tier that reaches more than 1 billion weekly active users. 1

The product decision is more consequential than the revenue headline. ChatGPT Ads uses the context of a current conversation to choose relevant ads. Depending on a user's country and settings, the ad system may also use context from the broader ChatGPT experience. OpenAI says ads remain labeled and separate from answers, advertisers cannot access private conversations, and users can control ad personalization. 1

OpenAI says more than 50 technology and measurement partners now support the platform. The company also says cost-per-click and outcome-optimized bidding account for most campaigns, while product feeds, geographic targeting, custom audiences, Pixel, and Conversions API support measurement. 1

The next checkpoint is whether advertisers can keep getting measurable outcomes as self-service access broadens. OpenAI says it plans to add more markets, formats, objectives, buying options, and measurement features. 1

Google gives websites a direct AI-search choice

Google's June announcement received an Aug. 31 update that says the new controls and insights have rolled out to all websites worldwide. The original release introduced a Search Console toggle that lets site owners decide whether their pages appear in and help ground responses in AI Overviews, AI Mode, and AI Overviews in Discover. 2

The choice has a clear tradeoff. A site that opts out of Google's generative AI features gives up traffic and impressions from those features. Google says the setting stays separate from ranking in conventional search results. 2

Search Console insights add impressions data, the pages that appear in AI responses, and the countries where those appearances occur. Google says AI Overviews now has more than 2.5 billion monthly active users, while AI Mode has passed 1 billion monthly users. 2

For publishers, the practical question has moved from whether AI search exists to how much distribution a site is willing to exchange for inclusion. Google's control makes that exchange explicit, while the new reporting gives site owners a way to see where AI features surface their pages. The next checkpoint is how publishers use the data and whether Google adds more control over the forms of reuse.

A new test separates multi-image hallucinations

The arXiv paper Fine-Grained Multi Image Object Hallucination Benchmark introduces MIOH, a benchmark for testing whether multimodal large language models describe objects accurately across several images. The paper was submitted on Aug. 31, 2026 by Joonki Min, Chaeyun Kim, Hyungwook Choi, Yejin Kim, Kihyun Kim, Yohan Jo, and Joonseok Lee. 3

MIOH breaks the problem into four tasks: checking whether an object exists, counting objects, identifying attributes, and locating objects. It also tests three reasoning patterns: comprehensive reasoning across all images, comparative reasoning between images, and selective reasoning over only relevant images. 3

The benchmark adds three controlled pressures: increasing the number of visual contexts, making perception harder, and introducing contextual bias. The authors evaluate 29 models, including GPT-5 and Gemini-2.5-Pro, and report distinct failure patterns across tasks and reasoning patterns. 3

The paper's interpretation changes where evaluators should look for the error. A model may recognize an object in one image and still lose the object's identity, count, attribute, or position while combining information across images. The authors describe that as an integration-stage limitation, alongside ordinary perceptual failures. 3

The next checkpoint is whether future model reports publish results on these separate tasks and pressures instead of one combined multimodal score. MIOH gives evaluators a way to ask which part of multi-image reasoning failed.

What to watch

Ad economics: OpenAI's next product changes will show whether a $1 billion annualized run rate can support broader access without weakening separation between ads and answers.
Publisher control: Google's worldwide rollout turns AI-search participation into a measurable distribution choice. Watch which pages gain impressions and which sites opt out.
Vision evaluation: MIOH's task breakdown should make it easier to compare models on object identity, counting, attributes, and position across different multi-image settings.