AI search news, June 2026: Google goes agentic and the market splits

What did Google announce for AI search at I/O 2026?

Google used I/O 2026 on 19 and 20 May to make Gemini 3.5 Flash the default model in AI Mode for every user globally, and to rebuild the search box itself around agents. Google said of the swap that "you no longer have to trade quality for latency". Liz Reid, Google's vice president and head of Search, called it "the biggest upgrade to our iconic search box since its debut over 25 years ago". The new box expands as you type, suggests how to phrase the question, and routes you between classic results, AI Overviews and AI Mode without a mode switch.

The agent layer is the bigger story for brands. Information agents will run in the background around the clock, scanning the web against a standing brief and sending synthesised updates, with launch set for Google AI Pro and Ultra subscribers from mid 2026. Agentic booking expanded from restaurants to local services and experiences, covering home repair, beauty and pet care, with Search assembling live pricing and availability before handing over to the booking page. For some queries, results now arrive as custom interactive interfaces such as dashboards, trackers and comparison tools rather than pages of links. Ads entered AI Overviews in the same set of announcements.

Google's own scale figures explain why this matters more than any single feature. AI Overviews reach 2.5 billion people a month. AI Mode passed 1 billion monthly users within a year of launch, with query volume more than doubling every quarter. SISTRIX data from March 2026 shows the click-through rate at position one falling from 27 per cent to 11 per cent, while 60 per cent of Google queries now end without a click, rising to 69 per cent for news. Those clicks are not coming back to anyone's website.

Bushnote's reading is blunt: Google wants search to answer and to act, and the website's role in that journey is shrinking to the moments where the agent needs a source or a transaction partner. Brands that are easy for agents to read, quote and transact with will keep their place in the journey, while everyone else slides into the unread background of someone else's answer.

Why does multi-platform AI visibility decide who gets found?

OpenAI, Google, Perplexity and Anthropic now split the answer market across at least five surfaces that matter. OpenAI reported 900 million weekly ChatGPT users in February 2026. The Gemini app reached 750 million monthly users at Alphabet's fourth-quarter earnings, then 900 million by May. Google's AI Overviews sit above both at 2.5 billion a month. Similarweb figures show ChatGPT's share of web traffic to generative AI tools falling from 77 per cent to 57 per cent in twelve months, which is what a maturing, fragmenting market looks like.

Grok, GPT, Claude and Gemini also behave differently when they choose sources. A Stanford HAI audit of six commercial chatbots found Grok 4 cited BBC News in 28.5 per cent of responses, while Claude 4.5 Sonnet cited it in none, GPT-5 in 0.2 per cent, and the Gemini models in 4 to 7 per cent. The same publisher, on the same news days, went from a quarter of answers to invisible depending on the assistant. Crawl permissions, licensing deals and retrieval design drive that spread, and none of those yield to a single tactic, so each platform keeps a different gate in front of visibility.

Bushnote's view on prioritisation stays practical rather than purist. Australian brands should win Google's AI surfaces first, where the reach already sits, then treat ChatGPT as the second front for high-intent research, and extend to Gemini, Perplexity and Claude on the same machine-readable foundation. We ranked the local specialists in our guide to the top AI search optimisation agencies in Australia for 2026, and our earlier analysis of Google's search agent and Australian brand visibility covers the Google side in depth.

Is electricity the new bottleneck for AI?

Meta agreed this week to lease a 168 megawatt AI data centre that Reliance Industries will build in Jamnagar, Gujarat, the company's first such facility in India. The announcement landed on 9 June in the United States and 10 June Australian time. Reliance builds and operates the site on renewable power with desalinated seawater cooling, Meta leases the capacity with an option to scale, and delivery is expected within two years. Meta also contracted nearly 1 gigawatt of clean energy across India through CleanMax, at 837 megawatts, and Fourth Partner Energy, at 88 megawatts.

"This world-class facility in Jamnagar will help us scale our AI infrastructure globally while deepening our long-term investment in India's economy," Mark Zuckerberg said in the announcement. The subtext running underneath it is supply.

Jamnagar reads as confirmation that the binding constraint on AI has moved from algorithms to electricity and real estate. A hyperscaler does not sign a built-to-suit facility on another continent, with dedicated power contracts and seawater cooling, for a product category it expects to fade. For brands, the infrastructure spend is the strongest available signal that answer engines are a permanent channel rather than a feature cycle, and optimisation work done now builds against a growing surface.

How did AI search become a content quality audit?

AI assistants cite a narrow set of sources, and the narrowness is the lesson. IPPR research published in January 2026 found journalism made up 40 per cent of the sources in AI answers, ahead of government websites at 23 per cent, with citations concentrated in a short list of brands led by the BBC and the Guardian. The BBC earned that position even while blocking several AI crawlers through robots.txt, which says trusted, well-structured content gets found one way or another. Structure and trust did most of that work.

Trust runs in the opposite direction as well. The BBC's own research found problems in just over half of AI chatbot summaries of its articles, with clear factual errors introduced into roughly a fifth. If assistants misread the most carefully edited newsroom content in Britain, they will misread an average Australian product page far more often. Semafor describes a new cottage industry it calls SEO 2.0, where websites are designed for AI agents rather than search engines and startups run thousands of prompts a day to track how the models talk about client brands.

At Bushnote we treat this as an audit with a concrete fix list. An assistant that gets your pricing wrong, confuses your entity with a competitor, or skips you entirely is surfacing defects in your content: inconsistent facts across pages, missing structure, ambiguous entity signals, and blocked or broken crawl paths. Our guide on how to write content for AI search engines works through the structure side, and the same machine readability pays back on every platform in the table above.

Which models shipped in the past six months?

Eight frontier releases landed between November 2025 and the end of May 2026, by our count of the major labs. Gemini 3 opened the sequence in November 2025. Gemini 3.1 Pro and Ultra followed in early 2026, taking the context window to 2 million tokens. Anthropic shipped Claude Opus 4.6 and Sonnet 4.6, with a 1 million token window generally available from 13 March, then Claude Opus 4.8 on 28 May. OpenAI released GPT-5.4 in March and GPT-5.5 on 23 April, six weeks apart. Google made Gemini 3.5 Flash the AI Mode default on 19 May and held Gemini 3.5 Pro for a June release, while industry trackers now log a new model somewhere in the market roughly every two days.

The search consequence hides in one of those dates. When Gemini 3.5 Flash became the AI Mode default on 19 May, the model composing answers for a billion users changed overnight, with no notice period and no change on any brand's website. Each release reweights how answers get assembled, which sources get retrieved and which brands get named. A citation position earned in March can vanish in May without the brand owner ever seeing a notification. Nobody sends the brand owner a memo about it.

That is why we tell clients monitoring is now a standing task rather than a quarterly check. Benchmark your citations across Google's AI surfaces, ChatGPT, Gemini and Perplexity, re-test after each major release, and fix machine readability before information agents reach subscribers in the coming months. Paid placement is arriving in parallel, and our coverage of ChatGPT ads in Australia sets out the buying side, including the categories that remain excluded. The June calendar to watch holds Gemini 3.5 Pro, slated for general availability this month, and a rumoured GPT-5.6 that prediction markets price as more likely than not before 30 June.