Google overhauls Search around AI Mode, muddles llms.txt

What happened

Per Search Engine Journal, Google rolled out its May 2025 core update in the same week it redesigned Search around AI at I/O, published the first usage data for AI Mode, and gave contradictory signals on whether llms.txt matters. Four moves, one week, one company. The compression is the story.

AI Mode is now a default destination inside Search, not an experiment. Google released its first usage numbers alongside the launch, signalling that the feature has moved past the pilot phase. On llms.txt, the proposed standard that lets publishers tell large language models what they can ingest, Google's public posture has shifted between "we don't use it" and "we may consider it." Google's John Mueller has been openly dismissive while other Googlers have left the door ajar.

For brands that have spent eighteen months tuning content for AI Overviews, the playing field just moved again. AI Mode is a different surface with different retrieval logic, and llms.txt is now a governance question rather than a technical one.

Why it matters for your brand

The first implication is that AI Overviews is no longer the ceiling of your AI search exposure inside Google. AI Mode is a separate generative surface with its own retrieval behaviour, its own citation patterns, and, based on the early data Google released, its own user base. If your visibility dashboards only track Overviews appearances, you are now measuring a shrinking share of the actual AI-mediated traffic Google sends. Financial services brands tracking citation share for terms like "private credit outlook" or "Basel III implementation" need a second column in the spreadsheet.

The second implication is about content depth. AI Mode is built for longer, multi-turn queries. That changes what Google's models reach for when assembling an answer. Short, optimised explainer pages have less gravity in a multi-turn context than substantive primary research, methodology notes, and named-expert commentary. For multilaterals and policy institutions, this is structurally good news: the UNDRR-style technical reports and CGAP-style field studies that underperformed in classic SEO tend to over-index in generative retrieval because they carry the kind of specificity models prefer when stitching answers. The reverse is true for brands that have leaned on thin thought-leadership posts. Those assets are losing retrieval weight quickly.

The third implication is llms.txt. Google's mixed signalling is not neutral. When the dominant search engine refuses to commit, every other AI vendor watches. Anthropic, Perplexity, and OpenAI all have different stances, and the absence of a Google endorsement means llms.txt will not become a universal standard in 2025. For industrial groups like HOLCIM or large standards bodies like ISO and IEEE, this matters because their legal and comms teams have been asking a binary question: do we publish an llms.txt file or not? The honest answer is that the file is currently a signal of intent to a handful of crawlers, not a control mechanism. Publish it if you want to declare a position. Do not publish it expecting enforcement.

The fourth implication is the core update itself. Core updates are now AI updates by default. Google's ranking systems and its generative retrieval systems share underlying signals, so a site that loses visibility in classic Search during a core update tends to lose citation share in AI Overviews and AI Mode at the same time. The diagnostic work that used to belong to the SEO team is now a brand-visibility question that belongs on the CMO's desk. If your organic traffic dropped in May, your AI citation rate almost certainly dropped with it, and the recovery path is the same: demonstrate expertise, primary sourcing, and topical authority at a level the model recognises.

The fifth implication is pace. Google compressed four major moves into one week. That cadence is the new baseline. Brand teams that review AI visibility quarterly are now operating at one-third the speed of the platform they depend on. Monthly is the floor. For comms leaders at institutions with long approval chains, the organisational implication is uncomfortable but unavoidable: someone inside the comms function needs decision rights on AI-surface response without routing through legal every time.

The signal in context

Google's I/O week marks the point at which generative search inside Google stopped being a feature and started being the product. AI Overviews launched in 2024 as a layer on top of classic results. AI Mode in 2025 is a parallel destination with its own retrieval stack. The two will coexist for a period, but the direction of travel is clear: Google is gradually rebuilding Search as a generative interface with classic blue links as a fallback, not the other way around. Every other major search and answer engine, from Bing to Perplexity to ChatGPT Search, is making a version of the same bet.

The llms.txt waffling is a smaller story but a more revealing one. It tells you that the major AI platforms have not yet agreed on how publishers will negotiate access, attribution, and exclusion. Until they do, brand visibility in LLM answers will remain a function of what models choose to retrieve rather than what publishers choose to expose. That asymmetry favours brands with strong existing authority signals and disadvantages everyone else. The work, for now, is to build the kind of content footprint a model wants to cite, because the opt-in mechanics are not coming this year.