AI Mode hit a billion users, Google released its first official AI search guide, and two Google teams shipped opposite instructions about llms.txt in the same week. Here is what healthcare and life sciences brands should actually do about it.
The short version
- AI Mode hit 1B monthly users. Google demoed background agents, a redesigned search box, and generative UI that builds custom interfaces inside results.
- Google's new AI search guide says AEO and GEO are not separate disciplines. They are SEO.
- Google Search Central says ignore llms.txt and chunking. The same week, Lighthouse shipped an audit that checks for llms.txt. Both can be true.
- For healthcare brands, the bigger story is generative UI and Information Agents. Both threaten click economics that pharma corporate sites have measured against for a decade.
- Our take: ship a minimal llms.txt, stop selling AEO and GEO as separate work, and double down on first-party expertise that AI cannot fake.
What happened, in one paragraph
On May 19, Liz Reid took the I/O stage and said AI Mode had crossed one billion monthly users. Five Search announcements followed: Gemini 3.5 Flash as the new default for AI Mode, a redesigned search box that handles longer queries and follow-ups, background Information Agents that monitor topics for users, agentic commerce through Universal Cart, and generative UI that builds custom interactive interfaces inside the results page. Four days earlier, on May 15, Google had quietly shipped its first official guide to optimizing for generative AI features in Search. Two days after the keynote, Google rolled out the second core update of 2026.
So in one week, Google published its clearest-ever guidance on AI search, demoed products that complicate that guidance, and started moving rankings under everyone. That is the real story of I/O 2026, and it is more useful than any single announcement.
We work almost entirely with healthcare and life sciences brands. Medical device manufacturers, biotech, clinical platforms, pharma corporate affairs. The questions we are getting right now are not abstract. They are: do we add an llms.txt file, do we restructure our content into shorter chunks, and is our SEO investment still worth the line item. Here is what we think.
The contradiction nobody is talking about
Google's Search team and Google's Lighthouse team published guidance pointing in opposite directions, in the same week, and Google has not reconciled it.
On May 15, Search Central published Optimizing your website for generative AI features on Google Search. The guide is unusually direct. It names four tactics you can ignore: llms.txt files, content chunking, AI-specific rewriting, and special Markdown versions of pages. The exact phrasing: you do not need to create new machine-readable files, AI text files, markup, or Markdown to appear in generative AI search.
Less than a week later, the Chrome team released documentation for new Agentic Browsing audits in Lighthouse. Buried in that documentation, the audit checks for the presence of a machine-readable summary at the domain root. Specifically, an llms.txt file. The same file Search Central just told you to stop worrying about.
The Search team is grading the answer. The Lighthouse team is grading the readiness. They are not the same test, and Google has not said which one matters more for which surface.
This is not a small documentation inconsistency. Search Central is telling you llms.txt does not influence inclusion in AI Overviews or AI Mode, the surfaces fed by Google's web index. Lighthouse is telling you llms.txt is part of what makes a site readable to agents, the things that will increasingly browse on a user's behalf. Different audiences, both Google.
John Mueller has been calling llms.txt the new meta-keywords tag for months, and he is probably right that no major model uses it as a retrieval signal today. But Lighthouse is the audit your client's IT and procurement teams pull when they ask whether the site is AI-ready. If llms.txt shows up red on a Lighthouse report, you are going to spend an hour explaining why Google told you to ignore the thing Google's other tool just flagged.
So what should you actually do?
Ship a minimal llms.txt file. Fifteen minutes of work, no downside, clears the Lighthouse audit. Do not do the heavier llms-full.txt lift, do not maintain bot-only Markdown versions of pages, and do not let anyone position this as a ranking play. It is hygiene for the agentic web, not an answer engine strategy.
Chunking is over. Structure is not.
The second tactic Google retired is content chunking. Breaking longer articles into 120-to-180-word sections so AI extractors can lift them cleanly into an Overview. Google's exact language: there is no requirement to break your content into tiny pieces for AI to better understand it. Google systems are able to understand the nuance of multiple topics on a page.
Danny Sullivan said the same thing in January. The new guide makes it official. But here is the trap we are watching agencies fall into: reading that line as permission to publish 3,000-word walls of text again.
It is not. Chunking-for-AI is over. Structure-for-readers is not. Most of the practices that got rebranded as AEO and GEO over the last 18 months, direct-answer H2s, short summary paragraphs, defined-term callouts, clear semantic hierarchy, were always good information design. The mistake was the framing, not the practice. Defend them on reader-experience grounds, not because an extractor wants them.
This matters more in our space than in a B2B SaaS blog. A clinician scanning a device indication, a patient looking for an eligibility criterion, an investor relations team finding a pipeline asset. They all need scannable, structured content for human reasons. The fact that AI also benefits is a side effect.
The five I/O announcements, ranked by what actually changes for healthcare brands
1. Generative UI. The announcement most SEOs are underweighting.
Google demoed AI Mode building custom dashboards and interactive visuals on the fly, inside the results page. Ask about a marathon training plan and Google builds a stateful tracker. Ask about black holes and it renders an interactive explainer. You cannot rank inside that interface. You can only be one of the sources Gemini pulls from to construct it.
For consumer health queries, this is the biggest threat to organic traffic patterns we have seen since featured snippets. Picture a user asking about a chronic condition. Google can now stitch together a personalized explainer that pulls from your clinical content, your patient stories, and your competitor's research, and never sends the user to any of you. Brand visibility becomes citation visibility. The page-view metric your CMO has reported against for fifteen years gets further away from real reach. Our guide to healthcare website redesign covers how we are advising clients to restructure measurement against this shift.
2. Information Agents. Continuous SEO is now a real category.
Google announced background agents that monitor things on a user's behalf even when the device is locked. Stock prices, listings, sports scores, news topics. The implication is straightforward: SEO has always been optimized for a single moment in time, the query a user types right now. Information Agents optimize for continuous tracking. If a patient sets an agent to monitor developments in their disease area, your content is either in that loop or it is not, and you may never see the user visit your site at all. The conversion happens in the agent's alert.
For pharma corporate affairs and patient advocacy teams, the question to ask procurement is simple: when an Information Agent surfaces "new developments in X" to a patient in your therapeutic area, what does it cite? If the answer is we do not know yet, that is the Q3 work. Our work with XDS Health is increasingly focused on exactly this question.
3. Personal Intelligence goes global
Personal Intelligence, the framework that lets Google pull from a user's Gmail, Photos, YouTube history, and Search history into responses, is rolling out in nearly 200 countries and 98 languages. It used to be a US-only beta for Pro and Ultra subscribers. Now it is the default for most of the world.
Two users with the same query will see substantively different results. The clinician whose Gmail is full of journal alerts gets a different answer than the patient whose Gmail is full of support-group threads. Anonymous-cohort SEO modeling stops being a complete picture. For HCP-targeted content especially, the question shifts from how do we rank to how do we get into the signals that mark a user as a clinician in the first place. That means real authorship, real institutional affiliations, real citation patterns in places Google trusts.
4. The new AI search box
Longer queries, follow-up questions, multimodal input. The change most users will notice first, with the least direct SEO implication. Query log analysis becomes more valuable. Head-term tracking becomes less useful. If your reporting is still built around a list of 200 priority keywords, this is the year to move to topic-cluster and entity-level reporting. Our HIPAA-compliant GA4 setup guide walks through the measurement side.
5. Universal Cart and AP2
Agentic commerce is real and rolling out, but it is the announcement least relevant to most of our work. The bulk of our portfolio is not transactional e-commerce. Worth tracking for the medical device and consumer health teams that are transactional. Not a Q3 priority for the rest.
The May 2026 core update is the noise behind the signal
Google started rolling out the second core update of the year on May 21. Two days after the keynote. No companion blog post, no stated goals, rollout window of up to two weeks. Marie Haynes connected the timing to Gemini 3.5 Flash now powering Search AI features. Other SEOs suggested it could target sites that have been over-optimizing for AI citations.
Whatever the actual target, the timing means any ranking movement you see in Search Console over the next two weeks will overlap with multiple variables. The core update, the Gemini 3.5 Flash rollout into AI Mode, the redesigned search box, and the new optimization guidance setting different expectations across teams. Your baseline is the weeks before May 21. Compare against performance after the rollout finishes. Wait at least a full week of post-rollout data before drawing conclusions. Anyone reporting clean attribution on this update before mid-June is guessing.
Why this hits healthcare and life sciences differently
Most I/O coverage you will read is written for e-commerce and B2B SaaS, where the entire stack from query to content to conversion lives in one company. Our work is different. We build for industries where:
- Clinical accuracy is non-negotiable. Generic AI-extracted summaries are a regulatory risk for pharma and a malpractice risk for clinical platforms. Non-commodity, expertise-backed content is not a nice-to-have. It is the only defensible position. See our work on FDA social media guidelines.
- Audience splits matter. HCP versus patient versus caregiver versus investor content has different evidentiary standards, different gating, and different visibility expectations. Personal Intelligence will only sharpen these splits.
- Most of our clients are not selling a product on the page. Universal Cart and agentic commerce are interesting and mostly not relevant. Information Agents, generative UI, and Personal Intelligence are. Read the coverage with that filter on.
- Authorship is becoming infrastructure. Google's optimization guide names first-hand experience and non-commodity content as the largest visibility levers. For our space, that translates directly to verified clinician authorship, clearly attributed institutional affiliations, and original data from clinical or scientific work. Not borrowed authority. This is the entire premise behind biotech brand building the right way.
- AI itself is part of the work now. Our applied AI studio, XDS.ai, and our medical device branding partner Hue Rx, both spend time on what AI can and cannot do credibly inside regulated content. The places it cannot are exactly where brands need to keep showing up.
Three things we are telling clients this quarter
Stop buying AEO and GEO as separate disciplines.
Google's own guidance says they are not. Agencies still selling "GEO services" as a line item separate from SEO are going to look exposed by the end of the year. The work is integrated. Reporting should be too.
Defend structure on reader-experience grounds.
Question-as-H2, answer-first paragraph, definition callouts. Good information design for clinicians, patients, and caregivers who skim. It happens to also help retrieval systems. Lead with the reader. The retrieval benefit is a bonus.
Build for the audit you cannot predict.
The Search/Lighthouse contradiction will not be the last one. Different Google teams will keep publishing guidance that points in different directions because they are optimizing for different surfaces. The defensible move is a clean technical foundation, real expertise on the page, and structured authorship metadata. The things that read well to every audit, human or otherwise, without needing a separate file or a separate strategy.
Work with XDS
XDS is a digital agency built for healthcare and life sciences. We design, build, and optimize websites and digital experiences for medical device manufacturers, biotech, clinical platforms, and pharma corporate affairs teams. If you are wrestling with what generative search means for your brand's visibility, get in touch. You can also see our services, our work, or our life sciences arm XDS Health.