What Google Is Actually Building. Anatomy of a Platform Strategy in the Age of the Agent
A forward-looking analysis, June 2026
For twenty-five years, Google sold a simple promise. You type a question, ten blue links appear, you click, you leave Google. The transaction was clear. The user got an answer, the publisher got a visitor, the advertiser got an auction, Google got a cut. This tacit contract funded half the open web. It is dying. And what is being born in its place, judging by the technical artifacts the company ships into its systems every week and by the announcements at Google I/O 2026, does not look like an improved search engine. It is an operating system for human intent. An infrastructure for delegated action. A layer of intermediation that slides between the user and everything else, and wants to keep them there.
The central question is no longer whether Google will survive the AI revolution. With 1 billion monthly active users for AI Mode, 2.5 billion for AI Overviews, and $110 billion in revenue in Q1 2026, the company is posting all-time-high search growth. The question lies elsewhere. What is Google actually building? What world is it making? And who, in the end, will capture the value?
I. Where we come from: an archaeology of the search engine
To understand the ongoing rupture, you need to dig down into the sedimentary layers of the web. In 1995, Yahoo organized information through human-curated directories, category by category. AltaVista, launched the same year, countered this editorial logic with a massive full-text index. The promise was new, the noise immense. In 1998, Sergey Brin and Larry Page published The Anatomy of a Large-Scale Hypertextual Web Search Engine, the paper that introduced PageRank. The idea was elegant: let the web graph evaluate itself, treat each link as a vote. For years, the company imposed the ten-blue-links paradigm as the neutral, almost democratic embodiment of access to knowledge.
The contract starts warping in 2007. On May 16, Marissa Mayer unveils Universal Search, which introduces verticals (images, videos, news, maps, books) into the main results page. Sergey Brin calls it “essentially the largest revision we’ve made in the past two or three years.” The SERP stops being an index and becomes a composite stage. On May 16, 2012, exactly five years later, Amit Singhal announces the Knowledge Graph under the slogan “things, not strings.” Google no longer just searches for character strings. It manipulates entities.
As the algorithm thickens (Hummingbird 2013, RankBrain 2015, BERT 2019, MUM 2021...), the share of zero-click queries explodes. Studies put the number near 60% of searches ending with no click.
In May 2023, Google I/O unveils the Search Generative Experience (SGE). A year later, SGE becomes AI Overviews. In March 2025, AI Mode launches under Liz Reid’s leadership. Within a year, AI Mode surpasses one billion monthly users. Google speaks of “AI search, through and through.” The 1998 contract has been silently torn up.
What is happening needs to be named: at each transition, Google disintermediated something. Yahoo’s directories first. Then the specialized publishers that Universal Search folded into the SERP. Then Wikipedia, absorbed by the Knowledge Graph. Then comparison and definition sites, absorbed by featured snippets. Today, it is the generic publisher being disintermediated. Every study shows Google Search traffic to publishers falling, hitting small publishers harder than large ones. Digital Trends saw its traffic collapse by 97% and laid off most of its staff. Nicholas Thompson, CEO of The Atlantic, told his teams to plan for Google traffic “dropping toward zero.” The phrase “Google Zero” has entered the vocabulary.
Google’s storytelling to publishers does not change. When the HCU rolled out, Google talked about “hidden gems.” Now it talks about “non-commodity content.” The wording shifts, but the narrative stays the same: a smokescreen. The reality today is that ranking has never meant so little. The classic execution levers (position, snippet, CTR...) are increasingly replaced by upstream signals (entity, freshness, personal context, anchoring in model weights...). SEO is now an exercise in influencing opaque systems that decide, downstream, whether you will be cited.
II. What decompiling Google’s apps reveals
To understand where Google is headed, listening to its keynotes is not enough. You have to open the binaries. Between February and May 2026, we decompiled and compared several versions of the Google App (AGSA), the Gemini app, AI Core, and Chrome Canary, totaling more than 135,000 reconstructed Java source files, analyzed class by class. The goal is not auditing for auditing’s sake. It is reading, in the code that arrives on phones before it is announced on stage, the architectural choices that betray the real strategy. We found four shifts whose scale exceeds what public announcements suggest.
Search tips into conversation
Between the Google App versions we compared (from v17.5 to v17.26), AI Mode-related files go from 42 to over 150, and associated UI resources from a few dozen to nearly 300. AI Mode is no longer an experimental feature tucked into a corner of the product. It is the skeleton of the new engine. In parallel, an “Ask” button appears at the bottom of the classic results page. We reconstructed it visually: the button, anchored at the foot of the screen, materializes the slide from query logic to conversation logic.
Google confirmed the turn at I/O 2026, calling its search bar redesign “the biggest change in twenty-five years,” now powered by Gemini 3.5 Flash. The old “Search...” label gave way to “Ask Google.” The code reveals the underlying mechanics of this “Intelligent Search Box.” Ten subtypes of Prompt Expansion now coexist in the suggestion pipeline. The framework distinguishes context-anchored reformulations (Anchor) from floating popup proposals (Overlay), zero-prefix suggestions triggered before the user types a single character, and prefix expansions activated mid-typing. This is no longer keyword autocomplete. It is a query co-writing system between user and model, which reformulates, extends, and anticipates intent in real time. The search bar is no longer a passive input field. It is a dialogue initiated by the machine.
The mechanism goes further than query reformulation. Chrome’s code contains a system of 28 contextual skills pushed from Google’s servers to the browser, page by page. Each skill is a complete package: a name, a category, and a 2-3 KB system LLM prompt. That prompt is the instruction set the model follows to act on the page’s content. On an e-commerce site, Chrome activates “Buying advice,” “Summarize reviews,” “Sizing help.” On Wikipedia, “Explain simply,” “Build timeline.” On Google Docs, “Executive summary,” “Correct grammar.” On a calendar, “Prep for today.” The browser does not just know where you are. It already knows what it can do for you, and it knows it page by page.
On the Chrome side, the same mutation is visible. In Chrome Canary 150, we identified five operating modes for the omnibox (Search, AI Mode, Image Generation, Canvas, Deep Search) where only one existed three months ago. The omnibox is no longer an input field. It is a multimodal router toward five types of intent resolution, each with its own suggestion pipeline and performance metrics.
The convergence does not stop at Chrome. In Google Maps, the “Ask Maps” entry point, which existed only as a single subtype in the previous version, deploys into a full conversational system: prompts, fallbacks, query reformulation. Maps becomes a full AI Mode surface. In Lens, “unified” layouts organize the fluid transition from visual search to AI Mode. Google also announced at I/O 2026 the ability to ask a follow-up question directly from an AI Overview and switch into AI Mode with preserved context. Every Google product becomes a bridge to AI Mode. Conversational logic is not a Search feature. It is a cross-cutting layer colonizing every surface, and it wants the user to never leave the agent’s orbit.
The agent wakes up
This is what the decompilation shows most clearly. In the previous version of the Google App, Agent Mode existed only as a trace: two pure telemetry files, no user interface, no human-facing label. Three months later, it is a complete product. Over 100 user-facing messages. Fourteen icons. An agent dashboard with search, filters, task creation and management. The product carries the name Gemini Spark, the 24/7 personal agent announced at I/O 2026, running in the cloud on dedicated virtual machines, including when the phone is off.
On the execution architecture side, two parallel systems coexist. The first, codenamed Dynamo in the code, is a remote web browser operated on Google’s servers: the agent drives pages, fills forms, makes purchases, all streamed as screenshots to the user’s phone. The user can take back manual control at any time, then explain to the agent what they changed. This is a human-machine feedback loop designed from the start. The second, codenamed Bonobo, is a local screen automation on the smartphone, built on Android accessibility services, which captures the display and interacts with installed apps as a human finger would. One operates in the cloud, the other on the device. Both converge toward the same objective: executing tasks on behalf of the user.
The consent strings extracted from the code are unambiguous: “it may do things like share your info or make purchases without asking.” And further: “Gemini saves remote browser data, like login details and remote code execution data.” This is no longer an assistant that answers. It is an agent that acts, in a browser you do not see, on a phone you are not touching, with your credentials.
The roadmap is in the code
A central enum named FeatureMode, which we extracted and fully documented, exposes 46 values, meaning 46 operating modes for Gemini. Some match stage announcements point for point: Deep Research, Agent Mode, Personalization, Canvas, Video Generation. Others have never been publicly mentioned. MARKETING_AGENT designates an autonomous marketing agent for businesses, a direct echo of the Google Marketing Live announcements on campaign automation. CHROME_WEB_UNIVERSAL_CART confirms that a universal shopping cart, integrated into Chrome itself, is under construction. WELLNESS_COACH points to health coaching. TRIP_PLANNER to travel planning. GEMINI_UNIVERSITY plus a LEARNING mode reveal an education offensive. MUSIC_GENERATION targets the Suno and Udio market. DESKTOP_INTELLIGENCE confirms expansion beyond mobile, consistent with Gemini Spark for macOS and the Googlebook ChromeOS-Android project. Add to these still-opaque codenames: SMITH, VESPER, AURA, HOTLINE, BLUEDOG, JANUS_TASKS, projects maturing but not yet documented.
Google’s agent will not be limited to search or productivity. It is designed to cover commerce, health, education, travel, marketing, entertainment. A systematic mapping of human intent, branch by branch, need by need.
Personalization becomes the default architecture
Last finding from our reverse engineering work, and the most consequential for anyone who depends on Google. In the on-device ranking signals we decoded (a weighted linear score across 57 signal identifiers), the weighting tilts heavily toward the temporal, the behavioral, and the personal: time of day, home-work commute, interactions over the last 7, 14, and 28 days, click-through rate maps, habit confidence index. On the server side, a new suggestion type PERSONALIZED_CONTEXTUAL joins existing re-ranking mechanisms (LLM_RERANK, HOBBES_CONTEXTUAL_RERANK). And AI Overview gets its own subtype in the suggestion pipeline: the system can now pre-identify, while the user types their query, that an AI Overview will be needed, and launch its pre-computation before the search is even submitted. The AI Overview is no longer an insert grafted onto the SERP at render time. It is anticipated upstream, within the autocompletion flow itself.
Personalization is no longer a veneer applied after the fact. It structures ranking from the first step. For SEO, the consequence is clear: there is no longer a reference SERP to observe. There are only billions of personalized SERPs that nobody, not the publisher, not the agency, not the regulator, can audit.
These artifacts, taken together, say what the keynotes leave implicit: Google is not just adding AI to its search engine. It is rebuilding its product around an architecture in which the results page is just one surface among many, and where the unit of value is no longer the query or the pageview, but the task completed by an agent on behalf of the user.
III. The on-device counterpart: from app launcher to intent router
This server-side architecture rests on an equally ambitious on-device counterpart, which RESONEO’s phone-future study, based on decoding 274 PCS manifests from the experimental AICore APK, maps precisely.
Gemini Nano went through four generations between December 2023 and April 2026. Version 4 (2 billion and 4 billion parameters, based on Gemma 4, native support for 140+ languages) marks the break that matters: native on-device function calling. The model no longer just produces text. It knows how to call tools. The chatbot becomes agent. On the phone. Without network.
This capability fits into a three-tier privacy cascade that Google formalizes. First tier: on-device, execution in Gemini Nano with zero data transmitted, latency of 30 to 50 milliseconds. Second tier: Private AI Compute, Titanium TEE hardware enclaves, stateless, with no logs, designed as an encrypted airlock between device and cloud. Third tier: cloud, Gemini Pro or Ultra, mobilized only when external data is needed. Firebase Hybrid Inference orchestrates the arbitration automatically, with a default policy of PREFER_ON_DEVICE.
The internal codenames are themselves a statement of intent. Demerzel (the secret robot-ruler from Asimov’s Foundation), Rosie (the maid from The Jetsons), Cortana (the AI from Halo), Bumblebee, C3PO, K2SO. These are the names behind Teaspoon (AI diary), Memora (on-device memory engine), Legion Kiki (memory generation via Google Search). When a company names its systems after the robot that quietly rules the galaxy for twenty thousand years, that is not engineering whimsy. It is a roadmap.
Android 17 (codename Aluminium, available June 2026) integrates agentic capabilities at the system level. The AppFunctions framework, announced in March 2026, lets every Android app expose its capabilities to an agent in natural language, via the Jetpack library androidx.appfunctions. This is, on mobile, the equivalent of the Model Context Protocol that Anthropic established on desktop and server. For apps that do not comply, Google deploys a UI automation framework that drives the graphical interface as a human would. We move from app launcher to intent router. The user no longer launches an application. They express a need, and the operating system decides which app executes. The icons on your smartphone will soon stop being the unit of attention.
On the silicon side, the Tensor G6 (Pixel 11, late 2026) confirms the same priority. The leak documented by WebProNews reveals a 1+2+4 CPU configuration (seven cores instead of eight), an IMG CXT GPU older than the G5’s, but an expanded main TPU and a nano-TPU dedicated to simple AI tasks. In plain terms, Google sacrifices a CPU core and the high-end GPU trend to free up silicon dedicated to neuromorphic processing. The stated bet is that by this generation of smartphones, on-device AI capability will be the differentiator, not 3DMark scores.
And then there is the striking fact that many analysts glossed over, but which changes the nature of the landscape itself. Apple’s Siri assistant voice will be powered by Gemini, with a model eight times larger than Apple Intelligence (150 billion parameters). Project Glenwood, Apple’s internal codename for vendor evaluation, reportedly chose Google over OpenAI and Anthropic. On Samsung, Gemini Nano is licensed for Galaxy AI. On Pixel, Gemini is native. The conclusion follows naturally: Google is no longer just a builder of AI products. It is the AI engine of the industry. The smartphone, whether built in Cupertino, Suwon, or Mountain View, becomes a gateway to the Google stack.
Gemini is no longer limited to the smartphone. The latest Google App embeds a full AndroidX XR runtime (OpenXR, eye tracking, hand tracking, foveal rendering), a complete AR pipeline for 3D (SceneViewer, Immersive Media Platform), and a dedicated SDK for connected glasses that can suggest commands from voice and visual context. On Bluetooth headsets, the “Hey Google” hotword onboarding strings now reference Gemini where they used to say Google Assistant. The substitution is underway, surface by surface, form factor by form factor. The agent is not just colonizing the phone screen. It is targeting your ear, your gaze, the space around you.
IV. From SERP to plumbing: the protocol strategy
The most misunderstood strategic pivot. Google no longer aims to own the interface. It aims to own the infrastructure. When one layer of the stack commoditizes, value migrates to another. The chat interface and the agent are commoditizing visibly. ChatGPT, Claude, Gemini, Perplexity, Grok converge. Value migrates toward the protocols through which these agents interact.
Four protocols, laid down in less than eighteen months, form the strategy. MCP (Model Context Protocol, opened by Anthropic in 2024), the standard grammar by which an agent calls tools. A2A (Agent2Agent), for inter-agent communication. UCP (Universal Commerce Protocol), launched in January at NRF by Google with Shopify, Etsy, Wayfair, Target, and Walmart, standardizing agent-merchant interaction from discovery to post-purchase. AP2 (Agent Payments Protocol), announced in September 2025 with more than sixty partners (Mastercard, PayPal, Stripe...), defining how an agent authorizes a payment on behalf of a human via cryptographically signed mandates. Google gave AP2 to the FIDO Alliance to guarantee its independence.
On April 24, 2026, Amazon, Meta, Microsoft, Salesforce, and Stripe join the UCP Tech Council. Eight months earlier, Amazon was blocking AI bots, suing Perplexity, and building its own proprietary systems. Today, Amazon sits on the governance body of the standard Google wrote. Ken Huang, in a widely read Substack essay, sums it up: “the protocol war is over. The commerce war is not.” Stripe, which had co-built the competing ACP with OpenAI, joins UCP while staying in ACP. Standardization around Google is becoming the default scenario.
When execution shifts from the human to the agent, value no longer sits with the brand that owns the buy button. It sits with the firm that owns the protocol and the trust infrastructure. AP2 demonstrates this: three signed mandates (Intent, Cart, Payment) solving the three problems of agentic commerce (authorization, authenticity, accountability). Whoever defines the legal contract for delegation controls the economy written above it. Google, managing the Shopping Graph (60 billion listings), Chrome (which CHROME_WEB_UNIVERSAL_CART turns into a universal cart), Wallet, and AppFunctions on Android, occupies four strategic layers simultaneously. This is an encirclement.
The universal cart is not a metaphor. Chrome’s code ships four blocks of classification rules that define how the browser identifies, in real time, e-commerce pages across eight categories: cart, checkout, confirmation, details, tracking, history, account, returns. The classifier uses three independent signals: regex on the URL path (/cart, /checkout, /thank-you), keyword matching in the page’s <title>, and semantic similarity via embeddings with a 0.9 threshold. A single signal suffices to trigger classification. Critical transactional pages (confirmation, tracking, returns) receive the highest relevance score. Chrome already knows whether you are in a purchase funnel, at which stage, and with what certainty. The classifier currently covers English only. No patterns for /panier, /commande, or /warenkorb. The rollout is US-first.
V. The disintermediation of the web itself
Let us put the question bluntly. If the agent acts, what is a website still for? If Dynamo runs a remote browser to buy on your behalf, if AppFunctions exposes app capabilities without opening their interface, if the query finds its answer in the SERP without a click, if the phone becomes an intent router that bypasses icons, then the web destination ceases to be the unit of presence. It reverts to an indexable resource, an information supplier whose compensation depends on the distributor’s generosity.
This is a change of economic regime. The advertising model of the open web is collapsing quietly. The signals pile up. Penske Media filed a 101-page antitrust complaint against Google in September 2025, accusing it of “cannibalizing” publisher traffic. The European Publishers Council did the same before the European Commission in February 2026. The DCN report on 19 premium publishers measures a median 10% decline in eight weeks in 2025, with weekly losses twice as frequent as gains. ChatGPT, the leading AI referrer, sent 1.2 billion visits to publishers between September and November 2025, up 52% year-on-year, but that represents less than 1% of total traffic. As AdExchanger put it: “The AI Search Reckoning Is Dismantling Open Web Traffic, And Publishers May Never Recover.”
Google knows this paradox well. The company needs a rich open web to train and ground its models. The grounding everyone talks about today is precisely the ability to cite fresh web sources in real time to reduce hallucinations. Destroying the editorial ecosystem means drying up the cognitive fuel for the models. Hence the hedging strategy: signing licensing deals (Google pays $60M/year to Reddit for access to all its content), maintaining a visible citation system in AI Overviews and AI Mode, multiplying reassuring statements (”we care passionately about the health of the web ecosystem”). But the numbers publishers have tell a different story. As Lily Ray of Amsive wrote, the shift will have a “devastating impact on the Internet.”
Google’s implicit calculation is probably this: if the traditional publisher dies, high-value content (original research, proprietary data, verified expertise, investigative journalism) will find other monetization models (subscriptions, licenses, B2B), and the long tail of generic content will be produced by LLMs, in a logic where Google captures aggregation value. This is, at bottom, every dominant intermediary’s dream: let marginal producers die and maintain a rent on the flows. But the calculation is precarious. It assumes premium producers will accept the intermediary’s terms. The Really Simple Licensing (RSL) movement, modeled on ASCAP and BMI, backed by Reddit, Yahoo, Medium, People Inc., demands a standardized compensation framework. No AI player has committed to it yet.
A moral, or rather political, question surfaces in the background. Is the open web infrastructure of general interest? If so, is its cannibalization by a private actor something that can be regulated? The Mehta ruling of August 5, 2024, in United States v. Google LLC, held that “Google is a monopolist, and it has acted as one to maintain its monopoly.” It found that Google “violated Section 2 of the Sherman Act.” The remedies phase, concluded by Judge Amit Mehta on December 5, 2025, nonetheless rejected the DOJ’s toughest demands: no forced divestiture of Chrome, no divestiture of Android, no ban on payments to Apple. Wedbush titled its analysis “Government Folds Like Cheap Suit.” The market applauded. Google lost on paper but won what mattered: it keeps its stack. The appeal is pending, with oral arguments expected in late 2026 or early 2027. Over a five-year horizon, antitrust does not appear capable of slowing the vertical unification underway.
VI. Personal data as moat and as shield
The second strategic pillar, and perhaps the most durable. No pure AI lab (OpenAI, Anthropic) can produce an agent that knows you as well as one from a company that holds your emails, your photos, your calendar, your location, your search history, your watched YouTube videos, your Google Pay transactions. Personal Intelligence, launched in early 2026, is the current expression of this advantage. And it is only an outpost. The double-lock architecture (CAN_SEE_YOUR_DAY_ENTRYPOINT), the cross-profile Calendar connector, the StreamGenerateYourDay RPC, the Memora and Legion Kiki components we identified in AICore, all converge toward a persistent personal memory, an individual knowledge graph on which the agent will reason with increasing precision.
This moat doubles as a regulatory shield. The privacy cascade (Gemini Nano on-device, Private AI Compute in stateless enclaves, cloud as last resort) lets Google claim a privacy-preserving position that its device-less competitors cannot credibly match. It is a double argument: it reassures the consumer, and it prepares the defense before the European Commission, the GDPR, the Digital Markets Act, and the AI Act. The most sensitive data is never transmitted in clear text. Inference happens in a hardware enclave or on the device. So the value extraction remains legally defensible. The positioning is solid enough that Anthropic, aware of the imbalance, launched a memory import tool from ChatGPT to Claude in 2025, in a forced portability logic. But Google’s advantage with Personal Intelligence will be hard for OpenAI or Anthropic to close.
Anonymous search is eroding. The SERP already does not serve the same result to two users in the same location. The PERSONALIZED_CONTEXTUAL ranking we identified is not an add-on module. It is the default architecture. For SEO, this is a tipping point: there is no longer a reference SERP to observe, and this is even truer with prompts.
VIII. The competitive landscape and the civilizational stakes
OpenAI bets on the frontier model and a hardware device with Jony Ive; mass production of the “AI agent phone” is announced for H1 2027, a smart speaker earlier in February 2027, against a backdrop of imminent IPO. By OpenAI’s official count, ChatGPT has “over 900 million weekly active users worldwide.” But OpenAI owns no OS, no browser, no payment infrastructure, no devices. Anthropic bets on safety as infrastructure, on MCP as open protocol, on Claude Code as a vertical moat. Apple, paradoxically, now outsources Siri’s brain to Gemini, placing it in an unprecedented position of dependency; its counter-offensive may come from Vision Pro and an in-house rebuilt agent, but the lost ground is considerable. Microsoft integrates Copilot into Office and Windows, and maintains a stake in OpenAI whose contractual future is complex. Amazon defends Alexa+ and Rufus, and ultimately joined UCP rather than stay on the sidelines. Meta plays on open source (Llama), Ray-Ban glasses, and social omnipresence.
Each player thus has a theory of the post-app, post-website world. Google has the most systemic theory. For the Mountain View company, the revolution is not a product disruption. It is a platform ascending to a higher level of generality. The search engine becomes the operating system for our intentions. Android and ChromeOS become an agent orchestrator. Chrome becomes a universal cart. The Shopping Graph becomes a global merchant index and Wallet becomes the agentic payment layer. AppFunctions becomes an execution grammar. Gemini becomes a cross-cutting layer powering Pixel, Samsung, and even Apple. The coherence is staggering. This is what makes it the most complete civilizational project of the moment.
What does this project mean for the citizen, for the very idea of a public web, for the concentration of informational power?
Three questions arise. Human agency first. When an assistant that knows you better than you know yourself proposes, buys, books, plans, preempts, where is the line between help and total delegation? The consent screen found in Agent Mode (”it may do things like share your info or make purchases without asking”) is a philosophical threshold. Google’s, Apple’s, and OpenAI’s promise is that the human stays in the loop. But experience shows that friction is the enemy of product, and that as trust builds, the loop stretches. The structural risk of a centralized AI future is not malice. It is the silent erosion of individuality.
The commons next. The open web was never a legal commons, but for two decades it functioned as a de facto one: a space where anyone could publish, be found, monetize modestly, and contribute to the collective information system. Consolidation around a proprietary agentic protocol, even when it calls itself UCP and is nominally open, shifts editorial sovereignty. Sir Tim Berners-Lee, in his book This Is For Everyone published in 2025, writes: “In the near future, our lives will be transformed by AI agents which will interact with the web and take actions to achieve specific goals.” He views the prospect with hope, but he also leads the Inrupt project and Charlie, a personal assistant that belongs to the user, not the platform. This counter-model exists. It is fragile.
Last, discoverability. If a brand or an idea is not recognized as an entity by the models, it does not exist in the intent economy. The shift from the attention economy to the intent economy changes the very nature of marketing, SEO, journalism. Informational diversity now depends on the diversity of sources that an agent deems worth citing. But LLMs converge statistically toward the most frequent, most structured, most linked sources. The risk of homogenization is structural.
IX. Implications for SEO, AI Search, and brands
Everything above draws a radically new optimization surface. The historical SEO triptych (technical, content, popularity) is not dead, but it becomes table stakes. The strategic layer moves elsewhere.
Brand as entity. In the ranking signals we identified, as in the on-device weightings, entity recognition outweighs textual relevance. A brand that is not a clean node in the Knowledge Graph, that lacks off-site authority signals (press mentions, Wikipedia, factual databases, structured schemas), does not exist as a citation candidate. Entity establishment work (claim, Google profile enrichment, Organization/Person schema, presence on sources that models cite heavily...) becomes the priority.
Third-party surfaces the models cite. Citation frequency across ChatGPT, Gemini, Perplexity, or Claude depends on brand awareness, frequency in training corpora, cross-platform presence, and freshness. For a brand, this means investing in the sources models consume: Reddit, YouTube, Wikipedia, specialized listicles, trade publications, transcribed podcasts, etc. The current GEO logic (fact density, entity density, citation readiness, answer-first structure) takes us back, paradoxically, to both the earliest rudimentary SEO techniques and the fundamentals of marketing.
Agent shopping and the Shopping Graph. With UCP, AP2, the Universal Cart, and FEATURE_MODE_CHROME_WEB_UNIVERSAL_CART, the transaction migrates into Google’s world. A brand whose catalog is not exposed via structured feeds conforming to UCP, whose inventory is not accessible to agents, whose reviews are locked inside its own site, will be invisible in the cart next door. The directive for every CTO should be to prepare for a 100% headless future: commerce logic must be decoupled from the visual front end, exposed via compliant API. The site is no longer the primary channel. It is one channel among many.
The erasure of measurement. With personalized SERPs, AI Overviews that do not distinguish their traffic in Search Console, probabilistic LLM citations, and agents operating in the background, the black box is total. Metrics must migrate from impression and click to share of citation and parametric visibility, anchored in model weights. Here too, it is a return to marketing fundamentals, and to spontaneous awareness studies, except you are no longer polling consumers but their intermediary agents.
AppFunctions and webMCP orchestration. For an app publisher or online service, exposing capabilities via AppFunctions on Android and via an MCP server on the web is no longer optional. If you are not exposed, you are not callable. This is the contemporary equivalent of the XML sitemap from 2005 or schema.org from 2011: an infrastructure layer that is invisible but decides whether you exist.
The browser that audits. Chrome’s code shows what the browser extracts for every page visited. Thirteen data types are actively collected: image compression, site information, salient images, JavaScript optimization, writing assistance, key points, content annotations, privacy policy detection, Gemini zero-state suggestions, read aloud, contextual skills, and shopping metadata. Four of these types are AI-related: key points (page summary for AI Overviews), zero-state suggestions (Gemini prompts displayed when the side panel opens), skills (the contextual actions described above), and shopping metadata (link to the universal cart). For every URL visited, Chrome queries Google’s servers and receives a structured response indicating which AI capabilities are available for that page. The browser no longer browses. It audits.
Brand as direct experience. The only lever Google does not control is the direct relationship with your audience. The pivot toward newsletters, communities, proprietary apps, subscriptions, events, is not a fad. It is the only possible fortification. Conde Nast already models Google traffic “trending toward zero.” That projection holds for every brand whose model depends on organic acquisition.
X. What is at stake in the next twelve months
Let us look at the possible trajectories.
The amplification scenario. Gemini 3.5 Pro arrives in June 2026. The Pixel 11 and Tensor G6 arrive in August. Android 17 (Aluminium) in June. Information Agents open to all Pro and Ultra subscribers this summer. Universal Cart goes live in Search and in Gemini. The inference cost of Gemini 3.5 Flash, already less than half that of competing frontier models, drops further. Sundar Pichai, in the I/O 2026 opening keynote, gave the order of magnitude: “In March we were processing half a trillion tokens a day internally across our AI developer tools... Now, we’re processing more than three trillion tokens a day.” At this pace, Google’s agentic stack will cross several adoption thresholds before any competitor has opened its equivalent.
The friction scenario. The appeal in the US antitrust case gives regulators fresh ammunition. The European AI Act imposes transparency obligations on autonomous agents. The first prompt injection attacks on AP2 cause publicized incidents. European publishers obtain a mandatory licensing status. Apple, frustrated by its dependence on Gemini, accelerates its sovereign model project. OpenAI launches its phone and captures the premium segment. Google’s stack remains dominant but stops being hegemonic.
The civilizational scenario. Massive use of intimate agents silently transforms practices. The early signs are already visible: average time spent on the open web is declining, subscription volumes for 24/7 assistants are rising, the generations entering the workforce have never known Google’s ten blue links. Brands that invested in entity, AI search, protocol exposure, and direct audience relationships will make it through. The rest will find themselves in the position of small publishers in 2025: plummeting traffic, an indeterminate future, a closing window of action.
None of these scenarios is exclusive. The most probable outcome is their combination.
To sum up, what Google is building, in one sentence
If we had to state in one sentence what Google is building, it would be this: a universal intermediation layer between human intent and its execution in the digital world, powered by a model that knows, a personal graph that understands, a protocol that handles transactions, an OS that routes, and a platform that progressively substitutes itself for the open web as the dominant infrastructure of knowledge and commerce. The search engine was only the prelude. The intent engine is the work.
This work is neither inherently bad nor inevitable. It poses a societal choice that goes far beyond the SEO community, one that legislators, citizens, publishers, entrepreneurs, and educators will have to articulate. For professionals in search, marketing, and digital strategy, this is not the moment for nostalgia about the ten blue links. It is the moment for clear-sightedness: understanding what is shifting, where value is going, and at which layers of the new stack you must now build. Reverse engineering the mechanisms the major players are assembling, patiently observing codenames, reading protocols closely, these are not academic exercises. For the next twelve to twenty-four months, they are the only truly reliable instruments for reading what Google is building, before the next keynote makes it official.
One fact technologists forget too quickly. Chess was the first domain AI conquered in full view of the world. On May 11, 1997, IBM’s Deep Blue defeated Garry Kasparov. Thirty years later, Chess.com has 250 million members and 20 million games played every day. AI did not kill chess. It made it more alive than ever. Players did not stop playing because a machine played better than they could. They kept going because what draws them to a game was never the perfection of the move. It is the opponent’s gaze, the shared tension, the pleasure of learning together.
The Harvard Study of Adult Development, the longest longitudinal study ever conducted on human life (85 years, three generations), reaches a conclusion almost embarrassingly simple for a civilization obsessed with performance: the factor that best predicts health at 80 is not income, not IQ, not genetics. It is the relationships people maintained at 50.
If the agent knows, if the agent acts, if the agent anticipates, one thing remains that no model will ever do: be someone to someone else. This is the message for brands, for publishers, for anyone building something in this new landscape. In a world where intermediation becomes total, the last inimitable asset is not content, not data, not the protocol. It is the relationship. The one you build, conversation after conversation, with the human beings who choose to give you their trust. The intent engine will optimize everything. Not the reason why, thirty years after the machine won the game, millions of humans keep starting a new one.
Aside with all the very rich and interesting insights from this article, there is a big flaw that I cannot but warn against : using references (pointers) like Demerzel, Rosie, Cortana and so on. Being a Sci-fi fan myself, I really don't think that it is a good idea to name and apply some past concepts invented in some books to nowadays technologies. Exactly like Isaac Asimov's famous 3 laws of Robotics for the robots. This was never meant to be used IRL. Misnaming things is a source of problems to come.