AI Mode runs multiple related searches in the time it takes you to run one. It returns a single synthesized answer before you've seen any of the underlying process. That's not an incremental upgrade to Google Search. It's a different product, and the post-search Google era it points toward runs on different rules than anything the web has operated under before.
Google's Search Senior Engineering Director Dounia Berrada describes the technique as "fan-out" in an interview posted on Google Blog in early March: the system identifies every sub-question implied by a complex query, fires them all simultaneously, reads through the results, and delivers one cohesive response with relevant links, all in seconds. The architecture pairs Gemini as the reasoning layer with the Search and Lens retrieval backend — the "brain" operating on top of the "library," as Berrada puts it.
That framing is the real story. Not because it's a clean break from what existed before, but because it shifts the decisions that shape what you see, what gets clicked, and what gets monetized from a visible ranked list into a process the user never observes. Visual AI Mode is the primary lens here by design: it makes the architecture concrete in a way that abstract talk of "AI search" doesn't.
What fan-out actually does
The clearest illustration is a furnished room. Point AI Mode's camera at one and, instead of prompting you to search for the lamp, then the rug, then the chair as separate queries, the system breaks the image into individual objects, runs visual searches for each element in parallel, and returns a synthesized response with links. Google's blog describes exactly this scenario: what previously required multiple individual searches now requires none.
The process is also conversational. Start with plain text "visual inspo for work outfits," get a set of results, then point to something specific: "show me more options like the second skirt." The system takes that selected image and begins the fan-out process from there, per the same Google interview. Each exchange re-anchors the retrieval process on whatever the user has just indicated.
The ambition Berrada describes goes well beyond object identification. The goal, as she frames it, is moving from "What is this one thing?" to "Explain this entire scene to me."
Reverse image search through Lens has existed for years. The meaningful departure in fan-out isn't the visual retrieval — it's the parallel decomposition step. The system infers what sub-questions a complex query implies and runs them without being prompted. The user no longer decides what to search for next. That's an architectural choice, not a feature addition.
Why the Gemini-plus-Search combination is harder to copy than it looks
The architecture has a specific structure: Gemini analyzes the image and decides which retrieval tools to invoke, while the visual search backend supplies what the reasoning layer draws from. Billions of indexed web results, sitting underneath a multimodal reasoning model that already has years of visual expertise built into Lens, per Google's account. Neither piece alone explains the competitive position.
LLM-native competitors have capable reasoning layers, and some have real-time web access. ChatGPT Search and Perplexity can do a version of what fan-out does. What they can't replicate quickly is the depth and specificity of Google's indexed retrieval, built across decades of crawling and user-intent signal at a scale no competitor has matched.
Google's bet, as Berrada describes it, is that Gemini's multimodal capabilities benefit directly from the visual expertise built into Lens over the years. A strong reasoning layer on deep retrieval, the argument goes, outperforms a strong reasoning layer on real-time crawling alone.
The DOJ antitrust remedies exhibit list, published last year, includes presentations titled "Exploring Gemini in Search" and "Search GenAI <> Gemini v3" from mid-2024, along with a Google Board of Directors Search update from the same period. A separate entry in the same DOJ record is an OpenAI internal presentation titled "Why are we solving 'Search'?" from June 2024.
To be precise about what that evidence actually establishes: those titles confirm these topics existed as formal internal documents at both companies during the period Google was defending itself in federal court. The contents of those documents, and any conclusions they reached, are not in the public record.
The counterargument to the infrastructure moat is real and shouldn't be dismissed. If the synthesis layer handles the final output, the marginal value of retrieval depth shrinks. A capable enough model with a fast enough index may approximate the advantage without replicating the underlying infrastructure. That question remains live and unresolved.
Which leads to the sharper problem. If the system now manages queries invisibly, if users see a synthesized answer rather than a ranked list, the practical question is who loses visibility and control when that happens.
What the post-search Google era means for users, publishers, and advertisers
The before-and-after looks different depending on where you sit.
Users get a genuine convenience gain. A visual search that previously required several separate queries and personal judgment about which results to trust now requires one submission. The tradeoff is opacity. The system no longer shows what was searched or what was ranked.
Berrada describes the design goal as building something that understands the "why" behind a visual query, inferring intent from context rather than just matching an image to a result, per the Google blog. That's a consequence of AI Mode's design specifically; it doesn't describe every Google Search surface. But within the fan-out model, the relevance judgment shifts toward the algorithm. That's an inference from the stated design goal, not a claim supported by usage data.
Publishers face a structural tension the mechanics make hard to avoid. Before, a page that ranked well for a query had a direct path to a click. After, the synthesis layer may draw on that page's content without directing the user to it. The publisher's work feeds the answer; the referral doesn't follow automatically.
Whether that's damaging or manageable depends on how widely AI Mode gets adopted and how synthesized responses handle attribution — neither of which is quantified in any public data. That Google was paying attention to the tension is at least partly documented: the DOJ exhibit list includes a presentation titled "Search (incl SGE) Publisher Controls" from April 2024, per the antitrust record. The title confirms a formal internal document on this topic existed. What it concluded isn't public.
Advertisers face a different structural problem. Keyword-based search ads have historically worked by attaching to terms the user supplied. In a fan-out response, that anchor doesn't exist in the same form.
The system decomposes intent and runs multiple sub-queries behind the interface, none of which the user typed. Where ads attach in that model, and at what rate, is genuinely unknown. The DOJ exhibit list reportedly includes a presentation titled "Search Ads/AI" from September 2024, entered into the antitrust record. Same caveat: the title confirms the document existed; the contents aren't public.
The honest summary: the mechanics of the shift are clear; the commercial consequences at scale are not. Anyone presenting confident click-through projections or publisher traffic forecasts right now is running ahead of the available evidence.
What has already changed, and what hasn't been decided yet
At the level of design intent, the transition is unambiguous. The sequence of typed queries, ranked results, and deliberate clicks is being absorbed into the system. Gemini decomposes the query, Lens and Search handle retrieval, and the user receives a synthesized answer. That's what Google describes, and it represents a genuine structural departure from keyword-based retrieval, not an incremental feature update.
The DOJ exhibit list adds useful context without resolving the commercial questions. Presentations on Gemini-Search integration, publisher controls, LLM impact on search usage, and advertising in AI-driven results were all entered into the antitrust record as formal internal documents from 2023 and 2024. Those titles tell you the topics were being worked on internally, during the same period the federal antitrust case was active. They don't tell you what was decided or what the outcomes were.
The next question worth watching is whether citation placement becomes the new keyword position. If publishers and advertisers can't reconstruct their value from click-through data, the logical response is to optimize for visibility inside synthesized answers rather than ranking in a list.
That would make appearing in the answer the goal, not ranking above the fold. Whether Google builds explicit tools for that, or whether it emerges as informal practice through prompt engineering and structured data, will shape how content gets produced and funded on the web for years. That's the problem the fan-out architecture creates that no one has solved yet.
Comments
Be the first, drop a comment!