Every AI search optimisation strategy rests on an assumption so basic it is never stated: your content has a chance of being retrieved.
For a significant category of queries, that assumption is wrong.
The Architecture Nobody Is Talking About
In March 2026, Baidu Search published an engineering paper co-authored by 24 practitioners documenting the internal architecture of a production AI search system running at industrial scale. This is not a research proposal or a theoretical model. It is a description of how a major AI search engine actually works, written by the engineers who built it.
The paper describes a four-agent collaborative architecture. A Master Agent assesses every incoming query and determines which of three agent configurations handles it. If the query is complex enough to need decomposed sub-tasks, a Planner Agent maps them into a Directed Acyclic Graph — a dependency-ordered sequence where some sub-tasks must complete before others begin. An Executor Agent runs each sub-task by invoking external tools: web search, calculators, external APIs. And a Writer Agent synthesises everything into the final response.
The paper formally adopts Anthropic’s Model Context Protocol (MCP) as the standard tool interface for the Planner-Executor pipeline. AI search at production scale is no longer a retrieval system with a language model attached. It is a reasoning system that invokes retrieval as one of several possible tools.
What makes this architecture consequential for brand strategy is what happens before any of it kicks in. The Master Agent’s first job is not to retrieve anything. It is to decide whether retrieval is necessary at all.
The Decision Before Retrieval
Every incoming query is classified by the Master Agent and routed into one of three configurations:
Writer-Only: Simple queries answered entirely from the model’s internalised training knowledge. No web search. No retrieval. No external data fetched. The model writes the answer from what it already knows.
Executor-Inclusive: Moderately complex queries requiring one piece of external information. A single retrieval step runs, then the Writer synthesises the result.
Planner-Enhanced: Complex queries decomposed into a directed graph of sub-tasks, each assigned a specific tool, executed in dependency order.
The paper gives an example of a Writer-Only query: “What is the name of Emperor Han-Wu?” The answer is in training data. Retrieval adds nothing, so no retrieval occurs. (source: Towards AI Search Paradigm.md)
The example is deliberately simple. But the principle extends to a much larger class of queries: any question where the Master Agent assesses that the model’s training knowledge is sufficient to generate a confident, accurate response. For many everyday brand queries — what a company does, where it sits in a market category, what it is known for — that assessment is made before a single URL is fetched.
What Writer-Only Actually Means for Brands
The two-stage citation pipeline — retrieve a pool of URLs, evaluate candidates, open selected pages, cite — applies only to queries routed to Executor or Planner mode. Writer-Only queries never reach the pipeline.
No URL is fetched. No content is read. No website is visited. If your brand is relevant to the answer, the only way it appears in the response is if it is already embedded in the model’s training knowledge.
This is not a fringe category. Simple factual queries — product categories, brand comparisons, general “what is X” questions — are precisely the everyday queries where brands most want to be represented accurately. They are the queries buyers ask early in the decision process, before they move to the specific comparison and evaluation questions that trigger multi-step retrieval.
The implication is blunt: a significant fraction of the queries that most matter to a brand are answered before any retrieval occurs. Content strategy, earned media, search rankings — none of it matters for those queries. The training-data channel is the only channel that exists for them.
The Evidence for How Deep the Bias Runs
Researchers at the University of Toronto tested this empirically. In a 2026 study of 1,000 queries across five AI systems, they ran perturbation experiments on brand ranking queries: randomising retrieved evidence, restricting the model to only provided snippets, swapping entity names across documents. For well-known brands — major automotive, consumer electronics, household names — the rankings barely moved.
The Kendall τ correlation between normal and perturbed rankings for popular entities was 0.911. Near-perfect consistency, regardless of what was done to the retrieved evidence. The model’s internal representation of those entities is so stable that retrieval barely influences it. (source: Navigating the Shift A Comparative Analysis of Web Search and Generative AI Response Generation.txt)
For niche entities, the same perturbations caused rank deviation to roughly double. Strip the retrieved evidence, and niche entity rankings become unpredictable. The model has no stable prior — it follows wherever the retrieval leads.
The citation miss rates make the same point in a different way. For popular brands in the automotive category, the study logged how often a brand appeared in an AI response without any supporting retrieved snippet:
- Toyota: 6% citation miss rate
- Honda: 3%
- Kia: 10%
- Chevrolet: 26%
- Cadillac: 58%
- Infiniti: 73%
For Infiniti and Cadillac, the majority of ranked appearances came without any retrieved evidence. The model is drawing on training priors and presenting the output with the same confidence as a fully evidenced response. (source: Navigating the Shift A Comparative Analysis of Web Search and Generative AI Response Generation.txt)
The AI’s position on Cadillac and Infiniti is not based on what any website currently says. It is based on whatever snapshot of those brands was encoded in training. If that snapshot is outdated or incomplete, no amount of current content investment will correct it until the next training run.
How AI Learns Without Citing: The Reddit Paradox
Understanding why training-mode representation is so stable requires understanding where it comes from. And here the data from the citation channel reveals something counterintuitive.
ChatGPT draws URLs from five distinct retrieval channels when it does browse the web. The citation rates by channel:
- Search: 88.46%
- News: 12.01%
- Reddit: 1.93%
- YouTube: 0.51%
- Academia: 0.40%
Reddit is pulled at massive volume — over 16 million data points in one study of 1.4 million prompts — and cited at a rate of under 2%. The research team concluded: ChatGPT uses Reddit extensively to understand topics, gauge consensus, and build background context, then cites more institutional sources when formulating its response. The phrase they used: “It learns from the crowd, then cites another institution.” (source: Why ChatGPT Cites One Page Over Another (Study of 1.4M Prompts).md)
This is the mechanism behind training-mode knowledge. The sources that shape what a model knows about a brand are not the same as the sources it cites. Community discussion — Reddit threads, forum posts, Quora answers, industry discussions — builds the model’s tacit understanding of what a brand represents and where it sits in a category. That understanding then governs the Writer-Only response.
A brand that exists primarily in press releases and well-structured product pages may have excellent citation-channel signals and almost no community-level signal. For Executor-mode queries, it may be cited correctly. For Writer-Only queries, the model has little to draw on.
The Adversarial Layer: Why Inconsistency Is Penalised
The Baidu paper describes a further architectural decision relevant to brands with fragmented training-mode presence. The system includes adversarial robustness training: a separate agent deliberately introduces noise, fabricated content, and misleading documents into retrieval sets during training — and the Generator is trained to remain accurate despite these attacks.
The effect is designed to protect against misinformation. The side-effect is structural: the robustness layer penalises inconsistency. A brand whose training-data profile is fragmented — different descriptions on different platforms, inconsistent entity naming, contradictory claims across sources — is more likely to be treated as unreliable noise by a system trained to reject exactly that pattern.
Adversarial robustness and brand consistency are the same problem, viewed from different angles. The AI system is not being selective about your brand in particular. It is applying a generalised filter for signal quality. Brands that pass the filter — consistent entity naming, corroborated claims, cross-platform coherence — are treated as reliable. Those that do not are filtered out, regardless of the volume of content that exists about them. (source: Towards AI Search Paradigm.md)
The Two-Channel Strategy
Most brand AI visibility work is focused on the citation channel: earning coverage in sources AI engines retrieve, ensuring pages are parsable and semantically structured, maintaining search rankings so content enters retrieval pools.
That work is necessary for queries that trigger Executor or Planner mode. For Writer-Only queries, it does nothing.
The practical split:
Citation-channel queries (moderately complex, multi-step, time-sensitive, comparison-focused): invest in fresh earned coverage, structured extractable content, search rankings, consistent publication in trusted sources. These queries trigger retrieval. What you publish today can influence the response in the next query cycle.
Knowledge-channel queries (simple factual, category-level, definitional, early-funnel): invest in Wikipedia presence, Wikidata records, consistent entity naming across platforms, long-cycle community discussion, and persistent authoritative mentions that will enter future training runs. What you do today has effects measured in months.
The two channels are not alternatives. They serve different populations of queries. A brand invested in only one is invisible in the other.
The Timing Asymmetry
The timing problem is where most brands get into trouble, because it is counterintuitive.
Citation-channel work can have visible effects within days. A new piece of coverage in a trusted outlet can enter the retrieval pool in the next query cycle. The feedback loop is short.
Knowledge-channel work compounds slowly. Training data has a cutoff. Future training runs will incorporate new information, but the cycle is measured in months to years, not days. Content published today does not update the model’s training-mode knowledge today.
This creates an asymmetric urgency. When a brand realises that Writer-Only queries are generating wrong or incomplete answers, the correction cycle is already months away. The brands with accurate, stable training-data representation built it continuously — not reactively, not in response to a problem that had already emerged.
The practical implication: knowledge-channel investment should start before you need it. If a brand is not actively building its training-mode presence, it is not holding steady. It is falling behind as models continue to train on an information environment the brand is not actively shaping.
The Invisible Query
The question for any brand is: what percentage of the queries relevant to your category are answered in Writer-Only mode?
There is no direct way to measure this from outside the system. But the architecture tells you where to look. Queries that are simple, factual, and stable — definitional questions, brand category placement, market positioning — are the strongest candidates. They are the questions buyers ask early in the decision process, before they reach the specific comparison and evaluation queries that trigger retrieval.
The query that never reaches your page is not the rarest kind. It may be the most common kind. And for most brands, it is the one nobody is optimising for.
Date: 2026-09-03
Leave a Reply