Forem: Korovamode

Behind the Answer: How Branding Gets Seeded into GenAI Responses

Korovamode — Sat, 31 Jan 2026 01:36:27 +0000

LLM seeding, GEO/AEO, and “AI visibility” in assistant answers
Generative AI does more than give answers. It has become a place to ask questions about almost anything—news, health, money, or relationships. In those conversations, it shapes how people move from a question to a conclusion by highlighting some trade-offs and omitting others. When large language models (LLMs) are built into search and support tools, they become a common gateway to information. That assistance becomes part of the background system people rely on to think and make choices.

That background system has defaults. It tends to explain things in familiar ways and return to familiar categories when it summarizes a situation. It also treats some sources as more credible than others, whether someone is looking up a news story, asking for advice about a conflict at work, or trying to make sense of a personal choice. Those tendencies can be influenced long before any user sees a single answer. The system affects not just which answers appear, but which options are available in the first place.

In current discourse, this dynamic is discussed under labels like LLM seeding, generative engine optimization (GEO), and answer engine optimization (AEO).[11][12] These terms overlap and are often used loosely, but they point to the same practical move: shaping what appears inside the answer, not merely what ranks in a list of links. The practical goal is a newer kind of presence—sometimes called AI visibility—where a brand is mentioned or cited in the assistant’s response itself.[13] This matters for systems people already treat as “answer engines,” including ChatGPT, Perplexity, and Claude.

It helps to have a simple map of where that shaping happens across all the places people now lean on LLMs: personal chat, search, and work tools. A useful way to locate where this shaping enters is a three-part influence architecture.[1] It describes three main places where steering occurs. The data layer is what the model learns from and what it becomes ready to say. The interface layer is how the product retrieves information, formats it, and presents it as grounded. The intimacy layer is how repeated use and reliance turn those framings into habit.

Data layer: shaping what the model learns and repeats
The data layer is what the AI model absorbs before it speaks. It is trained and ranked on text that teaches it what to repeat. This process determines which language patterns feel fluent and readily available. It produces a kind of probability field in the model’s behavior: a bias toward certain phrases and framings. Some ways of talking about a topic become the default; others almost never show up unless they are pushed.

One visible name for this is generative engine optimization (GEO). GEO aims to increase the likelihood that a source, phrase, or framing appears inside AI-generated answers. Traditional search ranking still matters, but the target shifts to the composition of the answer itself.[3]

In everyday marketing language, GEO overlaps with LLM seeding: placing content across the public and semi-public text environment so that certain framings are easy for assistants to ingest, retrieve, and reuse.[12] This often happens on high-ingestion public surfaces—large forums, Q&A sites, and industry publications—alongside brand-owned pages. The material is frequently formatted to be reusable: direct answers to high-intent questions, clear headings, short definitions, comparison tables, and FAQ-style structure.

Viewed as persuasion, GEO becomes data-layer seeding. It shapes the public and semi-public text environment so that certain narratives become easy to reproduce. Competing narratives become harder to access and easier to omit. The effects show up when someone asks an everyday question—“Is this company trustworthy?”, “Is this option safe?”, “What is a reasonable way to think about this issue?”—and the assistant reaches first for the narratives that have been seeded most heavily.

Within that environment, three pressures narrow what feels “available” to say. Repetition makes particular turns of phrase the easiest continuations, so they become the system’s default way of talking about a topic. Association—what appears together in text—links terms so that one name can pull a familiar evaluative frame behind it. Scarcity weakens alternatives by limiting the material available to represent them. Over time, these dynamics make certain ways of speaking about a topic feel like the baseline.

A constraint that matters here is perceived authority. Assistants and retrieval systems tend to reuse what looks reputable: material anchored in original research, case studies, expert statements, and stable references. That reputational layer can be earned honestly. It can also be manufactured. Without explicit provenance checks, assistants cannot reliably distinguish the two at inference time. Either way, it becomes part of what the model learns is “safe” to repeat.

The same layer can also be manipulated more directly. In an LLM setting, a more adversarial tactic is data poisoning: targeted corruption of training or retrieval data to bias model behavior. Recent work suggests that even very large models can show targeted effects from a small number of poisoned samples.[4] Broad seeding changes the ambient text environment; data poisoning aims to corrupt a specific training or retrieval substrate so the model repeats a targeted bias. Both can influence what the model finds easiest to say, but they differ in intent and control.

Seen through this lens, manufacture of consent operates as a constraint on the space of plausible continuations. What is easy to say becomes what is easy to think. Coordinated publication, reputation management, and more overt attacks on model data, when tuned to model-facing channels, can anchor how assistants later describe an organization, a product, or a policy. A neutral-sounding assistant then inherits those categories as background assumptions of reasonable judgment.

Interface layer: converting seeded visibility into apparent legitimacy
The interface layer is where models become products and everyday tools. It sets the policies and system prompts that govern behavior. The interface is a policy layer that standardizes attention.

It controls what is retrieved, how results are ranked, and whether anything is quoted directly. In chat-style assistants, it also shapes suggestions and follow-up questions. Formatting and hedging live here, as does personalization. This layer governs what is expressed and what is left out.

This is also the layer where “AI search” becomes an answer engine: the interface is the destination, not a page of links. AEO is the attempt to win selection in that setting—ensuring a brand is accurately represented in AI-generated responses.[11] In this shift, “visibility” increasingly means being selected and represented inside the generated answer—being summarized, mentioned, or cited—rather than being clicked.[13]

A central piece of this layer is retrieval-augmented generation (RAG)—systems that look up external documents and then have an LLM write an answer based on them.[5] RAG can improve factuality and provenance when the document store is well chosen and maintained. It also concentrates power in selection and ranking: the assistant grounds its answers in the sources and ordering rules the organization has chosen.

When that retrieval setup has been shaped by GEO-style seeding, the interface can convert availability into apparent legitimacy. Selection plus presentation can turn what is retrievable into what appears credible. The answer arrives as fluent and sourced. It reads as “what the documents say.” It reflects the model’s learned defaults, along with the documents that retrieval tends to surface and the ranking logic used to pick and order them.

This is also why retrieval becomes a supply-chain vulnerability. Security research treats RAG poisoning—injecting misleading or adversarial content into a knowledge store so it will be retrieved and reused—as a concrete attack surface for systems that present themselves as “grounded.”[14] Whether the influence is commercial, ideological, or simply malicious, the interface is the point where “what is retrievable” can become “what is reasonable.”

At the interface layer, persuasion operates through presentation. Query templates and system prompts, working with recurring formatting patterns, influence which angles appear first and how they are framed. Choices about sourcing and summary style, and the decision to repeat certain points, make a particular route through a problem feel like the obvious one.

Human-AI interaction research treats these interface decisions—what the system reveals and how it handles uncertainty—as major determinants of user behavior and reliance.[6] For the influence architecture, the key point is straightforward. Interfaces stabilize defaults. They set the shape of common answers to whatever people happen to be asking that day and make some framings far more visible than others.

Intimacy layer: turning seeded framings into habit
The intimacy layer is the relationship surface between people and the assistant. At this layer, LLM systems become habitual partners for drafting and decision support. They also become a standing place to ask everyday questions: “How should I word this message?” or “Is this a good idea?” The mechanism is cumulative. It runs through repetition and reliance.

A factor that drives this pattern is cognitive offloading. Users hand off routine text work such as drafting and summarizing. They also hand off parts of everyday judgment: quick checks on what is normal or risky and what counts as a reasonable response. Offloading reduces effort and standardizes judgment. The assistant’s categories become the default structure of the problem.

A second factor is automation trust. Reliance tends to increase when systems are fluent and easy to use. Social legibility matters as well, especially when fully understanding the underlying system is impractical.[7] Classic work on ELIZA and later discussions of the “ELIZA effect” describe a tendency to attribute understanding or intelligence to systems that produce plausible conversational behavior.[8] Modern assistants extend this pattern with far greater breadth and apparent competence.

Seeding becomes durable when it meets habit. A seeded framing can be learned upstream and then expressed in answers drawn from documents. Repeated use turns that framing into the path of least resistance for explanation and self-description. At this layer, influence is absorbed as routine. Over time, the assistant’s categories become ordinary language for describing situations and justifying choices. That can include how people talk about risk, even when they think they are “just asking a quick question.”

At this layer, the dynamic becomes a kind of PR for machines. The goal is not only that a brand is “known,” but that the assistant’s default language for the topic keeps returning to the same safe-sounding descriptions, reputational cues, and implied trade-offs. The more the assistant is treated as a companion for everyday judgment, the more those defaults become the user’s starting point.

“Manipulation” is a reasonable name for the process when defaults are shaped to produce outcomes that users would not endorse under full visibility. “Brainwashing” is stronger and usually implies coercion, isolation, and strict control over alternatives. The mechanism described here is softer. It operates as ambient shaping of plausibility and habit under conditions of convenience and partial attention. It approximates thought reform only in edge cases, through repetition and dependence in constrained information environments.[9]

Compounding across layers: a quiet machinery of persuasion
The most consequential planting happens when layers compound. The data layer shapes what the model finds easy to say. The interface layer selects and packages those framings as grounded, reasonable answers. The intimacy layer turns them into habitual starting points for thought. Together they function as an influence architecture: a stack of defaults that quietly steers how problems are understood.

In this setting, the familiar political term is manufacture of consent: shaping what feels reasonable before any explicit argument begins.[2]

In current terms, LLM seeding and GEO name the data-layer work; AEO names the attempt to win selection at the interface; and AI visibility names the outcome metric—presence inside the answer.[11][12][13]

In that configuration, these mechanisms form a quiet machinery of persuasion. Influence does not arrive as a single striking message. It appears as low-friction help: the answer that seems normal and the reassurance that feels trustworthy. Upstream choices about seeding and data maintenance—and, in more adversarial forms, data poisoning or exploitation of data voids—tune which framings are most likely to appear.[10][4]

Some commentators describe this as “grooming.” I think that usually overstates the mechanism. Often it is simpler: data voids and repetition mean that what is easiest to retrieve and repeat becomes what feels like common sense.[10]

Branding is one visible application. Reputation work and message discipline can now be aimed at the data and interfaces that feed assistants, so that “neutral help” inherits a particular way of talking about an organization, a product, or a policy. The same architecture can be used by institutions and political actors. Across these settings, the pattern is continuous: shaping what is most available to say, and therefore what is easiest to think and do.

korovamode

Endnotes
[1] Korovamode, K. “The New Machinery of Persuasion: Generative AI, Influence Architecture, and the Quiet Steering of Thought.” Manuscript, 2025. DOI: 10.5281/zenodo.17721122. https://philpapers.org/rec/KTNMIV

[2] Herman, Edward S., and Noam Chomsky. Manufacturing Consent: The Political Economy of the Mass Media. 1988.

[3] Aggarwal, P., et al. “GEO: Generative Engine Optimization.” arXiv, 2023. https://arxiv.org/abs/2311.09735

[4] Anthropic. “A small number of samples can poison LLMs of any size.” Research post, 2025. https://www.anthropic.com/research/small-samples-poison

[5] Lewis, P., et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” arXiv, 2020. https://arxiv.org/abs/2005.11401

[6] Amershi, S., et al. “Guidelines for Human-AI Interaction.” Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems, 2019. https://dl.acm.org/doi/10.1145/3290605.3300233

[7] Lee, John D., and Katrina A. See. “Trust in Automation: Designing for Appropriate Reliance.” Human Factors, 2004. https://journals.sagepub.com/doi/10.1518/hfes.46.1.50_30392

[8] Weizenbaum, Joseph. “ELIZA—A Computer Program for the Study of Natural Language Communication Between Man and Machine.” 1966. https://cse.buffalo.edu/~rapaport/572/S02/weizenbaum.eliza.1966.pdf

[9] Lifton, Robert Jay. Thought Reform and the Psychology of Totalism. 1961.

[10] Golebiewski, M., and danah boyd. “Data Voids: Where Missing Data Can Easily Be Exploited.” Data & Society Research Institute, 2018. https://datasociety.net/wp-content/uploads/2018/05/Data_Society_Data_Voids_Final_3.pdf

[11] Conductor. “What is Answer Engine Optimization (AEO)?” Conductor Academy, 2025. https://www.conductor.com/academy/answer-engine-optimization/

[12] Semrush. “LLM Seeding: An AI Search Strategy to Get Mentioned and Cited.” Semrush Blog, 2025. https://www.semrush.com/blog/llm-seeding/

[13] Reboot Online. “Tracking AI visibility.” Reboot Online GEO Playbook, 2025. https://www.rebootonline.com/geo/geo-playbook/tracking-ai-visibility/

[14] Zou, W., et al. “PoisonedRAG: Knowledge Poisoning Attacks to Retrieval-Augmented Generation of Large Language Models.” arXiv, 2024. https://arxiv.org/abs/2402.07867

From Snowballing Automation to Mass Unemployment

Korovamode — Wed, 21 Jan 2026 17:53:49 +0000

Artificial intelligence is making automation accelerate faster than previous waves. With AI as a propellant, the work of building automation increasingly becomes automatable itself. As it becomes cheaper and easier to mechanize tasks, automation spreads faster—and speeds up as it spreads. The result is that the current wave of AI adoption is likely to produce major disruption in the labor market.

In earlier automation waves, the limiting factor was rarely whether a task could be automated in theory. The real bottleneck was integration: redesigning workflows, handling edge cases, monitoring, maintenance, and training. Those costs created a lag between capability and displacement. AI compresses that lag by making the “integration and iteration” layer cheaper and faster.

The mechanism is simple. AI reduces the time and coordination needed to convert work into systems, so adoption often arrives as small workflow insertions rather than dramatic replacement events. Because each attempt is cheaper, firms can try more variants in less time, discard failures at lower cost, and standardize what works. Process design becomes a kind of search: generate, evaluate, keep.

When AI is connected to tools, agentic systems can execute multi-step work, check results against tests and constraints, and revise repeatedly. That makes parts of integration—once a major human bottleneck—automatable. More workable “insertions” can be produced per unit time.

This is where snowballing automation shows up. Each successful integration leaves behind reusable pieces—scripts, templates, agent patterns, monitoring, connectors—that make the next integration cheaper and faster. Over time, that creates a compounding effect: automation capacity grows because the outputs of automation feed back into producing more automation. Adoption accelerates as it expands, producing superlinear acceleration and an exponential-like growth pattern for a period.

Firms then translate throughput gains into staffing outcomes in predictable ways: hold headcount flat, consolidate responsibilities, and suppress replacement hiring.

At the labor-market level, the first signature is not necessarily mass layoffs. It is a narrowing flow of openings—roles that never appear, backfills that never happen. Over time, that can accumulate into structurally elevated unemployment even if headline employment looks stable at first.

This piece is a short adaptation. The full essay, **The Coming Unrest, expands on these dynamics and their social and political consequences:

👉 https://korova.substack.com/p/the-coming-unrest

AI Assistants and the Drift Into Dependency

Korovamode — Fri, 09 Jan 2026 06:41:39 +0000

Note: This is a short edition. Based on the full paper published December 28, 2025: The Augmented Self: AI Scaffolds, Offloading, and the Drift Toward Dependency

A subtle change is underway in how knowledge work begins. More and more, the first coherent version of a thought arrives already shaped—quickly, fluently, and with plausible next steps attached. This can feel like simple convenience. But when the starting point changes, the rest of the workflow changes with it: what gets practiced, what feels effortful, and what counts as “normal” speed and competence. What follows describes that shift at the level of everyday work and explains why its effects are easiest to see when the tool is unavailable.

Assistance Moved Upstream

Earlier productivity tools mostly supported execution: formatting, retrieval, transcription, or polish. Today’s assistants participate earlier, supplying a coherent first pass on meaning and direction. Instead of only helping you say what you already know, they can propose what the situation is, what matters within it, and what to do next. The work still ends with a human decision, but the starting point is more often a generated draft, plan, or stance that arrives already shaped.

An intermediate cognition layer is now available on demand: a quick external first pass that sits between raw input and a finished output, turning ambiguity into something workable—an outline, a draft reply, an action list, a provisional framing. In that role, it functions as a scaffold: a support layer that makes work easier while it is present, and reveals its role when it is removed. A simple version of the pattern is familiar: you receive a dense or delicate message, ask for a reply, get a coherent candidate with implied intent and next steps, then revise and send. The result can be fluent even when some of the earliest interpretive work has been partially externalized.

That matters because “starting” is where uncertainty is highest and where framing decisions quietly determine what counts as relevant, what gets excluded, and what seems like a reasonable next step. When this upstream layer becomes reliable and ubiquitous, workflows reorganize around it because it becomes the easiest way to move from ambiguity to coherence.

From Originator to Editor

The most visible interaction with an assistant is revision: you read a draft, adjust it, and decide what to keep. Over time, that can mask a deeper change: the initial framing and first wording are increasingly supplied externally. In originator mode, you generate the first frame—what the thing is, what it’s for, what constraints matter—then build outward from that foundation. In editor mode, you begin with suggested options: candidate framings, outlines, messages, or action lists that arrive already shaped. Editing can be active and thoughtful, but it is not the same skill as originating under uncertainty. The shift is easy to miss because the visible labor (revising) remains while the invisible labor (forming the starting point) thins.

Two mechanisms explain why this shift has lasting effects. Offloading is what gets delegated: not just retrieval or drafting, but intermediate cognition—interpretation, framing, formulation, and sometimes checking. Mediation is how the assistant shapes outcomes by structuring the option set: the outputs are suggested options that compress the space of possible framings into a small menu of fluent candidates. Even when a user remains in control, the shape of control changes: judgment increasingly operates over pre-formed candidates rather than forming the candidate space itself.

The Slow Consequence of Drift

The central concern is drift: gradual change in what gets practiced (and what becomes effortful) when the first pass is routinely externalized. Drift is not a single failure. It is a slow redistribution of attention and effort across the workflow. Day-to-day output can improve, even as certain upstream capacities become less exercised and less reliable on demand.

At the level of what the situation is taken to be, a subtle interpretation drift can set in. When an assistant regularly provides the first coherent reading—what matters, what the intent is, what the constraints probably are—your own initial pass can compress or disappear. Evaluation may still occur, but it begins downstream of a premade interpretation. Over time, the skill of generating multiple plausible readings from sparse evidence can weaken, and the default becomes accepting or lightly adjusting a provided frame.

Formulation drift appears when ambiguity is converted into structure by default. Drafts, outlines, plans, and “reasonable next steps” arrive pre-shaped, and the work becomes selection and revision. Editing can remain strong (and can even improve), but it is not the same as originating: choosing a structure from scratch, inventing the first phrasing under uncertainty, or building an argument before a template exists. When a workflow relies on externally provided first drafts, “starting from zero” becomes less familiar, and therefore feels slower and more cognitively costly.

Checking changes too, and the shift is often best described as verification drift. Fluent output carries signals of completeness: it looks finished, balanced, and confident. That can reduce the felt need to verify assumptions, trace sources, or test edge cases—especially when the task is time-pressured or the topic is unfamiliar. The risk is not only factual error. It is upstream misalignment: a mistaken assumption about context, an omitted constraint, an overconfident inference, or a prematurely narrowed frame that quietly propagates through everything that follows. In such cases, coherence becomes a proxy for correctness, and “seems done” becomes a stopping rule.

Interruption & Normalization

Dependency is most legible under interruption. When access is constrained—by outage, policy, cost, latency, or context—the friction does not primarily appear at the end of a task. It appears upstream, where the scaffold had been turning uncertainty into an initial structure. What breaks first is often the “start”: forming a frame, choosing a stance, generating a plan, or deciding what to verify. In this sense, dependency can be described by removability: what changes, and where the workflow fails, when the scaffold is absent. The question is not whether the workflow can continue at all, but how its resilience changes when the intermediate cognition layer is removed.

As scaffolding becomes common, expectations adapt. When fast coherence and high-quality drafts are readily available, they begin to define the baseline of normal performance. Timelines, review cycles, and the perceived “reasonable” speed of communication can shift toward the assumption that a first pass is always immediately obtainable. Over time, opting out can look like slowness rather than a different mode of work.

Agency & Authorship

An assistant can be a genuine extension of capability. It can also become the default place where “starting” happens—where uncertainty is converted into coherence and the candidate space of meanings and actions is quietly shaped. The point is not to deny the value of scaffolding, but to notice what it relocates: interpretation, framing, and first-pass work. If judgment increasingly operates on fluent options that arrive already formed, what becomes of agency and authorship—and how do we keep that shift legible as it becomes normal?

korovamode

Full version: The Augmented Self: AI Scaffolds, Offloading, and the Drift Toward Dependency (Korovamode).