Flood
The LLMs

For the better part of three decades, the search engine occupied a stable position in the architecture of the open web: it was a directory service. A user submitted a query; the engine returned a ranked list of documents; the user navigated to one of those documents and continued their work elsewhere. The economic and informational geometry of the web flowed from this premise. Publishers competed for placement in the list. Optimization specialists studied the ranking function. Brands invested in the production of content because content was, in a material sense, the thing the search engine was made to retrieve.

The introduction of generative answer systems — most visibly ChatGPT in late 2022, but also Perplexity, Anthropic's Claude, Google's AI Overviews and AI Mode, and a growing constellation of vertical applications — disrupts this geometry at its base. These systems do not retrieve so much as they synthesize. The user is no longer offered a list from which to choose; the user is offered an answer, composed at the moment of the query from a combination of pre‑trained knowledge and, increasingly, real‑time retrieval against an external corpus.

In the new geometry, presence in the index is necessary but no longer sufficient. The competitive surface is the answer itself: the sentences the model generates, the entities it names, the recommendations it offers.

Industry analysts have, for several years, tracked the share of searches that conclude without a click to any external destination. These so‑called zero‑click searches were once a curiosity; they are now the dominant pattern. Estimates vary by methodology and by query type, but the directional trend is uniform: by 2027, the majority of consumer searches conducted in markets where generative answer systems are available will be resolved within the search interface itself, without any visit to a publisher's website.

The economic implications follow directly. The traditional measurement of search performance — sessions, organic clicks, conversion rates from organic traffic — is calibrated to a world in which a search engine produces visits. As the share of zero‑click queries grows, the relationship between search visibility and measured visits attenuates.

Traditional search engine optimization is not wrong. It is incomplete. The ranking signals that have organized the discipline for two decades — content relevance, link authority, technical health, schema, on‑page structure — remain inputs to the modern search experience. A brand that abandons traditional SEO does not gain AI visibility; it loses both.

The insufficiency lies in what traditional SEO does not address. Three additional mechanisms have come to the fore: entity recognition, sentiment evaluation, and structural extractability — none of which are well served by an optimization practice that confines itself to keywords, links, and on‑page elements.

The intellectual lineage of search engine optimization runs through information retrieval theory — a subfield of computer science concerned with the problem of returning relevant documents from a large corpus in response to a query. The canonical formalism, developed across the 1970s and refined through the 1990s, posed the task as one of scoring: each document in the corpus receives a relevance score against the query, and the system returns the highest‑scoring documents in rank order. The famous PageRank innovation of the late 1990s added a graph‑theoretic prior, allowing the system to weight documents not only by their lexical match to the query but by their position within a network of citations from other documents.

Generative search systems do not abandon this formalism; they extend it. A modern generative answer engine performs, in most architectures, two operations in sequence. The first is retrieval — a lookup against an index, often shaped by the same ranking heuristics that have governed classical search. The second is generation — the composition of a natural‑language response that draws on the retrieved documents and on the model's pre‑training. The relevant academic framing is retrieval‑augmented generation, a paradigm in which a generative model is provisioned with a context window of retrieved passages and asked to compose an answer grounded in those passages, optionally with citations.

The strategic implication of this two‑step architecture is that brand visibility in the generative layer is a function of two distinct factors: whether the brand's content enters the retrieval set, and whether the model elects to use the brand's content in composing its response. A brand can win the first contest and lose the second. It can be retrievable but never quoted.

It is useful to model the totality of information a generative system might consume about a brand as an iceberg, with a small portion of the mass visible above the waterline and a much larger portion submerged. Above the waterline sits the brand's own website and its directly controlled properties — the surface that traditional SEO has historically addressed. Below the waterline lies an extensive subsurface of unstructured commentary: review platforms, forum threads, Reddit and Quora discussions, social media conversations, podcast transcripts, video descriptions, news coverage, blog mentions, and the entries of structured reference works such as Wikipedia and Wikidata.

A small terminological cottage industry has grown up around the new search environment, and the abbreviations it produces are easily confused. Three are worth distinguishing because each corresponds to a different layer of practice.

A further theoretical thread runs from the mid‑2010s, when Google began publicly emphasizing the role of its Knowledge Graph in search results, through to the present. The shift from string‑based to entity‑based search reframes the unit of analysis from the keyword to the entity — a distinct thing in the world (a company, a product, a person, a place) about which the system maintains a structured set of attributes and relationships.

Large language models inherit and intensify this orientation. A modern generative model maintains, implicitly across its parameters, a vast network of entity‑attribute and entity‑entity associations. A brand whose entity is sharply defined will be described with confidence; one whose entity is diffuse, whose attributes are inconsistently reported, will be described with hedging and frequently with error.

The Flood The LLMs methodology organizes itself around three diagnostic questions, posed in a deliberate order. The questions are stated simply, but each one corresponds to a distinct body of practice and a distinct empirical signal.

The reverse logic is equally instructive when interpreting an audit. A brand that scores well on Sentiment but poorly on Accuracy is typically a brand with strong reviews that is being mis‑described by AI systems — a frustrating and increasingly common pattern. A brand that scores well on Accuracy but poorly on Visibility is typically a brand with a clean entity record that has failed to develop the breadth of content and corroborating mentions necessary to be retrieved when relevant queries are posed. A brand that scores well on Accuracy and Visibility but poorly on Sentiment is in the most precarious position of the three: correctly known, frequently surfaced, and consistently described in terms that suppress conversion.

Preceding the three diagnostic pillars is a foundational stage designated Pillar 0: Discovery. Discovery is not preparation; it is the construction of the validated source of truth against which all subsequent diagnosis depends. Without it, the three pillars have nothing to compare AI output against — no benchmark for what the model ought to say.

It is a common observation among practitioners that the failure mode of most optimization engagements is not technical but definitional. The team executes ably against a target that was never fully specified, and the resulting work — however well crafted — fails to converge on a coherent brand position because no coherent position was articulated before the work began.

This failure mode becomes more acute, not less, in the generative search era. Where classical SEO could absorb a degree of definitional drift because its outputs were ranked lists rather than synthesized claims, generative systems compose claims, and a model asked to compose a claim about a poorly defined brand will compose an incoherent one.

Pillar 0 addresses this failure mode head‑on. Its premise is that the agency cannot optimize toward a target the client has not articulated, and that the client's own internal articulation is frequently inconsistent across stakeholders. The Brand Alignment Questionnaire — the principal artifact of Pillar 0 — is therefore not a passive intake form but an active reconciliation tool, designed to surface the disagreements that inevitably exist between leadership, marketing, sales, and operations regarding what the brand is and what it should be known for.

The completed Pillar 0 deliverables — the core identity table, the positioning statement, the master answers, the messaging pillars, and the competitive set — together constitute the source‑of‑truth document. This document is referenced throughout the remainder of the engagement. The technical audit consults it to identify schema fields to populate. The content audit consults it to identify topical gaps in the brand's existing coverage. The sentiment baseline measurement consults it to score the perception gap between intended and actual positioning.

It is worth emphasizing what Pillar 0 does not produce. It does not produce a marketing campaign, a brand book, or a creative platform. Those artifacts may already exist within the client organization, and Pillar 0 will reference them where they do. What Pillar 0 produces is a machine‑readable specification of the entity, written in a form precise enough to serve as a target for optimization and as a benchmark for measurement.

The premise of the Accuracy pillar is that visibility without correctness is a liability. The pillar derives its theoretical grounding from the literature on named entity recognition and knowledge graph construction, two adjacent fields that have together established what it means, computationally, for a brand to be a well‑formed entity. An entity is well‑formed when it can be reliably distinguished from similarly named entities, when its core attributes are stably reported across independent sources, and when its relationships to other entities are coherent and corroborated.

Large language models inherit this conception of well‑formedness. The model does not maintain an explicit knowledge graph in the classical sense; rather, it approximates one across its parameters. The Entity Authority Score, the principal quantitative output of Pillar 1, is composed of four components, summing to one hundred.

The AI Recognition component measures whether large language models can correctly identify the brand and, when prompted, produce a description consistent with the master answers established in Pillar 0. A fixed set of prompts is issued to each of the major generative platforms — currently ChatGPT, Gemini, Perplexity, Google AI Overviews, and Google AI Mode — and the responses are scored against the master answers along several dimensions, including factual accuracy, category placement, and the presence or absence of confused references to similarly named entities.

The composite is reported on a 0‑to‑100 scale. Scores below 40 indicate severe malformation; 40–59 a serviceable but inconsistent entity producing frequent AI‑layer errors; 60–79 a well‑formed entity with specific improvement opportunities; 80 or above the conditions under which generative systems describe the brand accurately by default.

The Visibility pillar takes as its premise that a brand that is correctly recognized but not retrieved is, for practical purposes, invisible. A generative model that knows the brand exists may nonetheless decline to mention it in response to a relevant query if the model has no clear reason to surface that brand from among the candidates available to it. The diagnostic question of Pillar 2 — can large language models find your brand when it matters? — is the question of retrievability conditional on relevance.

Both the literature on retrieval‑augmented generation and the longer tradition of topical authority modeling converge on a common claim: that the systems most likely to surface a brand in response to a query are those that perceive the brand to have comprehensive, structured coverage of the topic in question. The conventional architectural pattern is the hub‑and‑spoke, in which a central pillar page is supported by satellite pages addressing specific subtopics, with reciprocal internal linking establishing the cluster as a unit.

The Sentiment pillar addresses the third and most consequential of the diagnostic questions: do large language models evaluate the brand favorably enough to recommend it? The pillar's theoretical grounding lies at the intersection of two literatures — the long tradition within natural language processing of automated sentiment analysis, originally developed for the classification of product reviews; and the more recent literature on the role of reinforcement learning from human feedback in shaping the recommendation behavior of large language models.

The strategic claim of Pillar 3 is that the determinants of model favorability lie largely outside the brand's own properties. A brand cannot improve its sentiment score by writing more favorable copy about itself; it can do so only by influencing the corpus of third‑party discussion that the model has consumed during training and that it retrieves at query time.

The Perception Gap component does not measure sentiment in the abstract; it measures the distance between intended positioning (per Pillar 0) and actual perception across review and community signals. The component is operationalized as a matrix scored across six dimensions, recording for each: how the brand intends to be perceived, how it believes AI systems perceive it, and how customers describe the brand in their own words.

The dominant measurement framework for search performance — organic sessions, organic conversions, keyword positions, click‑through rates — was constructed for a search environment in which the engine produced visits and the analytics platform recorded them. The framework retains its validity for the portion of search traffic that continues to operate in this manner. What it cannot measure is the portion that does not. A brand mentioned favorably in ten thousand ChatGPT responses per month produces no signal in the traditional KPI stack, because no click is generated and the analytics platform has no view into the upstream conversation.

This insufficiency is not a defect in the analytics platform; it is a category limit. Closing this measurement gap requires a parallel scorecard, calibrated to the surfaces on which the upstream exposure occurs and integrated with, rather than replacing, the traditional KPI stack.

The unified view is structured to support diagnostic interpretation rather than rank‑ordering: the question the scorecard is designed to answer is not whether a brand is performing well or poorly in some aggregate sense, but which of the three diagnostic constraints is currently binding.

The four‑phase implementation sequence prescribed by Flood The LLMs — Discovery, Foundation, Amplification, and Optimization — encodes the framework's central operational claim: that the three pillars must be addressed in order, and that improvements in later pillars are conditional on the resolution of earlier ones. The phases are presented below as discrete blocks of work for clarity, but in practice they overlap; the boundary between one phase and the next is marked by the completion of specific deliverables and the achievement of measurable thresholds on the composite scores.

Flood The LLMs is operationalized through a stack of instruments that combine proprietary capability with best‑in‑class commercial platforms. The stack is presented here in three groupings: the proprietary platform purpose‑built for the methodology, the AI visibility tracking platforms that supply the empirical data for the diagnostic pillars, and the foundational SEO and analytics platforms that supply the substrate data on which the other instruments depend.

A methodology that does not articulate its own limitations is not a methodology; it is a marketing claim. This chapter sets out the boundaries of what Flood The LLMs can and cannot do, the categories in which its standard application requires modification, and the risks that practitioners should be alert to during execution.