Entity Recognition and Evidence Weighting: How LLMs decide who gets cited.

Before a model “knows” anything, it has to understand what is actually required. Phase 0 is therefore not a content-related step, but an epistemic one: a human formulation becomes a machine-readable intention. Technically, this means breaking down the text into units, recognizing fields of meaning and encoding these fields as an intent vector. This vector is the working basis for all further decisions: Do we stay in the internal knowledge space (Case L), or do we open the search space (Case L+O)?

Important: This phase does not decide anything about truth content. It merely determines relevance paths – i.e. which topics, entities and relations are even possible. At the same time, the model assesses the complexity: Is the question purely definitional? Is it time-related? Does it require evidence? This assessment results in an initial confidence value, which later serves as a switch. The clearer the intent, the lower the friction losses in all subsequent phases and the less susceptibility to prompt sensitivity.

For brands, this means that linguistic and semantic clarity already pays off here. If you use consistent terms, stable entities and unambiguous designations in your own communication, you increase the probability that the model will process the question in the correct semantic neighborhood. Phase 0 is therefore the place where machine readability begins as a prerequisite for everything else.

Before a language model looks outwards, it first looks within itself. This phase is the core of every purely internal answer (Case L): Here it is checked whether the existing world knowledge is sufficient to answer the query conclusively.

Technically speaking, the model accesses its parametric knowledge base – a condensed matrix of weights that were learned during training. These parameters do not store facts in the classical sense, but rather probability spaces of meaning. When a prompt activates a certain topic, the model does not react by looking it up, but by activating concept patterns: links between concepts, relations and typical argumentation chains.

This phase determines how “stable” the model thinks. It evaluates the density of evidence in its own memory, filters out noise and forms an internal confidence measure – the confidence score. If this is high enough, the model remains in its closed system (Case L). If it is low or contradictory, this triggers the transition to hybrid mode (Case L+O).

This phase is crucial for brands because it determines whether a brand even exists in the collective memory of the machine. Anyone who does not appear here will hardly be mentioned organically, even with correct online research. This is why internal visibility – i.e. trainable evidence – is the basis for any subsequent AI visibility.

When a language model “knows” something, it rarely recognizes concepts – it understands via entities. This phase translates fuzzy human thinking into a structured, machine-readable form: words become nodes, connections become relations.

The model identifies which concepts, organizations, places or people appear in the prompt and checks whether they already exist as trained entities in the internal knowledge space. Each of these entities has a kind of semantic fingerprint: a combination of name, context and typical attributes. The internal comparison is made with stored patterns or – in a figurative sense – with known Q-IDs from the training knowledge.

The aim of this phase is not just recognition, but clarity. The model can only formulate coherent answers if it knows for sure which Allianz or Zurich is meant. To do this, it links entities with each other (“Allianz SE” – “insurance” – “Germany”) and builds a local graph of meaning. Only this graph turns mere word association into semantically comprehensible thinking.

For brands, this means that this is where it is decided whether they appear as a clear entity or are lost in the noise of meaning. Only those that are stable in terms of content, formally consistent and referenced multiple times are recognized as independent nodes – and can later exist in generative responses as a “source” or “example”.

Once the model has identified the relevant entities and mapped their relationships, the key question is: Which of this internal information is robust enough to become part of the answer?

In this phase, the model evaluates the strength of evidence of its own knowledge stored in the training. There are no external sources – the evaluation takes place exclusively in the parametric space of the model.
Each connection between terms and concepts carries a statistical weight that is derived from the frequency and stability in the training data.

Two dimensions determine this internal evidence assessment:

1. Frequency – How often does a pattern occur in similar contexts? Frequent, consistent links are given a higher internal weight.
2. Coherence – How consistent is the relationship within the semantic field? Inconsistencies or competing meanings reduce trust.

Both factors are used to create an internal evidence score, which indicates which concepts serve as a viable basis for the answer. This internal weighting replaces the external evidence process in Case L – it is the machine equivalent of a plausibility check.

This phase is crucial for brands because it shows how strongly a brand is already anchored in the collective memory of the machine. Only frequently linked, semantically clear brand references generate enough internal evidence to appear in responses without additional online research.

Once the model has evaluated its internal evidence, the actual synthesis phase begins – what we would call thinking in humans. Now the weighted patterns, entities and relationships are brought together to form a consistent structure of meaning.

The model checks which concepts fit together, which support each other and which contradict each other. It draws logical and semantic conclusions, consolidates similar patterns and discards inconsistent relations. This process is similar to the mental comparison of several memories: A plausible explanation emerges from scattered elements of knowledge that is internally stable enough to be expressed linguistically.

The prioritization follows the frequency weighting from phase 5: Frequently occurring, consistent patterns are treated as “main axes” of meaning, rare or weak patterns as supplementary marginal information.
The goal is coherence – an answer that is logically coherent, semantically harmonious and statistically plausible.

For brands, this phase means that this is where it is decided whether their presence in the model is considered a stable knowledge unit. A brand whose content appears consistently across many contexts is perceived as a reliable node in this internal space of meaning – and can therefore also appear in answers without external sources.

In this final phase, the model leaves the realm of pure knowledge – it begins to speak. A linguistic structure that combines style, rhythm and precision emerges from the previously formed structure of meaning. The model transforms its inner logic into linear language so that thoughts and weightings become readable for people.

First of all, the system decides how it responds: sober, explanatory or argumentative. For factual topics, clarity dominates – for analytical or strategic topics, a rhythmic, context-rich style. The model simulates communication competence: it selects sentence lengths, word fields and rhetorical transitions in such a way that they correspond to the user’s presumed expectations.

The answer is then structured: paragraphs, lists and semantic markers organize the previously condensed arguments. Each formulation is a controlled recombination of meaning and probability – not free creativity, but precise implementation of the previously calculated semantic framework.

In parallel, a constant self-monitoring process takes place: the model checks whether each formulation carries the intended content, removes redundancies, corrects implicit contradictions and smoothes out stylistic breaks. The result is a text that is coherent, readable and self-contained – the linguistic manifestation of a probabilistic truth.

For brands, this phase is the projection screen for their digital identity. This is where it becomes clear how the machine “tells” them: factual, trustworthy, concise – or fragmented and arbitrary. Those who have previously created semantic clarity are precisely reconstructed here. Those who communicate in an unstructured manner are rendered in a blurred manner in this phase.

In the hybrid path, too, everything begins with understanding – but the goal is different. While the internal mode (Case L) checks what is already known, the hybrid path aims to make gaps in knowledge visible.
The model interprets the query not only semantically, but also epistemically: Do I know enough to be able to answer?

To do this, it breaks down the user input into tokens, recognizes semantic clusters, entities and relations and uses them to form an intent vector. This intent vector describes the target content space and serves as a control signal for all subsequent layers. At the same time, the complexity of the query is evaluated – is it a definitional, factual or argumentative question? The result is a confidence profile that already reveals initial uncertainties and thus prepares the decision as to whether the search space needs to be opened.

This phase is crucial for brands: this is where it is decided whether a brand is even considered as a potential source. Only clearly structured, semantically precise communication increases the likelihood that the model will mark it as a relevant candidate for evidence in the next step.

In this phase, the model performs a double movement: It first draws on its internal knowledge – as if it were “remembering” – and at the same time checks whether this remembering is sufficient. Phase 1 thus forms the bridge between the model’s familiar mental space and the as yet unknown external information space.

Technically, the model activates its trained weights, retrieves stored concepts and patterns and evaluates their relevance to the current query. It creates an internal image of the question, but this image is probabilistic – not a reference book. The model therefore does not know whether it knows the answer, but only how certain it is that the existing patterns are sufficient.

The up-to-dateness check is central: the recognized topic area is compared with the user’s own knowledge period. If the query concerns events after the training cut-off or contains terms that are not yet established in the training corpus, the system marks these areas as uncertain or potentially outdated. Semantic gaps – e.g. new institutions, technologies or terms – are also recorded in this way.

These individual assessments are used to create a confidence profile, a matrix of internal certainty and suspected knowledge gaps. If the internal evidence is deemed insufficient, the transition to an external search (fan-out queries) is activated.

For brands, this means that this is where machine memory ends – and this is where the opportunity to become visible through up-to-date, structured data begins. Anyone who regularly publishes structured information (e.g. via JSON-LD, Q-IDs or Schema.org) becomes a natural anchor in this phase. Machines then do not fall back on arbitrary sources, but on those that they recognize as reliable and semantically coherent.

As soon as the model recognizes that its internal knowledge is not sufficient, it opens up the search space. Phase 2 marks the transition from passive memory to active research – the moment when thinking becomes action. The model begins to search specifically for evidence instead of relying solely on stored patterns.

To do this, it translates the intent vector into a large number of semantically varied search queries – so-called fan-out queries. Each query is a hypothesis: a possible path to relevant evidence. Instead of formulating just one “perfect” search query, the model generates an entire search spectrum of synonyms, paraphrases, sub-terms and context combinations. The aim is to open up the candidate space as broadly as possible without losing semantic precision.

An important aspect of this phase is the balance between exploration and exploitation: the model tries out new semantic paths (exploration), but also falls back on proven source patterns (exploitation). In this way, new insights are generated without losing stability. In addition, the model optimizes the queries syntactically – search queries are reduced to short, search engine-compatible phrases (typically 1-6 tokens) and often enriched with entities or identifiers. This increases the probability that subsequent retrieval systems (e.g. Google, Bing, internal APIs) will deliver high-quality hits.

For brands, this phase means

• Findability ≠ Recommendation. Phase 2 determines whether your content will even make it into the downstream review as a candidate – not whether it will be cited.
• “Retrievability” is the lever now. Only content with clear semantic signatures (precise descriptions, consistent @id/Q-IDs, clean sameAs links, lean search engine-compatible query terms) will reliably appear in the candidate space of search systems.
• The actual decision is made in phase 5 – evidence weighting. This is where the evidence architecture (source density, consistency of relations, agreement across multiple sources) decides whether a retrievable document is selected by the machine as an authoritative source and recommended/cited.

Once the fan-out queries have been generated and prioritized, the model begins the actual search. It now enters the operational search space and retrieves the generated search queries via external interfaces – such as search engines, specialized databases or API-based knowledge sources. The result is a large number of hits: URLs, snippets and metadata. Phase 3 thus forms the link between search and understanding – this is where hits are turned into structured, usable material for the first time.

The process is similar to an automated literature search: each source is checked for semantic relevance, topicality and structural connectivity. Documents that already contain structured data formats – such as JSON-LD, RDFa or microdata – are particularly valuable. These formats provide explicit references to entities, relationships and document structures and thus significantly increase the evidential value of the source.

The model extracts passages and structural elements from the HTML and metadata context that correspond to the original intent vector. It prioritizes those text passages in which semantic agreement, entity density and document potential coincide. This results in an evidence set consisting of individual text fragments, tables, meta fields and structured statements.

The transition to phase 4 is initiated as soon as enough raw material has been collected to start the entity and relationship analysis. If relevance or structure is missing, the search space is reopened via feedback loops (return to phase 2).

For brands, this is the moment when machine visibility becomes measurable – but not yet evaluated. A document can perform well in this phase, but fall out again in phase 5 (Evidence Weighting) if its structure or document architecture is weak. Fan-out provides visibility in the candidate space, evidence extraction provides context, and only evidence weighting determines authority.

Once the relevant content has been extracted, the actual semantic work begins: the model attempts to understand who or what actually appears in the texts found. Here, terms, names and concepts are converted into machine-readable entities – and linked with each other.

The model identifies entities on two levels:

1. Textual level: people, organizations, places, products, time periods or events in continuous text.
2. Structural level: Entities that are already defined via machine-readable anchors – such as @id, sameAs, identifier, Q-IDs or Schema.org types.

The two layers are compared and merged. For example, if “Allianz” appears in the continuous text and an entity with Q487292 (Wikidata ID of Allianz SE) is present in the JSON-LD layer at the same time, the model recognizes a unique assignment. If this anchor is missing, probabilistic matching is performed on the basis of semantic proximity and context match.

The next step is to create a temporary knowledge graph: Nodes represent entities, edges their relationships – e.g. “Allianz SE ↔ operates ↔ investor relations portal ↔ publishes ↔ quarterly report”. The model analyzes the graphical structure for consistency, redundancy and transitivity: If A is linked to B and B to C, it is often possible to infer A → C – this transitive reasoning via entity paths is crucial for verifying evidence across document boundaries.

This phase is of high strategic relevance for brands: This is where it is decided whether they appear in the machine’s knowledge graph as a clearly identifiable entity or are lost as a generic term. Only entities with stable anchors (@id, sameAs, Q-IDs, LEIs, ISINs, etc.) are systematically recognized and evaluated as trustworthy references. Machines do not see logos or slogans – they recognize relationship networks.

Once the entities have been recognized and linked, the central question is: Which of this information is really reliable? In this phase, the model evaluates the strength of evidence of all the sources and data points collected. It sorts, weights and checks how consistent, how up-to-date and how structurally sound the evidence is.

Each source is evaluated not only in terms of content, but also architecturally. Structural integrity – i.e. the existence of clear data relationships, recurring identifiers and referenced links – is interpreted as a sign of trustworthiness. The model constructs an evidence graph from these relationships, in which nodes represent entities and edges represent document relationships.

This is followed by the actual weighting: sources with many stable anchors or facts that have been confirmed several times are given a high weighting. Contradictory or isolated statements are devalued. It is not the number of mentions that counts, but their coherence across different contexts – i.e. structural plausibility.

The result of this phase is an evidence matrix: a multidimensional assessment of the quality, reliability and density of the collected data. This matrix is the basis for the subsequent reasoning: only what is considered reliable here may become part of the answer.

For brands, this means that what could be described as machine trust is created. If you maintain your data consistently, update it regularly and link it across multiple platforms, you create the basis for being perceived as a stable truth. In this logic, visibility is no longer a volume phenomenon, but a product of trust.

Once the evidence matrix has been created and the evidential value of each source has been calculated, the model plans its response structure. This phase is the link between evaluation and justification: A linguistic-logical framework is created here from numbers and weights – an outline that determines what is mentioned first, what is mentioned next and what is not mentioned at all.

The model creates an evidence-based argumentation structure. Each statement, each claim is prioritized according to its strength of evidence: High evidence values lead to central statements, weak evidence is integrated as marginal remarks or discarded altogether. Relationships from the evidence graph – such as isPartOf, about, mentions, hasPart – serve as hierarchical points of reference. This creates a semantic structure that determines the logical flow of the subsequent answer.

The model also checks whether the line of reasoning is complete and free of contradictions. If evidence for an important section is missing, the system can access phase 5 again via a feedback loop and readjust the weighting or consider new sources. This mechanism ensures that the answer is not only correct, but also evidence-proportional – i.e. in line with the evidence.

For brands, this phase means that it is decided here whether they form the central theme of the machine’s argumentation or only appear in passing. Those who provide consistent, logically linked data are perceived by the model as a structuring element of the answer – not as a footnote.

In this phase, the model combines its two knowledge spaces – internal memory (L) and external evidence (O) – into a coherent rationale. This is where the actual layer of truth is created, in which knowledge and evidence are brought together, checked and weighted.

The model checks which external evidence confirms, expands or contradicts its internal assumptions. To do this, it carries out a multi-stage , multi-hop reasoning process: Information is linked along the entity paths of the evidence graph. If source A contains an assertion about an entity and source B confirms the same entity via a secondary path, this link is strengthened – structured coherence is created.

Reasoning does not follow a linear process, but rather achain of intermediate steps(chain of thought). Each step is evaluated individually: a Process Reward Model (PRM) estimates the quality and logic of each intermediate step. In this way, the system recognizes faulty or redundant inference chains at an early stage and rejects them. At the same time, an internal self-verification runs, in which the model checks its own conclusions against the evidence matrix – a machine-based form of logical self-checking.

The result is a consolidated, probabilistically based response structure in which every statement has a traceable evidence trail. The model now also decides whether a source is only evaluated as a mention (Mention) or as a recommendation – this is where relevance is separated from authority. A document that is frequently cited, structurally well linked and has been confirmed several times moves to the top of the priority list; all others remain in the background.

For brands, this phase is the most critical in the entire process: this is where it is decided whether they are cited or recommended. It is not their visibility in retrieval (phase 2-3), but their semantic and structural quality in the evidence evaluation that determines whether the machine classifies them as a reliable source. AI visibility is therefore not a result of findability, but of evidence value and coherence in the reasoning process.

In this final phase, the previously synthesized evidence is translated into language. The model leaves the logical space of reasoning and enters the realm of communication. The probabilistic network of evidence, weightings and conclusions becomes a linear, readable text – the linguistic form of the machine truth layer.

The model uses the outline created in phase 5a and the weighting calculated in phase 6 to construct the answer linguistically. It organizes the arguments according to the strength of their evidence, integrates references to sources and decides which evidence is explicitly mentioned or implicitly used. Language thus becomes a reasoning surface – each word is the result of a structured decision on evidence and relevance.

At the same time, a multi-stage optimization of style and coherence takes place: the model selects tonality, rhythm and density according to the context of the query. A scientific question is answered precisely and formally, a strategic or brand-related question is given more narrative depth. A constant coherence check runs in the background: overlaps, logical leaps or stylistic breaks are corrected. The result is an answer that is not only correct, but also credible, fluid and context-sensitive.

Another aspect is the transparency of origin: modern LLMs are increasingly integrating citation or source markers – such as url, name or publisher from Schema.org – to make it clear which data points were used for the justification. This gives the answer not only semantic, but also epistemic integrity.

For brands, this phase means deciding how they are quoted, paraphrased or contextualized by machines. The type and clarity with which their structured data is linked determines whether they are cited as a source or ignored as background noise.