The Phenomenology of Search: From Similarity to Reasons for Trust

Similarity can find a source. It cannot verify one.

An embedding can place two passages close together because their language is related. That can help a search system find material about the same subject even when the wording differs. But closeness in a vector space does not tell us whether either passage is accurate, well supported, or worth trusting.

That distinction is the central problem in LLM-mediated search. Language models learn patterns from language, and language carries traces of how people describe the world. Those patterns can help systems find and combine relevant material. They do not, by themselves, turn descriptions into evidence.

The distinction matters to anyone trying to understand or influence search. If a system finds a page because it resembles the wording of a question, that is evidence of relevance to the system. It is not yet evidence that the page deserves to shape the answer. To see why, we need to separate what representations capture from what a search system does with them.

A project-management comparison page gives this sequence a shared setting. A search system has to find it and assess its claims. An editing assistant may later recommend changing its title and remember the outcome. Visitors arrive with different questions, which the site may try to interpret. At each stage, the system constructs an account of something it does not observe completely.

Part 1 of 3: representation and reasons for trust. The shared question is how a system can form, retain, and revise interpretations without mistaking them for reality. This article asks what an answer’s sources warrant. Memory & Agency asks how that support survives over time. The Geometry of Intention applies the distinction to interpretations of human goals.

What a learned representation can preserve

An embedding is a list of numbers that represents a word, passage, or other item. A system can compare these lists using a measure of similarity. If language about related subjects tends to appear in related contexts, a model can learn patterns that help it match different expressions of a similar idea.

This is the basic insight behind distributional representations. In work by Tomas Mikolov and colleagues, word vectors learned from text captured particular syntactic and semantic relationships. The result shows that useful structure can be learned from patterns of language. It does not show that a vector contains the full meaning of a word, or that language preserves every relationship in the world.

A mathematical sketch names the stages of mediation:

• W: aspects of the world people experience
• C: how people represent those aspects
• L: how people express them in language
• E: a learned representation of language

The proposed path is W → C → L → E. It asks whether a relationship in the world can remain recognizable after people perceive it, describe it, and a model learns from those descriptions. It does not specify how all cognition or language-model training works. Nor does it prove that distances in E reproduce distances in C. That would require us to define the relationships and distances, then test how well they carry across each step.

Writing the stages as mappings, f: W → C, g: C → L, and h: L → E, makes the proposed composition h ∘ g ∘ f explicit. The notation tracks transformations; it supplies no preservation theorem. A distinction can be lost when an experience is described, and a later model cannot recover it merely by estimating the text’s patterns more accurately. Conversely, a collection of descriptions can expose relationships no individual author noticed. Both possibilities belong in the research question.

A neuroscience study offers one bounded example of such a test. Goldstein and colleagues recorded activity in the inferior frontal gyrus of three participants as they listened to a podcast. They found common geometric patterns between representations derived from brain activity and contextual language-model embeddings, and tested whether the correspondence could predict held-out words.

This is evidence of a specific correspondence under the study’s conditions. It does not validate a complete theory of how the world, thought, language, and models relate. It does show that representational similarity can be studied with defined data and tests rather than asserted as a metaphor. That is a useful limit as well as a useful result.

Why the model learns from representations of representations

Following Sergey Levine’s analogy to Plato’s cave, we can think of language as a shadow of experience, and a model trained on language as learning from shadows of shadows. The image is useful because it makes mediation visible. A written account of an event is shaped by what someone noticed, measured, remembered, and chose to say.

The image also has limits. Language is not merely a copy of private experience. It includes measurements, shared conventions, disagreements, records, and descriptions built from other descriptions. Some writing gives us a trace back to an observation or an experiment. Some writing repeats a claim with no independent support. A model can learn patterns across both.

Aristotle’s De Interpretatione offers an older account of the relation between words, thought, and things. In its opening, Aristotle connects spoken words with affections of the soul, and those affections with things. That is a philosophical framing for the chain of mediation, not evidence about modern language models. It helps pose the enduring question: when a description is related to what it describes, what makes that relation reliable?

Here, “second-order” describes a model’s learned representation of descriptions that already represent experience or observations. Language is the intermediate record in this analogy; the model learns from that record. This is a way to mark mediation, not a universal count of cognitive levels or a claim that all models have only textual inputs.

Search has to do more than match meaning

A language model, a search index, and a generated answer can be parts of one system, but their roles differ. Training changes a model’s parameters. Retrieval selects material from a collection in response to a query. Ranking orders candidate results. Answer generation uses available material to produce a response. Assessing a source asks whether its evidence supports the claim at hand.

A concrete architecture makes some of these boundaries visible. In Lewis and colleagues’ retrieval-augmented generation study, a neural retriever accesses passages in a Wikipedia index, and a language model generates text conditioned on retrieved passages. The paper evaluates this retrieval-and-generation approach on knowledge-intensive tasks. It does not establish a general method for judging the credibility of web sources.

This gap is where relevance and credibility part ways. A page can closely match a query and still provide little evidence. Signals such as citations, expertise, publication history, and recent updates can help a person or system decide what to inspect. Each is a clue, not a substitute for checking whether a source supports the claim being made. Several credible-looking signals can surround a poorly supported statement.

The same distinction applies to generated answers. Retrieving a source does not guarantee that the answer represents it accurately. A system could find a careful study and turn a limited finding into a universal claim. To evaluate the answer, we need to trace its claims back to sources and see whether the evidence warrants them.

Credibility also belongs to a social practice

Geometry describes relationships between representations. Credibility also depends on how communities establish and challenge claims. Robert Brandom’s account of conceptual content and discursive practice connects meaning with inferential relations and with commitments and entitlements in social practice. In plain terms, making a claim involves what follows from it and what reasons support making it.

This offers a question for search beyond whether two passages resemble one another. What is the source committed to, what supports that claim, and what would count against it? The application is an editorial interpretation of Brandom’s philosophy, not a claim that a search engine implements his theory.

For the comparison page, a claim that a title change improved clicks should lead to a measurement, a time window, and an explanation of what else changed. An institutional name or a chain of citations may help locate expertise. It cannot supply those missing observations. Sources can also inherit standing from one another while all depending on the same original report.

“Authority” here means epistemic standing: a reason to take a source seriously about a particular question. It differs from permission to access data or act on someone else’s behalf. Expertise is specific, can be challenged, and can survive a corrected error; apparent prestige and repeated agreement are much weaker substitutes for inspecting the relevant reasons.

What philosophy adds to the practical question

The cave and Aristotle help us notice the layers between experience and language. Two further philosophical questions clarify what those layers leave open: how descriptions connect to the world, and what we mean when we say a system understands.

Stevan Harnad’s symbol grounding problem asks how symbols get meaning beyond their relations to other symbols. Harnad addresses classical symbol systems, so applying his argument directly to a neural language model would need care. The grounding question still matters for search: a hundred pages repeating one claim do not amount to a hundred independent observations. A traceable measurement gives us something else to examine, though its methods and interpretation can also be mistaken.

Kant’s distinction between phenomena and noumena concerns the conditions and limits of human cognition. His terms do not map directly onto text and external reality, and a language model is not the subject of Kant’s argument. Used as an analogy, Kant helps us ask what a system can know from the form of access it has. A model given a summary should not be treated as if it had inspected the underlying evidence.

These philosophical approaches do different work. The cave makes mediation visible; Aristotle relates words, thought, and things; Harnad asks about grounding; Kant raises questions about the conditions of access; Brandom brings reasons and social accountability into view. They do not jointly prove a theory of language models. They let us distinguish successful linguistic performance from the further question of what kind of understanding it demonstrates.

For search, the immediate test remains concrete: can the system find relevant information, represent its source accurately, and preserve the evidence’s limits in its answer?

What to test before making optimization claims

A publisher can make a claim easier to assess by showing what was observed, how it was measured, when it happened, and what the result does not establish. That is good practice for readers. Whether it improves visibility in a particular LLM-mediated search product is a separate question that needs evidence.

A bounded experiment could test that question by changing one feature at a time. For example, hold the underlying claim steady while comparing a version that makes its terminology clearer with a version that makes its evidence and limits easier to inspect. Then measure separate outcomes:

1. Did the system retrieve the page for relevant questions?
2. Did it distinguish the page’s evidence from its claims?
3. Did its answer accurately represent the source and retain its qualifications?
4. Did these results hold across repeated runs, different query wording, and changes in candidate order?

The test would need to record the system, model version, prompts, collection, and date. A result for one fixed setup would describe that setup. It would not prove that the page will be surfaced reliably across other systems or future versions.

Multimodal inputs and interactive search offer further questions to test. Does a system that can inspect a chart catch a mismatch between a result and its written summary? Does retrieving a primary source improve the answer? Does asking the user to clarify the question reduce mistaken matches? More inputs and more steps can help, but each has to earn its place through evaluation.

From representation to memory

The result of this first article is a claim linked to its support: the source and version, the relevant observation or argument, the conclusion it warrants, and its limits. A representation helps locate that record; it does not replace the record.

The comparison page’s editor may then use a source to decide which title to try. That creates a new kind of evidence: an intervention, its circumstances, and an observed outcome. Part 2, Memory & Agency asks how to preserve both the original reason for trying it and the evidence produced afterward. A source claim and a locally tested lesson belong to the same history without becoming the same thing.

References

• Aristotle, De Interpretatione, part 1, for the relation between spoken words, thought, and things.
• Kant, Critique of Pure Reason, for the limits of cognition discussed here by analogy.
• Mikolov et al., Distributed Representations of Words and Phrases and their Compositionality, on learned syntactic and semantic word relationships.
• Goldstein et al., Alignment of brain embeddings and artificial contextual embeddings in natural language points to common geometric patterns, on a bounded comparison of brain and language-model representations.
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, on combining retrieval with generation.
• Brandom, Conceptual Content and Discursive Practice, on inference, commitments, and entitlement in discursive practice.
• Harnad, The Symbol Grounding Problem, on grounding symbols beyond their relations to other symbols.