How LLMs Store You: Parameters, Embeddings and Your Name

People imagine an AI model as a giant filing cabinet with a folder per person. The reality is stranger and more useful to understand: you are not stored anywhere. You are implied everywhere.

There is no file with your name on it

A large language model contains no database, no records and no lookup table. It is a single enormous function: text goes in, a prediction of the next token comes out, and everything it "knows" lives in the billions of numerical parameters that shape those predictions. When a model correctly says what you do for a living, it did not retrieve a fact. It completed a pattern, because during training the text of the web made that completion statistically likely. This distinction sounds philosophical. It is not. It is the difference between editing a record, which you cannot do, and shifting a probability, which you can. Everything practical in Person Engine Optimization follows from that shift in mental model.

Tokens: what your name looks like on the way in

Before a model processes anything, text is chopped into tokens, sub-word chunks from a fixed vocabulary. Common names ride through as one or two familiar tokens. Uncommon names get sliced into fragments; a surname the tokenizer has rarely seen might become three or four pieces the model must learn to treat as a unit. Neither situation is a verdict, but each has a consequence. A common name is easy for the model to read and easy to confuse with the thousands of other people sharing those tokens. A rare name is unmistakably yours, but the model only learns to treat the fragments as one person if it sees them together, spelled identically, many times. Variant spellings, stray initials and inconsistent forms literally split you into different token sequences, which is the mechanical reason the consolidation work in structuring your identity for machines matters at all.

Embeddings: your name as a position in space

Once tokenized, your name becomes an embedding: a vector, a long list of numbers, that positions it in a space where distance means relatedness. Words and names that appear in similar contexts sit near each other. If your name consistently shows up in text about maritime insurance litigation, its representation drifts toward that neighborhood, near the topics, the institutions and the other names of that world. Ask the model about your field, and the names positioned deepest in that neighborhood are the ones that surface most naturally.

This is the geometry behind a rule PEO practitioners state as strategy: you become where you are mentioned. Every article, podcast and citation that places your name beside your specialty nudges your position. Scattered mentions across unrelated topics leave you floating between neighborhoods, near everything, retrievable for nothing.

The geometry also explains a pattern that frustrates generalists: the consultant who does strategy, fundraising, hiring and product advisory, and is documented doing all four, occupies no neighborhood deeply enough to surface in any of them. The machine is not punishing range. It is doing arithmetic on context, and diluted context produces a diluted position. You do not have to narrow what you do; you do have to narrow what the public record says you do, at least until the association is strong enough to carry extensions.

Parameters: where the knowing actually lives

The associations themselves are stored in the model's parameters, the billions of weights adjusted during training. Facts about you are not written anywhere; they are smeared across huge numbers of weights that each encode a tiny fraction of many patterns at once. Three properties of this storage explain most of what people find baffling about AI and their own reputation:

• It is frequency-weighted. A pattern seen once barely registers. A pattern seen hundreds of times across independent sources becomes something the model states with confidence. One flagship profile in a major outlet matters less, mechanically, than the same three facts repeated across fifty ordinary pages.
• It is frozen at the cutoff. Once training ends, the weights ship and do not change. Corrections, career moves and new work wait for the next model, unless they arrive through live retrieval, the second door described in training data vs retrieval.
• It has no confidence meter attached to truth. The model knows what is likely, not what is verified. Fluent wrongness is native to the architecture.

How does an LLM know who I am?

Put the pieces together and the answer to this question has three tiers. If you are heavily documented, the model recognizes your name from training and completes accurate patterns about you from memory. If you are lightly documented, the model may recognize the shape of your name but hold associations too weak to say anything specific, so it either goes vague or goes wrong. If you are undocumented, you exist for the machine only through retrieval: whatever a live search surfaces at question time is the whole story. Research on generative engines backs the intuition that what gets surfaced can be deliberately influenced; the study that coined Generative Engine Optimization, Aggarwal et al. on arXiv, measured how content changes shift visibility in generated answers. And the audience deciding against these tiers is no longer niche: ChatGPT alone reached roughly 900 million weekly active users by February 2026, per Similarweb's generative AI statistics.

Attention: why the context window beats the weights

One more mechanism completes the picture. At question time, a transformer's attention layers weigh everything currently in the context window, the conversation plus any retrieved documents, and that in-context material dominates whatever the weights vaguely remember. Feed the model a paragraph stating your current role and it will use that paragraph, even if its training memory holds the stale version. This is the architectural reason retrieval can override memory: a fetched page about you is, for the duration of that answer, the loudest signal in the room.

The practical consequence is enormous for anyone who is not famous. You do not need to win the weights to win the answer; you need the material that enters the context window to be yours, current and correct. Every answer-shaped page you publish is a candidate for that window. Every stale bio floating around the web is a rival candidate. The window is small, the competition is per-question, and the engine fills it with whatever the live web serves best.

Why hallucinations about people happen

A hallucination is not a malfunction; it is the architecture doing its job without enough signal. Asked about a thinly documented person, the model still must predict the next token, so it completes with what is plausible for someone with your kind of name, your kind of title, your neighborhood in the embedding space. Plausible-for-your-category is how you get credited with a book you never wrote or a university you never attended. Same-name collisions are the same failure at higher volume: two humans sharing one token sequence, their documented lives blended into a single statistical silhouette. The defense is not outrage; it is signal. Dense, consistent, machine-readable facts about the real you, so the model never has to improvise, and structured markup like Person schema so retrieval systems can verify which human the tokens refer to.

The five storage facts, and the move each one implies

Notice what is absent from that table: tricks. There is no prompt to inject, no hidden tag that rewrites the weights. The architecture only responds to the shape of the public record, which is why the work looks less like hacking and more like disciplined publishing, the network-building described in the network signal, done consistently until the statistics tip your way.

Pretraining versus fine-tuning: two different layers of memory

Not every stage of building a model touches your name the same way. Pretraining is the enormous first pass, the model reading a huge slice of the public web and adjusting its weights to predict text generally. This is where the bulk of factual association forms, including whatever the model has absorbed about you, because pretraining is the stage with the volume needed to make a pattern statistically reliable. Fine-tuning comes after, on a dataset many orders of magnitude smaller, and it mostly reshapes behavior: how the model follows instructions, formats answers, and stays within safety boundaries, rather than what facts it holds about the world.

The practical consequence is that a person almost never gets a fresh, accurate association added at the fine-tuning stage just because a company wants the model to behave better. Your factual footprint is set overwhelmingly by what pretraining read, which loops back to the same lever covered throughout this piece: the volume and consistency of your presence on the crawlable web at the time a model's pretraining data was collected. Fine-tuning occasionally does touch specific facts when a narrow, deliberate dataset targets them, but treating fine-tuning as a way to fix your name's associations is treating a style adjustment as a fact correction, and it is not built for that job.

Quantization and model updates: what survives compression

Shipping a large model often means compressing it, a process called quantization that stores weights at lower numerical precision to save memory and run faster. Compression is not neutral to what a model remembers. Strong, heavily repeated associations, the ones reinforced by many independent mentions during training, have redundancy built in and mostly survive the precision cut. Weak associations, the kind built from a single mention or a handful of inconsistent references, were already fragile before compression touched them, and quantization is exactly the kind of lossy process that erodes a fragile signal first.

This gives thin documentation a second way to fail, beyond simply not surviving the ranking and disambiguation stages that decide who gets recommended. Even inside the model's raw memory, being mentioned once is a worse position than being mentioned consistently, because the associations that get built thin tend to also be the ones degraded hardest when a lab compresses a model for deployment. Density is not just a bid for the model to notice you the first time. It is insurance against the model quietly forgetting you later.

A name collision resolving across training snapshots: a walkthrough

Say two people share a name, hypothetically call them Alex Rivera, one a structural engineer and one a marketing consultant. Neither is well documented when an early model trains, so that model's snapshot of "Alex Rivera" is a thin, undifferentiated blend of both, and asking it what Alex Rivera does produces a vague or inconsistent answer.

• Snapshot one: Both Alex Riveras are lightly mentioned online. The model's association is a blur, no clear neighborhood, low confidence either way.
• Snapshot two: The engineer starts publishing structured, dated case work and picks up a few citations from industry sources. The next training pass reads more of the engineer's material than the consultant's, so the shared token sequence drifts, slightly, toward the engineering neighborhood.
• Snapshot three: The engineer has kept a consistent name, a Person schema block, and several independent mentions, while the consultant's footprint stayed thin and scattered. This training pass reads a lopsided ratio of evidence, and the model now answers "Alex Rivera" questions with structural engineering context by default, occasionally still surfacing the consultant only when the query supplies enough extra context to break the tie.

Nothing in that progression required either person to do anything to the model directly. Each snapshot simply re-read a fresh slice of the web and rebuilt the association from whatever evidence existed by that cutoff date. That is the entire lever available to anyone dealing with a name collision: outdocument the collision, consistently, before the next snapshot trains.

What this means for managing your name

The professionals who stay accurately represented in AI answers are running probability campaigns, whether they call it that or not: one name, one story, repeated across every source a crawler might read, refreshed often enough that retrieval always finds a current version. The ones who stay misrepresented are usually not victims of the machine; they are thinly documented, inconsistently named, or contradicted by their own old bios, and the model is faithfully reflecting that mess back at them. The weights do not care which kind you are. They just count. If you want to know what the count currently says about you, and what it would take to move it, that is exactly the baseline our service builds first.