Large language model (LLM)

A large language model is a neural network with a very large number of parameters, trained on large amounts of text, that generates language by repeatedly predicting the next token. Anthropic's glossary calls LLMs "AI language models with many parameters that are capable of performing a variety of surprisingly useful tasks", trained on vast text data and able to answer questions and summarise information.

Everything else about how these systems behave, the fluency, the confidence and the particular ways they get facts about your company wrong, follows from that mechanism and from what was done to the model after pretraining.

How does an LLM work, step by step?

Tokenisation. Text is cut into tokens before the model sees it. Anthropic describes tokens as words, subwords, characters or bytes, and documents that a Claude token is about 3.5 English characters. Your product name may be one token or five. See tokens.

Pretraining. The model learns to predict the next token over a huge unlabelled corpus. Anthropic notes that pretrained models "are not inherently good at answering questions or following instructions": the objective is prediction, not helpfulness. Machine learning is the wider field; this is the specific technique.

Fine-tuning and RLHF. Further training, including reinforcement learning from human feedback, turns the predictor into an assistant by ranking outputs and pushing the model toward the preferred ones.

Inference. At request time the model works inside a context window, its working memory for that request. Anthropic documents 1M-token windows on several current Claude models and warns that accuracy and recall degrade as the window fills.

The architecture underneath is the Transformer, from Attention Is All You Need (12 June 2017), which dropped recurrence for attention alone and reported 28.4 BLEU on English-to-German and 41.8 on English-to-French after 3.5 days on eight GPUs.

How does a sentence about your brand get into an answer?

By one of two routes, with opposite economics.

Through the weights. Something in the training corpus shaped a tendency. You cannot edit it or date it, and it comes with no citation. The model can even produce plausible-looking URLs it never fetched. See AI training data and knowledge cutoff.

Through retrieval. A page was fetched during the request and placed in the window. That passage exists, can be cited, and changes the moment you edit the page. This is the only route with a feedback loop short enough to work with, which is why RAG matters more to a publisher than model choice does.

Do different AI assistants answer the same question the same way?

No, and the difference is measurable. Attensira reads the same tracked questions on ChatGPT and Perplexity every day for our own site. Over the 30 days to 4 October 2026, the two diverged on evidence more than on rates.

ChatGPT named Attensira in 99 of 2,349 answers (4.2%) and cited attensira.com in 154 (6.6%). Perplexity named us in 98 of 2,175 (4.5%) and cited us in 85 (3.9%). Every one of those namings came from the three tracked questions that contain our name; the other 58 questions named us in none of 4,326 answers. The figure at the top of this page sets the citation rates beside the starkest difference: a ChatGPT answer carried 33.5 citations on average, a Perplexity answer 10.0.

Our mention rates on the two were close. The evidence was not. ChatGPT read widely and leaned hardest on OpenAI's and Google's own documentation; Perplexity cited a third as many pages per answer, spread across vendor blogs, LinkedIn and review sites. Same questions, different retrieval, different evidence. A single answer from one assistant tells you about that draw on that engine, not about "AI".

That is also why Attensira tracks the product surface a person uses, with a stable id such as chatgpt, rather than a model version: tracked models keep a continuous history when a vendor swaps the model underneath.

What goes wrong when LLMs answer about brands?

• Fluent wrongness. The training objective rewards likely text, not true text. A confident wrong sentence is the base behaviour when nothing grounds the request. See AI hallucination.
• Fabricated citations. A URL generated from the weights looks exactly like a retrieved one. A link is not evidence a page was read.
• Stale claims. A parametric answer reflects a corpus frozen months earlier; prices and positioning you changed last quarter can persist.
• Context rot. Filling a large window is not free; a retrieval step that dumps everything can do worse than one that selects well.
• Non-determinism. Anthropic states that even at temperature 0 identical inputs can produce different outputs. One answer is one draw, which is why any rate needs its sample size.

What does Attensira do with LLM answers?

It treats them as observations to count, not quotes to screenshot. Attensira asks your buyers' questions on every engine your plan reads (the models documentation lists which engines each plan includes) at a fixed daily cadence, keeps every answer verbatim, and computes mention and citation rates over runs with the sample size attached. A change is reported only when it passes a significance test, as its methodology documents. Then the crew does the part a dashboard leaves to you: it finds the pages the answers lean on and writes the fix, which waits for your approval before it ships.