LLM parametric knowledge: what models remember about your brand

Parametric knowledge is what a large language model knows without looking anything up: facts stored in its weights at training time. When ChatGPT, Claude or Gemini answers from memory, that answer is parametric. AI visibility tracking shows who gets named when search is on. This guide is about the memory that was already there before the search ran, and what you can do about it.

Most AI visibility advice covers the pages an assistant fetches, because those show up as citations. Before any search runs, the model already holds a picture of the category, a short list of brands, and a sense of which ones it trusts. That picture was fixed months before the model shipped. It decides what gets searched, how the results are read, and what is said when search never runs.

Parametric, contextual and retrieved knowledge

An answer can come from three places. The research term for the last two together is non-parametric knowledge, from the 2020 paper that introduced retrieval-augmented generation. A 2019 paper had already asked whether language models could serve as knowledge bases with no external data at all. This guide keeps the three-way split, because you can influence what gets retrieved and you generally cannot influence what a user pastes in.

How a fact gets into the weights

The useful recent evidence is Physics of Language Models, which trained models on synthetic biographies and counted how often each fact appeared. Capacity was about two bits per parameter once a fact had been seen enough times, so space is not the constraint. Exposure is. At about a thousand appearances, recall was strong. At about a hundred, it fell sharply. A brand described a few dozen times is in the weak-storage regime. The goal is to raise the exposure count of a few facts, not the word count of pages about you.

Two other results matter. Junk text diluted capacity, and a source label (a domain-style token) mostly restored it: the model learned which sources were worth storing. And a single fixed sentence, repeated, taught the model a string it could continue but not answer questions about. The same facts in many wordings became extractable. Syndicating one boilerplate description a hundred times is the first case. A hundred writers describing the same product is the second. A 2026 study, Predictable Confabulations, found the same shape on real models: larger models remember more of the long tail, and topic frequency still predicted recall across every size they tested.

Why most brands are barely there

Nikhil Kandpal and colleagues counted, for questions in standard benchmarks, how many pretraining documents mentioned both the subject and the answer. On the 176-billion-parameter BLOOM model and TriviaQA, accuracy rose from about 25 percent at roughly 10 relevant documents to above 55 percent at 10,000. Deleting those documents and retraining caused the drop, so the link was not just a correlation. Their conclusion was that retrieval, not scale, was the realistic route to long-tail accuracy.

When Not to Trust Language Models found the same pattern with Wikipedia page views as a stand-in for how well known an entity is. Accuracy tracked popularity. Larger models gained almost all of their improvement on popular entities. A handful of brands in a category are known cold. A few dozen are recognised when named. Everyone else has, at best, a weak association between the name and the category. An AI visibility tracker shows the symptom: the brand is missing from the recommendation list. Parametric knowledge is the slower of the two causes.

What the model does when it barely knows you

A 2025 paper from OpenAI on why language models hallucinate argues that error on rare facts is bounded below by the singleton rate: the share of facts seen exactly once. A fact seen once gives the model almost nothing to generalise from, and the training objective rewards a confident guess over "I don't know", because most benchmarks score a wrong answer and an abstention the same. A thin footprint produces a plausible description assembled from the category and the name. An old pricing page gets quoted as current.

On recommendation questions the effect is sharper. The model does not abstain. It lists the brands with the strongest associations. A brand seen a thousand times in that context beats a brand seen fifty times. Each slot you do not hold is held by a competitor, and the model's answer then flows into the next round of web content. Citation tracking cannot explain that gap, because there is no citation.

Reported cutoff, effective cutoff, and the wait for release

Dated Data found that the date a model demonstrably knows things up to is often months earlier than the cutoff the vendor publishes. Common Crawl snapshots mix pages fetched in earlier months, and deduplication can keep an older near-copy of a page. Graphite's October 2026 analysis put the gap from cutoff to release, for models shipped since July 2026, at about four and a half months for OpenAI, two and a half for Anthropic, and four and a half for Google. Content published today can show up in a model you can use in six to eighteen months, and only if it was crawled, survived filtering, and appeared often enough.

Why this still matters when the model searches

Search does not overwrite the weights. On commercial comparison prompts, Graphite saw ChatGPT citation rates rise from about 57 percent in early August 2026 to nearly 90 percent by late September. Across all prompt types, Similarweb's figure for US ChatGPT responses that include any citation is about 7 percent. The rest are answered from memory. Fewer citations per answer also means fewer visits from the answers that do search, which is the pattern in why AI referral traffic is falling.

When search does run, the model writes the queries. Graphite labelled about 957,000 fan-out queries from about 23,000 non-branded comparison prompts: nearly 60 percent contained a brand name, and nearly a quarter used a site: operator. If the model has no association between your brand and the category, it will not write a query that finds you. You can see the trace in Search Console. In October 2026, queries that showed llmscout.co in Google's AI features included "how does bot.spot compare to profound and peec ai for tracking ai brand visibility?" and "profound peec ai llm pulse ai visibility platform comparison". Nobody types those. They are fan-out queries, and the brand names were chosen before the search ran. Those impressions are what AI overview tracking reads from an export. The tools in the queries are compared in Profound vs Peec AI.

Retrieved pages are then judged against the prior. ClashEval found GPT-4o abandoned a correct stored answer for a wrong retrieved one in about 60 percent of direct contradictions, with a context-bias rate around 30 percent against about 16 percent for Claude Opus. Models resisted more when they were confident, and when the retrieved value was far from what they believed. Adaptive Chameleon or Stubborn Sloth found a confirmation bias when evidence was mixed: models leaned on the passages that agreed with them. Graphite's test of 3,115 brand and question pairs found brands that appeared on 15 percent or more of cited pages, but were weakly known with search off, reached a median search-on visibility of only 11 percent.

The short version. A search-on answer is a negotiation between what was fetched and what the model already believed. Citation work argues with the first party. Parametric work changes the second.

What gets cited is what gets learned

LLM Scout's citation study, AI cites lists, not homepages, covered 63,500 modelled citations. List-style pages (best-of, comparison, category and directory listings) were 81 percent of ChatGPT citations and 75 percent of Claude citations. Homepages were the least cited type, at 7 percent and 4 percent. A homepage that says "we are the leading platform" gives a model one boast. A comparison page that says "A, B and C all do X, and here is how they differ" is a fact the model can store and later repeat. The same bias almost certainly applies to training, not only to citations. Older URLs are what assistants often cite, and removed pages can keep describing the company months later.

How to measure what a model remembers

Use the API, attach no tools, and pin a model version. The consumer apps can search without saying so. Ask three kinds of question, at least 50 times each at default temperature for the list questions: what the brand is (recognition), which brands are in the category (association), and which brand a described buyer should use (recommendation). A brand the model describes accurately but never volunteers needs mentions next to the category phrase. A brand it has never heard of needs foundational descriptions on sources a training crawler will read.

Run that probe on release day. The search-on side moves every week, which is what a buyer actually sees. LLM Scout records which brands are named and which pages are cited. The free AI visibility check is a company email, five ChatGPT prompts, no card. Pro is $39.99 a month for 25 ChatGPT prompts. Enterprise is $99.99 a month for 100 prompts across ChatGPT, Claude, Gemini and Perplexity. Personal inboxes such as Gmail are not eligible for the free account. See how the free account works.

Which bots build the weights

Blocking a training crawler keeps you out of future models. Blocking a retrieval crawler keeps you out of citations. Many sites have done the first and left the second open. A brand that wants to be remembered wants both open on pages that describe what it sells. A September 2026 analysis of 4,242 robots.txt files found about 8 percent fully blocked GPTBot, 7 percent ClaudeBot and 6 percent Google-Extended. Check the CDN bot settings as well as the file: a firewall block never reads robots.txt. Training crawlers generally do not run JavaScript, so a page whose text arrives only after a script is an empty shell. Confirm captures in index.commoncrawl.org.

Five patterns in a no-search answer

The same failures show up across brands. Each one points at a different problem in the corpus.

Mentions that survive deduplication

Training pipelines drop exact copies and near-duplicates. A press release on two hundred wires becomes about one document. Mentions count when they are different words, on different domains, by different authors. Ahrefs, across about 75,000 brands, reported a correlation of about 0.66 between branded web mentions and visibility in Google's AI Overviews, against about 0.22 for backlinks. BrandMentions found that brands repeatedly visible in AI answers had a footprint that was 70 to 80 percent third-party, and that about half of those mentions carried no link. Seer Interactive found a brand mentioned in an answer was also cited about 53 percent of the time, against about 11 percent when it was not mentioned. The mentions that matter often have no link, live on other people's sites, and sit next to the words that describe the product.

Five facts, one name, one category phrase

Write the five facts you want every model to hold. For LLM Scout they are: it is an AI visibility tracker; it records whether ChatGPT and other assistants name a brand; it has a free account on a company email; it is used by marketing teams; it is made by a small team. Then check every page that describes you. A page that contradicts a fact should be fixed or corrected. Use one spelling. "LLM Scout", "LLMScout" and "llmscout.co" split the exposure count. Put the buyer's phrase, "AI visibility tool", next to the name more often than any internal label. Put a date in any sentence that states a price or a plan limit.

Sentiment only exists where the brand is named. If a model never mentions you, there is no sentiment score to report, and a tracker that invents one is making up data. LLM Scout leaves that field empty. Old complaints stay in the weights until a later document, with a date, says the problem was fixed.

Where this sits next to SEO and AEO

Parametric knowledge optimisation, which Graphite abbreviates to PKO, is the layer under SEO and under answer-engine work. A comparison page on a trusted site can earn a citation this week and raise the exposure count for the next training run. The difference is what you measure, and how long you wait. For the citation layer, see what answer engine optimization is and SEO vs AIO.

What to ignore

• llms.txt does not write the weights. No major vendor has committed to reading it for training or retrieval.
• Schema markup is not what the model trains on. Pipelines keep the visible text. Put the facts in the sentences.
• One consumer-app answer is not a measurement. Search may have run. Fifty API runs, tools off, is a measurement.
• Bigger models will not know everyone. Scale helps well-covered entities more than rare ones.

What to do this week

1. Ask GPT, Claude and Gemini through their APIs, with no tools, what your company does.
2. Check CDN bot settings and server logs for GPTBot, ClaudeBot and CCBot. Leave them open on marketing pages.
3. Write the five facts and make your own pages state all five, with dates on prices and plan limits.
4. Find the ten third-party pages most likely to be in training data, and ask for corrections where they are wrong.
5. Start weekly tracking of the buyer prompts, so the next model release has a before.

Measure the half you can see every week

Parametric knowledge moves once per model release. What ChatGPT says to buyers moves every week.

Get the free AI visibility check. Company email, five ChatGPT prompts, no card. Pro is $39.99 a month. Enterprise is $99.99 a month. Plan limits are on pricing.

Glossary

Parametric knowledge

Information in the weights, available with no external source. Also called parametric memory.

Exposure

One occurrence of a fact in training data. The count predicts whether the fact is stored.

Knowledge cutoff

The date after which no training text was included. The effective cutoff is often earlier than the one the vendor states.

Deduplication

Removal of exact and near-duplicate documents before training. Syndicated copies are mostly removed here.

Query fan-out

One user question rewritten as several search queries, from what the model already knows.

Training crawler

A bot that may collect text for training: GPTBot, ClaudeBot, CCBot, and the Google-Extended token.

Retrieval crawler

A bot that fetches pages for search or at answer time: OAI-SearchBot, ChatGPT-User, Claude-SearchBot, Claude-User, PerplexityBot, Googlebot.

Third-party figures are as of 28 October 2026. Vendors change crawlers often, so check the linked pages for the current position.

Answers

What is parametric knowledge in a large language model?

Parametric knowledge is the factual content stored in a model's weights during training. It is what the model knows without reading anything at answer time. The term comes from the 2020 paper that introduced retrieval-augmented generation, which contrasted the model's parametric memory with a non-parametric document index.

What is the difference between parametric and non-parametric knowledge?

Parametric knowledge lives in the weights and was fixed at the training cutoff. Non-parametric knowledge arrives from outside at answer time, either in the user's prompt or through retrieval such as web search. Both shape an answer; the parametric side also decides what gets searched for and how retrieved pages are read.

Is parametric knowledge the same as RAG?

No. RAG, retrieval-augmented generation, is the technique of fetching documents and giving them to the model before it answers. RAG supplies non-parametric knowledge. The model's parametric knowledge is what RAG is supplementing. Web search in ChatGPT, Claude and Gemini is RAG applied to the live web.

How do I check what ChatGPT knows about my brand without search?

Use the API, attach no tools, pin a model version, and ask three kinds of question: what the brand is, which brands belong to its category, and which brand it would recommend to a described buyer. Run each at least fifty times and count how often your brand appears. The consumer app is unreliable for this because it can search without telling you.

How long does it take for new content to affect an LLM's parametric knowledge?

Months at minimum. Content must be crawled, survive filtering and deduplication, fall before a vendor's data cutoff, and then wait for the model to train and ship, which took roughly two to five months after cutoff for frontier models released in the second half of 2026. Realistically, plan for six to eighteen months and for continuous rather than one-off work.

Should I block GPTBot and ClaudeBot?

Not on pages that describe your company and products, if you want models to recommend you. GPTBot and ClaudeBot are the training crawlers; blocking them signals that your content should be left out of future models. Block them, if at all, only on private or paid areas. Leave OAI-SearchBot, ChatGPT-User, Claude-SearchBot and Claude-User open as well, since those control citations rather than training.

Does Google-Extended affect my Google rankings or AI Overviews?

Google states that Google-Extended has no effect on inclusion or ranking in Search. It controls whether content Googlebot has already fetched may be used to train and ground Gemini. AI Overviews draw on the ordinary Search index.

Why does an LLM still name competitors when my page is in its search results?

Because the model wrote the search queries from what it already believed, and because it reads retrieved pages through that belief. Research on knowledge conflict shows models favour evidence that agrees with their prior and discount evidence that contradicts it, especially when they are confident. Graphite's 2026 data found brands that were well represented on cited pages but weakly known to the model reached a median search-on visibility of only eleven percent.

Does LLM Scout measure parametric knowledge?

LLM Scout tracks what buyers are shown each week, with search on. The free account and Pro both use ChatGPT: five prompts on a company email, or 25 prompts on Pro at $39.99 a month. Enterprise is $99.99 a month for 100 prompts across ChatGPT, Claude, Gemini and Perplexity. The no-search probe in this guide is separate: you run it through the vendors' APIs with tools turned off. Use the two together. The probe at each model release, the tracker every week in between.