Parametric Knowledge Optimization

AEO Should Include Parametric Knowledge Optimization (PKO)

By default, LLMs answer using their parametric knowledge. Parametric knowledge is the information encoded in the LLM weights during training. It is the model’s fuzzy recollection of the internet.

LLMs often have access to a web search tool they can use when they need additional or up-to-date information. Relevant content retrieved from the web is added to the context when the model is responding to the prompt.

We provide a more detailed description of how LLMs use search in our previous paper Do Not Force Search in Prompt Tracking.

Much of Answer Engine Optimization (AEO) is devoted to optimizing for citations. We argue that parametric knowledge optimization (PKO), which aims to improve how the underlying model describes and recommends a brand by influencing the training set, is just as important.

LLMs Use Search Frequently, But Not Universally

Recently, we have observed an increase in citation rates, suggesting more frequent search use. In a fixed set of entity-comparison prompts tracked in Graphite’s internal tool, the percentage of logged-out ChatGPT responses containing a citation rose from 56.6% in the week of August 3rd, 2026 to 88.5% in the week of September 28th, 2026. Over the same period, logged-out Gemini citation rates rose more modestly, from 79.0% to 81.2%. Still, search is not universal, and these rates have changed over time and could go down.

In contrast, Similarweb found that across all prompt types, only 7% of U.S. ChatGPT responses have citations, suggesting that use of the web search tool is rare in general. Although the datasets and collection methods differ, these results suggest that the use of search is much more frequent in entity-comparison prompts.

Note that brands can also appear in answers to educational, planning, and problem-solving prompts, where users are not explicitly asking for product recommendations. Search rates may be lower for these prompts.

As we argued in our recent paper, we should not force search when tracking prompts because the resulting citations may not represent what users see.

Parametric Knowledge Shapes Responses Even When LLMs Use Search

LLMs generate fan-out queries, and they often seed them with candidate answers. Across 1.17 million ChatGPT fan-out queries issued in response to 22,953 non-branded entity-comparison prompts collected from August to early October 2026, 22.7% contain a “site:” operator, and 59.6% contain a brand (as labeled by Claude Haiku 4.5). Including a brand in a fan-out query increases the chance that pages featuring that brand will be retrieved. In this way, parametric knowledge can shape the answer through the fan-outs.

Early evidence suggests that parametric knowledge also influences how a model uses search results. Using the data from our prior paper (which has details on data collection), we compute entity visibility (the percentage of responses that contain an entity, often a brand) with and without search, using the same prompts. For the search setting, we also measure the mention rate, the percentage of cited pages (pooled across responses) that contain a substantial mention of each entity. We group entities by how often they appear on cited pages: < 5%, 5–15%, ≥ 15%.

A 15% mention rate seems small, but we only consider substantial mentions of a brand, and not every cited page is a listicle. In our data, 51% of cited pages have a substantial mention of zero or one brand, and only 25% have substantial mentions of at least 5 brands.

Across all three mention-rate bins, search visibility rises with parametric knowledge visibility. Even entities mentioned on at least 15% of cited pages reach only 11% median visibility with search when their visibility without search is in the lowest band, 0–25%.

Search visibility also rises with mention rate. This suggests that both parametric knowledge and mentions in cited pages are important.

We observe the same pattern in two independent replications. In the first, we test 100 prompts through the API, explicitly disabling or enabling search. In the second, we test 200 prompts from the Graphite tracking tool over a different three-week period. In both cases, search visibility rises with parametric knowledge visibility and mention rate, with comparable effect sizes.

Prior academic research also shows that LLMs sometimes rely on their parametric knowledge even when conflicting information is provided in the context.

In future work, we plan to study this relationship in more detail.

PKO Is More Challenging Than Optimizing for Citations

LLM Training Datasets Are Vast and Not Public

Frontier LLMs are trained on enormous amounts of data. Epoch AI estimates that GPT-5 was trained on at least 30 trillion tokens. In contrast, a model using search usually retrieves tens of pages and cites 5-10. When optimizing for citations, one document can have an effect, whereas one document is unlikely to have an effect for parametric knowledge optimization.

Additionally, AI labs do not make their training datasets publicly available, so we do not know which webpages are included.

LLMs Cannot Reliably Identify Which Training Documents Influence a Response

When a model uses search, it can cite pages from the search results to support its response. In contrast, LLMs generally cannot identify the specific training documents that influence a response.

New Content Takes Months to Affect Parametric Knowledge

New models released by OpenAI, Anthropic, and Google are typically trained on data that is several months old. The latest date when pages are included in the training set is called the knowledge cutoff. We use the term knowledge cutoff gap to mean the amount of time between the knowledge cutoff and the release date. The knowledge cutoff gap is the minimum time-to-impact for newly published content.

For Anthropic, we use the training-data cutoff rather than the earlier reliable knowledge cutoff to make results comparable across providers.

We calculate gaps from elapsed days, using 30.44 days per month. Month-only cutoffs are treated as the last day of the month.

The following figure shows knowledge cutoff gaps for models released since 2022.

Knowledge Cutoff Gaps Are Decreasing

Some models released in the first half of 2026 still had knowledge cutoff gaps of over a year. For the models released since July 2026 in our dataset, gaps are six months or less across all three providers.

Early Strategies for PKO

Although attribution is difficult and the feedback cycle is long, a technical understanding of model training suggests several approaches to parametric knowledge optimization. Below are some ideas we are exploring.

Measure When New Models Are Released, and Be Aware of Knowledge Cutoff Gaps

When new models are released, supplement standard prompt tracking with parametric knowledge visibility tracking. Use APIs with web search disabled to make sure responses are based only on parametric knowledge. Collect at least fifty responses per prompt. This process is only needed once per pinned model version, as the response distribution will not change as long as the prompts, system instructions, and sampling settings remain fixed. Therefore, we recommend collecting a larger sample once per model version.

In addition to measuring visibility, ask the model about the brand’s products, capabilities, and intended customers. This helps distinguish between a model that lacks accurate information about the brand and one that understands it but rarely mentions it.

When optimizing, be aware of knowledge cutoffs and gaps. Do not expect impact from PKO in two weeks when the knowledge cutoff gap is five months.

Allow AI Labs’ Crawlers to Access Published Content

AI labs use crawlers to collect training data. Blocking these crawlers prevents them from collecting new content from a site.

Make sure that sites used for publishing allow access to the AI labs’ crawlers. Note that there are generally different crawlers for collecting training data and indexing content for web search.

Google-Extended controls permission for Gemini training; content is collected by Google’s existing crawlers. We include Common Crawl because AI labs likely use its data for training.

Allowing access in robots.txt alone is not sufficient. Confirm that CDN or firewall bot-management rules are not blocking crawlers. Serve key content in server-rendered HTML, since training crawlers may not execute JavaScript.

Get Mentioned on Many Unique Pages and Sites, But Avoid Duplicate Content

To influence a response generated using search, publishing a single page may be sufficient because the number of pages cited by the model is generally small (5-10). In contrast, the training set is massive, so influencing parametric knowledge will likely require many mentions.

Additionally, training data preparation includes deduplication of pages and may also limit the number of pages from a single domain.

Therefore, for PKO, we recommend targeting many mentions across many distinct pages, while avoiding near-duplicate content. Comprehensive owned content is an important part of this strategy, but we also recommend earning mentions on a variety of third-party sites.

Prioritize Sources with Evidence They Are Used in Training

We generally do not know which sites are included in the training data, but there are some clues. Reddit has data partnerships with both Google and OpenAI, and Google explicitly identifies model training as a use of that data. Wikipedia is a disclosed training source for OpenAI. Publishers also have agreements covering model training. For example, Axel Springer’s partnership with OpenAI includes training on content from publications such as Business Insider.

We recommend prioritizing mentions on these sites and in these publications.

Common Crawl is likely used in training, and its public index can be used to check whether pages from a site have been collected.

Ensure the Brand Is Described Accurately and Consistently Online

Parametric knowledge is determined by the text in the training set. If a brand is described inaccurately online, for example, with outdated pricing or an incomplete feature set, the model may learn and repeat those inaccuracies. We also hypothesize that inconsistency reduces visibility, since post-training encourages models to avoid confident claims when there is uncertainty.

We recommend auditing how the brand is described on its own site and third-party sites, and correcting outdated or inaccurate information. Use consistent brand and product names, and clearly describe the products’ capabilities and intended customers. Include dates when publishing information that changes frequently, such as pricing and features, to make clear when it applies.