Graphite’s research on Parametric Knowledge Optimization, or PKO, makes a useful distinction for anyone pursuing visibility in ChatGPT: appearing on pages ChatGPT cites does not necessarily mean the brand will be mentioned in the answer.
The premise is that ChatGPT may rely on two information layers. One is live web retrieval, which can produce cited pages. The other is parametric knowledge, meaning information derived from the model’s training. Graphite argues that this training-derived knowledge can influence the queries ChatGPT creates, the brands it considers, and the way it uses retrieved material.
That does not make citations unimportant. Rather, it suggests that citation-focused answer engine optimization is incomplete when the goal is brand inclusion or recommendation in AI-generated answers.
What Graphite’s PKO research found
In its October 16, 2026 Parametric Knowledge Optimization research, Graphite examined 3,115 entity-question pairs across 57 questions. The data was collected in August 2026 and analysed by October.
The study categorised brands by the share of cited pages containing substantial brand mentions. It then compared visibility across three cited-page mention-rate tiers: under 5%, 5% to 15%, and 15% or more.
Graphite reported a consistent pattern: brands with stronger PKO visibility also tended to achieve greater visibility when ChatGPT used search. In other words, a brand’s apparent presence in the model’s training-derived knowledge was associated with better outcomes even in answers that involved live retrieval.
| Cited-page mention-rate tier | Graphite’s reported pattern | Practical interpretation |
|---|---|---|
| Under 5% | Lower search-driven visibility relative to stronger PKO signals | Being absent or lightly represented on cited pages may coincide with limited brand visibility. |
| 5% to 15% | Visibility increased as PKO visibility increased | Retrieved-page presence is only one part of the observed relationship. |
| 15% or more | Higher PKO visibility corresponded with higher search-driven visibility | Brands should consider both retrievable evidence and broader online representation. |
The table describes the direction of Graphite’s findings, not a universal performance benchmark. The research does not establish that a particular percentage of cited-page mentions causes a brand to be recommended, cited, or clicked.
Why cited pages and brand mentions can diverge
A cited source can support factual claims in an answer without making one company central to the response. A brand may appear briefly, be one option among several, or not be named at all even if a cited page discusses it.
Graphite’s explanation is that the model may have already formed an internal view of relevant entities before it retrieves web pages. Its analysis included 957,000 ChatGPT fan-out queries to examine how often brands appeared in prompts and whether those prompts led to cited pages.
This matters because a model-generated query can shape the retrieval set. If a brand is included in an early query, it may receive a different opportunity for exposure than a similarly relevant competitor that is not included. Conversely, earning a mention on a page that later appears as a citation may not be enough if the model does not treat the brand as a relevant entity for the user’s question.
Related reporting has documented cases where ChatGPT writes search queries containing brand names before retrieving pages. Search Engine Journal’s August coverage described this behaviour, reinforcing the importance of what happens before retrieval.
PKO is a complement to SEO and AEO, not a replacement
PKO is best understood as an analytical framework for a difficult visibility problem. Traditional SEO still helps a site be crawled, indexed and discoverable. Content quality, technical accessibility, clear entities and authoritative coverage can also help a page become useful retrieval material for AI search systems.
But Graphite’s research suggests that practitioners should not equate citation acquisition with complete AI visibility. A business may need to assess whether its identity, products, category and differentiators are described consistently across the wider web, not just on its own domain or a handful of target articles.
That creates a longer planning horizon. Changes to a website can affect crawlable content relatively quickly, while changes to training-derived model knowledge may take months to be reflected in future model training or updates. The timing and source composition of proprietary training corpora are not transparent, so no publisher can reliably trace a particular page to a later model response.
A practical PKO-oriented playbook
For smaller businesses and marketing teams, the useful response is not to try to manipulate an unknown training corpus. It is to improve the quality and consistency of the public evidence that describes the brand.
Start with these priorities:
- Audit core brand descriptions. Check whether the company name, offering, audience, geography and category are accurately and consistently described on the website, major profiles and legitimate third-party references.
- Make entity relationships explicit. Help readers and systems understand what the brand does, which products or services it provides, and how those relate to the problems customers are trying to solve.
- Build useful third-party representation. Seek accurate coverage and mentions where they are editorially appropriate. The aim is not volume for its own sake, but a coherent record across relevant sources.
- Publish evidence that can be retrieved. Clear documentation, original research, detailed product information and well-structured explanatory content can support live-search visibility alongside broader brand understanding.
- Measure citations and mentions separately. A cited domain, a cited page, a named brand and a recommendation are different outcomes. Tracking them as one metric can hide important gaps.
- Set realistic feedback expectations. Website improvements may be visible in search sooner than any effect associated with model training. Avoid treating short-term changes in AI answers as proof of training-data influence.
The discipline here is consistency rather than certainty. Teams can improve the signals they control, but they cannot directly verify whether a given source has entered a model’s training data or determine exactly how much weight it receives.
If you want to turn these observations into a repeatable review process, create a GEO Search Leads account. It can help you assess your website’s GEO readiness and identify optimisation opportunities around clear, accessible brand and content signals. Use those findings alongside your SEO and citation monitoring to distinguish technical discoverability issues from broader gaps in how your business is represented online.
Frequently Asked Questions
What is Parametric Knowledge Optimization (PKO)?
PKO is Graphite’s term for optimising the training-derived information that may shape how an AI model understands and surfaces brands. It complements work on live-search citations rather than replacing it.
Does a ChatGPT citation guarantee a brand mention?
No. Graphite’s research found that brands can have limited mentions even when cited pages contain brand references. A citation supports information in an answer, but it does not guarantee that a particular brand will be named or recommended.
Can a business know whether its content is in ChatGPT’s training data?
Not reliably. Training-data provenance is opaque, and large proprietary corpora make it difficult to trace the influence of a specific page or mention on a model response.
How quickly can PKO work affect AI answers?
Graphite notes that content can take months to affect parametric knowledge. The timing depends on model training and updates, which are not fully visible to publishers or marketers.
Do Graphite’s findings apply to every AI search platform?
No. The study concerns specific ChatGPT prompt sets and logged-out responses. Results can differ by prompt type, model, search behaviour and platform.
Conclusion
Graphite’s PKO research is a reminder that AI visibility is not reducible to a citation count. Strong technical SEO and retrievable content remain essential, but brands should also build a clear, consistent and evidence-based presence across the web. Because training data and model behaviour remain opaque, PKO should guide better representation and measurement, not create false certainty about how any individual answer is produced.
Top comments (0)