New analysis from growth agency Graphite should make performance marketers rethink what AI visibility actually means. When ChatGPT ran comparison prompts with search switched on, brands it rarely named without search stayed rare—even when those brands were heavily present on the pages ChatGPT cited.
Graphite looked at 3,115 entity-question pairs across 57 comparison prompts collected in August. Visibility was the share of answers that mentioned a brand or product. Among brands substantially mentioned on at least 15% of cited pages, median search visibility was just 11% when ChatGPT had named them in fewer than a quarter of its no-search answers.
Search results get filtered through training
The pattern held across every group Graphite tested. Brands that earned more no-search mentions also earned more search mentions. The study ran two replications: one with 100 API prompts where search was explicitly enabled or disabled, and one with 200 prompts tracked over a different three-week window.
Graphite sells AEO services, so its call for PKO deserves that context—but the directional evidence is worth attention. The agency’s conclusion is that both factors matter: how often a brand shows up on cited pages and how strongly the model already associates it with a category. That is what Graphite calls parametric knowledge optimization, or PKO.
Think of it this way: ChatGPT is not reading search results from scratch. It is matching a trained consideration set against the pages a search retrieves. If your brand is not already in that set, a citation alone may not drag it into the answer.
What background searches reveal
Graphite also analyzed about 957,000 fan-out queries from 22,953 non-branded comparison prompts collected from August to early October. None of the prompts named a brand. Yet 59.6% of the background queries contained a brand, and 22.7% used a site: operator.
That means ChatGPT often injects brand names into its own searches before generating an answer. Suganthan Mohanadasan observed a similar pattern in a smaller SEJ test: in 21 of 27 conversations, ChatGPT’s first search query included brand names he had not typed. If a brand appears in a fan-out query, Graphite says, that raises the chance pages featuring it get retrieved.
Citation rates are rising too. In Graphite’s fixed comparison prompts, the share of logged-out ChatGPT responses with citations jumped from 56.6% in the week of August 3 to 88.5% in the week of September 28. Gemini went from 79.0% to 81.2% over the same windows. But Graphite cautions: a citation does not prove a search happened. As of May, ChatGPT could add inline links without showing search evidence.
What to do about it
The practical shift is to measure and build brand presence at the model level, not only the citation level.
- Create a no-search baseline. When a new model ships, run at least 50 responses per prompt through the API with search disabled.
- Keep training crawlers open. Check whether CDNs or firewalls block AI training crawlers.
- Plan for lag. Graphite estimates a minimum two-to-five-month gap between a model’s knowledge cutoff and release before newly published content can appear.
Google-Extended is a useful control point for Gemini: it can restrict training and grounding without affecting Google Search rankings. The caveat is that AI labs do not disclose exactly which pages went into training, so your no-search baseline is the best proxy.
None of this proves that improving citations or training mentions will automatically lift search visibility. But it strongly suggests the old AEO playbook of ‘get cited’ is incomplete. Brands that want generative answers need both search presence and model memory.
Source: Search Engine Journal
