A brand that ChatGPT rarely names from its own knowledge tends to stay rarely named after it searches the web, even when the pages it retrieves discuss that brand. That is the central finding of a new analysis from Graphite, reported by Search Engine Journal, and it complicates the common assumption that AI visibility is mostly a matter of getting into the sources.
The pattern is a close cousin of what we call a ghost citation, where a brand's own page sits in the source list while the answer recommends someone else. Graphite measured mentions of brands across the cited pages rather than whether each brand's own page was cited, so its numbers describe the broader version of the problem. Its explanation for why it happens is useful either way, and it suggests that much of the fix sits outside a brand's own website.
What the Graphite Study Found
Graphite found that a brand's visibility in ChatGPT answers without search predicted its visibility with search at every level of cited-page coverage. The study used 3,115 brand and question pairs across 57 comparison questions, collected from logged-out ChatGPT in August 2026, and measured visibility as the share of responses that named a brand or product.
For answers that used search, Graphite also measured how many of the cited pages substantially mentioned each brand. The headline figure comes from the brands in the lowest band of no-search visibility, named in 25% or fewer of ChatGPT's answers without search. Among those brands, the ones substantially mentioned on at least 15% of cited pages reached a median visibility of only 11% once search was on. These brands appeared in the material ChatGPT retrieved, yet the median brand was named in only about one answer in nine.
| Measure | What Graphite tracked | Why it matters |
|---|---|---|
| Visibility without search | How often ChatGPT named the brand when answering from its own knowledge | A proxy for what the model learned in training |
| Cited-page mention rate | Share of cited pages that substantially mentioned the brand | Shows whether the brand was in the material the model read |
| Visibility with search | How often ChatGPT named the brand once it searched the web | The outcome buyers see |
The study has limits that belong in any reading of it. Graphite describes its methods as exploratory, and the core dataset covers one assistant, one collection period, and logged-out answers only. Search Engine Journal notes that the write-up does not describe controls for brand awareness, which could drive both kinds of visibility at once, and that the data shows a relationship without showing that raising either factor would increase mentions. Graphite sells an optimization service built on this thesis, which does not make the numbers wrong, although it is a good reason to treat the explanation as a working hypothesis.
Why a Brand in the Sources Can Still Go Unnamed
Graphite's explanation is parametric knowledge, which is the information a language model stores in its weights during training, as opposed to what it retrieves from the web while answering. Graphite's argument is that this stored knowledge shapes answers twice, first by influencing which searches the model runs, and then by influencing how it weighs the pages those searches return.
The first effect shows up in the searches themselves. In Graphite's dataset of roughly 957,000 fan-out queries, generated from 22,953 comparison prompts that named no brand, about 59.6% of the queries ChatGPT wrote contained a brand name anyway. A model that already associates a category with a handful of names tends to go looking for those names, which gives the brands it already knows a head start before any page is read.
The second effect is harder to see, although Graphite points to earlier research showing that language models sometimes favor what they learned in training even when the material in front of them says something different. A page that compares five vendors may supply a price point or a feature list for the answer, while the recommendation itself still leans toward the vendors the model already trusts. That is how a brand can end up supporting the answer while a competitor gets the credit in the text.
What Our Own Index Shows
Our own AEO Index data shows an important exception to the pattern. In our analysis of 9,046 AI answers across four industries, one commercial debt collection agency was never named by any of the three engines we tested with web access turned off, yet it ranked first in its category on our blended index score, which combines what models say with web access off and what they say once they search. Its own website was the most-cited domain in the entire category.
The difference lies in the category. In categories where buyers rarely discuss purchases publicly, as in commercial debt collection, a model that searches finds little third-party material and reads what the companies have published themselves instead. In categories where buyers discuss their options openly, the model has far more third-party material to learn from, and the brands with a strong presence in that material start every answer ahead. Graphite's median describes the typical case, while our data suggests that a thin public conversation creates an opening for brands whose own sites are the strongest available source.
One brand is not a trend, although the practical reading is that the importance of training data may vary by category. A brand competing in a crowded consumer or software category will likely need to work on how it is described across the web, while a brand in a niche B2B category may get further, faster, by making its own site the definitive source.
How to Earn Named Mentions in AI Answers
Earning named mentions in AI answers requires work on two fronts: the third-party record a model learns from and the pages it retrieves while answering. These are the steps we prioritize, drawing on Graphite's recommendations and our own audit work.
- Measure both layers: Track how often AI assistants name the brand with web access turned off as well as on, since the no-search number shows what the model has learned and the search number shows what buyers see. Our AEO Index separates the two layers, and Graphite recommends checking each new model release, because what a model knows changes with every training run.
- Earn mentions across many independent sites: Graphite's evidence points toward breadth, meaning substantive mentions spread across many distinct pages and publications, with near-duplicate content adding little. This is the thinking behind our Brand Consensus work, which builds consistent third-party coverage of a brand in the places a category's buyers and journalists already read.
- Open the door to training and retrieval crawlers: Confirm that training crawlers such as GPTBot, ClaudeBot, Google-Extended, and CCBot can reach the site, along with the retrieval crawlers assistants use while answering. We have found all major AI crawlers blocked at the CDN level during a client audit, which silently removes a brand's own pages from both layers at once.
- Keep the brand description consistent and current: Use one brand name, one clear category description, and dated pricing and capability details across the brand's own site and the profiles it controls, so the model learns a single, accurate version of who the brand is.
- Make the brand's own site the best source in these categories: Where buyers rarely discuss their options in public, build the comparison pages, specifications, and expert explainers that a model will retrieve because nothing better exists. Our guide to why most B2B brands are invisible in AI search covers the entity and content signals that help a brand's own pages get extracted.
- Plan for the lag: Graphite measured a gap of roughly two to five months between a model's knowledge cutoff and its release for recent models, which it treats as a minimum time to impact, so third-party coverage earned this quarter may not appear in what models know until well into next year. Retrieval work on the brand's own site can move faster, which is one reason to run both fronts together.
The Bottom Line
A citation shows that a brand's content reached the answer, while a named mention shows that the model trusted the brand enough to recommend it, and Graphite's data suggests the second depends heavily on what the model learned before anyone typed a prompt. For most brands, that makes third-party coverage and consistent naming part of AI visibility work rather than a separate PR exercise, while brands in quiet categories have a faster route through their own sites.
To find out how often AI assistants name your brand, with and without search, talk to our SEO, AEO, and GEO team. To connect those mentions to pipeline and revenue, see how Galileo maps AI-assisted journeys.