ChatGPT omits brands despite citing pages that name them

Graphite's comparison-prompt study, reported by Search Engine Journal, contains an uncomfortable number for any brand that assumes a citation equals a mention: pages ChatGPT pulled into its answers frequently named a brand even when ChatGPT's own generated text did not. The retrieval worked. The attribution did not. For marketers who treat "we got cited" as the finish line, that gap is the whole story.

The mechanism is unglamorous but important. ChatGPT, in both logged-out and search-enabled modes, pulls source pages that mention a given brand, then writes a comparison answer that leaves the brand unnamed anyway. Graphite's data shows this was not random noise. Brands that fared badly in logged-out mode (no search invoked, answer drawn from model weights and whatever it retrieves) tended to fare badly in search mode too. Weak presence in the training-and-retrieval layer predicts weak presence in the generated layer. Citation is necessary but not sufficient.

Why "being in the sources" is not the same as "being in the answer"

Most brand-visibility tracking built for the AI-search era still measures the wrong thing. Tools count citations: how often a company's domain shows up in ChatGPT's source list for a category of prompts. That is a proxy for presence, not for outcome. The outcome that matters commercially is whether a prospective buyer, reading a comparison of five vendors, sees your name in the paragraph that the model actually writes. Graphite's findings say those two measures diverge often enough to matter, and that the divergence is not evenly distributed. Some brands get named roughly in proportion to how often they are cited. Others are cited at a reasonable clip and still get talked around, summarised into "other options" or omitted from the named shortlist while a competitor fills the slot.

This is a synthesis problem, not a retrieval problem, and the two require different fixes. If a brand is undercited, the answer is more and better third-party coverage, the usual SEO-for-LLMs playbook: structured content, clear entity definitions, consistent naming, placement on the sites ChatGPT already trusts for that category. If a brand is cited but not named, the problem sits downstream, in how clearly and how repeatedly the brand's identity is asserted on the pages the model reads. A page that mentions a company once, in a list of twelve vendors, with no distinguishing claim attached, gives the model little reason to surface that name specifically when it compresses twelve mentions into three. A page that states a specific claim, "X processes settlement in 90 seconds, Y in four hours," gives the model a fact worth repeating with the name attached.

The sectors where this bites hardest

Comparison prompts are where B2B purchasing decisions increasingly start, and they are exactly the prompts Graphite tested. Financial-services brands competing on compliance infrastructure, industrial groups competing on supply-chain software, philanthropic funders being compared by grant-seekers researching "which foundations fund X": these are all categories defined by comparison, not single-brand lookup. A multilateral institution like CGAP or UNDRR rarely wins a query where it is the only entity being searched; it wins, or loses, in the paragraph where a model stacks it against three similar institutions and decides which names are worth including by name versus which get folded into "several other organisations also work on this."

The asymmetry Graphite found, where weak logged-out performance correlates with weak search-mode performance, suggests this is not a transient quirk of one model update but a structural feature of how these brands exist in the broader information environment ChatGPT has already absorbed. A brand with thin, generic, uncredited coverage going into its training data starts from a deficit that live search citation does not fully correct. That argues for treating LLM visibility as a cumulative asset built over quarters, not a variable to tune per prompt.

The practical implication is that audit tools measuring only citation rate are telling B2B marketers half the story and, for the brands that need it most, the more comfortable half. A citation count that looks healthy can mask a naming problem that is actively costing shortlist presence. Brands serious about this should be testing actual generated answers against actual competitor sets, not just checking whether their domain shows up in a sources panel, because Graphite's data says those two numbers can and do move independently.