Most AI search demand is unclaimed. Here is what that costs brands.

Kevin Indig ran the numbers on 1,094 AI search categories and found that only 15.2% have a brand ChatGPT names consistently enough to call a clear owner. Search Engine Journal carried the analysis. The other 84.8% are contested at best, invisible at worst. For most industries, the territory has not been claimed.

That figure deserves more attention than it has received. The conventional response to AI search has been to chase citations: get mentioned in the sources a model pulls from, and you will get named in the answer. Indig's data punctures this assumption. Citations barely predict which brand surfaces in a ChatGPT response. The model is doing something more complex than counting footnotes, and that complexity works against brands that have optimised purely for traditional SEO signals.

What the model is actually rewarding

When citations do not reliably predict presence in an answer, the candidate explanations are few. One is that ChatGPT draws heavily on patterns embedded during training, not just on retrieval at query time. A brand mentioned frequently, authoritatively, and in contexts that signal category relevance across millions of documents will have a stronger prior in the model's weights than a brand that earns a single high-authority citation. Recency matters less than the depth and consistency of the signal laid down over time.

The second explanation is topical coherence. A model asked about, say, trade finance solutions does not simply fetch the most-linked page. It looks for an entity whose identity is legibly associated with that category across multiple surfaces: editorial coverage, industry publications, peer discussion, structured data, forum mentions. Brands that have built this kind of distributed presence get named. Brands that have built a strong website and a thin off-site footprint do not.

This has direct consequences for the organisations TCE works with most. A multilateral institution whose policy expertise lives almost entirely in PDF reports and a .org website has a very thin representation in the kinds of text on which large language models were trained. A global industrial group whose brand coverage concentrates in trade press rather than broader editorial faces the same problem. Neither citation volume nor domain authority, the metrics that structured web SEO for two decades, map neatly onto the signals that determine LLM presence.

The contested 85%

The open territory in Indig's data is not a single blank field. It contains categories where multiple brands are named inconsistently, categories where the model defaults to generic descriptions rather than any named player, and categories where the named brand shifts with slight variations in how the question is phrased. All three present different challenges.

Inconsistent naming is the most tractable: the category has demand, and the model has some relevant associations, but no brand has been reinforced strongly enough to win it reliably. This is where a sustained content and earned-media programme can move the dial in months rather than years.

Generic answers are harder. If ChatGPT describes what a category does without naming anyone, the model either lacks strong brand associations or has been trained to treat the category as a commodity. Owning this type of category requires building the kind of thought leadership that turns a generic answer into a specific one.

Prompt-sensitive naming is the most instructive of the three. If a brand surfaces when the question is phrased one way and disappears when it is phrased another, it has a shallow presence: real but narrow, easily displaced. Broadening the range of contexts in which a brand is discussed and cited is the remedy, which means more editorial formats, more sectors covered, more angles taken.

For financial services brands and policy institutions, prompt sensitivity is a particular risk. Their audiences ask highly specific, technical questions. A brand may be well-represented in answers to one formulation and absent from functionally identical alternatives. Mapping that variability, rather than checking a single prompt and declaring success, is the actual measurement problem.

What the window means in practice

Indig frames this as a closing window, and the framing is correct in one important sense. The categories that are unclaimed today will not remain unclaimed indefinitely. The brands that build consistent, multi-surface presence in a category now will benefit from the same incumbency advantage that early Wikipedia editors and early .com registrants enjoyed. LLMs do not neutrally resurvey the web each time they are updated; they adjust at the margins. A brand embedded in the training signal has a compounding advantage over one trying to enter later.

The implication is less about urgency for its own sake and more about sequencing. Brands that spend this period waiting for cleaner measurement frameworks, or for AI search to "stabilise," are conceding ground that will become progressively harder to recover. The 89% is an opportunity, but opportunities have a cost basis, and this one is rising.