Where US airlines produced a settled top three, industrial machinery fractures: the five engines overlapped just 0.34 (airlines: 0.64), agreed on the #1 only a third of the time, and filled most of the list with companies that aren't American. This category is molten — the answer hasn't set.
Each row is observed surfacing — how often a company entered the AI-mediated consideration set across a bounded battery of buyer questions — not a ranking. The heatmap maps citation territory: the small set of companies the engines repeatedly surface for each question. The question is not "who is best?" — it is "who appears when the buyer asks?"
Observed surfacing, not endorsement. These pages sit inside the same Content / Retrieval / Trust architecture as the rest of our working papers on AI-mediated buyer discovery.
The engines have converged on a small canonical set, reinforced by broad corroboration, accumulated entity clarity, and a large shared retrieval surface. On-page changes alone are unlikely to displace the top tier; the work shifts to defending edge cases.
The answer has not set. Engines disagree run-to-run, adjacent or foreign firms enter the answer set, and retrieval-surface quality decides surfacing. Companies can still claim citation territory with answer-shaped, entity-clear, machine-retrievable content.
Industrial machinery is the calibration example of a molten category (inter-engine overlap 0.34). US airlines is the frozen one (0.64).
Hover or tap any engine cell above to see the real prompts behind that number.
Scope & caveats. Rows show US companies (the index roster); the foreign firms the engines surfaced are reported in the finding above, not charted. This cut covers all five engines (ChatGPT, Claude, Gemini, Perplexity, Grok), web-grounded, with company names canonicalized from auto-extracted output.
Run the same instrument on industrial machinery and the airlines structure collapses. Overlap falls by nearly half (0.34 vs 0.64), only Caterpillar is a true consensus pick, and the engines disagree run-to-run even with themselves. The answer hasn't set — it's molten.
When no canonical answer exists, the engine improvises from whatever it can retrieve. That is why "leading US industrial companies" fills with Komatsu, Siemens, ABB, and Schneider: the global industrial corpus is far denser than the US-specific one, so the engines appear to retrieve on topical density and quietly drop the "US" qualifier. It is not a fact about who is American — it looks like a retrieval effect you can watch happen in the data.
In a molten category, the lower rungs of the ladder decide who gets named. With corroboration unsaturated, surfacing appears up for grabs — and it goes to whoever is most retrievable (an AI crawler can reach and parse you), most cleanly resolved as an entity, and most answer-shaped for the buyer's question. Those are precisely the things a company controls — and precisely what a diagnostic measures.
Define your content territory before it sets. Because the surface hasn't solidified, the opening is to publish the answers your buyers are actually asking, make them retrievable and entity-clear, and stake a claim while the field is still forming. In a frozen market the retrieval surface is already drawn and defended; in a molten one it is being drawn right now — and whoever maps it first tends to get named when it hardens. Hover any cell above to see the prompts behind each number — the retrieval surface, made legible.
Research publication based on sampled AI outputs collected on 2026-05-27. Findings reflect observed outputs in this sample and are not statements of company quality, ranking factors, or business performance.