We built this glossary on coining our own terms. Does AI search actually reward it?

The eighth GEO Glossary dispatch. (The sixth failed its evidence gate and was retired unpublished; we number by our project ledger, so the gap is deliberate.)

This one turns the instrument on ourselves. A core selection rule for new practitioner terms in this glossary is to coin the term for a concept that no vendor or paper has canonized, write in that empty territory, and kill a candidate at the gate if its meaning is already contested by other definers. (Established and academic terms are separate editorial classes; the rule governs what we coin, not everything we publish.) It is, stated plainly, a falsifiable claim about how AI engines pick sources, and we had never checked it against our own accumulated data. Fourteen rounds in, we did.

What we measured

Since early June we have run the same frozen prompts on the same five engines (ChatGPT, Perplexity, Claude, Copilot, Gemini), recording for each probe whether the engine cited us. The panel rotates rather than re-probing every term every week: a fuller round alternates with smaller core-watch and new-term slices, so a given term is re-probed on a schedule. We tag every term by the territory it occupies: empty-coined (a term we introduced into vacant space), academic-import (a paper concept brought into practitioner language), established (a vendor or industry owns the canonical definition), and contested (multiple definers compete, split into with-an-angle and bare). Across fourteen clean rounds that is 1,762 probes over 90 distinct tagged terms, a median of about 15 probes per term, not the 6,300 a fixed 90-by-5-by-14 grid would imply.

Two honest constraints on the read. This is one site's citation behavior, not a law of AI search. And the rates are term-weighted: for each term-engine pair we take the share of that term's probes in which the engine cited us, then average those term-level rates within each territory, so every term counts equally regardless of how many times it was probed. Every rate below is per engine; the "All" column is the unweighted mean of the five engine rates, not a pooled probe rate, because the engines do not behave alike.

What the territories earned

The term-weighted rate at which each territory was cited, any position in the answer (n is the number of distinct terms in that class):

Empty-coined is our highest-cited territory and contested-bare the lowest, a headline ninefold gap (38 percent versus 4 percent). But that headline conflates two things: territory and editorial effort. Our empty-coined entries are the ones we pour the most work into; our contested-bare handful (four terms, so read that row as directional) are terms we barely invest in and mostly intend to kill. A cleaner, more quality-comparable comparison is established versus contested-bare, both non-flagship classes: 15 percent versus 4 percent, closer to fourfold. So the territory effect is real, but its size is uncertain, somewhere between roughly fourfold (holding effort more constant) and ninefold (not), and this data cannot separate territory from quality. What survives cleanly is the direction, and its shape: an open, differentiated definitional role is cited far more often than re-entering a contested space, and academic-import at 28 percent sits high for the same reason. The gradient is really about definitional role, not the bare act of coining.

The first complication: the edge is per-engine, and its biggest number is borrowed

The rule quietly assumes coining is a general edge. It is not. Compare each engine's coined rate to its established rate, as a percentage-point difference (the ratio is fragile on the near-zero denominators, so it is secondary):

The gap is largest on Gemini (+43 points) and Claude (+33), and it essentially vanishes on ChatGPT (+2). Two things make that more pointed than a simple "engine-conditional" caveat. First, ChatGPT is the largest of these engines by usage, and it is precisely where coining buys nothing: a rigorous established-term page is cited about as often as a coinage, because ChatGPT gates citation by query intent (definitional and factual questions get citations, comparative and procedural ones less so), not by who owns the term. Second, the biggest number in the table, Gemini's, is substantially Google rank in disguise. We found in an earlier dispatch that Gemini's preference for our coinages tracks our Google ranking for them, an empty-coined term is usually the top Google result by default, and that when a competitor outranks us on Google, Gemini switches to the competitor. So the honest reading is not "coining is a huge lever on the framing-transmitting engines." It is narrower: the edge is real but concentrated on two engines, one of which (Gemini) is largely reflecting Google rank, while on the highest-traffic engine it is flat.

The second complication: monopoly is not what earns the citation

Here is the finding that surprised us most, because it undercuts the rule's own logic. The rule says: avoid contested space, because competitors dilute you. So within our coined terms, the ones with no competing definer should be cited across more engines than the ones facing competitors. They are not. We scored each of our twelve empty-coined terms for how many direct competitors an engine had available, and correlated that with cross-engine breadth (how many of the five cite us). The correlation is essentially zero (Spearman rho +0.13). Our most contested coinage, cite-ability, has four direct same-concept competitors that engines have demonstrably cited, and it is still cited by four of five engines. Citation match rate, whose phrase collides with an unrelated established meaning, is cited by all five. At twelve terms this is weak evidence: a near-zero correlation in so small a sample cannot rule out a moderate effect. It can say we found no relationship, not that none exists.

What that shows is bounded, and worth stating precisely because it is our own strategy on the line. Competitor presence being uncorrelated with breadth means monopoly is not what protects the citation: an engine that has a rival definition available still includes our page. But it does not establish what does. Deference to a clean definition is one candidate; editorial quality and search demand are others, and this test cannot separate them, because our coined pages are also our most-worked pages. There is a further limit: we counted how many competitors exist, not whether they outrank us, and our own earlier work suggests rank, not count, is the variable that flips a citation. The phrase-collision result points the same way, and belongs here rather than in a footnote: the coined terms whose phrase collides with an established meaning (and are therefore searched more) are cited more broadly, which looks like demand, not monopoly and not coining, driving breadth. The safe conclusion is the negative one: within our coined terms, having competitors present did not cost us the citation.

Reconciling with the terms that "lose"

If competition does not cost us the citation, why did a recent round show cite-ability losing to a competitor on one engine? Because "losing" and "not being cited" are different axes. On the inclusion axis (is our page in the answer at all), cite-ability is still broadly cited. What it loses, on one engine, is the top slot: the competitor takes the first-cited position while we remain in the answer. A contested coinage can be included on most engines and still surrender prominence on some. When practitioners say a term "went contested," they usually mean the prominence contest, not that the citation vanished. Our data says the citation mostly does not vanish; the ranking rotates.

A caveat on a tempting claim from the same data: it is true that none of our twelve coined terms shows fewer citing engines now than two months ago. But we measure breadth cumulatively (engines that have ever cited a term in a clean round), and a cumulative count can only hold or rise, so "none lost breadth" is partly an artifact of the metric, not evidence of stability. A rolling per-round measure would test stability properly, and we do not have a clean one yet. The weaker thing we can honestly say: no coined term has fallen to zero current citation.

What this means if you run a content site

• Coin the practitioner term for a concept no one has canonized, and write the clean definition. At the territory level this is the highest-cited class we measure, though we cannot fully separate the territory from the extra effort these pages receive.
• Do not expect it to work uniformly. The coined edge is large on Gemini and Claude and roughly zero on ChatGPT, the highest-traffic engine, where a rigorous established-term page is cited just as readily, and the largest edge (Gemini) partly reflects Google rank rather than any preference for coinages. Pick the move by the engine you care about.
• Do not over-read competitor presence. A rival defining the same concept did not, in this sample, remove us from the answer; it competed for the top slot. Write the cleaner definition and you tend to stay cited even when you are not first.

Limits, honestly

This is one site, our own, audited against our own tags, so the territory rates are a property of this glossary and its niche, not a universal citation law. The territory effect is confounded with editorial effort, which we cannot separate here. The within-coined test rests on twelve terms and is an exploratory correlation, not a fitted model; it measures competitor count, not competitor rank, and its near-zero result cannot rule out a moderate effect. Breadth is measured cumulatively, so its stability is partly definitional. And "any position in the answer" folds a first-cited and a buried citation together; measured as first-cited only, the ranking of territories holds but the absolute rates drop. So the honest headline is narrow: on this site, citation rate varies sharply by definitional territory, the association is real but its size is uncertain and highly per-engine, and we cannot yet say coining itself causes the advantage rather than rank, query intent, or the effort these pages receive. The way to answer that is a multivariate test, territory alongside Google rank, term age, competitor rank, and query type, which becomes possible once we hold thirty to fifty coined terms and a longer panel. That is the next step, and this dispatch is the baseline it will be measured against.