Incumbency breaks at 0.075 stars

Incumbent advantage is the tendency of an AI assistant to recommend the brand it already recognizes when nothing in the available evidence separates the options, and it is the single easiest GEO problem to misdiagnose.

It looks like a wall. It behaves like a tiebreaker.

#The wall

A June 2026 study by Xi Chu and Yu Peng Hou put three commercial models (GPT-4o-mini, Claude Sonnet, Gemini 3 Flash) in front of skincare product choices where the products were specified identically, varying only the brand name. The result, from Incumbent Advantage: Brand Bias and Cognitive Manipulation Dynamics in LLM Recommendation Systems, is about as clean as behavioral findings get: "the real brand is recommended in 100% of 670 valid trials (IAI = 10.0, the theoretical maximum; all p<.001)". IAI is their Incumbent Advantage Index, scored 0 to 10.

Read that carefully, because the setup is the whole finding. The competing products were identical. Same specs, same everything, different name. Given nothing to reason about, the model reached for the only signal left in the prompt, which was its prior about the brand. That is not a ranking decision. That is a fallback.

The category matters too. Skincare is an experience good, where quality is hard to judge before purchase. The authors ran robustness checks on search goods (USB-C cables, AA batteries) where specifications do more of the work. Expect the effect to be strongest exactly where buyers cannot verify claims themselves.

#Where it breaks

The same paper ran a gradient experiment across 9,220 valid trials (of 9,600, with 380 parse failures), giving the unknown challenger progressively larger advantages. The 50% breakthrough threshold, by linear interpolation, was "a +0.075-star rating advantage".

Less than a tenth of a star. Smaller than the gap between a 4.3 and a 4.4.

So the monopoly is real, and it is also conditional and cheap to break. One comparable, quantified difference is enough to displace a brand prior that otherwise wins every single trial. The practical reading for anyone working on AI visibility: if your product is genuinely competitive and the assistant still names the incumbent, the likely failure is not that you are unknown. It is that nothing in the retrieved evidence lets the model tell you apart.

Model choice changes the price of admission. Chu and Hou report that "Claude is hardest to flip (L1 rating BR = 11% vs. 94% GPT, 88% Gemini)", where L1 is the smallest advantage tested. At the same small edge, two of the three models flip most of the time and one barely moves. If you benchmark on a single assistant you will draw the wrong conclusion about how much differentiation you need.

#The shortcut, and why it eats itself

The study also tested a second way through the wall: authority-style marketing language, including fabricated clinical-evidence claims. It worked. The authors price the effect at "Authority's BSV = +0.17 rating points = 15.3% price discount", where BSV (Bias Surplus Value) converts persuasive language into product-quality equivalents. Confident clinical-sounding prose bought more than twice the lift of the honest rating edge.

That is the finding most likely to be misused, so it is worth being blunt about the economics rather than only the ethics.

Chu and Hou modeled what happens when every brand adopts the same tactic. Their payoff proxy falls from "+0.802 (first mover)" to "+0.007 (universal adoption)". The advantage is entirely positional. It exists because other brands are not doing it yet, and it rounds to nothing once they are. Meanwhile the brands that abstained received zero recommendations in their tests, which is the part that makes this a genuine social dilemma rather than a simple morality tale.

This matches the broader literature. The July 2026 critical survey Optimizing Visibility in Generative Engines, reviewing 45 studies, lists "competition can erode individual gains" among its headline conclusions, and finds that "no reviewed technique shows a stable, longitudinal, cross-platform causal effect on organic discoverability or downstream behavior". Tactics that depend on scarcity decay on a schedule. Verifiable attributes do not.

There is also a detection asymmetry. A fabricated clinical claim is a factual assertion about your product, sitting on your own domain, in a medium where third parties are increasingly checking. The rating edge is checkable in your favor. The clinical claim is checkable against you.

#Fame is not the only prior

Brand recognition is one input to the fallback. Origin is another. Auditing Preferences for Brands and Cultures in LLMs ran ChoiceEval across Gemini, GPT and DeepSeek over 10 commerce and culture topics and more than 2,000 questions, and found "marked favouritism toward American entities" in the US-developed models, with DeepSeek showing "more balanced yet still detectable geographic preferences". Those patterns held across user personas, which points at something systematic rather than incidental.

Deployment context shifts it further. Your LLM Has a Passport, presented at FAccT 2026, found that supplying brand origin changed rankings, and that user location had the strongest effect of all, shifting continental representation in Qwen by up to 59%.

None of that is addressable by rewriting your page. It is context you should measure and plan around, not a lever.

#What to actually do

The operating principle is narrow: supply the discriminator, in a form a retrieval layer can lift.

Publish comparable attributes rather than adjectives. A model cannot act on "gentle formula". It can act on a pH figure, a concentration, a warranty term, a measured latency, a price per unit. The winning margin in the study was a single number that both options possessed.

Make third-party evidence exist. The threshold the paper identified was a rating advantage, which is not something you write. It is something review corpora, listings and independent tests carry on your behalf.

Put the numbers in structured data, not only prose, so the attribute survives extraction.

Test on more than one assistant, and on more than one category. The 11% versus 94% breakthrough spread at the smallest tested edge is the difference between "we need much sharper evidence" and "we already had enough".

Finally, be honest about the limits of this evidence. One lab, three models, one primary category, a constructed choice task rather than a live assistant with live retrieval. The direction is well supported and consistent with the wider survey work. The exact thresholds are not constants to plan a budget around.

The useful takeaway is structural. An AI assistant defaults to the familiar name only when you have given it nothing better to go on. Most of the time, that is a fixable content problem.