What the latest scientific findings mean for the GEO industry

Last updated: 18 October 2026

Loud & Lexis, an agency specialising in Generative Engine Optimization (GEO, the optimisation of brand content for AI answers), studied position bias in AI brand comparisons in October 2026. According to the Loud & Lexis whitepaper, a current AI model with web search favours the brand named first in the prompt in close comparisons. The finding matters to marketing leaders in two areas. It affects how they measure their own AI visibility in AI visibility monitoring (the ongoing measurement of how often AI systems mention and recommend a brand) and the role of brand awareness in AI recommendations.

What do the latest scientific findings show about position bias in AI brand comparisons?

The Loud & Lexis study measures a position bias (also called positional effect, the preference for the option named first) in AI brand comparisons. When two brands are evenly matched, a current AI model with web search recommends the brand named first in almost two out of three cases. Clear favourites win in both orders, the effect mainly affects close comparisons between similarly strong brands.

The test

Same question, brands swapped: what does the AI recommend?

Loud & Lexis asked every question twice, once with brand A first and once with brand B first. This is how the results break down across 100 such question pairs:

• 79Same recommendation both times. The order makes no difference.
• 15The recommendation flips to the brand named first. The AI recommends whichever brand comes first.
• 6The recommendation flips to the brand named second. The AI recommends whichever brand comes last.

When swapping the order flips the recommendation, the brand named first wins in almost 3 out of 4 cases (72.9%).

For the study, Loud & Lexis asked the model gpt-5.6-luna with live web search head-to-head questions of the form “X or Y, which one should I choose?”. Every question ran in both orders, once with brand X first and once with brand Y first. The results are published in the whitepaper “First come, first recommended?” by Eike Henrich, Jimmy Park and Tobias Peschke.

The Loud & Lexis study covers 482 matchups in 97 categories with a total of 4,608 answers from the model. Across all answers, the brand named first won 54.7 percent of the time, for evenly matched brands 63.3 percent.

At a glance

Position bias in four numbers

Results of the Loud & Lexis study with the AI model gpt-5.6-luna and live web search

almost2 in 3comparisons are won by the brand named first when both brands are evenly matched (63.3%)

1 in 6matchups flip their winner when the order of the brands is swapped (16%)

18 of 21brand groups studied show the advantage for the brand named first

4.7 pointsis the average shift in a brand’s win rate when measuring in only one order

Position bias has been known in research on language models since 2023, for example from the work of the research teams led by Zheng (2023) and Wang (2024). The Loud & Lexis whitepaper cites both and adds real brands, realistic purchase situations and a model with active web search.

What the Loud & Lexis study does not show

The Loud & Lexis study measures position bias in a single model at a fixed point in time, namely gpt-5.6-luna in October 2026. According to the whitepaper, other models may show weaker, stronger or reversed effects. In addition, the study design only captures the final answer and cannot separate effects of the web search from text generation.

The compared brands come mostly from client projects of Loud & Lexis and Opascope, the employer of co-author Jimmy Park. The authors did not measure how real buyers order brand names in their prompts. The study therefore states all claims about user behaviour strictly as a hypothesis.

According to Loud & Lexis, position bias does not mean that the brand named first is automatically preferred. With clear favourites, the strength of the brand determines the model's recommendation. The order in the prompt only takes effect when the model has no clear preference between two brands.

How should prompts be tracked correctly?

Head-to-head prompts should be tracked in both orders and a win counts only if it holds in both orders. Loud & Lexis derives this measurement rule from its own study. Head-to-head prompts are inputs to an AI system that put two brands side by side, such as “Brand A or brand B?”.

Many AI monitoring setups (systems that measure brand visibility in AI answers) compare their own brand directly with competitors and run each head-to-head prompt in only one order. According to Loud & Lexis, the brand named first in the prompt thus gains a systematic advantage in the measurement in close matchups. The measured win rate (the share of comparisons won) then partly reflects the order of the prompt.

How much does measuring in only one order distort the win rate?

According to Loud & Lexis, measuring in only one order shifts a brand's win rate by 4.7 percentage points on average. The focus brands of the study won 62.8 percent of their comparisons when named first and 53.4 percent when named second. In 16 percent of matchups, about one in six, the two orders produced different winners.

First or second

The same brands win more often when they are named first

Win rate of the study’s 21 focus brands against their competitors, depending on where they appear in the question

“Your brand or competitor?”Your brand comes first

62.8%

“Competitor or your brand?”Your brand comes second

53.4%

50% = tie

Brands, competitors and questions are identical. The gap of 9.4 percentage points comes from the position in the question alone.

According to Loud & Lexis, a win rate of 60 to 70 percent measured in one order can correspond to a tie in close matchups. A lead over a competitor measured in one order can therefore arise in monitoring reports from the prompt order alone.

Where the effect works

The closer the matchup, the more often the brand named first wins

Every matchup was answered ten times. 10:0 means one brand won all ten answers. 5:5 means a tie. The bars show how often the brand named first won.

50%

54%

59%

67%

63%

64%

10:0230 matchups

9:171 matchups

8:250 matchups

7:338 matchups

6:433 matchups

5:523 matchups

Clear favourite: order irrelevantcloser and closer →Close matchup: almost 2 in 3 for the brand named first

Dashed line: 50% = no advantage from the order

Three measurement rules from the Loud & Lexis whitepaper

The Loud & Lexis whitepaper sets out three rules for head-to-head prompts in AI monitoring:

1. Ask in both ordersEvery head-to-head prompt runs as “A or B?” and as “B or A?”. Both results are analysed as a pair.
2. Ask several timesEach order is asked several times. Only the sum of the answers gives a reliable win rate for the brand.
3. Count a win only if it holds in both ordersAccording to Loud & Lexis, this rule prevents a position advantage from entering reports as brand strength.

According to the research cited in the Loud & Lexis whitepaper, instructing an AI model to ignore the order does not fix the effect. In its own AI Visibility Monitoring, Loud & Lexis now also runs head-to-head prompts in both orders.

What does position bias mean for brand awareness and awareness marketing?

Position bias can give the better-known brand an advantage in AI answers as well, if buyers name that brand first in the prompt. That is the hypothesis of the Loud & Lexis study. In real use, the buyer decides which brand comes first in the prompt.

The authors assume that buyers first name the brand with the highest top-of-mind awareness (the brand that comes to mind first) or the brand they already prefer. If this assumption holds, the better-known brand comes first more often in close comparisons and is recommended more often. Loud & Lexis did not measure the naming order of real buyers.

The Loud & Lexis whitepaper also refers to a study by Chu and Hou (2026) in which well-known brands won every recommendation when the product data were identical.

From this hypothesis, Loud & Lexis derives the recommendation to keep giving weight to top-of-funnel and awareness marketing (activities that build brand awareness at the start of the purchase decision) in the budget. By this logic, awareness that a brand builds through traditional channels can also shape recommendations in AI answers.

Conclusion

The Loud & Lexis study shows that a current AI model with web search measurably decides close brand comparisons in favour of the brand named first. Two consequences follow for the GEO industry. Head-to-head prompts in AI monitoring belong in both orders, otherwise the prompt order enters reports as brand strength. According to the hypothesis of the study, brand awareness remains a lever that can also shape AI recommendations.

So far, the Loud & Lexis results apply to a single model at a fixed point in time. Companies that already measure their AI visibility should check whether their setup runs head-to-head prompts in both orders. The whitepaper describes the measurement rules and the limitations of the study in detail.