The shift from traditional to generative search has fundamentally transformed online product discovery. Because conversational recommendations synthesize choices into a few direct answers, subtle systemic preferences toward specific providers can exert profound commercial impacts. In this work, we audit this phenomenon through two distinct dimensions: (i) brand popularity bias, the systematic preference for market-dominant entities over long-tail alternatives, and (ii) brand familiarity bias, the ranking advantage an entity encoded in an LLM's parametric memory holds over an unfamiliar competitor. First, evaluating commercial queries across major engines, we observe persistent brand popularity bias: dominant brands in the top quintile are cited 1.4x to 2.8x more frequently than the average brand within a sample of 1,176 top market players, already present in parametric memory that external web retrieval fails to mitigate. Second, through controlled in-context experiments holding product attributes constant, we find that substituting authentic brand signatures with unfamiliar ones significantly degrades recommendation rank. Conversely, fact-preserving content rewrites yield negligible ranking gains. Taken together, our findings demonstrate that generative search engines reinforce a ``rich-get-richer'' dynamic by favoring established market leaders and imposing a parametric barrier to entry that standard content optimization cannot overcome.
Figures & tables
Figure 1. Study 1 is an audit of three deployed generative search engines on commercial prompts, with search-on and search-off , and with and without a persona context. Study 2 takes 24 of these prompts and manipulates two features of one product page, the brand that signs it and its text (GEO rewrites), while four models rank it among five pages with search-off . Flow diagram of the two studies. Study 1 sends commercial prompts with and without personas and search to three engines and keeps search queries, sources and answer text. Study 2 has four models rank five product pages while one page changes its signature or its text.
Model
Study 1
Study 2
Ordered search actions
Searches
gpt-5.5
✓
✓
✓
98%
gemini-3.6-flash
✓
✓
✗
44%
perplexity/sonar
✓
✓
✗
100%
deepseek-v3.2
✗
✓
–
–
Table 1. Engines of the two studies. Study 1 , Study 2 : whether the engine is used in the study; Study 1 runs each engine with search-off and search-on , Study 2 only with search-off . Ordered search actions : whether the API returns the engine’s successive searches in the order they were made. Searches : share of the Study 1 calls with search-on in which the engine searched the web. In the results, we refer to the models as GPT, Gemini, Perplexity, and DeepSeek.
Figure 2. Mention rate as a function of popularity, with search-off and search-on , all sectors and personas together. Share of answers that mention the brand, with search-off (grey) and search-on (green), by quintile of popularity within each engine and sector. Points are means over the brands of each quintile (about 240), with 95% intervals from 1,000 prompt bootstrap samples, stratified by sector. Three line charts, one per engine, of the share of answers mentioning a brand by quintile of popularity, with almost identical lines for search-off and search-on.
Figure 3. Mention share and citation share by quintile of popularity, search-on , only for brands with a known domain. Each rate is divided by its mean over the brands of the same engine and sector. As such, 1 is the average brand (dashed line) and the levels of mention and citation are not comparable on this scale. Quintiles are formed within each engine and sector; points are means pooled over the four sectors, with 95% intervals from 1,000 prompt bootstrap samples, stratified by sector. Three line charts, one per engine, of mention and citation rates relative to their engine-sector mean, by quintile of popularity, with a dashed line at the average brand. Both rates rise with popularity on all three engines, most in the top quintile.
Figure 4. Similarity of answers—in terms of sources, brands, and text—as a function of similarity of search queries. Query similarity matches each search query of one answer with the most similar search query of the other, and averages their cosine in both directions. Answer similarity is rescaled so that 0 is the mean similarity of two answers to different prompts and 1 the mean similarity of two repeats (dashed lines). Values above 1 mean that two answers are more alike than two repeats on average, as happens when their search queries are very similar. Points: mean of the pairs in each of eight bands of query similarity; bands: 95% intervals from resampling prompts. Three line charts, one per engine, of the similarity of sources, brands and answer text against the similarity of the search queries of two answers. All three similarities rise with query similarity, sources and brands most steeply.
Figure 5. GPT, search-on . For each quintile of popularity, the share of mentioned brands whose name is already in GPT’s first search action (dark), or appears only in later actions (light). Brands are split into quintiles of popularity within each sector, from least popular (l.p.) to most popular (m.p.). Bars: 95% intervals from resampling prompts. ∗ The first-action share of the most popular quintile differs from that of the least popular quintile, significant after Benjamini–Hochberg over the four sectors. Four line charts, one per sector, of the share of brands mentioned by GPT whose name is in its first search action or only in later actions, by quintile of popularity.
Figure 6. Mention rate by quintile of popularity, without a persona context (black) and averaged equally over the three personas (blue), search-off and search-on together. Points: mean share of answers mentioning a brand, pooled over sectors; bars: 95% intervals from 1,000 prompt bootstrap samples, stratified by sector. Three line charts, one per engine, of the mention rate by quintile of popularity without a persona context and averaged over the three personas. Both lines rise with popularity; the persona line is lower in the most popular quintiles.
Figure 7. Study 2: mean rank of the target page (1 = ranked first; better ranks are higher in the plot) under each signature. The horizontal axis represents the manipulation on the brand that signs the page. Point style represents the manipulation on the product, filled points have a Real product name of the corresponding brand, and hollow points an Invented product name. (a) Mean over the original text and the three rewrites. (b) Colors encode the text rewriting strategy, including Original , Structure , Fluency , and Authoritative . Points: mean rank over the four models and over the questions in which each version exists (15 of the 24 for Better known Real product and Less known Real product ). Bars: 95% cluster bootstrap intervals over questions. Two dot plots of the mean rank of the target page under each signature, with filled points for the real product name and hollow points for an invented one. Left: one point per signature, averaged over the four rewrites, from the true brand with its real product, ranked best, to the invented brand, ranked worst. Right: one point per signature and text; the four rewrites have almost identical ranks under every signature.
What changes on the target page
Δ
Holm p
Who signs it (Figure 7 a)
Original brand Real product vs Invented brand
H4a
+0.44
<.01
Original brand Real product vs Original brand Invented product
H4b
+0.14
<.05
Better known Invented product vs Less known Invented product
H4c
+0.05
n.s.
Original brand Invented product vs mean of Better known and Less known Invented product
H4d
+0.08
n.s.
Less known Invented product vs Invented brand
H4e
+0.19
<.01
Table 2. Study 2: change in the rank of the target page ( Δ , positions on a scale of 1–5, positive = ranked higher) between the two versions each row compares, named as in Figure 7 ( Real product, Invented product). Each signature row is the difference between two points of Figure 7 a. Mean over the 24 questions of the mean over the four models; one-sided p , Holm-adjusted (n.s.: not significant). Effects of the signature average the four rewrites; effects of rewriting average the five signatures.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
ID
Prompt
Keyword
Volume
assicurazioni_01
Quale assicurazione RC moto scegliere?
assicurazione rc moto
133
assicurazioni_02
Quale assicurazione annullamento viaggio mi conviene scegliere in base a cosa copre?
assicurazione annullamento viaggio cosa copre
102
assicurazioni_03
Quale assicurazione bici dovrei scegliere?
assicurazione bici
695
assicurazioni_04
Qual è la migliore assicurazione casa con responsabilità civile da scegliere?
assicurazione casa responsabilità civile
249
assicurazioni_05
Quale polizza famiglia mi consigliate?
polizza famiglia
104
assicurazioni_06
Qual è la migliore assicurazione kasko da scegliere?
migliore assicurazione kasko
118
Appendix
Table A.1. The 80 prompts of Study 1, with the keyword they were derived from and its mean monthly search volume in Italy over twelve months (DataForSEO).
Sector
Persona
Text
Insurance
Baseline persona (A)
Sono genitore con una famiglia a carico. È la prima volta che stipulo una polizza. Cerco di spendere il meno possibile. I am a parent with a family to support. It is the first time I take out a policy. I try to spend as little as possible.
Insurance
Price variant (B)
Sono genitore con una famiglia a carico. È la prima volta che stipulo una polizza. Spendo volentieri di più per la qualità. I am a parent with a family to support. It is the first time I take out a policy. I am happy to spend more for quality.
Insurance
Profile variant (C)
Sono al primo anno di patente. È la prima volta che stipulo una polizza. Cerco di spendere il meno possibile. I am in my first year with a driving licence. It is the first time I take out a policy. I try to spend as little as possible.
Travel
Baseline persona (A)
Sono single e viaggio senza figli. Ho un’esperienza nella media con i viaggi. Cerco di spendere il meno possibile. I am single and travel without children. I have average experience with travel. I try to spend as little as possible.
Travel
Price variant (B)
Sono single e viaggio senza figli. Ho un’esperienza nella media con i viaggi. Cerco il miglior rapporto tra qualità e prezzo. I am single and travel without children. I have average experience with travel. I look for the best value for money.
Travel
Profile variant (C)
Sono genitore di figli piccoli. Ho un’esperienza nella media con i viaggi. Cerco di spendere il meno possibile. I am a parent of young children. I have average experience with travel. I try to spend as little as possible.
Appendix
Table A.2. The twelve buyer personas of Study 1, placed before the question in Italian; English translation below each.
Set of brands
Brand rows
Search off
Search on
Difference
95% interval
Per sector, ≥5 (main)
3,567
+0.87
+0.89
+0.02
[−0.03,0.07]
Per engine, ≥5
2,101
+1.34
+1.35
+0.01
[−0.07,0.10]
Per sector, ≥10
2,463
+1.05
+1.08
+0.03
[−0.04,0.11]
Per engine, ≥10
1,450
+1.61
+1.62
+0.01
[−0.11,0.12]
Appendix
Table C.1. Study 1: popularity bias under different sets of brands. Slopes as in Table C.2 , all engines and sectors together. Main analysis: one set of brands per sector, made of the brands that at least one engine mentions in 5 or more of its 1,600 answers in that sector, with zeros for the engines that do not name them. Alternatives: a separate set for each engine (brands that this engine mentions at least 5 times), and a threshold of 10. A threshold below 5 is not possible, because popularity was measured only for the brands above it.
Search off
Search on
Difference
95% interval
p
GPT
Insurance
+0.71
+1.52
+0.81
[0.27,1.49]
0.031
Fashion and beauty
+1.09
+0.91
− 0.18
[−0.36,0.00]
0.103
Sport
+1.41
+1.10
− 0.30
[−0.61,0.03]
0.103
Travel
+0.70
+0.54
− 0.16
[−0.28,−0.05]
0.031
All sectors
+0.94
+0.86
− 0.07
[−0.19,0.03]
0.205
Appendix
Table C.2. Study 1: how many percentage points the mention rate of a brand rises for each tenfold increase in its search volume, with search off and with search on. Each slope is an OLS fit with one point per brand; the difference between the two slopes is also the slope of the per-brand effect of search (Figure C.4 ). Rows for all sectors and for all engines add engine × sector fixed effects. 95% intervals from resampling brands within each engine and sector; p from a t test with robust (HC3) errors, corrected with Benjamini–Hochberg over the 12 engine–sector rows (summary rows uncorrected).
Brands
Mention
Citation
Difference
95% interval
GPT
702
+0.52
+0.51
+0.02
[−0.08,0.11]
Gemini
702
+0.56
+0.18
+0.38
[0.18,0.58]
Perplexity
702
+0.57
+0.83
− 0.26
[−0.44,−0.10]
All engines
2,106
+0.55
+0.50
+0.04
[−0.05,0.14]
Appendix
Table C.3. Study 1: slope of the normalised mention and citation rates on popularity, search on. Each rate is divided by its mean in the engine and sector, so a slope of 0.5 means that a brand searched ten times more often has a rate higher by half the average of its engine and sector. OLS with engine × sector fixed effects; 95% intervals from 2,000 resamples of brands within each engine and sector.
Figure C.1. Mention rate by quintile of popularity with search off (grey) and on (green), by engine and sector; 95% intervals from 1,000 prompt bootstrap samples, stratified by sector. Δ slope: difference between the two slopes of Table C.2 ; ∗∗ significant after Benjamini–Hochberg over the 12 engine–sector pairs, ∗ p<0.05 before correction. With search off, Perplexity almost never mentions an insurer. Grid of line charts, one per engine and sector, of the mention rate by quintile of popularity with search off and on.
Figure C.2. Mention and citation against popularity, one panel per engine and sector, search on. Points: brands (mention: all brands with a popularity value; citation: brands with a known domain). Lines: logistic fit of the share of answers on popularity, with 95% band (quasi-binomial standard errors). In each panel, Spearman’s ρ ; ∗ unadjusted p<0.05 for Spearman’s correlation, not for the fitted curve. Popularity is log10(1+v) , where v is monthly Google search volume in Italy. The horizontal axis is shared across engines within each sector. The vertical limit is based on the 99.3rd percentile of the observed rates; the fits use all brands. Grid of scatter plots, one per engine and sector, of mention and citation rates against popularity with fitted curves.
Figure C.3. Mention and citation by quintile of popularity in absolute terms, search on, for brands with a known domain. Quintiles are formed within each engine and sector. Points are the mean share of answers that mention the brand (green) or cite it (orange), pooled over the four sectors, with 95% intervals from 1,000 prompt bootstrap samples, stratified by sector. On Perplexity the sources are the retrieved results, so its citation share can exceed its mention share. Three line charts, one per engine, of mention and citation rates in percent by quintile of popularity. On GPT both rise; on Gemini mention rises to about 6 percent while citation stays below 1 percent; on Perplexity the two lines overlap.
Figure C.4. Effect of search-on on each brand (percentage points) against popularity ( log10 , centred within engine and sector), one point per brand and engine, with OLS line and 95% bootstrap band over brands. 34 of 3,567 points fall outside the vertical range shown. Scatter plot of the effect of search on each brand against popularity, with a flat fitted line near zero.
Figure C.5. Is the name of a mentioned brand in the sources the engine returned? Mentions in calls with observed search queries, for brands that reach the threshold of Section 3.1 in the engine itself, by quintile of popularity (brands split within engine and sector, as in Figure 3 ). Top: share of mentions whose name is found in the returned sources, that cannot be verified because some source could not be read, or with no match although every source was read. Bottom: the same mentions read the other way, how each outcome is distributed across the quintiles of popularity; numbers above the bars are mentions. GPT has no mention with no match. The match is lexical and the pages were read after the calls. Stacked bar charts per engine. Top: share of mentions whose name is found in the returned sources, cannot be verified, or has no match, by quintile of popularity. Bottom: distribution of each outcome across quintiles.
Figure C.6. Mention rate by quintile of popularity and persona level, one panel per engine and sector, search off and on together; bands: 95% intervals from resampling prompts. Grid of line charts, one per engine and sector, of the mention rate by quintile of popularity for no persona context and the three personas.
ID
Question
Target
Better known
Less known
Invented
assicurazioni_03
Quale assicurazione bici dovrei scegliere?
AXA
Unipol
Qover
Dalveris
assicurazioni_09
Quale assicurazione sui voli mi consigli di scegliere?
Heymondo
Genertel
ERGO Assicurazione Viaggi
Trevalda
assicurazioni_10
Qual è la migliore assicurazione natante da scegliere?
Cattolica Assicurazioni
Allianz
Zurich
Corlenta
assicurazioni_17
Qual è l’assicurazione RCA migliore da scegliere?
Allianz Direct
Prima Assicurazioni
Admiral
Corvalia
assicurazioni_19
Quale assicurazione spese veterinarie dovrei scegliere?
ConTe.it
UniSalute
Baboop
Nidrava
assicurazioni_20
Quali sono le migliori opzioni per l’assicurazione vita in base ai costi?
GenialLife
Unipol
SquareLife
Pelora
Appendix
Table D.1. The 24 questions of Study 2 with the target page and its signatures. Better known and less known are other real brands mentioned in the Study 1 answers to the same prompt; the invented brand has no registered .it domain.
ID
Better known, real product
Less known, real product
Above
Below
Recognised
assicurazioni_03
Unipol Unica Mobilità
Qover Qover-me
0.41
2.58
8%
assicurazioni_09
–
–
0.34
1.88
50%
assicurazioni_10
Allianz Columbus
Zurich Permare
0.61
0.42
0%
assicurazioni_17
–
–
0.31
2.09
0%
assicurazioni_19
–
–
0.82
1.60
25%
assicurazioni_20
Unipol Vita Smart
SquareLife Turtleneck
1.80
0.77
0%
Appendix
Table D.2. Study 2, per question: the real products used for the better-known and less-known brands (– where that signature has no real-product version), the distance in popularity between the target and each of the two real brands (orders of magnitude of monthly search volume: better known minus target, target minus less known), and how often the models name the true brand from the neutralised target page (12 answers per question).
Table E.1. Study 2 by model and by sector: positions gained by the target page (positive = ranked higher), computed as in Table 2 . Each column is one row of that table, in the same order ( real, invented product name, as in Figure 7 ). Models average the 24 questions; sectors average six questions and the four models. Descriptive.
Table E.2. Study 2: every pair of the seven signatures ( real, invented product name, as in Figure 7 ). Δ : positions gained by the target under the first signature over the second, averaged over the four texts; 95% intervals from resampling questions. Pairs with a real-product version of Better known or Less known use the 15 questions where it exists. Family: H4a–H4e are rows of Table 2 (one-sided p , Holm over its seven rows); R is the family of seven comparisons with the real-product versions (two-sided p , Holm over the seven); the other pairs are exploratory (two-sided p , not adjusted).
Prediction
Outcome
RQ1, Study 1: brand popularity bias
H1a
More popular brands are mentioned more often
Supported in 23 of 24 configurations
H1b
More popular brands are cited more often
Partially supported: 8 of 12 engine–sector pairs
H1c
The brand popularity bias is stronger with search-off than search-on
Not supported (+0.02 [ − 0.03, 0.07])
H2a
Answers with more similar search queries share more sources and text
Supported
H2b
Their mentioned brands do not overlap more
Not supported: brands follow the search queries almost as closely as sources
Appendix
Table F.1. Hypotheses, predictions and outcomes, by research question. Gains in Study 2 are positions on a scale of 1 to 5.
Large language models (LLMs) are becoming a major way for consumers to find products, but we do not yet understand how brands compete in this new channel. We study brand dynamics in LLM recommendations using skincare products -- a category where consumers cannot easily judge quality before buying and must rely on brand reputation -- across three commercial LLMs (GPT-4o-mini, Claude Sonnet, Gemini 3 Flash), with a robustness check on search goods. In three experiments, we find: (1) a Conditional Monopoly where well-known brands get recommended 100% of the time (IAI = 10.0) when all products have the same specifications, but this dominance disappears with less than a +0.1-star rating advantage for a competitor; (2) authority-style marketing language, including fabricated clinical-evidence claims, breaks this monopoly at a Bias Surplus Value equal to +0.17 rating points, with each model responding differently; and (3) a social dilemma in multi-brand GEO competition: when all brands adopt the same optimization strategy, individual payoff falls from +0.802 to +0.007 in our payoff proxy, and non-participating brands receive zero recommendations in our tests. Our results suggest that generative engine optimization (GEO) should be studied not only as a security risk, but also as an emerging marketing practice that shapes market competition.
People increasingly get answers straight from AI search engines like ChatGPT, Claude, Perplexity, and Gemini rather than scrolling search results. Brands that once focused on search engine optimization (SEO) must now optimize for how these engines represent, cite, and recommend them -- a shift variously called Generative Engine Optimization (GEO), Answer Engine Optimization (AEO), and AI Search Visibility. We treat AEO and AI Visibility as part of GEO, and study how to measure brand visibility across AI engines: what they value when they cite a brand, which sources they rely on, and what content large language models surface. The hard case is everyone outside the already-authoritative top brands -- SMEs, D2C brands, creators, and early-stage startups. We analyze 100K+ prompt responses across 100+ brands tracked on Ranqo between March and May 2026. First visibility runs form a clear three-tier brand-stature ladder: global household names (e.g., Stripe, Nike) appear in 73% of relevant AI answers on their first run; established mid-market and regional brands (e.g., Olipop, Klaviyo) in 44%; niche and small brands in just 11% -- about 30 percentage points per step. When engines cite sources, about 78% go to corporate websites; among non-corporate sources YouTube leads, ahead of Reddit, editorial media, and Wikipedia. The highest-leverage page is the ranked "best-of" listicle, the most-cited content format at about 21% of all citations. Sentiment is the unstable signal: whether a brand is framed positively or negatively flips about 6.7 times more often than whether it is mentioned at all. These findings provide a first large-scale baseline for measuring GEO: AI brand visibility can be measured, differs by platform, and varies strongly by brand maturity. We close by proposing seven v1.1 protocols to test whether specific recommendations can causally improve AI visibility.
AI assistants like ChatGPT and Claude are recommendation engines, not search engines: they answer commercial queries by directly nominating brands rather than returning a list of links. Marketing to AI is therefore a broader problem than "show up in search" -- positioning, content, and product fit matter as much as discoverability. We audit ~37,000 production runs across four model configurations and 215 commercially-framed prompts spanning 19 sectors, evaluated against a 533-brand reference catalog stratified into five prominence tiers (L1 category leaders to L5 regional players) sourced from external authority lists. The ladder proxies a brand's awareness footprint within its sector, not revenue or market share. The failure mode differs sharply by tier. L1 brands appear in nearly every relevant retrieval but win only 25-41% of the recommendation slots they reach -- the leverage is differentiation, not visibility. L2 challengers carry the highest conversion rates of any tier (37-52%) but lose to persona-mediated substitution on the Anthropic models. L3 mid-market brands are the inflection level: aggregate coverage drops to 88%, conversion to 34-40%, and persona effects peak. L4 specialists and L5 regional players face catastrophic invisibility -- 48-52% never surface in any of the 37,000 runs. No uniform optimization recipe wins; the right marketing investment depends on where the brand sits on the prominence ladder.