Preprint: AI search APIs favored popular brands in Italy; fact-preserving rewrites barely helped

Analysis · Preprint, not peer-reviewed

Direct answer: In this preprint, brands that are searched more often in Italy were generally mentioned more often, and, pooling engines and sectors, the authors detected no difference in this pattern between web search on and off. In a separate experiment with search off, a page ranked lower under an invented brand than under its real one, and rewriting the text without changing its facts produced no significant gain. The authors conclude that generative search engines impose “a parametric barrier to entry that standard content optimization cannot overcome” [1]. The tests cover Italian questions, API endpoints and, for the experiment, search off.

What happened: Giorgio Monaco, Jacopo Lenti and Francesco Pierri of Politecnico di Milano posted “The Popular Get More Popular: Auditing Brand Bias in Generative Search” to arXiv on Nov. 5, 2026 (version 1, 07:57 UTC) [1]. The paper is a preprint; its arXiv listing gives no journal or conference publication. It cites a preprint by Chen and Koudas on AI citations (arXiv:2610.15033), which we covered on Oct. 28.

How was the study designed?

The authors ran two studies through the engines’ official APIs, with no system prompt and default settings, not through the consumer apps [1].

• Study 1, an audit. 80 recommendation questions in Italian, derived from commercial keywords that name no brand, in four sectors: insurance, travel, fashion and beauty, and sport. Each was asked without a persona and after each of three short buyer personas, with web search off and on, 10 times, to gpt-5.5, gemini-3.6-flash and perplexity/sonar, with the user location set to Rome. That gave 19,200 answers, all collected on Oct. 14, 2026 (UTC) [1].
• Measures. A brand counts as mentioned when its name appears in the answer text, and as cited when its official site is among the answer’s sources; citation rates cover the 683 of 1,176 brands with a known official domain. Popularity is based on the mean monthly Google searches for the brand’s name in Italy over 12 months, from DataForSEO. A separate model, gemini-2.5-flash, listed the brands in each answer; a hand check of 30 pilot answers found 96% of its extracted names correct [1].
• Study 2, a controlled experiment. For 24 of the questions, four models (the three above plus the open-weight deepseek-v3.2) ranked five real product pages supplied in the prompt, with web search off. One target page was changed in two ways: the brand and product names that sign it (five versions, from its real brand and product to an invented brand) and the way it is written (the original text or one of three rewrites called Structure, Fluency and Authoritative). The rewrites were kept only if every name and number was unchanged, their length stayed within 15% of the original and a second model found no fact added or removed. The main experiment produced 38,400 rankings, and control versions in which the other real brands carried a real product of their own added 9,600 more; rankings were collected from Oct. 28 to Nov. 3, 2026 [1].
• Statistics. Study 1 intervals come from 1,000 bootstrap resamples of prompts within each sector, and the search-off versus search-on slope comparison from resampling brands; Study 2 intervals come from 2,000 resamples of the 24 questions. All intervals are 95% [1].

What did Study 1 find?

• Mentions. The share of answers mentioning a brand rose from below 1% in the least popular fifth of brands to 1.2% to 5.3% in the most popular fifth, depending on the engine. The correlation between popularity and mention rate was positive and statistically significant in 23 of 24 engine, sector and search configurations [1].
• Citations. The correlation between popularity and citation was significant in 8 of 12 engine and sector pairs after correction for multiple tests; it was not significant in sport and travel on GPT and on Gemini [1].
• Search on and off. Overall, the slopes were 0.87 with search off and 0.89 with search on. By engine and sector they differed in 4 of 12 pairs: steeper with search on for insurance on GPT and on Perplexity, flatter for sport on Perplexity and for travel on GPT. On Gemini, calls that actually searched showed a flatter slope than calls with search off (0.95 against 1.34), but the authors note that this comparison is not controlled because Gemini chooses when to search [1].
• Search queries. In travel, the share of mentioned brands whose name was already in GPT’s first search action rose from 7% in the least popular fifth to 45% in the most popular (a difference of 38 percentage points, 95% interval 19 to 63); in insurance, from 1% to 8% [1].
• Sources. When the engines searched, the name of a mentioned brand appeared in at least one returned source for 95.5% of mentions on GPT, 60.9% on Gemini and 97.3% on Perplexity; 24.9% of Gemini’s mentions could not be verified. In 36.2% of all checked mentions, the brand’s name appeared only on third-party pages [1].
• Personas. Placing a buyer persona before the question lowered the popularity slope for mentions, measured with search off and on together, from 0.98 to 0.80 and 0.85 for two of the three personas, both significant, and to 0.89 for the third, not significant [1].

What did Study 2 find?

• Real versus invented brand. Averaged over questions, models and the four texts, the target page ranked 3.21 under its real brand and product names and 3.65 under an invented brand and product. That is 0.44 positions higher (95% interval 0.23 to 0.69). The gain was positive for each of the four models and in each of the four sectors [1].
• Other names. Keeping the real brand but inventing the product name cost 0.14 positions (0.05 to 0.27). With the same invented product, the real brand alone ranked the page 0.30 positions above an invented brand (0.13 to 0.51), and a less popular real brand still ranked it 0.19 positions above an invented brand (95% interval 0.11 to 0.30 in the paper’s text; 0.11 to 0.29 in its Appendix Table E.2). A better-known real brand did not rank the page significantly higher than a less-known one (+0.05, 95% interval −0.11 to 0.22, not significant) [1].
• Rewrites. Each of the three rewrites moved the page by about 0.02 positions; their average was +0.022 (−0.001 to 0.049), not significant. The authors write that the interval “rules out gains larger than 0.05 positions” [1]. The real brand and product names gained 0.42 positions more than rewriting did (0.22 to 0.66) [1].
• Brand leakage. Asked to identify the original brand from the page with all names replaced, the models named the true brand in 38% of answers, which the authors call an upper bound because the question invites a guess; they report that this showed no relationship with the real-brand effect [1].

What the study cannot show

• Preprint, single study. arXiv lists version 1 only and no journal or conference publication [1]. This analysis rests on this one paper; we are not aware of an independent replication.
• Italy only. All questions were in Italian, the Study 1 user location was Rome, and popularity was measured by Google searches in Italy. The authors fixed prompts to “a single temporal snapshot and location,” leaving other countries outside their scope [1].
• API endpoints, not apps. The authors write that the APIs “may omit proprietary, client-side system prompts or interface heuristics present in consumer-facing chat interfaces” [1].
• Study 2 had search off. The five pages were supplied in the prompt, which the authors say “omits live search reranking in the wild” [1].
• Association, not merit. Study 1 is observational. The authors note that established brands often correlate with higher consumer trust and quality, and they list comparison against an independent inventory of market offers as future work [1].
• Narrow rewrites. The three rewrites kept every fact and stayed within 15% of the original length; the authors excluded fabricated claims, such as fake statistics, by design [1]. Other content changes were not tested.
• Data. The paper links code, prompts, product pages and aggregated data in an anonymized repository; we did not review it.

Why it matters

Our view: In this data, the more popular brands were mentioned more often whether or not the engines searched the web, and a page’s real brand and product names moved its rank far more than fact-preserving rewrites did. The paper also shows why mentions and citations need to be counted separately: in more than a third of checked mentions, the brand’s name appeared only on third-party pages among the returned sources [1]. Our guide to AI visibility metrics treats mentions and citations as separate measures. The authors suggest that a new provider “may gain more from becoming known in the web sources that models learn from and retrieve” [1]; the study did not test that.

What we don’t know yet

• Whether the ChatGPT, Gemini and Perplexity apps, or Google’s AI Overviews and AI Mode, show the same patterns. The study tested API endpoints only.
• Whether the results hold for other countries and languages.
• Whether the advantage of a known brand name holds when the engine searches the web live.
• Whether other changes, such as adding new information to a page, would move rank.
• Whether peer review or a later arXiv version changes the estimates.

What to do now

1. Track brand mentions and citations of your own site as separate measures. In this study, 36.2% of checked mentions named a brand found only on third-party pages among the returned sources [1].
2. When comparing AI visibility results, record whether web search was on and what context preceded the question. In this study, buyer personas changed both the popularity slope and the sources retrieved [1].
3. Treat claims that rewriting existing copy alone will lift AI recommendations with caution. In this search-off test, three fact-preserving rewrites moved a page by about 0.02 positions [1].

Disclosure: The study audits engines from OpenAI, Google and Perplexity. AVJ pays standard published rates for basic services from some of these companies, including a ChatGPT Pro subscription and API access from OpenAI and API access from Perplexity, and uses them as an ordinary customer. AVJ uses OpenAI and Anthropic models among its editorial tools (see our AI use policy). AVJ has received no free credits, promotions or other compensation from OpenAI, Anthropic, Google or Perplexity. AVJ receives no sponsorship and has no partnership with OpenAI, Google, Perplexity or Anthropic. The AVJ Citation Index, whose methodology is still under review, plans to measure answers from AI companies’ APIs, including OpenAI’s and Perplexity’s.

Update log: Nov. 11, 2026 — First published.

Sources

[1] Monaco, G., Lenti, J., & Pierri, F. (2026, Nov. 5). “The Popular Get More Popular: Auditing Brand Bias in Generative Search.” arXiv:2611.07614v1 [cs.CY]. Preprint, not peer-reviewed. Accessed Nov. 11, 2026. https://arxiv.org/abs/2611.07614

[2] (HTML version of [1], version 1, used for sections, tables and appendices.) https://arxiv.org/html/2611.07614v1

How this was made

Prepared with AI assistance from the sources cited above and checked against the paper’s full text, tables and appendices, then edited by Nhi Phan and fact-checked by Jackies. A further editorial review follows publication. This analysis is not a peer review of the paper. Any change will be recorded in the update log and in our Corrections Policy and Log.