Today we are publishing Refinea Analysis, the public observatory that measures how frequently Italian brands get cited inside AI-generated answers. The leaderboards refresh every night at 03:00 UTC, the methodology is documented end to end, and access is free for everyone.
The Italian market did not have a public standard for this question. ChatGPT, Gemini, Perplexity and Google AI Overviews have captured a meaningful share of discovery traffic over the past eighteen months. Similarweb puts ChatGPT alone at roughly 79% of global Gen AI web traffic. In Europe the picture closed in March 2025, when Google extended AI Overviews to Italy, Germany, France, Spain and seven more countries (Search Engine Land, 2025).
In this environment, asking whether a brand ranks on Google is no longer the right question. The right one is how often a given company gets cited by AI engines when a user asks something commercial. Refinea Analysis answers that question with a method anyone can verify.
What Refinea Analysis is
Refinea Analysis is a public ranking that measures, industry by industry, how often a brand gets cited inside answers generated by mainstream AI models. The headline metric is the AI Visibility Index (AVI), the share of AI responses in which a brand is actually mentioned across all prompts evaluated for that industry.
The observatory lives inside Refinea, the platform companies use to optimize their presence in conversational engines. Publishing the leaderboards for free is a deliberate stance: a market with no measurable standard produces opinions instead of decisions. A public benchmark forces the conversation onto common ground.
Two Italian industries are live at launch:
- Italian SaaS business software (TeamSystem, Zucchetti, Aruba and other category vendors)
- Italian fintech and neobanks
Additional verticals are on the roadmap and get added based on qualified demand from companies in the sector.
The methodology, in four steps
Methodological rigor is what separates Refinea Analysis from earlier attempts at this kind of measurement. Every step is documented and applied identically across every industry we track.
01. Prompt intelligence
The prompt panel is not invented at a desk. Refinea continuously aggregates real-world search demand from premium data providers, the queries users actually type when looking for products in a given industry. Three refinement layers sit on top of that raw signal: machine learning, semantic embeddings and clustering. The output is a curated panel of high-intent commercial prompts, validated against a database of more than one million real queries.
This choice is the first defense against bias. Measuring AI visibility on prompts constructed in isolation produces noise dressed up as signal, because it rewards hypothetical scenarios in which a brand would like to be cited rather than the ones in which its category actually gets interrogated by the market.
02. Multi-run sampling
Every prompt in the panel gets submitted to the target model ten times per day, in independent runs. The logic is purely statistical. Generative models produce non-deterministic answers, so a single query does not represent the system’s behaviour. Thousands of weekly responses allow citation frequency to be estimated with a precision comparable to a peer-reviewed sampling study.
03. Brand extraction
Extracting brands from AI answers is less trivial than it sounds. Refinea combines two recognition layers: deterministic matching and machine learning. An alias graph collapses sub-brands into a single canonical entity, while a curated blocklist eliminates false positives at the source (institutions, common nouns, infrastructure references that share a name with a vendor).
The underlying principle is consistent with the academic literature on GEO. The foundational paper GEO: Generative Engine Optimization presented at SIGKDD 2024 showed that the mere presence of a keyword is not enough to attribute a citation. You need a system that recognizes entities inside their narrative context.
04. Aggregation
A brand’s AVI is the share of AI answers in which it gets cited. It is a standard binomial estimator, comparable over time and independent of the panel composition. This means an AVI of 0.34 measured in March is comparable to an AVI of 0.41 measured in May for the same industry, even when the number of prompts changes.
Choosing a classical statistical estimator instead of a freshly minted proprietary metric is a transparency choice. Anyone who knows how to compute proportions can replicate the check.
Why a national standard matters
Traditional SEO rankings measure a world that is disappearing. Zero-click searches have surpassed 58% in the United States and 59% in Europe according to SparkToro’s analysis on Datos data. For a brand selling complex products, a growing share of evaluation happens inside an LLM chat window and never shows up in Google Search Console metrics.
Italy needs its own benchmark for two concrete reasons. The first is linguistic: prompts in Italian trigger partially different answers than those in English, both because of training data composition and because of how retrieval behaves across languages. The second is market-specific: vendors like TeamSystem or Satispay are relevant entities for AI in Italy but routinely ignored by English-language benchmarks.
The Italian business software market alone is worth tens of billions and growing double digits (Osservatorio Software Polimi). Neobanks operating in Italy are now forty, between Italian-headquartered entities and passported operators (Osservatorio Fintech Polimi). Sectors of this size deserve dedicated measurement.
How to read the leaderboard
Each industry page on Refinea Analysis shows four elements.
The AVI chart reconstructs the historical trajectory of every brand across every tracked model. It lets you read the trajectory, not just the snapshot. A rise in AVI followed by a plateau tells a different story from continuous growth. The two require different operational interpretations.
The leaderboard ranks brands by current AVI. Vendors carrying the “Tracked” badge are part of the industry’s reference panel, the others surface organically from AI responses. The delta versus the prior period signals who is gaining or losing ground.
The prompt examples display some of the questions actually submitted to the models. They make the type of intent the observatory measures unambiguous.
The last-update timestamp is explicit on every chart. It marks the exact snapshot moment and makes the measurements verifiable over time.
What to do if your brand does not appear
There are two possibilities. The first is that the brand simply does not get cited by models for any of the prompts in the active panel. In that case the gap itself is information: it tells you that your brand’s conversational category is held by competitors. The second is that the brand gets filtered by the blocklist designed to eliminate false positives on ambiguous entities.
In both cases, writing to hello@refinea.io is worth it. The reference panel gets extended periodically based on qualified requests from sector vendors.
The broader positioning
Refinea Analysis is a public manifestation of the work Refinea does privately for its customers. The difference between observatory and platform is one of scope: the observatory shows AVI across entire industries, the platform shows AVI for a single brand on prompts built from its real customers, not from sector samples.
For the broader framework, we have published our operational guide to Generative Engine Optimization and a summary of how GEO differs from traditional SEO.
In the months ahead we will publish monthly reports on the most mature industries, with qualitative analysis of the most significant movements. The stated goal is one: give the Italian market a measurable reference point for talking about AI visibility with data instead of opinions.