Methodology
AI Search Visibility in Italy

Descriptive, observational and panel-based

This research does not estimate the preferences of the Italian population and is not a market-share study. It describes the behavior observed in a predefined query panel during a specific time window.

The same panel is evaluated across four AI surfaces. The surfaces are not treated as interchangeable: each retains its own observed response and source semantics.

A panel fixed before the final analysis

The research uses 500 frozen Italian queries: 480 main-panel queries and 20 branded controls. The main panel spans ten sectors and includes different informational and commercial decision contexts.

Freezing the panel before the final analysis prevents the query set from being retrofitted to observed outcomes. Branded controls remain analytically separate from the main-panel metrics.

The published panel can be inspected directly in panel-500.csv.

Three complete, separate runs

The full panel was collected in three separate runs during September 3–5, 2026. Keeping the runs separate makes short-window variability visible instead of averaging it away at collection time.

The observation unit is query × surface × run. With 500 queries, four surfaces and three runs, the study contains 6,000 final observations.

For the primary main-panel brand metrics, the 20 branded controls are excluded, yielding 480 × 3 = 1,440 observations per surface.

Conservative commercial-brand families

Brand identification works on normalized commercial-brand families. Aliases are handled conservatively to reduce accidental matches and to keep a clear distinction between a brand, a product term and a generic category term.

Commercial brand family

The unit used for visibility metrics is the normalized brand family, not every literal string variation appearing in an answer.

Query seeding guard

A brand is considered “unprompted” only when the corresponding commercial brand family was not already present in the query. The metric therefore avoids counting a brand simply because the user explicitly asked about it.

Public brand definitions and denominator rules are documented in brand-metrics-v1.2.json.

Primary metrics

Unprompted Commercial Brand Visibility

The percentage of eligible observations in which at least one commercial brand appears even though that brand family was not seeded in the query. For an individual brand, the denominator excludes observations whose query already names that brand.

Brand agreement and divergence

For the same query/run unit, the observed brand sets are compared with Jaccard similarity. Divergence is defined as (1 − Jaccard) × 100. A divergence of 0% means identical sets; 100% means no overlap.

Google AI Overview denominator

AI Overview is reported on two bases: the full eligible observation base, and the conditional base restricted to observations in which the AIO generative surface actually appears. These denominators answer different questions and are not interchangeable.

Source analysis is separate from brand analysis

Attributed domains are normalized and compared separately from commercial brands mentioned in answer text. A cited source is not treated as equivalent to a brand mention.

The aggregate source-divergence comparison covers ChatGPT, Gemini and Google AI Overview. AI Mode is excluded from that aggregate because provider citation metadata was not consistent across all three runs.

This is a comparability guardrail, not an inference about AI Mode quality. AI Mode remains part of the brand analysis.

Checks built around the research boundaries

• Frozen query panel and declared control set.
• Separate runs rather than hidden averaging at collection time.
• Conservative alias normalization for commercial brands.
• Per-metric denominator definitions, especially for AI Overview.
• Separation of brand mentions from attributable source domains.
• Public machine-readable summaries and methodology artifacts.
• Explicit exclusion of non-comparable AI Mode citation metadata from aggregate source divergence.

What this research does not measure

It does not measure market share, revenue, conversion, user preference in the Italian population or causal commercial impact. It also does not imply that the observed patterns will remain stable as AI products, models and interfaces change.

The three runs provide a short-window view of observed variability; they are not a longitudinal time series and no formal statistical independence between runs is assumed.

AI outputs are stochastic and platform behavior can change. The study should therefore be read as a versioned observation of a declared panel and time window.

Researcher and product relationship

Telescop Research designed, collected and analyzed the study. Telescop also develops products and services related to AI visibility and decision intelligence. This relationship is disclosed so readers can evaluate the research with the appropriate context.

The public methodology, frozen panel and DOI-backed release are intended to make the analytical perimeter inspectable even though the full production pipeline and raw response corpus are not published.

Dataset, manifest and public verification files

How to cite the dataset and methodology

Dataset: Gentian Hajdaraj. (2026). AI Search Visibility in Italia — Dataset e metodologia v1.0 (Version v1.0) [Dataset]. Zenodo. https://doi.org/10.5281/zenodo.22640375

English methodology page: Gentian Hajdaraj (2026). Methodology — AI Search Visibility in Italy 2026. Telescop Research. https://telescop.it/en/research/ai-search-visibility-italy/methodology/

The English editorial page does not create a new dataset version. The DOI continues to identify the frozen v1.0 deposit.