Randstad appears in 96% of ChatGPT responses, with a 90%–98% interval, alongside Gi Group and Adecco. This range indicates the estimate’s precision; the recorded responses show different mention frequencies for Umana, MAW and Orienta when comparing Milan with Naples.
Refinea’s study covers 100 responses collected on 13 September 2026, split between Milan and Naples, with 50 responses per city; intervals for overall shares use the Wilson method at the 95% confidence level to express uncertainty around observed frequencies.
How we measured
Refinea submitted 25 distinct questions to ChatGPT in Italian, based on real keywords with measured search volumes. The questions contained no brand names and did not ask the engine to adopt a persona when answering.
Each question was repeated 2 times for each city, keeping the wording identical between Milan and Naples. The comparison therefore includes 100 ChatGPT responses, all collected on the same day, 13 September 2026.
The city was the local information supplied to the engine for this comparison; the results describe that setup without establishing the user’s physical location. The study covers ChatGPT alone and does not support extending these observations to other AI search services.
For each agency, we counted responses mentioning it without ruling it out as an option, then separated the results by city. The overall share uses all 100 responses, while each city’s share uses the 50 responses associated with that location.
A response can include several providers, so individual agencies’ shares do not represent mutually exclusive portions of the same total. An agency’s presence in a response also does not imply that ChatGPT presents it as preferable to the alternatives.
Replicating the measurement exactly also requires the full question texts and the method used to supply the city. Those details are not reproduced here, so this description explains the comparison’s design without guaranteeing an identical replication.
For broader context on the platform behind this study, Refinea describes its approach in why Refinea is different.
The results
The overall sample shows a leading group comprising Randstad, Gi Group and Adecco, whose uncertainty intervals overlap. We therefore present these names together, without turning differences in observed frequencies into a definitive ranking.
| Agency | Responses out of 100 | Observed share | 95% interval |
|---|---|---|---|
| Randstad | 96 | 96% | 90%–98% |
| Gi Group | 92 | 92% | 85%–96% |
| Adecco | 92 | 92% | 85%–96% |
The table reports Refinea’s measurement of 100 ChatGPT responses collected on 13 September 2026, with intervals calculated using the Wilson method.
Each interval accompanies its percentage to make measurement uncertainty visible, rather than presenting the result as an exact, stable value. Their overlap supports reading these providers as a leading group, without declaring a winner based solely on observed shares.
The city comparison shows different frequencies for Umana, MAW and Orienta, despite using identical questions on the same day.
| Agency | Milan, across 50 responses | Naples, across 50 responses |
|---|---|---|
| Umana | 60% | 26% |
| MAW | 20% | 2% |
| Orienta | 16% | 32% |
These frequencies also come from Refinea’s study on 13 September 2026, separated into the 50 responses collected for each city. For city shares, we report observed frequencies only, without claiming that these differences reliably extend beyond the collected sample.
Umana and MAW appear more frequently in responses associated with Milan, while Orienta appears more frequently in responses associated with Naples. This observation describes ChatGPT’s recorded behaviour without identifying the mechanism responsible for the differences between cities.
What this means when choosing a provider
For a marketing manager preparing a provider shortlist, these findings suggest practical checks before using AI responses to inform decisions.
- Compare responses associated with the cities where you operate, keeping the question identical and recording which agencies appear in each location. The Umana figures show why an overall view alone leaves out differences relevant to a local selection.
- Separate mention frequency from service verification by asking providers about geographical coverage, sector specialisation and how they deliver their services. The measured shares describe what ChatGPT presents, while these details document whether an agency meets your company’s requirements.
- Open the sources cited in responses and check which passages actually support each provider’s inclusion in the proposed shortlist. Save the question, response and checking date so the comparison remains understandable when a later query returns different names.
The official websites of Randstad, Gi Group and Adecco provide direct references for exploring the services offered by the leading group. These links support commercial checks on providers and are not the source of the percentages, which come from Refinea’s measurement.
For institutional checks on employment agencies, the starting point is the website of Italy’s Ministry of Labour and Social Policies. Checking an agency’s authorisation forms part of provider selection and remains separate from its frequency in AI responses.
For agency marketing leaders, monitoring should also distinguish generic questions from the local conditions used in the comparison. Our guide to Generative Engine Optimization (GEO) provides broader context on AI visibility, without automatically explaining the differences observed in this study.
Limitations
The scope covers 25 questions, written in Italian and repeated within the single collection day of 13 September 2026. The responses capture that sample and date, without describing mention frequencies during other periods of the year.
The cities considered are Milan and Naples, so the comparison does not represent every area served by the agencies mentioned. The questions come from real searches but do not necessarily cover every hiring need, sector or professional role.
Every collected response cites at least one source, but this does not establish how completely we can observe the engine’s research process. Intermediate searches are the queries ChatGPT runs to gather information before composing the response shown to the user. For this specific city comparison, we do not quantify how many responses expose those searches or which steps remain unobservable.
The published intervals accompany the leading group’s overall shares, while the local comparison remains a description of recorded frequencies. The measurement concerns names appearing in ChatGPT responses and does not measure market share, service quality or hiring outcomes.
Frequently asked questions
Why did you choose these questions?
The 25 questions come from real keywords with measured search volumes, grounding the comparison in searches that have actually been observed. Leaving brand names out avoids directly suggesting providers in the question, without making the sample representative of every business requirement.
Why did you use only ChatGPT?
The measurement examines ChatGPT’s responses when it receives identical questions associated with Milan and Naples. The scope therefore keeps the same service throughout the comparison and produces no findings that automatically apply to other AI engines.
How should I read an interval?
Randstad’s 96% describes the observed frequency, while 90%–98% expresses the estimate’s uncertainty under the method used. The 95% confidence level describes how reliably the procedure captures the underlying frequency across repeated studies, without guaranteeing that the next collection will fall within the same range.
If you are assessing your agency’s visibility, we can compare the commercial questions and cities relevant to your market.