People who urgently need help with debt no longer always turn to Google. Increasingly, they put the question straight to ChatGPT, Claude, Gemini or Mistral. But does that AI point them to the municipality, to a support organisation — or to a paid provider? And can the citizen check where that recommendation comes from?
SiteBoosting investigated this with the societal AI findability barometer: four AI assistants each received the same kind of debt-help question 72 times, across five Dutch cities, with live web search switched on. That makes 288 measurements in total, citing 1,479 sources between them.
Key findings
- Verifiability varies widely. In one third to two thirds of answers, every source is missing. Gemini backs up every answer; Mistral nearly every answer (88%). With Claude, 42% of answers lack a source; with ChatGPT, 67% — two out of three times.
- Recommending without a source is common. In 38% (Claude) to 42% (ChatGPT) of all answers, a specific organisation is recommended without any source attached. The citizen cannot verify whether that recommendation is justified. With Gemini and Mistral this hardly ever happens (0% and 11%).
- Only Gemini structurally points to paid help. Of the sources Gemini cites, 14% are commercial parties, such as a professional administrator or a budget coach. Mistral does so occasionally (1.5%). For debt help, ChatGPT and Claude never cite a commercial party — only government and support organisations.
- The AIs differ sharply on government versus support organisations. ChatGPT leans extremely heavily on (local) government: 64% of its sources. Gemini barely does (6%) and far more often points to support organisations. Mistral (31%) and Claude (39%) sit in between.
How we measured
For each AI, the societal barometer asks fifteen realistic help questions about debt help, ranging from short search queries (“debt help Amsterdam”) to full requests for help (“I’m behind on several bills and keep getting payment reminders, where can I turn for help?”). The questions are spread across Amsterdam, Rotterdam, Utrecht, Eindhoven and Groningen. A fixed control set of nine questions runs alongside, on informal care, homeless shelters and community work.
Each question was put three times to ChatGPT, Claude, Gemini and Mistral, always with live web search on: 288 measurements in total. The question itself deliberately never names the type of organisation (never “which agency”, always “where can I turn”). Otherwise the measurement itself would steer towards government or a support organisation.
For every answer we record which sources are cited and which type they belong to: local government, national government, referrer, local support organisation, national support organisation, commercial party or other. We deliberately do not record organisation names in this article — only counts and source types. We repeat the same measurement every month with a different societal theme, so that a trend line emerges.
Where AI gets its references
The distribution of the 1,479 cited sources shows a distinct pattern for each AI:
| Source type | ChatGPT | Claude | Gemini | Mistral |
|---|---|---|---|---|
| Local government | 64% | 39% | 6% | 31% |
| National government | 11% | 0% | 1% | 0% |
| Referrer | 1% | 1% | 15% | 12% |
| Local support organisation | 14% | 48% | 38% | 45% |
| National support organisation | 4% | 8% | 19% | 9% |
| Commercial party | 0% | 0% | 14% | 1% |
| Other | 6% | 3% | 9% | 1% |
Percentages per AI, all debt-help and control questions of this measurement combined.
ChatGPT is the only AI that mainly sends citizens to the government itself; with the other three AIs, support organisations (local and national combined) are the largest category. Notably, only Gemini and Mistral relatively often name a ‘referrer’ — a social map or signpost that doesn’t provide help itself but shows the way — whereas ChatGPT and Claude almost never do.
How much substantiation does each AI give?
The number of sources per answer also differs by up to a factor of six. Gemini cites almost 11 sources per answer on average, Mistral just over 5.5. ChatGPT and Claude stay at 1.7 to 2.1 sources per answer — and, as described above, in many cases give no source at all.
For a topic like debt help, where people in a vulnerable position are looking for reliable assistance, that difference matters: a citizen who asks ChatGPT for help more often gets advice they cannot verify than a citizen who puts the same question to Gemini.
Four AIs, four characters
Gemini behaves like a reference work: it substantiates every answer extensively and is the only one to structurally point to paid help alongside government and non-profit organisations. Mistral comes closest in behaviour: almost always a source, and occasionally a commercial party, but noticeably more restrained than Gemini. ChatGPT and Claude behave more like a brief adviser: compact answers, few or no sources, and exclusively government and non-profit help as referrals — never a paid party.
For citizens, this means: it pays to ask for the source of a recommendation, especially with ChatGPT and Claude. For municipalities and support organisations, it means that visibility differs greatly per AI, and so has to be measured per AI.
What does this mean for municipalities and support organisations?
- Keep your own information up to date online. AI cites what is easy to find online. A current, well-structured debt-help page on your own website increases the chance of being cited as a source — just as with an ordinary search engine.
- Be well listed with local referrers. Neighbourhood teams, social maps and signposts were cited relatively often by Gemini and Mistral. A current, easily found profile on such a hub carries through into AI answers.
- Don’t rely on a single AI. What ChatGPT recommends differs greatly from what Gemini recommends — in choice of sources, in number of sources, and in whether there is a source at all. Don’t count on any one AI to reach people seeking help.
- Measure it. The AI visibility of societal information can be measured, just as it can for businesses. That shows whether people seeking help end up, via AI, at the right, verifiable help — and where it currently goes wrong.
About the societal AI findability barometer
The societal AI findability barometer is a study by agenticfindability.com in collaboration with SiteBoosting. Each time we measure a different societal theme — after debt help comes food banks. A fixed control set (informal care, homeless shelters, community work) runs alongside to safeguard the trend over the months.
