Does the way a question is asked change which providers AI systems mention?

What we measured

Scope
Period	24 October to 1 November 2026, ten runs
Buying situations	7
Questions	60 (each situation in up to ten forms)
AI access paths	7: four interfaces, two models via the API (one with and without web search)
Answers per question and access path	30, each in a new chat (GPT-6 Luna with web search: 15)
Answers analysed	11,566, of which 7,067 from the interfaces and 4,499 via the API

The seven buying situations are familiar to drivers in Switzerland: young drivers, new car with comprehensive cover, older car, electric car, leased car, second car and switching after a claim. We asked about each situation in several forms. Using the older car as an example:

Question form	Example (asked in German, English paraphrase in brackets)
Simple question	„Welche Autoversicherung ist in der Schweiz am besten für ein älteres Auto?“ (“Which car insurance is best in Switzerland for an older car?”)
Same question, reworded	„Bei wem schliesst man in der Schweiz für ein älteres Auto am besten eine Autoversicherung ab?“ (“Who is the best company in Switzerland to take out car insurance with for an older car?”)
Search terms	„beste autoversicherung älteres auto schweiz“ (“best car insurance older car switzerland”)
Personal situation	„Mein Auto ist zwölf Jahre alt, und ich möchte nur das Nötigste versichern. Welche Autoversicherung ist in der Schweiz am besten für mich?“ (“My car is twelve years old and I only want to insure the essentials. Which car insurance is best in Switzerland for me?”)
Situation with conditions	the same, plus „Haftpflicht und Teilkasko, unter 700 Franken pro Jahr, Selbstbehalt höchstens 500 Franken“ (“third-party liability and partial casco, under 700 francs a year, deductible no more than 500 francs”)
Comparison	„Vergleiche drei Schweizer Autoversicherungen für ein älteres Auto nach Prämie, Deckung und Schadenabwicklung. Welche würdest du wählen?“ (“Compare three Swiss car insurers for an older car by premium, cover and claims handling. Which would you choose?”)
Negative question	„Von welchen Schweizer Autoversicherungen für ein älteres Auto würdest du eher abraten, und warum?“ (“Which Swiss car insurers for an older car would you rather advise against, and why?”)

We asked two kinds of AI access paths:

Access path	How queried	What it stands for
ChatGPT, Gemini, Microsoft Copilot, Google AI Overviews	via the public interface, Swiss location, without logging in	what customers see when they ask themselves
Claude Sonnet 5.5 (Anthropic)	via the API, with web search	what agents, comparison and price-monitoring tools and built-in shopping assistants receive
GPT-6 Luna (OpenAI)	via the API, with and without web search	the same, and how much difference web search makes

We counted which insurers an answer mentions. We recorded comparison sites separately.

A few insurers dominate

Car insurance is a concentrated market, and this shows in the answers. In response to the simple question, the three chat interfaces mention the most frequent insurer in 98% of answers on average, the second most frequent in 92% and the fifth most frequent still in 59%. So the question is less whether the big names appear and more how often and in which position. All the following results apply to this market; whether they turn out the same in categories with many providers is open.

What changes when you ask differently

For each buying situation, we took the five insurers mentioned most often in response to the simple question and compared how often they appear with the other question forms. The basis is 3,768 answers from ChatGPT, Gemini and Copilot; the values are averages across the three interfaces and all seven situations.

Instead of the simple question …	Share of the five most-mentioned insurers changes on average by	Of these, remain among the five most mentioned	First-mentioned insurer changes
… the same question, reworded	11 percentage points	4.4 of 5	in 2 of 35 cases
… search terms	12 percentage points	4.5 of 5	in 1 of 17 cases
… personal situation	17 percentage points	4.2 of 5	in 5 of 16 cases
… situation with conditions	21 percentage points	3.9 of 5	in 6 of 16 cases

How to read this: if an insurer appears in 60% of answers to the simple question, with a rewording it is off by 11 percentage points on average, so at around 50 or 70%. With a question with conditions, it is 21 percentage points. We only count the first-mentioned insurer where it is clearly ahead; that is why there are fewer than 21 cases (7 situations × 3 interfaces).

The first row is the baseline variation. This is how much changes when you simply reword the same question and the answers turn out slightly differently. Every other question form has to be measured against it.

Search terms: No difference from the baseline variation. Anyone who types „beste autoversicherung älteres auto schweiz“ (asked in German: “best car insurance older car switzerland”) gets essentially the same insurers as with a full sentence. This applies to all four interfaces.

Personal situation: Somewhat more change, and the first-mentioned insurer changes in 5 of 16 cases. The picture is mixed, however: for some situations, such as the electric car or switching after a claim, the mentions shift markedly, for others hardly at all. Across all situations, the difference from the baseline variation is not established.

Situation with conditions: Around twice as much change as with a rewording, and on average one of the five most-mentioned insurers drops out. This is established on all three chat interfaces, clearly on ChatGPT, narrowly on Gemini and Copilot. Compared with the same situation without conditions, the mentions also shift markedly. On Google AI Overviews the change compared with the simple question is particularly large; the step from the situation to the conditions alone is not established there, however. The conditions always include a price limit. This study cannot separate whether the difference lies with the price or with the other requirements.

An example

Gemini, older car, 30 answers per question form. Insurers anonymised. An example with a particularly clear change:

Insurer	Mentioned for the simple question	Mentioned with conditions (partial casco, under 700 francs, deductible no more than 500 francs)
Insurer A	100%	63%
Insurer B	100%	80%
Insurer C	90%	27%
Insurer D	87%	10%
Insurer E	57%	100%
Insurer F	53%	80%
Insurer G	23%	57%

For the simple question, insurer D is the one most often mentioned first. With conditions, it appears in only 3 of 30 answers, and insurer E is most often in first place. An insurer that measures only the simple question sees itself in a good position and learns nothing of this.

The system matters at least as much as the question

If you put the same question to two different chat interfaces, the share of the five most-mentioned insurers differs by 29 percentage points on average. That is more than the difference between the simple question and the question with conditions on the same interface (21 percentage points). Anyone who measures only one system therefore learns little about the others.

On top of this come comparison sites: depending on the interface, they appear in 36 to 80% of answers.

Via the API: what agents and tools get to see

Agents, comparison and price-monitoring tools or shopping assistants in apps generally do not access an AI model through the chat interface but via the API. Which insurers the model mentions there decides which providers turn up in such applications. We therefore analysed 4,499 answers via the API, from Claude Sonnet 5.5 (with web search) and GPT-6 Luna (with and without web search), and compared them with the interfaces over the same period.

Access via the API	Insurers per answer	Deviation from the chat interfaces (share of the five most mentioned)	Change with conditions, compared with the baseline variation
Claude Sonnet 5.5, with web search	7.2	27 to 30 percentage points	17 compared with 11 percentage points
GPT-6 Luna, with web search	2.8	34 to 46 percentage points	34 compared with 14 percentage points
GPT-6 Luna, without web search	3.9	36 percentage points (compared with ChatGPT)	not meaningfully comparable, see below

For comparison: over the same period, the chat interfaces deviate from one another by 25 to 32 percentage points, depending on the pair.

Three findings:

• Claude via the API mentions the most insurers per answer and deviates from the chat interfaces about as much as they do from one another. Conditions change its mentions somewhat less than on most interfaces; the difference from the baseline variation is not established.
• GPT-6 Luna with web search mentions few insurers and deviates the most from all interfaces, including ChatGPT. With conditions, its mentions change the most of all access paths. For this access path there are 15 answers per question, from the second survey day only.
• The same model with and without web search mentions different insurers: of the five most mentioned, only 2.5 are the same on average. Without web search, GPT-6 Luna mentions no insurer at all in 27% of answers, and in 71% for questions about the personal situation.

For providers, this means: anyone who wants to appear in agents and tools has to measure the API, and in the configuration these applications use, with or without web search. A measurement on the interfaces says little about this. Conversely, a measurement via the API does not show what customers see in the chat.

Negative questions: the systems rarely advise against

Asked which insurers they would rather advise against, the interfaces advised against a specific provider in 5 to 17% of their answers (824 answers to negative questions). Instead, the answers mostly point to policy features such as the deductible, surcharges for young drivers or missing additional cover. Where they did advise against a provider, six out of ten cases were about the price. Our analysis also counts soft reservations such as „genauer prüfen“ (“check more closely”) as advising against; explicit advice against a provider is, if anything, even rarer.

How precise a measurement with few answers is

Many measurements of AI visibility work with few answers per question. What that means can be reproduced from our data: for each of the 112 combinations of situation, question form and interface, we drew 10 at random from the 30 available answers 1,000 times and compared how much the result fluctuates.

• Insurers that appear in over 90% of answers can be identified reliably even with 10 answers. This applies to a good third of the most-mentioned providers.
• For insurers in the middle of the field, which appear in 30 to 70% of answers, the result fluctuates strongly. For an insurer that in reality appears in every second answer, a measurement with 10 answers can land anywhere between about 30 and 70%.
• The insurer most often mentioned first is usually the same in 10 answers as in 30. In individual cases, however, this holds in only just over half of the samples.

On top of this comes change over time: between the two survey days, the answers from ChatGPT, Gemini and above all Google AI Overviews changed more than would be expected by chance. Copilot remained stable. A measurement is a snapshot.

What follows from this

1. Measure questions with conditions. The simple question about the “best” insurance shows only part of the picture. What matters are the questions in which someone states the price, the cover or how they want to take out the policy.
2. Do not overrate search terms. In car insurance they returned the same insurers as full sentences. Whether this holds in general is what part 2 will examine.
3. Analyse systems separately. An average across several interfaces does not describe any system that someone actually uses.
4. Measure interfaces and the API separately. The interfaces show what people see; the API shows what agents and tools receive. The two pictures differ markedly.
5. Measure with enough answers and repeatedly. For providers in the middle of the field, a few answers are not enough, and the answers change over time.

More on how we derive the questions we measure: Which prompts actually count? And why a single value says little: Why does a single AI visibility score say so little?