The Selection Changes Insurance Rankings in Google's AI Overviews

By Øystein Kalseth

The plan was one list. We counted in two ways and got two different ones.

In our brand catalog, source occurrences for seven Norwegian insurance domains in Google's AI overviews are displayed. The dataset is not thematically limited, and answers to questions on entirely different topics may therefore be included. We wanted to see what happened to the distribution when the selection only included questions containing “forsikr”. Both counts were conducted in the same coordinated extraction round.

The Figures

The table provides the number of registered AI responses citing the domain at least once for each domain. The shares are calculated based on the sum of hits for the seven domains in each selection. Since one response can cite multiple domains, it can result in multiple domain–response hits. The sum therefore does not correspond to the number of unique responses.

The difference in shares says something about how the limitation affects the distribution. It says nothing about development over time. The selections overlap and have different percentage bases, and the shares can therefore not be added together.

The unfiltered measurement is limited in the same way as the current catalog. The figures in the table come from a new extraction on October 8, while the catalog may state a different extraction date.

How many unique responses are behind the hits?

We found 2,314 registered AI responses with “forsikr” in the question. 1,627 of them, or 70.3 percent, cite at least one of the seven domains. The remaining 687 cite none of them, and which sources these responses use, we have not investigated.

Altogether, the 1,627 responses yield 3,379 domain–response hits, an average of 2.08 per response.

What shifts

Among the responses citing tryg.no, 69 percent have “forsikr” in the question. For storebrand.no, the share is 30 percent and for gjensidige.no 34 percent.

With the limitation, Tryg moves up from fourth to third place. The share of domain–response hits rises from 13.3 to 21.0 percent, i.e., 7.7 percentage points between the selections. Storebrand falls from third to fourth place, and the share drops from 17.1 to 11.8 percent.

We have seen examples of questions without “forsikr”, but the thematic distribution has not been mapped. When we write “questions without forsikr”, we therefore only mean that the character string is missing. The question could well be about the same topic. The figures also do not explain why Storebrand receives these hits.

In the unfiltered measurement, Fremtind has the lowest share, 2.8 percent. With “forsikr”, Fremtind gets 4.4 percent, and then JBF is lowest with 3.3 percent. Fremtind's insurance is sold through banks and associations. The low share for fremtind.no may therefore be because relevant source citations go to the distributors' websites. This has not been tested. The figures for fremtind.no do not necessarily provide a full picture of how visible Fremtind's products are.

What we measured, and how

Each company is measured as its own domain, including subdomains. We counted the registered AI responses in Google's AI overviews that list the domain as a source at least once, and calculated the share of the sum for the seven. The extraction was limited to Norway and Norwegian Bokmål, with data from DataForSEO LLM Mentions. The analysis is based on registered responses in the provider's dataset. Our own live searches were not part of the study.

We are measuring source citations to the domains. A response that mentions If, but does not cite if.no, is therefore not included in the count.

The limitation is a pure text match. The question must contain “forsikr”, which captures words like forsikring (insurance), forsikre (insure), bilforsikring (car insurance), and forsikringsselskap (insurance company). The questions are not hand-picked and validated as general insurance questions. Questions about life insurance and pension insurance may thus be included, while relevant questions about, for example, kasko (comprehensive coverage) may fall outside. As a next step, we will define and validate a broader set of questions within the category.

The extraction was done with 14 API calls in 22 seconds on October 8, 2026. We checked that the counts with and without the text filter together provided the baseline for each domain, that the two counting methods yielded the same result, and that a final control call yielded the same figures as at the start. The checks revealed no discrepancies. They show that the extraction was internally consistent but do not document that the dataset covers all relevant responses from Google.

This is not market share. The measurement shows which of the measured domains are listed as sources.

What the measurement does not say

• It applies to Google's AI overviews. ChatGPT, Perplexity, and Gemini are not included in the measurement.

• The date of extraction does not say when the responses were given. The responses in the dataset are registered and updated at different times.

• All shares are calculated within the seven measured companies and no others. A company outside the group may have more citations than each of them, without it appearing here. Furthermore, the fact that 687 responses do not cite any of the seven shows that the group does not capture everything.

• We count responses that cite the domain. The number of links does not matter. If a response links to if.no seven times, it counts as one.

See the full catalog

The brand catalog gathers the published results, the stated extraction date, and an explanation of the metrics. The current methodology description is documentation version 1.1. See insurance in the brand catalog.