Plenary meeting notes - 4.02.2026

[InKyungChoi]

Statistics Norway’s presentation

Presentation slides here

• Google's AI Overview (AIO) mode was launched, this marks a paradigm shift in the world's largest search engine. Users increasingly get answers to the questions without clicking a single link.
• In the experiment, keywords that had higher search volume from the SEO tool Ahrefs were used.
• Note that search engine results based on personal context, so tested both API and manaully in incognito mode
• Big question - does good ranking in traditional search engine translate to high visibility in AI overview?
• How to access visibility: if we are mentioned in the first para; If we are mentioned in bullet list; If yes, with or without link; If we are mentioned in the source list; If yes, what position?
• Create a scoring formula based on UX principles
• Hypothesis: SBB’s good ranking in traditional search engines leads to high visibility in AI overviews? => no guarantee. Top position in search results gives you 50% chance in appearing in AI overview. Similar findings from Ahref studies.
• What content does Google prefer in AI overviews? => Larger share of articles in AIO (78%) than in SERP (53%). => Tables on stat pages is difficult for google to interpret so not being used as materials. Google tends to prioritize articles over stat pages.
• Are AIO correct? => statistics presented surprisingly often correct, but 13 cases where old data is used (potential explanations: difficulty reading our tables, multiple sources pointing to older numbers; newer content may not be visible enough).
• AIO changes frequently.
• Main takeaways: establish benchmark for AI overview visibility, new benchmarks; investigate whether table readability in google can be improved, increase use of schema markup to make it easier for AI to understand the content; Evaluate tools for monitoring visibility; Top ways to ensure your content performs well in google’s AI experiences on search

Insee’s presentation

Presentation slides here

• SEO not enough
• Key challenges – ensuring that NSO content is discovered, correctly interpreted and properly cited.
• A recent French survey showed that almost half who access info via AI tools also check source links.
• What do AI conversational agents know? 1) during training phase: Wikipedia (Insee is looking at wiki pages on stat and check if the pages are correct) and Common Crawl (so need some work on indexation in common crawl); content ingestion (well-structured pages and editorial consistency); 2) usage phase (often access the web): SEO (search engine indexes) plus direct page fetching (performance and accessibility matter, if it takes 4-5 secs to load, it won’t go) [maybe this is what happens in Norway’s presentation]
• Insee strengths for SEO and GEO: 1) editorial practices (clear titles); 2) strong institutional authority (e.g., backlinks from authoritative sources)
• Google bots are powerful, so what was good for google is not enough for other bots (e.g., other crawlers). Need to make automated crawling reliable (e.g., sitemaps), ensure content is readable client-side without javascript
• Improve automated understanding of content (semantic html tags, minimal metadata such as schema.org)
• Indicators to monitor SEO/GEO. Traditional tools (search console), verification of actual crawler activity on websites, traffic coming from conversational agents, AI visibility measurement tools – how to measure answer quality and citation on a defined set of prompts -> potential to evolve into a shared open-source tools, subject to interest of multiple NSOs. But above all, need reference set of prompts to ensure comparability

Q&A and Discussion

• Comparing with authoritative competitors like international orgs like IMF, WB, OECD?
• The way people google is different from what they use chatgpt
• How the testing of table done? => we just asked chatgpt (paid version) if they can read it
• similar experience as others, if you ask questions, it generally goes to articles or publication. Vast majority of requests come from machines via API and last few months, we are getting a lot of garbage requests that suspiciously look like LLM (?), AI goes to our article, but somebody somewhere are trying to read our data via API
• We are getting a lot of machines trying to ping our website. More traffic from our website, less our search. Since on cloud, it costs more money.
• we have to understand search architecture, they differ a lot, we need to understand that first. Search readiness.
• We have huge demand on API, and there are a lot of statistics that are provided only in this way.
• It would be great to have this scaled up and automated

Useful resources shared

• Content optmization
• AI Search Architecture Deep Dive: Teardowns of Leading Platforms