Measuring AI visibility: source pool first, content second.

An AI answer can name a brand without linking its website as a source, and it can use an expert article as a source without recommending the vendor. In the measurement for AFS-Magazin, the French answer to the CRM question shared only 3 percent of its sources with the German one. Yet all three languages named the same big vendors, including Salesforce and HubSpot. Different sources, same brands. A report that rolls mention, source and recommendation into one number hides exactly that, and it hides what you can work on.

In the measurement, German questions about CRM and project management software got the same sources as English ones, including comparison portals and forums, without a single German domain. In categories like these, a German article on your own website buys no visibility, because the system never pulls in German sources. For ERP for machine builders, software for trade businesses and supplemental dental insurance, the AI search found German sources of its own. What makes the difference is whether a category has a dense body of German-language sources. Naming the market changes the picture. The same CRM question got no German source without a market reference, 6 of 17 with “in Germany” and 13 of 19 with “GDPR-compliant, from Germany”. Even the English question naming the market pulled in 7 of 17 German sources.

The test from the guest article is the tool for this. Write four questions the way customers would ask them: two without a market reference and two with one, for example “in Germany”. Put each question to a system with live web search and record the cited sources, not the brands named. The brands are the outcome; the sources are what you can work on. Then count the share of domains with German-language content. Below 20 percent, the source pool is dominated by English. Above 50 percent, German source work is the right lever. Two weeks later, run the same measurement again.

If a category sits in the English source pool, the way in runs through the directories and portals cited there. In the measurement these were Capterra, TrustRadius and PCMag again and again, plus discussions on Reddit. A complete, well-kept profile there counts for more than the best German blog post. In globally traded categories the big international brands win, and a German CRM vendor competes with Salesforce on the general question. Its way into the answer runs through a narrower angle: industry, jurisdiction, company size or region. With “GDPR-compliant, from Germany” added, German CRM vendors appeared among the sources that were missing from the general version. In the German source pool, expert content with documented claims works.

A Google ranking has a position you can track. An AI answer has none, and its sources can change between two requests. One question I asked twice came back with identical sources, which for API queries may also be down to caching. What tells you something is the share of answers in a series that name, link or recommend a brand. The first series is a baseline. If the domain list changes a lot on the repeat, the category needs more runs before anyone acts on it. For ongoing monitoring, every four weeks with the same query set is enough, and new questions run as a separate series.

Every query needs its wording, language, platform and date on record, with mention, linked domains and recommendation kept separately. The AFS measurement covered 15 queries on a single system, Perplexity, from September 9 to 11, 2026, through its API and with no location signal. I did not test ChatGPT, Gemini or Google's AI Overviews, which draw on other indexes. Austria and Switzerland were not part of the measurement. Most monitoring tools also collect their data without a location signal. If the gap to location-aware queries turns out to be large, a good part of today's GEO reporting has a measurement problem. The hreflang attribute controls which language version shows in classic search results. I know of no evidence that it shapes which sources a language model picks, so I wouldn't sell it as a GEO measure.