AI describes Finland's largest B2B companies mostly from sources they didn't write, whether you ask in English or in Finnish.

Which outsiders do the describing depends on the language. In English, the most cited source is a German supplier directory, ensun.io. It was among the sources for 71 of the 100 companies on 12 October, against 61 for the companies' own websites (78 against 62 in September). Counted on the four engines that answered in both languages, it was a source for 77 companies in English and only 22 and 25 in Finnish, and Finnish sources took its place. The companies' own websites didn't gain a bigger share either way.

What this means in practice

A sales director at a Swedish manufacturer is building a supplier shortlist for a new line. She opens an AI assistant and asks who the leading suppliers in Finland are. The answer is fluent and confident, names four companies, and says a sentence or two about each.

She doesn't check where those sentences came from. Very few people do. If she did, she would find that for most of the companies on her list, the description was taken from a German directory page, an English Wikipedia article and a lead generation database. The company's own website, with the positioning someone worked on for a year, is often not among the sources at all.

The decision about who gets the first call is made on that text. Nobody at the supplier saw it, because it never appears in their analytics: no visit, no referral, nothing to track. This lands on whoever owns the brand story, and in most companies that person doesn't know it is happening.

Who AI reads

Read across the rows and two things stand out. The top three don't move. And below them, the layer is being reshuffled from one month to the next: two sources that were barely cited in early September were climbing three weeks later, and by mid October the American vendor's pages are the fourth most cited source of all. Finding 4 is about those two.

A source counts once per company, however many engines cited it and however many times. With Perplexity included, the American vendor's pages were cited for 37 of the 100 on 22 September and 51 on 12 October. The chart sticks to the four engines asked on every date.

Asked in Finnish

Companies each source was cited for, on the four engines that answered every Finnish question on the day (Claude, ChatGPT, Gemini and Perplexity). Google AI Mode is added below.

Counted as a share of all sources, the companies' own websites were 5.1% in English and 4.4% and 4.5% on the two Finnish dates. The German directory mostly drops out, and the American vendor falls to half or less.

Google AI Mode tells the same story. With it added as a fifth engine, the directory is still cited for 77 companies in English against 22 and 25 in Finnish, Finnish Wikipedia for 8 against 44 and 47, and the companies' own websites are 5.8% of sources in English against 4.7% and 5.1% in Finnish. On its own, Google AI Mode used the company's own website for 46 companies in English, and for 30 and 39 in Finnish. What comes in is Finnish: Wikipedia in Finnish, press releases, the technology industry association, an equity research firm, the national broadcaster, a business information service, a Finnish IT supplier directory, and student theses. Who gets named, and who gets named first, hardly changes between the two languages. What changes is who supplies the description, and it still isn't the company.

Why a directory wins

ensun.io is a supplier search engine run by a small German company. It lists companies by category and country, and it has a page for every combination: one for crane manufacturers in Finland, one for paper manufacturers in Finland, one for software consultancies in Finland, and so on, down to regions and cities.

So when an engine is asked "who are the leading paper manufacturers in Finland?", there is a page whose title is that question. No Finnish company has that page. A company's website is about the company, not about its category, and that is reasonable. It just isn't what the question looks like.

We found no sign that the directory set out to win in AI answers. It is a small company, mostly bootstrapped, and nothing we read about it mentions AI visibility as a goal. Its page structure happens to match the shape of the question. That is enough.

What makes this a finding rather than a curiosity is how few people read it. Semrush estimates that ensun.io gets fewer than 200 visits a month from Google searches in Finland (October; about 70 in early September). Its Finnish category pages rank for one to eight search terms each. By every measure a search team looks at, it doesn't exist here. In English-language AI answers about Finland's largest companies, it is the most cited source we found.

In Finnish answers the same shape wins, with different owners: Finnish Wikipedia's article on an industry, a trade association's member list, a business information service's company listing. Pages about the category, or about the company, written by someone else.

Doing well in Google and being used by AI are no longer the same thing. A company can check its search rankings every week and never see the source that describes it to buyers.

When AI does read your site

Across all answers, 6% of the sources were the company's own website on 12 October (136 of 2,118), 7% on 22 September (148 of 2,114) and 8% on 4 September (157 of 1,998). It went down a little on each date. Each site is counted once per answer, so a page cited three times in one answer counts once. Asked in Finnish, on the four engines that answered in both languages, the share was 4.4%, against 5.1% in English on the same engines.

In software and IT services it was under 4% on every date, the lowest of the seven sectors, and 2.4% in October. Industrial machinery was the highest on every date, at 9 to 11% in September and 9% in October. The sector pages have the rest.

And when the engines did use a company's own site, they mostly read the front door. Of the own-site citations whose page we could identify (175 on 4 September, 152 on 12 October):

• about half were the About page or the homepage, on both dates
• a product or solution page about one time in eight in September, one in six in October
• a news item fewer than one time in ten in September, about one in seven in October
• a customer case: two citations in October, none in September

So the answer about a company is built mostly from its most general text, the paragraph that says what it does and where. The product pages, the customer cases and the news, where the company says what makes it different, were used a little more in October, but they are still a small part of what AI quotes.

Asked as a buyer, the engines go one level deeper. Finding 5 shows where they end up.

The layer underneath

Below the directory and Wikipedia, the sources are lead generation databases, contact data services, and sites that rank agencies and suppliers by category. Their business is selling access to companies, or selling a place on a list. They describe Finnish companies because that is their product.

In software and IT services, the answers rest almost entirely on this layer, with Goodfirms, Clutch and TechBehemoths among the most cited sources. Some of the sources are hard to take seriously as authorities on Finnish IT. On 4 September three engines cited a blog post on an Indonesian higher education fair's website as a source for eight Finnish software companies. A Vietnamese outsourcing firm's blog post was cited for eight. Both were still being cited on 12 October.

In English, for 29 of the 100 companies the directory was cited and their own website never was, on any engine, on 12 October (31 on 22 September, 28 on 4 September). For 24 companies, their own website wasn't cited on any of the three dates, by any engine.

Some of the describing is done by companies you compete with. On 12 October, for 31 of the 100, another listed company's own website was among the sources for their category, most often a direct competitor's. For 15 of them, their own website wasn't cited at all that day (10 and 11 on the September dates). In two neighbouring categories, one listed company's site was cited six and seven times in the answers about its rivals. We added this count after the third date.

None of this is hidden. Every source is listed under the answer, for anyone in the company who looks.