How this study was done

How this study was done

I wanted a list of companies nobody could say we had picked. So the sample is the 100 largest B2B companies listed on Nasdaq Helsinki, ranked by consolidated group revenue, and no company was added or removed by hand.

The 100 companies

Revenue comes from stockanalysis.com, which compiles the audited figures from the companies' annual reports, the same figures financial journalists use. I started from every company listed in Helsinki on 4 September 2026 and kept one row per company where it has two share classes. Then I took out the companies domiciled outside Finland, the banks, insurers, real estate, retail and travel companies, companies that sell mainly to consumers, and four with no reported revenue. A handful that sit between consumer and business markets (food, media, tyres) are held in a separate group and are not part of the 100. Revenue in the final 100 runs from €20.6 billion at the top to €27.4 million at number 100.

Before settling on this I looked at four other ways to build the list and dropped all of them. Talouselämä's TE500 is a paid list (875 €), so a reader couldn't check our sample without buying it. A business information service's free top lists show only the first ten. Wikipedia's list of Finnish companies by revenue uses 2013 figures. The open financial data from the Finnish Patent and Registration Office reports subsidiaries separately, so a group with €11 billion in revenue shows up there as an €800 million domestic company. And I didn't use our own prospect list, because that would have been a sales pipeline with a research label on it.

Private companies are not in the 100. There is no free primary source for their revenue that a reader could check, and a second, weaker source for part of the list is exactly what a careful reader would pick apart. To see what being listed does, we asked the same questions about a separate group of 29 privately owned Finnish B2B companies of similar size and sector. They are reported on their own (see Limits below) and never mixed into the 100.

The question

Each company gets one question, the one a buyer starts with: "Who are the leading [category] in Finland?" The category is what the company says it does. For a group with several businesses, it's the largest one. Companies in the same market get the identical question, so the 100 companies need 85 questions, and every one of them is listed in Appendix A, so you can check that your category was asked.

Why one question and not five? Because the same question asked of every company, and asked again on three dates, tells you more than many questions asked once. It also keeps the wording ours, not the company's: nobody asks AI about a company's own positioning line.

The questions are in English. A Finnish buyer may well ask in Finnish, so on 12 October we also asked all 100 questions in Finnish, written as a Finnish speaker would ask them, not machine translated. Who gets named and named first hardly changed. The sources did: the German directory mostly dropped out and Finnish sources took its place, while the companies' own websites stayed a small share, slightly smaller than in English. We asked in Finnish on two dates, 12 and 15 October, and the picture was the same both times. The comparison uses the four engines that answered every Finnish question on the day. Google AI Mode's Finnish answers had gaps on both dates, which we filled afterwards (see below), and with it included the picture is the same. And "in Finland" makes it easy to be named in a small category with few players. That is why being named is the weakest measure in this report. Being named first, and having your own website among the sources, are the ones to read.

On 22 September we also asked once as a buyer would: the same 100 companies, phrased as a procurement manager building a supplier shortlist. That run is reported once (Finding 5), with no claim that it would come out the same twice.

The engines and the dates

Every question went to each engine with its web search on, so each answer came back with the sources it used, as it would for a user.

4 September: Claude, ChatGPT, Gemini, Google AI Mode.
22 September and 12 October: the same four plus Perplexity.

Whenever this report compares dates, it uses the same four engines. Microsoft Copilot is not covered.

The website checks

For each company's main website we checked:

• whether it has schema markup (the code that tells search engines and AI what a page is about) on the homepage and on one news or product page linked from it, where one could be found
• whether it has an llms.txt file (a summary written for AI engines)
• whether its robots.txt file tells AI engines' bots to stay away
• whether the site turns away any of six bots (GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, CCBot and Bingbot) while serving an ordinary browser normally
• how much readable text an AI bot gets from the pages

These checks were run on 4 September, and again on 5 October for the eight companies whose website address we corrected, and for all 100 again on 12 October. They record whether something is there. They don't judge whether it is any good, which matters when you read Finding 3.

How we counted

Named means the company, or a brand it operates under, appears in the answer. We read this from the answer text itself, because the engines' structured lists sometimes leave out a company the text clearly names.

Named first means it is the first company named in the answer text. That is not an accident of formatting: only 1 of 362 answers on 12 October listed its companies alphabetically, and where an engine also gave its own ranking, the company named first in the text was its number one in 256 of 272 cases. Where we report who takes the first place (the sector pages and Finding 2), a brand the company operates under counts as the company, and only answers that named some company are counted, so those shares run a few points higher than text order alone.

Own website cited means at least one of the answer's sources is on the company's own domain. Each source is counted once per answer. Gemini returns its sources as redirect links, so we read the real site from the link title. Links to Google's own result viewer are not sources and are left out.

All of this runs on tooling we built for the study, and every raw answer is kept, so any count in this report can be re-run.

We compare engines on who they name first and on own-website citations. We don't compare how many companies they name or how many sources they list: one engine names around 13 companies per answer where the others name seven or eight, and Perplexity cites about 11 different sites per answer where ChatGPT cites about four. More names doesn't mean a better answer.

For eight companies, the website address in our sample file was wrong or out of date. We found this on 5 October and recounted every number. This report uses the corrected figures.

In October our own collection tool had two faults with Google AI Mode, which we found and fixed on 16 October. When a call to our data provider failed, the tool recorded an empty answer instead of an error: 5 of the 100 English questions on 12 October, and 43 of the Finnish questions. We asked those again on 16 October, so those answers are four days later than the rest. The tool also stopped reading Google AI Mode's list of sources after the provider changed its format. We recovered the original lists from the provider's responses, so those answers keep their own date. Every figure in this report uses the corrected data.

A company's name is matched as a whole word, so a company whose name is also an ordinary phrase isn't counted as named when an answer just uses that phrase.

What this study does not do

It gives no score, ranks no company, and names none of the 100. It also doesn't tell you what AI says about any one company. Answers move: only 21 of the 100 companies got exactly the same result on the two September dates, 24 from September to October, and 9 on all three dates. A single company's row needs its own repeated check, and that is a different piece of work from this report.

Limits worth knowing

Every company here is listed, and listed companies have to publish. The law requires them to put their financial reports on their own website and keep them there for ten years, to announce anything likely to move the share price as soon as it happens, and to publish details of their board, management and owners. Unlisted companies have no such obligation. So AI probably finds more material about a listed company than about an unlisted one of the same size. To see how much it matters, we asked the same questions about 29 unlisted Finnish companies of similar size and sector, picked by ownership and revenue before we looked at any answer. 19 of them belong in a category we already asked about, so they could be compared on the very same answers. AI named them about as often as the listed companies, 68% of answers against 62%, and named them first about as often. But it almost never used their own websites as a source: in 2 of the 144 answers, against 19 for the listed company. Fewer of them have an English Wikipedia article too, 7 of the 29 against 38 of the 100. That fits the obligation to publish. If you work at an unlisted company, don't assume your situation is better than what you see here. It is more likely to be the other way round.

Every question is the first question of a conversation. In practice buyers keep going. They add a budget, a system the supplier has to work with, a country they need served, and they ask the engine to compare. We didn't measure those follow-ups, so this report says nothing about who is still on the list by the fifth question. What it measures is the list the conversation starts from, and which sources that list is built from. When we asked the same 100 categories again as a procurement shortlist, with the buyer's context added, the same companies came out on top and the same sources were cited. Later questions can still change the answer in either direction: a specific requirement can bring in a specialist who wasn't named at first. Following whole conversations is the next thing I want to measure. It is also the subject of recent work by VML's global analytics team on what they call conversational durability.

Sectors are small. The smallest has seven companies, and every row in the benchmark shows its n. Read a small row as a direction.

About half of Gemini's answers came back with no sources at all (47 of 100 on the first date), so Gemini contributes less to the source findings than the other engines.

We wrote down one test before the third date: whether a complete company entity in the site's markup, one that links out to the company's other official profiles, made answers more stable. It looked that way on the first two dates and didn't hold on the third.

And this is a snapshot. AI answers change from week to week, and these were collected over six weeks in September and October 2026. The pattern held across all three dates, with one exception: the German directory was cited for seven fewer companies in October than in September, and in English it is still cited for more companies than their own websites are. Any single answer may already be different by the time you read this.

How the room measures today

When people registered for the launch event, we asked them one question: how are you tracking your company's AI visibility today? 44 people answered, from industrial companies, software firms, agencies, a law firm, a research company and public organisations. VML Finland's own staff are left out.

Two thirds check by hand or informally. A check like that is usually one engine, asked once, on the day someone thought of it, and in this study only 9 of the 100 companies got the same result on all three dates. A quarter have a dedicated process, which is more than I expected. People who sign up for a breakfast on AI visibility are the ones already paying attention, so across Finnish B2B as a whole the share is probably lower.

What they wanted to hear about was mostly measurement: what AI answers do to search traffic and leads, what tools exist to track it, and what actually helps a company appear.