Having schema markup didn't decide whether AI used a company's website. Most of the markup says almost nothing about the company.

The gap is a few points, and it ran the other way on the first date. With 32 companies in the smaller group, a few points is two companies. And only 18 of the 100 use their markup to state any fact about themselves (see "What the markup says, and who reads it"), so this compares companies with mostly empty markup against companies with none.

What this means in practice

Somewhere in most of these companies there is a closed ticket that says the structured data was added to the website. The developers did what was asked, the validator came back clean, and the work was reported as done. Nothing in this study suggests it changed who gets cited. What went with the answer here was what the rest of the web says about the company, and whether its own content gives an engine something worth quoting. That isn't an IT failure, and the IT backlog won't solve it.

Deciding what the markup should say about the company, which story the content should tell and who outside the company should be telling it is a marketing and communications decision. In most companies that decision hasn't been given to anyone. The technical work wasn't wasted either: a third of the companies have no markup at all, and six turn away at least one AI engine's bot.

The technical picture

What we found on the websites of the 100 companies on 12 October:

• 31 have no schema markup on the pages we checked (the homepage and one news or product page linked from it).
• 84 have no llms.txt, a summary file written for AI engines. About half of the 16 files we did find don't follow the format, or are empty, so an engine reading them gets little out of them.
• 6 turn away at least one AI engine's bot while serving the same page to an ordinary browser. One of them turns away all six bots we tested. One more company asks OpenAI's bot to stay away in its robots.txt file.
• 96 serve readable text to an AI engine's bot on the homepage. Rendering, the problem of a site that only shows its content after code runs in the browser, was not a common issue in this group.

The checks on 4 September gave almost the same picture: 32 without schema markup, 85 without llms.txt, the same six turning a bot away. The sections "Finding 3: The technical layer" and "Readiness against outcome" use those earlier checks, so that all three dates are compared against the same set of websites.

By sector, 12 October:

Software has almost the most complete markup in the study and the lowest own-website citation. Industrial machinery has more than a third of companies with no markup and the highest. Sorted by markup, the sectors don't line up with the outcome at all.

Readiness against outcome

Schema is one item. To check that we weren't just looking at the wrong one, we scored every company on ten things an AI engine needs from a website, one point each, using the same checks on all 100 (the 4 September website checks):

Most companies scored between 4 and 7. The score doesn't follow company size.

It didn't predict anything. Companies whose own website was cited averaged 5.4 on the first date, and those whose website wasn't cited averaged 6.0. On the second date, 5.6 and 5.6. On the third, 5.7 and 5.4, a gap that chance alone would give about one time in two. The same held for being named first on every engine, and for being named at all. Grouped into bands, there was no steady pattern from one band to the next: on the third date the highest band did best, but the lowest band, seven companies, did better than the two middle ones.

So readiness can be measured, but in this data it didn't predict the answer.

One item in this table we wrote down in advance to check again: a named author on the page. In October it went with being cited again, 73% of companies with one against 49% without, after 71 against 53 on the second date and hardly any difference on the first, so it is a lead worth following, not yet a finding.

What the markup says, and who reads it

Whether a site has markup is a blunt measure, so we looked inside it, on each company's homepage and, where we could find them, an about page, a product page and a news item.

Most of it says almost nothing. 74 of the 100 have some markup on those pages, but for 17 it only describes the page itself, the kind many website platforms add on their own. 54 describe the company, and only 18 of those state a single fact about it: an address in 13, a founding year in 5, the number of employees in 4. None says who owns the company, who runs it or what it sells for. 32 link to the company's other profiles, almost all to LinkedIn, and none to Wikidata, the open database much of the web's company information is built from. On the product pages we found, 5 of 45 describe the product in markup.

So the result on "Finding 3: The technical layer" isn't a test of what well-made markup can do. In this group, hardly anyone has made it. Of the 100, 82 either have no markup or have markup that tells a machine nothing about the company beyond its name.

We also tested what the engines see when they read a page. We took three of the study companies' homepages, picked a value that appears only in each page's markup, and asked each engine three times to open the page and quote it. ChatGPT, Gemini and Perplexity read the pages and quoted the visible headline every time, but none of them could see the markup in what it had fetched. Claude, through the tool we used, couldn't open a given page, and Google AI Mode didn't answer.

When an engine opens a page, it reads the text. Markup works further up: Google says it uses markup to understand pages in its own index, and ChatGPT's search draws on Bing's. It is also how a machine reads a price, a specification or whether something is in stock, which matters most when an AI agent compares suppliers. We didn't test product questions or buying agents, and with product markup on 5 of 45 product pages, there would have been little to test.

What the engines repeat has to agree everywhere. One company's markup gives an office address in one city; every engine that answered named its real headquarters in another. Markup that says nothing, or says something the page and the rest of the web don't, can't help. The work is making the markup, the visible page and the outside sources say the same thing about the company. In the fact check, the 18 companies whose markup states a fact had fewer wrong facts in the answers, a lead we didn't plan to test (see "Where the errors are").

What does go with the outcome: Wikipedia

Of everything we could measure about a company, one thing went with the outcome on all three dates: an article about the company in the English Wikipedia.

94 of the 100 have an article in the Finnish Wikipedia. 38 have one in English. On 4 September the engines cited the English Wikipedia 104 times and the Finnish one 5 times. Those were English questions. Asked in Finnish on 12 October, the Finnish Wikipedia was a source for 33 of the companies.

The clearest difference is in being named first. The difference in own-website citation is smaller, shrank on the second date and widened again on the third.

Large companies are more likely to have an English article, so we compared within size. Among the 50 largest companies, those with an English article were first on all four engines about six times in ten, and those without fewer than one time in five, on all three dates. Among the smaller 50 only 12 have an English article, and there the difference was small and went both ways: about the same on the first date, a few points ahead on the second, behind on the third. So the clear version of this finding is about the larger companies.

This is a correlation, not a cause. An English Wikipedia article is also a sign of everything around it: international business, press coverage in English, people outside the company paying attention. That is the point. The things that went with the outcome in this study are about what others write about a company, not about how its website is built.