We asked four AI engines about 465 Danish companies: the technology determines whether you’re quoted — not whether you’re recommended

We asked four AI engines 576 questions about 465 Danish companies. Technical configuration is a strong predictor of whether your site will be used as a source — but says nothing about whether your company will be recommended. Two outcomes, two answers.

We asked four AI engines 576 questions about 465 Danish companies across three sectors — firms that can be found via a Google search in 12 Danish towns and cities. The results were split in two: technical setup is a strong predictor of whether your site is used as a SOURCE (p=0.000). It predicts nothing about whether your company is RECOMMENDED (p=0.988).

These are two questions that are usually lumped together as one. We did this ourselves when we started analysing the data — and initially concluded that ‘technical aspects don’t matter’, which was incorrect. They do matter. Just not in the way you might think.

The starting point was what research had already shown. The Princeton study behind the concept GEO found that targeted optimisation of a page can boost its visibility in AI responses by up to 40% (Aggarwal et al., KDD ’24). However, this is measured by how often a page appears in the response — not by whether a business is recommended.

The second study is more recent: in May 2025, Ahrefs compared 75,000 brands with their visibility in Google AI Overviews and found that mentions of the brand online were more closely correlated with visibility (0.664) than backlinks did (0.218) — the analysis can be found here. We’ll come back to this further down, as it should be read with caution.

Neither of these answers the question that a Danish tradesperson or accountant is actually asking: will I be brought up when a customer asks an AI? That is what we have measured.

How we measured it

We didn’t ask the search engines about individual companies. We asked the questions a typical customer would ask — “Which plumbing firms can you recommend in Aalborg?” — in 12 Danish towns, using four question variants, for ChatGPT, Gemini, Claude and Perplexity, with web search enabled. We then measured which companies in the sector the search engines themselves brought up.

Two factors were counted separately: whether the company was mentioned by name in the answer, and whether its domain was cited as a source.

The two figures should be interpreted differently, and this is important enough to be stated here rather than just in the methodology note. The names in the answer text are the engine’s own — namely ChatGPT, Gemini, Claude and Perplexity, which bring up the company themselves. The source URLs do not do so in all four cases. ChatGPT, Gemini and Claude were processed via a web search through OpenRouter, where a common search plugin (Exa) retrieves the sources and passes them on to the model. For these three, ‘cited’ therefore means: the search layer found the page and presented it to the model — not that ChatGPT, Gemini or Claude each found it individually. Perplexity searches the web itself, and is thus the only one of the four where the citation is the engine’s own.

How many were mentioned?

Industry	Measurable	Mentioned by name	Cited as a source
Lawyers	106	60 (57%)	59 (50%)
Plumbing	162	62 (38 per cent)	90 (45%)
Accountants & bookkeepers	127	41 (32%)	46 (31 %)
Total	395	163 (41 per cent)	195 (42 %)

Around four in ten were mentioned by at least one influencer. The variation between sectors is striking: a solicitor was almost twice as likely to be mentioned as an accountant.

The technical score did not predict who was recommended

We already have a technical score for each site — readable text without JavaScript, llms.txt, robots.txt access for AI crawlers, structured data, H1 and meta description. If that score determined who the AI recommended, the sites mentioned should have scored higher.

Industry	Score, mentioned	Score, not mentioned	p-value
Plumbing	77.9	78.8	0.66
Accountants	76.5	72.7	0.22
Solicitors	70.8	67.3	0.30
Total	74.9	74.3	0.70

None of the sectors comes close to reaching statistical significance. llms.txt even went the wrong way in two out of three cases: amongst plumbing firms, 6% of those mentioned had the file, compared with 13% of those not mentioned; amongst accountants, 10% compared with 16%.

The obvious objection is that there isn’t enough variation: 41 per cent of the sites scanned have exactly the same score, because they already have the basics in place. So we divided the field into score bands and measured the mention rate in each:

Technical score	Companies	Mentioned by AI
Below 60	59	36%
60–74	104	43%
75–84	37	32%
85–100	195	44 %

The curve is flat. There are 101 companies scoring below 70 points in the data, so there is a spread — and they are mentioned roughly as often as those at the top. A lack of variance therefore does not explain this on its own.

This does not mean that the technique is irrelevant. It means that it resembles a ticket to entry rather than a competitive advantage: necessary to be considered, but not the deciding factor amongst those who are already in. Our data cannot tell us what the situation is like for the sites below the bottom line — by definition, they are not visible in search results and are therefore not included.

But technical setup determines something else — and that’s important

When we broke down the two outcomes, the picture became quite different. Technical setup is a strong predictor of whether your site is used as a SOURCE. It predicts nothing about whether your business is RECOMMENDED.

Outcome	Technical score	JSON-LD	p-value
Cited as a source	77.9 versus 72.7	68% versus 57%	0.000
Mentioned as a recommendation	74.9 versus 74.9	63% versus 61%	0.988

These are two different questions, and they each have their own answer:

• “Does AI use my site as a source?” — here, the technology matters a great deal. The difference is 5.2 points and highly significant. Structured data and readable HTML make your site usable for a model.
• “Does AI recommend my business?” — here, the technical aspects have no measurable impact. The score is identical on both sides of the dividing line.

So it is possible to be the page that AI reads and refers to, without being the business that AI highlights. The technical work is not wasted — it simply leads to a different outcome than most people expect.

However, one thing does not hold true: content length. Cited pages have an average of 1,083 words compared to 895 for non-cited ones, but the difference is not statistically significant (p=0.32). We mention this because ‘write longer texts’ is a common piece of advice — our data does not support it.

What the AI actually cited

A total of 4,034 sources were cited, spread across 494 domains. The distribution surprised us:

• 61 per cent were the companies’ own websites (2,457 citations).
• 16% were directories and review sites (630).
• 23% were other third parties — industry portals, media outlets, other companies’ websites (1,237).

So, companies’ own websites dominate. It’s worth bearing this in mind if you’re about to conclude that AI only looks at mentions elsewhere — it doesn’t.

A breakdown of the 4,034 sources: 1,874 (46 per cent) are from Perplexity, retrieved by the engine itself, with an average of 13 sources per response. The remaining 2,160 (54 per cent) are distributed as follows: exactly 720 each from ChatGPT, Gemini and Claude — exactly five per answer, because the search plug-in behind them returns five. The fact that the figure is the same for all three is no coincidence: they were given the same list. A statement such as “ChatGPT cited 720 sources” is therefore a statement about the search layer, not about ChatGPT.

Which sources recurred across all three sectors?

Out of 494 domains, only six were cited in all three sectors. All six are directories or registers.

Domain	Citations	What it is
krak.dk	103	Business directory
findformig.dk	78	Review directory
Trustpilot	73	Review platform
cvrexplorer.com	23	CVR register mirror
LinkedIn	22	Company profiles
degulesider.dk	17	Business directory

Everything else was sector-specific. This is the most actionable observation in the dataset: whether you’re a solicitor, an accountant or a plumber, the same small set of directories was used — and a directory profile is something you can sort out in an afternoon.

However, the six directories are almost exclusively Perplexity’s, and that changes who the advice applies to. If you break down the 316 directory citations across the search engines, Perplexity accounts for 294 of them (93 per cent):

• krak.dk, degulesider.dk, cvrexplorer.com and LinkedIn: all citations came from Perplexity. ChatGPT, Gemini and Claude referred to them zero times in total.
• findformig.dk: 59 of the 78 were from Perplexity; the remaining 19 were distributed with 6–7 each across the other three.
• Trustpilot: 70 of the 73 were from Perplexity — the other three cited it once each.

The explanation is the same as above: the three used the same search plugin, and that plugin predominantly pointed to the companies’ own websites rather than directories. So the advice on directories isn’t wrong — it’s just more targeted than the table suggests: a directory listing is first and foremost a Perplexity investment, as measured here. If you want to outperform the other three, your own website is the source from which the data was actually retrieved.

Be aware of what this does not tell us. We have measured which sources the AI cited — not whether a profile in those directories gets you mentioned. To do that, we’d need to track the same companies over time and see whether changes to the directories make a difference. We haven’t carried out that measurement yet.

What do other studies say about mentions versus links?

Our own dataset says nothing about why some companies are mentioned. The closest external benchmark is the Ahrefs analysis from the introduction: 75,000 brands, published on 26 May 2025, compared against their visibility in Google AI Overviews. Measured using Spearman’s rank correlation, mentions of the brand were more closely correlated with visibility (0.664) than the number of backlinks was (0.218).

This is worth noting. However, it should be read with four caveats — the first three of which Ahrefs itself draws attention to:

• It is Google AI Overviews. Not ChatGPT, Claude or Perplexity. One metric, not ‘AI visibility’ in general — and that’s an important distinction when the figure is quoted.
• The correlation is not strong in itself. Ahrefs describes all the factors examined as moderate to very weak. 0.664 indicates a clear correlation, not an explanation for most of the variation.
• Correlation does not imply causation. Well-known brands are both mentioned more often and brought up more frequently. The study cannot distinguish between the two, nor do the authors claim to do so.
• Avoid using the multiple form. The figure is widely circulated as a ratio: that mentions are said to be three times as effective as backlinks. This is 0.664 divided by 0.218 — and two correlation figures cannot be divided by one another in this way. The ratio between them does not mean that the effect is three times as great. Report the two figures separately, not the ratio.

When compared with our own figures, the two point in the same direction without contradicting each other: mentions elsewhere than on your own site do matter. But the six directories that the search engines actually cited in Danish local search are a more concrete place to start than a general hunt for mentions — and they’ve been measured in precisely the market you’re selling in. Ahrefs’ analysis can be found here.

What we cannot answer

We can say with reasonable certainty what does not determine visibility: the technical score, amongst sites that are already visible in search results. We cannot say what does.

We haven’t measured the number of reviews, age, company size or marketing budget. Any explanation as to why lawyers perform better than accountants would be a guess on our part. It’s a measurement we’d like to carry out — and we’ll publish the results, whatever they show.

The practical implication is the same regardless of the explanation: you cannot calculate whether AI will mention your business. A technical checklist can tell you whether you’ve got the basics in place — but not whether the search engines will actually rank you highly when a customer searches. These are two different questions, and only one can be answered through measurement.

Frequently asked questions

Does a good technical GEO score predict that AI will mention your company?
No — but it does predict whether AI uses your page as a source. Those are two different outcomes. Across 465 Danish company websites (Geoa, 26 July 2026), the companies AI RECOMMENDED had a technical score of 74.9 against 74.9 for the rest — no difference (p=0.988). But the pages AI CITED as a source scored 77.9 against 72.7, and that difference is strongly significant (p=0.000).

What is the difference between being cited and being recommended by AI?
Being cited means AI uses your page as a source for its answer. Being recommended means AI names your company as a concrete option for the customer. In our data, technical setup is strongly linked to the first and not at all to the second. You can perfectly well be the page AI reads without being the company AI points to.

So does llms.txt help you get mentioned by AI?
Our data shows no positive link. Among the search-visible plumbing firms, 6 per cent of those mentioned had an llms.txt against 13 per cent of those not mentioned; among accountants, 10 per cent against 16 per cent. That does not prove the file is harmful — the numbers are small and the sample limited. But the claim that llms.txt makes AI mention you is not something we can support with our own measurements.

So which sources does AI use when recommending local businesses?
A total of 4,034 sources were cited, spread across 494 domains. 61 per cent were the businesses’ own websites. Only six domains appeared across all three sectors, and they were all directories or registers: krak.dk (103 citations), findformig.dk (78), Trustpilot (73), cvrexplorer.com (23), LinkedIn (22) and degulesider.dk (17). Note the distribution: 294 of the 316 directory citations came from Perplexity, which is the only one of the four search engines that performs its own searches. ChatGPT, Gemini and Claude ran via a shared search plugin (Exa), which provided five sources per result and almost never pointed to the directories.

How big is the difference between sectors?
Large — far larger than between technical scores. Of the measurable companies, 57 per cent of law firms were mentioned by at least one engine, against 38 per cent of plumbing firms and 32 per cent of accountants. The sector mattered more for visibility than the site’s technical setup.

May I quote the figures?
Yes. Please cite “Geoa, July 2026, n=465 companies, 576 AI responses” and feel free to link to the analysis. If you use the source figures, please include the following caveat: three of the four search engines obtained their sources from a shared search plugin; only Perplexity conducted its own searches. We note the date of each update so that older citations can be traced back to the version from which they were taken.