Cited or recommended? The two AI outcomes most people confuse

Your website may be cited as a source in an AI response — and your company may be recommended by name. Two very different outcomes: in our analysis of 576 AI responses, technical readiness strongly predicted the first outcome (p=0.000) and not at all the second (p=0.988). Here’s the difference, and how to determine which outcome is the most profitable for you.

There are two completely different ways of ‘being visible in AI’, and they are constantly confused: your website may be cited as a source in an AI response — and your company may be recommended by name. We’ve analysed 576 genuine AI responses: technical AI readiness strongly predicts the former, but not the latter at all. If you confuse the two, you’ll end up buying the wrong service and measuring the wrong metric. Here’s the difference, the figures behind it — and how to determine which of the two outcomes is worth the investment for your specific business.

What’s the difference between being cited and being recommended?

Ask an AI a question such as “how do I choose an accountant for my online shop?”, and the answer might feature you in two different roles. It might cite your guide on the subject as a source — your content provides the knowledge, and your name appears in the list of sources. And it may recommend your business — “a firm such as [name] is an obvious choice” — where you are not the source, but the candidate. The same answer, two very different outcomes: one provides authority and clicks, the other puts you straight onto the customer’s shortlist.

Why is this difference important? The figures behind it

The citation side is the well-documented one: the Princeton/Georgia Tech research (KDD ’24) measured up to a 40 per cent increase in visibility in generative engines by optimising content — cited sources, statistics, clear structure. This is the kind of visibility that technical work can demonstrably drive: being read and used as a source.

The recommendation side points in a different direction: 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 (correlation 0.664) than backlinks did (0.218). An important caveat: this is correlation, not a proven cause — but the trend is hard to ignore.

And then there is our own measurement, which is the reason this article exists: in an analysis of 576 AI responses with web search enabled — four AI engines, twelve Danish cities, four unbranded customer queries, measured against 465 real companies across three sectors (our own study, July 2026) — the companies that were cited had significantly higher technical scores than the rest (77.9 versus 72.7) and more frequently featured structured data (68% versus 57%). The companies that were recommended, by contrast, were technically identical to those that were not: 74.9 versus 74.9. Statistically: p=0.000 for citation, p=0.988 for recommendation. The full dataset and methodology are set out in the analysis of what determines whether AI mentions your company.

Why does almost everyone conflate the two outcomes?

Because both are marketed under the same name: “AI visibility”. Most tools display a single figure, most guides provide a single checklist, and most offers boil down to “get into ChatGPT” — without distinguishing between being read and being chosen. The consequence is predictable: companies buy technical optimisation and expect recommendations, or chase publicity and wonder why their content is never cited. Both approaches are valid enough — they simply drive different outcomes.

At the same time, our own survey shows just how much the industry influences the recommendation outcome: of the 576 responses analysed, lawyers were recommended by name in 57 per cent of the responses, plumbers in 38 per cent and accountants in 32 per cent — differences that are far greater than any technical differences between the firms. How often the AI actually points to specific firms therefore depends first and foremost on the type of questions your sector receives.

Which outcome is most profitable for your business?

The answer depends on how your customers ask their questions — not on what is technically the best approach. A three-part rule of thumb:

• Customers who do their research before making a choice (“how do I do this”, “how much does it cost”, “what should I look out for”): this is where the money is. If your guide becomes the go-to source, the customer will come across your name whilst forming their decision — this is where the advisor, the specialist and the online shop leave their mark.
• Customers who ask for a name (“recommend one”, “who’s best in [town]”): this is where the recommendation money lies — typically local services, where the customer wants a shortlist, not a textbook.
• Most SMEss have both — but in varying proportions. Run your own 5-question test and see what type your actual customer questions are: this will tell you which outcome to measure and prioritise first.

Which work drives which outcome?

The citation track is documented and within your control. Crawler access, content readable without JavaScript, structured data and citable, answer-first content — these are the signals that distinguished the cited sites from the rest in our measurement, and these are the ones the audit examines. One figure from the dataset that often comes as a surprise: of the 4,034 sources cited by the AI responses, 61 per cent were the companies’ own websites. The role of source is therefore not reserved for portals and directories — your own site is the largest source of citations, provided it is readable.

Frankly, the recommendation track hasn’t been cracked — by anyone. Our data doesn’t explain it: the industry provided the most explanation, whilst the technology offered none. And three popular strategies didn’t hold up when we tested them against the dataset: “write longer texts” (no significant difference, p=0.32), “create an llms.txt and you’ll be mentioned” (no support in our figures — create it for the right reason instead) and “get listed in directories” (the correlation cannot be determined without measurement over time). The ‘mentions hypothesis’ — that reviews and mentions online drive recommendations — is supported by the Ahrefs correlation above, but it is not a proven formula. The honest conclusion: be sceptical of anyone selling you a guaranteed path to AI recommendations, and instead measure whether what you’re doing actually moves your figures.

How do you find out where you stand on both counts?

Measure the two outcomes separately — that’s the whole point. You can assess the foundation for citation in 60 seconds using the free test (can the AI engines read, understand and cite your site?). You can test the recommendations side using your own customer queries — either manually as a spot check, or systematically week by week, with Geoa asking the same questions and showing whether you’re mentioned, who’s mentioned instead, and whether the trend is going your way. There’s one thing a checklist can never tell you: whether you’re being recommended. Only a measurement can do that.

Frequently asked questions

Can technical optimisation get AI to recommend my business?

Not according to our measurement: the recommended companies were technically identical to the non-recommended ones (74.9 versus 74.9, p=0.988). Technical work opens the door to being cited as a source — it’s documented and valuable — but it doesn’t in itself make you the business that the AI singles out.

So is technical GEO a waste of time?

No — it simply shifts the other outcome. Citation was strongly correlated with technical readiness in our data (p=0.000), and 61 per cent of the cited sources were the companies’ own websites. If your website cannot be read, you cannot be the source — no matter how good your content is.

So what determines who gets recommended?

It hasn’t been settled — not even by our own analysis. The industry provided the most explanation (lawyers 57 per cent, plumbers 38 per cent, accountants 32 per cent — figures here), and the ‘mention’ hypothesis is supported by Ahrefs’ correlation analysis, though without a proven causal link. That’s why the measurement is more important than the formula: test, monitor progress, and be sceptical of overconfident promises.

How reliable are the figures from the 576 responses?

The measurement covered 576 AI responses with web search enabled, distributed across four AI engines, twelve Danish cities and four unbranded customer queries, measured against 465 real companies in three sectors; the differences have been statistically tested (p=0.000 for citations, p=0.988 for recommendations). The methodology, limitations and what the figures cannot answer are clearly set out in the analysis itself.