Your shopping cart is empty!
AI visibility tools count citations, but a citation is not a recommendation
Publication date: 19.07.2026
Popular tools for measuring AI search visibility count how often you are cited — and present that as a success metric. The problem is that a citation and a recommendation are not the same thing: a brand can be mentioned as a source while the user is advised to go to a competitor.
How big the gap is
Jeff Oxford tested 20,000 ChatGPT responses and found a correlation between citation and recommendation of only 0.4. Lily Ray found listings cited 323 times that were excluded from the final recommendation in 69% of cases. Add instability: according to Rand Fishkin, you need to ask a model around 1,500 times to get two identical answer lists in a row — while dashboards report from a single run.
According to Search Engine Journal, BrightEdge found that 91% of cited URLs appear in only one system — citations do not travel between different AI platforms. Visibility in ChatGPT says almost nothing about visibility in Gemini or Perplexity, and a single blended "AI visibility" number hides exactly that difference.
Why answers differ every time
A model does not pull a ready-made list from a database — it assembles the answer from scratch each time. The wording of the prompt, earlier turns in the conversation, whether web search is enabled, even word order all change the result. Two consecutive runs will almost certainly return different sets of brands, and that is a property of the technology, not a broken tool.
One practical conclusion follows: a single run proves nothing either for or against you. Only a series of identical measurements over time carries meaning — and that is where a change in your share of answers becomes visible.
What to measure instead of raw citations
| Metric | What it shows | How to collect it |
|---|---|---|
| Presence Share | how often you appear in answers | a fixed prompt set, run on a schedule |
| Recommendation Share | how often a mention becomes advice | manual labelling: source or recommendation |
| Brand Accuracy | whether AI describes the company correctly | fact check: year, region, products, competitors |
| Link to action | whether a visit or enquiry followed | AI referrals in web analytics |
| Answer composition | how many brands appear and how wording shifts | storing full answers over time |
How to build the measurement without expensive tools
A minimum system takes half a day to assemble by hand:
- write 20–30 prompts from real customer questions, not from your Google keyword set;
- run them on a schedule — weekly, in exactly the same wording;
- store the full answer, not just the fact of a mention: that is what shows change later;
- mark in a separate column whether you were cited or recommended;
- check factual accuracy monthly: founding date, region of operation, service list.
Brand accuracy deserves separate attention: models take it from what is written about you across the web and on your own site. That is the same work as brand SERP optimisation — consistent company data across every source.
The metric worth tracking
The number of mentions on a dashboard stays an impressive but mostly uninformative figure — it's only weakly tied to whether a model actually recommends the brand to a real person. Until the industry settles on a single AI-visibility metric, what says something is recommendation share and actual AI referral traffic — the two numbers that can be tied to money.
