Key takeaways
- In most categories a small group of vendors is named in nearly every answer, and the rest are named far less often. Every count is in the figures above.
- The assistants do not agree. A vendor can be named in most of one assistant's answers and in few of another's, so a figure for one assistant says little about the others.
- Every figure rests on a small sample and carries its interval. Read the figures as directions, not as a ranking.
When a buyer in the UK asks an AI assistant which software to consider, which vendors does the answer name? We put the same buyer questions, category by category, to each assistant through its API and counted which of a fixed list of vendors each answer named. This page sets out every figure, how it was counted and what it cannot tell you.
The study is about the answers, not the products. Being named often is not the same as being the right choice, and nothing here is a judgement of any vendor's product.
What we counted
A vendor counts as named in an answer when the answer names its product, one of its listed aliases, or an edition, plan or feature of it. The company's own name counts where it clearly stands for the product in that category. A company named only as an owner or parent does not count, and neither does a separately sold product that shares the brand.
Each figure reads "N of M": M is the answers from that assistant that were stored and classified for the category, and N is how many of them named the vendor. Every figure in the explorer is one assistant's answers only. Only the category shapes and the finding of a vendor named in every answer combine the assistants, and they show an overall pattern rather than a single measurement, because the assistants search differently and answer differently.
Every figure carries an interval, a range that says how high or low the true rate could plausibly be. Answers to the same question are not independent, so the interval is a cluster bootstrap that resamples whole questions rather than single answers, at the conventional ninety-five per cent level. Where every question gives the same result, as when a vendor is named in every answer or in none, resampling shows no spread, so the interval is the Wilson interval over the number of questions. Bounds are written as decimals between zero and one and rounded outward, so a range on this page is never narrower than the stored one. Reading an AI visibility figure honestly explains why we report figures this way.
How the answers were collected
9 categories, 10 buyer questions in each, every question put to each assistant 2 times
Collected from 26 Oct 2026 to 27 Oct 2026
| Assistant | Answers planned | Answers stored | Stored but not classified |
|---|---|---|---|
| Gemini (API, Google Search) | 180 | 175 | 0 |
| GPT (Azure OpenAI API, Bing search) | 180 | 173 | 0 |
| Perplexity (API, Sonar Pro) | 180 | 180 | 5 |
API answers from three assistants
The questions are written as a buyer in a UK mid-market company would ask them, and each category's questions were the same for every assistant. Each assistant answered through its API with its own web search: Gemini with Google Search, GPT through Azure OpenAI with Bing search, and Perplexity through its Sonar Pro model. Each category tracks a fixed list of vendors chosen before any answer was collected. A vendor that is not on the list may appear in answers, but it is not counted here.
Not every planned answer was stored. Sometimes an assistant declined to answer, and in the category collected first some answers stopped at the spending limit set for that run. A few answers were stored but the model that classifies answers declined to classify them, so they count in no figure. The explorer lists these counts under each category's figures. Missing answers are disclosed here, not retried.
Before publication, every counted mention was checked by two AI models working independently, and every mention either model doubted went to Answer Position for a decision. Mentions that both models judged to be a different product or company were removed, and the figures were recomputed. The answers our reporting rule holds back for review were reviewed by an AI model, and Answer Position approved each change it proposed. This was a model check with our decisions on every removal, not a human reading of every answer.
These are answers from the assistants' APIs. The consumer apps can answer differently, because they carry a person's history, location and settings. We have not yet compared these figures with answers from the apps, and each category's figures in the explorer say so.
What this study cannot tell you
- It is small. Each figure rests on the answers to one category's questions from one assistant. The intervals are wide, and two vendors whose intervals overlap may not differ at all.
- It is a snapshot. Assistants change their models and their search often. The same questions asked next month may name different vendors.
- The questions are ours. Different questions, or the same intent worded differently, can name different vendors. The questions were written before collection and not changed afterwards.
- Named is not recommended. An answer can name a vendor to recommend it, to compare it or to warn against it. This study counts only whether a vendor was named.
- It covers the tracked list only. A vendor outside a category's list is not counted, however often it is named.
The UK B2B technology AI visibility report
The full report of our UK B2B technology study: the vendors Gemini, GPT and Perplexity name in each category, every count with its interval and the method.
Free with your work email
Request the download