Key findings
- ChatGPT names providers most often: for about 53% of all queries studied, and for about 85% of searches for a provider or product.
- Gemini and Perplexity are at about 37% and 36%, the Google AI Overview at about 26%.
- For provider searches, Google often shows no AI Overview at all (in about 57% of cases). When it appears, it names providers in about 74% of cases.
- Google ranking matters in AI too: positions 1 to 3 are cited about 3.8 times as often as positions 4 to 10 in the AI Overview, about 3.7 times in ChatGPT and only about 1.7 times in Perplexity.
- For about 35% of queries none of the four services names a provider, and only for about 9% do all four.
Which businesses does AI recommend, and when at all?
A business can only be named in an AI answer if the AI names providers for that question at all. Many AI visibility measurements count how often a brand appears. They do not show whether the question was suitable. An AI that answers a knowledge question with an explanation recommends nobody, however good a website is.
So we asked: for how many real search queries does an AI name specific businesses, shops, brands or products that people can choose, buy, book or contact? And does that differ between ChatGPT, Gemini, Perplexity and the AI Overview in Google Search?
AI comparison: the result per service
ChatGPT names specific providers for about half of all queries, the Google AI Overview for only about a quarter. Gemini and Perplexity are in between.
How often does AI name specific providers?
Share of all queries studied in percent, weighted by how common each question type is. The dark line shows the 95% range.
The differences are not chance. We asked every query in all four services and compared them in pairs. ChatGPT named providers for 171 queries where Perplexity named none, and the reverse happened only 24 times. Between ChatGPT and Gemini the ratio was 153 to 42. The smallest difference is between Gemini and Perplexity, and it is still clear.
When the services name providers, they often name several: ChatGPT and Perplexity name three or more in about half of their answers, Gemini in about a third, the AI Overview in about a quarter.
queries bring no provider recommendation in any of the four services. Only for about 9 out of 100 do all four services name providers.
Google: often no AI Overview, but providers when it appears
For searches for a provider, Google mostly shows no AI Overview at all. For knowledge questions it almost always appears. That is the opposite of what businesses would expect.
Google AI Overview: does it appear, and does it then name providers?
Real Google Search from Austria, per question type.
For provider searches this means: in more than half of them, classic ranking still decides who is found. But when an AI Overview appears, it names specific providers in about three out of four cases. Whoever is missing there is visibly missing at the very top. For knowledge questions the AI Overview is almost always there but rarely recommends anyone. How this affects clicks is shown in our study on Google AI Mode in Austria.
The type of question decides
Whether an AI names providers depends far more on the type of question than on the topic. We assigned every query to one of four question types:
People who want to buy, book or hire something, such as “X in your town” or “buy X”.
Questions about costs, prices and fees, such as “how much does X cost”.
Questions about the best option or differences, such as “X or Y” or “best X”.
Knowledge questions, explanations and guides, such as “how does X work”.
Share of answers with specific providers, per question type
The darker, the more often the service names providers.
| Question type | ChatGPT | Gemini | Perplexity | Google AI Overview |
|---|---|---|---|---|
| Find a provider or product | 85% | 54% | 61% | 32% |
| Price and cost | 55% | 44% | 31% | 39% |
| Comparison and choice | 39% | 34% | 31% | 28% |
| Knowledge and how-to | 22% | 19% | 12% | 18% |
For searches for a provider or product, all four services name specific names most often, ChatGPT almost always. For knowledge questions all of them rarely do. For price and comparison questions the services differ more: here ChatGPT names providers much more often than Perplexity.
What this means for an AI visibility metric
How many of a website’s own queries trigger a recommendation at all differs many times over from website to website. A visibility rate without this denominator is hard to compare.
Our queries come from the Search Console of websites. For 31 websites at least 20 of their queries were in the sample, 46 at the median. For each we calculated the share of its queries for which a service names specific providers.
How many of a website’s own queries trigger a recommendation
31 websites with at least 20 queries. Coloured bar: middle half of websites. Thin line: middle 80%. Dark mark: median.
Values as a table
| Service | 10. | 25. | Median | 75. | 90. |
|---|---|---|---|---|---|
| ChatGPT | 12 | 22 | 50 | 75 | 83 |
| Gemini | 5 | 9 | 45 | 64 | 75 |
| Perplexity | 3 | 9 | 28 | 63 | 77 |
| Google AI Overview | 9 | 13 | 24 | 41 | 49 |
This is not an effect of small numbers. If queries are assigned to websites at random, the gap between the lower and upper quarter for ChatGPT stays below 16 percentage points in 95% of cases. We measured 53. Websites with many recommendation questions in ChatGPT almost always have them in Gemini and Perplexity too (rank correlation 0.94 and 0.91), in the AI Overview only partly (0.56).
A worked example. Two websites are each named ten times in ChatGPT across 100 checked questions. For website A a share like the lower quarter of our data triggers a recommendation, for website B a share like the upper quarter:
| Example based on our distribution | Website A | Website B |
|---|---|---|
| Questions checked | 100 | 100 |
| Questions naming providers | 22 | 75 |
| Mentions of the own brand | 10 | 10 |
| Unadjusted visibility | 10% | 10% |
| Visibility in questions naming providers | 45% | 13% |
Assuming all ten mentions fall into answers with a provider recommendation.
Both websites look equally visible at 10%. But website A is in almost every second recommendation, website B only in about every eighth. A needs to cover more questions where recommendations happen. B already has these questions and is named too rarely there.
How reliable are AI recommendations?
The same question does not always bring the same result, and the Google AI Overview varies most. We asked about 90 queries per service three times.
Same result when asking the same question three times
Share of queries where three runs decided the same way every time whether providers were named. About 90 queries per service.
For the AI Overview, whether providers are named changes for about three in ten queries, partly because it appears once and not the next time. A single sample therefore says little. Reliable statements need several measurements per question and a share instead of a yes or no.
Where does AI get its sources? The role of Google ranking
A website in positions 1 to 3 of regular Google Search is linked as a source in the Google AI Overview in about 50% of cases, in positions 4 to 10 in only about 13%. For ChatGPT the gap is similar at a lower level. Perplexity depends least on Google ranking.
For this we captured the regular Google results for every query at the same moment and checked which of these websites the AI answer links as a source.
How often a website is cited as a source, by position in Google Search
The same query at the same moment. Share of websites in each position that the AI answer links as a source.
Values as a table
| Position | Google AI Overview | ChatGPT | Perplexity |
|---|---|---|---|
| 1 | 68% | 30% | 36% |
| 2 | 48% | 18% | 27% |
| 3 | 33% | 12% | 23% |
| 4 | 20% | 10% | 23% |
| 5 | 15% | 7% | 19% |
| 6 | 14% | 4% | 16% |
| 7 | 13% | 6% | 13% |
| 8 | 7% | 3% | 15% |
| 9 | 8% | 3% | 16% |
In the Google AI Overview about 81% of linked sources come from the top 10 of the same search, and about 87% of AI Overviews link at least one website from positions 1 to 3. In ChatGPT about 40% of sources come from the Google top 10, in Perplexity only about 23%. Perplexity draws on many sources that are not on Google’s first page. Google describes which content its AI features use in its official documentation for site owners.
Other measurements find lower values for the AI Overview. Ahrefs found in early 2026 that only about 38% of cited pages rank in the top 10 of the same search, down from 76% in 2025. We count at website level, Ahrefs at page level, and our queries are German-language and at least three words long. The relationship itself holds: in the largest pre-registered study so far, a page in positions 1 to 3 has about a 7.8 times higher chance of being cited than one in positions 11 to 30 (Lee 2026).
What other studies show
Previous studies mostly use self-written prompts and measure in the US. Our measurement uses real search queries from Austria and separates the Google AI Overview from Gemini.
- BrightEdge (2025): tens of thousands of purpose-written prompts from the US. AI Overviews name the most brands per answer on average, and the services recommend different brands for about 62% of questions.
- DEJAN: a fixed prompt that explicitly asks for brands, up to 100 repeats. Stresses that single answers vary.
- Atziry (this analysis): 873 real queries from Austria, asked unchanged, in four services. Measures whether an AI names providers on its own.
Our result does not contradict BrightEdge. There, only answers that appeared were counted. We also count the many searches where Google shows no AI Overview at all. When it is shown, it often names providers in our data too.
How to test your brand’s AI visibility
Distinguish between all relevant queries and those where the AI recommends specific providers. The first value shows how often you appear overall, the second how often you are included when recommendations happen.
- 01Pick real questionsUse your customers’ queries from Search Console instead of invented prompts, and mark those for which an AI names providers. The other questions stay in the overall rate because you can be cited as a source there.
- 02Measure each service separatelyChatGPT, Gemini, Perplexity and the Google AI Overview behave differently. A good position in one says little about the others.
- 03Measure several timesAsk every question several times and evaluate the share in which you are named. A single measurement is too uncertain, especially for the Google AI Overview.
- 04Track three metrics per serviceShare of questions naming you as a provider, share with your website as a linked source, and both again only for questions where the AI names providers or sources at all.
- 05Build your Google rankingFor many provider searches Google shows no AI Overview, so regular search decides. In the AI Overview and ChatGPT, websites in positions 1 to 3 are cited about 3.8 times and 3.7 times as often. For Perplexity, ranking alone is not enough.
How we measured
We entered real search queries unchanged into four services and had an AI model classify the answers. A second, independent AI rating serves as a cross-check.
- Queries: random selection of real queries from Google Search Console, asked by people, focus on Austria, three words or more. About 220 queries per question type; overall values are weighted by how common each question type actually is.
- Privacy: only queries that occur on at least three different websites, without company, brand or personal names. We do not name a single query from the data, nor topics or industries.
- Services: ChatGPT and Gemini via the public websites without login, the AI Overview in real Google Search with location Austria, Perplexity via the API with a neutral instruction (German, Austria, cite sources). No input hints at brands or providers.
- Classification: an AI model judged every answer: does it name at least one specific business, shop, brand or product as an option? Names that appear only as a source do not count. A second, different AI model rated 200 random answers without knowing the first verdict. 188 of 200 agreed (95% range 90 to 97%). The first rating was clearly wrong in 3 of 200 answers. This is not a human review.
- Weighting: the weights are the shares of question types among all 20,678 queries the sample was drawn from. Counting every query equally, ChatGPT names providers for 51.0% instead of 52.6%, Gemini 38.3 instead of 36.8%, Perplexity 34.1 instead of 35.6% and the AI Overview 29.7 instead of 26.1%. The order stays the same.
- Match with Google: for every query the first ten regular Google results at the same moment, compared at website level with the source links of the AI answer. For ChatGPT only answers with links count.
- Statistics: ranges as 95% bootstrap over queries, per question type as Wilson intervals. Differences between services tested in pairs.
| Question type | Weight | Queries in the sample |
|---|---|---|
| Find a provider or product | 44.1% | 226 |
| Price and cost | 7.4% | 229 |
| Comparison and choice | 3.6% | 214 |
| Knowledge and how-to | 44.9% | 204 |
| Total | 100% | 873 |
Data download
All figures of this analysis are available as CSV files.
ki-anbieter-je-dienst.csvPer service and question typeShare of answers with specific providers and share with an AI answer, per service and question type, with 95% range.ki-anbieter-je-frageart.csvCited as a source per Google positionShare of websites in positions 1 to 9 of regular Google results that the AI answer links as a source, per service, with 95% range.ki-zitiert-je-google-platz.csvDistribution per websiteShare of a website’s own queries that name providers, as a distribution across 31 websites (10th, 25th, 50th, 75th and 90th percentile). No values of single websites.ki-anbieter-je-website-verteilung.csvRepeat measurementShare of queries where three runs decided the same way whether providers were named.ki-anbieter-wiederholung.csvLicence: CC BY 4.0. You may use the data freely, including commercially. Please name “Atziry” as the source with a link to this page. The files contain aggregated shares only, no queries and no answer texts.
