How does Google's AI find its information? AI Mode tested from Vienna
On two evenings in October 2026 we asked Google's AI Mode twelve everyday Vienna questions, each one ten times, and recorded every source. We also read what Google, court records and tool vendors say about the AI in Google Search. On questions about businesses, AI Mode relies mainly on Google Business Profiles.
🇩🇪 Diese Studie auf Deutsch lesenAsk Google for a dentist, a restaurant or an estate agent in Vienna and AI Mode no longer gives you a list of links. It gives you an answer with names. For businesses, what matters is where those names come from. So we analysed 119 AI Mode answers: who gets named, how often, and which sources Google shows alongside.
On questions about businesses, AI Mode shows mainly Google Business Profiles as sources, and only those when the question names a place, while the businesses it names change by half from one answer to the next.
Key findings
- With a place in the question, only Google profiles. On the local questions that name a place (restaurant, dentist, vet), 154 of the 154 sources AI Mode showed were Google Business Profiles (24 answers with captured sources). On the buying and specialist questions (tax advisor, dental implant, estate agent) it was 82% of 194 sources, on local questions without a place 73% of 192.
- No place given, Google still answers for Vienna. All 30 answers to the questions about an emergency vet, a place to eat tonight and a plumber named Vienna businesses or districts.
- Information questions name no businesses. On dental crown costs, estate agent commission and registering a trade, Google named not a single business, and no Google profile appeared as a source. Of 267 sources in 30 answers, 36% came from authorities and chambers, 33% from business websites (their advice pages), 11% from YouTube and social media (only on the trade registration question), 10% from media and guides and 9% from directories.
- No Reddit. Reddit appeared nowhere: not as a source in any of the 101 AI Mode answers with captured sources or the 24 AI Overviews shown, and not in the text of any of the 119 answers.
- About half changes. Two answers to the same question shared 47 to 48% of the named businesses on average (47.1% on day one, 48.3% on day two, 47.7% across both days), almost exactly what we measured in August (47.4%). Of 120 named businesses (questions 1 to 9), 27 appeared in at least 8 of 10 answers and 47 only once. There is a stable core: for the emergency vet, 5 businesses in 10 of 10 answers; for estate agents, 4 firms in 9 of 9; for tax advisors, 2 firms in 10 of 10. On the vet practice question, by contrast, no practice appeared in more than 5 of 10 answers.
- AI Overviews read more widely. In our logged-in measurement, the AI Overview appeared for all 12 questions on both days. Its sources: business websites 33%, Google profiles 24%, media and guides 15%, directories 14%, authorities 11%, social media 4%.
- “Deep Search” on two questions only. On the restaurant question in 9 of 10 answers, on the dental implant question in 10 of 10, never otherwise.
- Same question, different AI, different names. For restaurants, Google (10 runs, 21 businesses) and ChatGPT (3 runs, 16 businesses, from our sister study) had 8 businesses in common; for tax advisors only 2 out of 10 each.
How Google's AI works
As of 13 October 2026
Before measuring, we summarise what is known about Google's AI in Search. We read what Google itself writes, what is in the records of the US antitrust case and what vendors of SEO tools have measured. Two features of Google Search are covered: AI Mode and AI Overviews. The Gemini app is a separate product and not part of this study.
1. Which models run
According to Google, AI Mode has run worldwide on Gemini 3.5 Flash as its default model since 19 May 2026.1 People who pay for Google AI Pro or Ultra can choose newer models, since 2 September 2026 for example Gemini 3.8 Flash.2 For AI Overviews, Google last named a model on 27 January 2026: Gemini 3.3 There has been no newer statement since.
According to Google, an automatic router sits above these models. Hard questions go to the strongest model, simple ones to faster models.4 According to Google, AI Mode and AI Overviews may also use different models and therefore show different answers and links.5 Since 13 October 2026, paying subscribers (Google AI Pro and Ultra) can also pick the new Gemini 4 Argon model in AI Mode.6 Our test account is a free one, so our measurement did not run on Argon.
Free AI Mode therefore very likely runs on Gemini 3.5 Flash. We infer this from Google's statements; it is not confirmed, and AI Mode does not display a model. Google has named five models for AI Mode in less than twelve months. Measurements therefore age quickly.
2. How one question becomes about a dozen searches
According to Google, a Gemini model breaks the question into subtopics and sends several related search queries at the same time. Google calls this “query fan-out”.7 According to Dounia Berrada, engineering director for Google Search, AI Mode runs about a dozen searches for one answer in the time it takes you to do one.8
Google says it takes the pages from the normal search index, with the same ranking, quality and spam systems as classic search. Google's own data is added, for example on products and local businesses.7 While the model is still writing the answer, according to Google it looks for further pages that support individual statements and links them.5
Google does not disclose which sub-questions are asked. For businesses this means a page can be cited because it answers a sub-question nobody typed in that way. According to Google, an AI Overview only appears when it adds something to normal search, which is often not the case.5
3. Full Google Search or FastSearch?
According to court records (US antitrust case), Google's head of Search, Liz Reid, testified under oath in spring 2025: to feed its Gemini models with web results, Google uses its own technology called FastSearch. It is faster than Search because it retrieves fewer documents, but of lower quality than fully ranked results. According to the court, Google does not use FastSearch for the normal results page.9
For AI Mode and AI Overviews the records say nothing of the kind. Google writes that both retrieve their pages from the search index using the core ranking systems.7 Reports that AI Overviews run on FastSearch go beyond the court text, which only refers to Gemini models.10
Our reading: AI Mode and AI Overviews rely on full Google Search. That is not certain. The statements date from before Gemini 3, and Google may have rebuilt things since. If the reading is right, pages that rank well in normal search are part of the pool AI Mode draws from. That alone does not get them cited.
4. Which sources get cited
A good position in normal search helps but is not enough. According to an analysis by Ahrefs (an SEO tool vendor) from March 2026, 37.9% of links in AI Overviews came from the top 10 of the same search, 31.2% from positions 11 to 100 and 31.0% from outside the top 100.11 In July 2025, according to Ahrefs, it was still about 76%. Because Ahrefs changed its method in between, only the direction is certain.
In AI Mode the overlap is even smaller. According to SE Ranking (an SEO tool vendor), about 14% of cited links matched the top 10 in June 2025.12 According to seoClarity (an SEO tool vendor), a result in position 1 is cited in AI Mode only about a quarter of the time, measured in September 2025 on US buying queries.13 The studies count differently and are not directly comparable.
On local questions the Google Business Profile plays a big role. According to Google, business profiles can help you appear in AI answers.7 According to BrightLocal (a vendor of local SEO tools), on local questions in the US, the UK and Australia about 53% of sources in AI Mode were business profiles, in AI Overviews about 67%.14 In the same study the share varies by industry: about 59% for vets, about 57% for dentists and 42% for lawyers, where specialist directories come into play more often.
Google also cites itself more and more. According to SE Ranking, google.com's share of all sources in AI Mode rose from 5.7% in June 2025 to 17.42% in February 2026, more than the next six domains combined. Of these links, 59% led to Google results pages and 36.1% to business profiles.15
Nobody has measured this for Austria so far. For Germany there are only rough figures: according to Sistrix (an SEO tool vendor), YouTube was linked in about 40% of AI Mode answers in October 2025, with a Google service in second place. Local questions were not included.16 Our own measurement below provides first figures for twelve Vienna questions.
5. What is available in Austria and what is not
AI Overviews have been available in Austria since the end of March 2025, in German and English. According to Google's launch announcement, they are shown only to signed-in users aged 18 and over.17 We found no newer statement on this. How often they appear depends on the question: according to Sistrix for just over 20% of German search terms,18 according to Ahrefs for only 7.9% of local searches in the US.19 According to Google, AI Mode has been available in German since 8 October 2025.20
There are no ads inside the AI answers here. According to Google's help page for advertisers, ads within AI Overviews run in only 12 countries and only in English, and Austria is not among them.21 Ads above and below the overview do run here. We found no sign of ads in AI Mode in Austria.
Some US features are missing entirely. “Ask Maps”, asking Google Maps questions with Gemini, has been available since August 2026 in over 150 countries according to the tech site 9to5Google, but only in English, and Austria is not named.22 According to Google, calls and bookings through AI Mode are offered only in the US.1 US guides that rely on these features do not apply in Vienna yet.
6. EU rules and opt-out
On 23 July 2026 the European Commission fined Google 460 million euros for favouring its own services in Search (Digital Markets Act, DMA). The Commission writes that it takes note of Google's proposals to apply the principles of the decision to AI Overviews and AI Mode as well.23 On 28 September 2026 Google filed actions concerning the DMA at the EU General Court.24 Separately, since 9 December 2025 the Commission has been examining whether Google uses content from publishers and YouTube for AI answers without paying and without a real option to refuse.25 No decision has been made yet.
According to Google, since 31 August 2026 every website worldwide has its own switch in Google Search Console: “Search generative AI control”.26 The default is “Include”. According to Google, a site set to “Exclude” disappears within one to two days from AI Overviews, AI Mode and the AI features in Discover. Normal ranking is not supposed to be affected. The switch applies to the whole property, not to individual pages.27 For almost every business, “Exclude” is the wrong choice.
7. What website owners can control, and which myths are wrong
According to Google, AI answers need no special optimisation; it is still SEO.7 What you can control is mainly whether a page is eligible at all. According to Google, that means:
- Googlebot may read the site, in robots.txt and at the hosting or CDN provider.7 Careful with Cloudflare: according to Cloudflare, since 15 September 2026 the “Block” setting for AI training also locks out Googlebot, and with it all of Google Search.28
- The page is indexed and may appear with a snippet. According to Google, “nosnippet” prevents content from flowing directly into AI Overviews and AI Mode.29
- The Search Console switch is set to “Include”.27
- Business Profile and Merchant Center are up to date. Google names both for local businesses and products in AI answers.7 Whatever is outdated there can end up in the answer.
After that, what counts is whether a page ranks well for the sub-questions and has paragraphs that clearly support a statement. Since 15 May 2026, Google's spam policies explicitly cover attempts to manipulate AI answers.30 According to Google, fake mentions across the web help less than people think.7
These myths are wrong:
- “Schema markup brings AI citations.” According to Google, structured data is not required for AI answers.7 Ahrefs (an SEO tool vendor) compared pages that newly added JSON-LD with similar pages without it. Afterwards they were cited 4.6% less often in AI Overviews and 2.4% more often in AI Mode, within the range of chance.31 Google has not shown FAQ snippets since 7 May 2026.32
- “An llms.txt file helps with Google.” According to Google, Google Search does not use llms.txt or other AI text files. They neither hurt nor help.32
- “Blocking Google-Extended takes you out of the AI answers in Search.” Google-Extended is an entry in robots.txt. According to Google, it has no effect on whether a website appears in Google Search.33 According to court records (US antitrust case), publishers cannot use it to remove themselves from AI Overviews either.9 Anyone who wants out needs the Search Console switch or “nosnippet”.
- “Google Ads get you into the AI answer.” Robby Stein, vice president of product for Google Search, said in an interview that paid ads do not influence recommendations in AI Mode.34
The 12 questions, word for word
We asked four kinds of questions, three each: local questions with a place, buying and specialist questions, local questions without a place, and pure information questions. We asked them in German, exactly as shown, including casual phrasing and lower case, the way people actually type. The English line below each one is our translation for this page. We had asked Google five of them word for word before (see Before and after), and two also went to ChatGPT in our sister study.36
Local questions with a place
Buying and specialist questions
Local questions without a place
Information questions
Method
Why signed in? We had planned to measure without signing in. In the first browser, however, Google showed a security check (CAPTCHA) on the very first request. We do not solve or bypass such checks. So we measured in our normal Chrome, signed in with our own Google account. This has two consequences: Google can personalise answers through the account's search history, and we see AI Overviews, which Google shows in Austria only to signed-in users.17
Procedure. For each of the 10 rounds we drew the order of the 12 questions in advance, so that no question always ran at the same time. Each question ran on its own in one tab, never several at once. After each answer, a small script in the browser saved the answer text and all source cards (domain and path, without parameters) to a file. This replaces the screenshots we had originally planned and has one advantage: every source is available as text. Source cards often appear seconds after the text, with “Deep Search” up to half a minute later. For day two we therefore extended the waiting time to up to 35 seconds. Even so, source cards are missing for 18 answers (8 on day one, 10 on day two). Their names count; they are left out of the source mix.
Replacement runs. When Google gave no valid answer, we logged it, did not count it and made a replacement run the same evening. This only happened on day two: 15 captures produced no valid answer, mostly Google's error page “Something went wrong and an AI response wasn't generated”, once a “Serverfehler” (server error) page and once an answer that broke off. We replaced 11 runs this way. On the vet practice and the estate agent questions, one run each needed a second replacement. For run 10 of the estate agent question, the replacement also returned only the error page, so this question has 9 runs instead of 10. After the many error pages in rounds 6 and 7 we waited 60 seconds between questions, after which they became rarer.
What counts as “named”. A business counts as named if it appears in the answer text. Businesses that only appear as a place card in the source list are counted separately and never reported as a recommendation. Chambers, authorities and social insurers (such as WKO, ÖGK, the municipal office) do not count as businesses. We merged different spellings of the same business, for example “Steirereck im Stadtpark” and “Steirereck”. Unclear cases were kept separate.
Source types. We assigned every source card to one of seven types, by hand and per domain:
- Google Business Profile: a card pointing to a profile in Google Maps
- Business website: sites of practices, firms, estate agents, restaurants and plumbers, including their advice pages
- Directory or portal: business directories, city and referral portals, such as firmenabc.at, ganz-wien.at or immoverkauf24.at
- Media, guides, advice: newsrooms, restaurant guides and advice pages of companies that do not offer the service asked about, such as zahnarzt-kosten.at or infina.at
- Reddit, YouTube, social media
- Authority, chamber, insurance: such as wko.at, gisa.gv.at or oesterreich.gv.at
- Other (did not occur)
The source mix is each type's share of all source cards in a question group.
Stability. We measure how similar two answers to the same question are with the Jaccard value: the number of shared businesses divided by the number of all businesses in the two answers. 100% means an identical list, 0% no overlap. We calculate it for every pair of answers within a question and take the mean, separately for pairs on day one, on day two and between the days. Information questions are excluded because they contain no businesses.
Ranges. For every mention rate (“in 8 of 10 answers”) we give the 95% Wilson range. It shows how uncertain a rate from 10 runs is: 8 of 10 is consistent with a true rate between 49% and 94%.
Location test. For the three questions without a place, an answer counts as Vienna-related if it names at least one Vienna business or one Vienna district.
AI Overview check. On both days we also entered each question once into normal Google Search (https://www.google.com/search?hl=de&q= plus the question) and recorded whether an AI Overview appeared and which sources it showed. Twice it only appeared in the replacement run.
Results
1. Where the answers come from: sources by question group
Next to each answer, AI Mode shows source cards. Depending on the kind of question, they look completely different.
| Group | Answers | Sources | Google profile | Business website | Directory, portal | Media, guides, advice | Reddit, YouTube, social | Authority, chamber |
|---|---|---|---|---|---|---|---|---|
| Local with place | 24 | 154 | 100% | 0% | 0% | 0% | 0% | 0% |
| Buying and specialist | 25 | 194 | 82% | 13% | 5% | 0% | 0% | 0% |
| Local without place | 22 | 192 | 73% | 20% | 3% | 4% | 1% | 0% |
| Information | 30 | 267 | 0% | 33% | 9% | 10% | 11% | 36% |
| All questions | 101 | 807 | 56% | 19% | 5% | 4% | 4% | 12% |
Share of all source cards in the group, answers with captured sources only. “Other” did not occur. Rounded, so totals may differ by 1.
On the questions with a place, every single source was a Google profile, namely the profiles of the businesses in the answer. On the dental implant and estate agent questions, Google likewise showed only profiles. For tax advisors, firm websites and the directory firmenabc.at were added. The plumber question was the reverse: there Google showed Google profiles in only one of six answers with sources, otherwise plumbers' websites, the city portal ganz-wien.at and advice pages.
This matches the direction of the research. According to BrightLocal, about 53% of sources in AI Mode were business profiles on local questions in the US, the UK and Australia.14 On our Vienna questions with a place it was 100%. The methods differ, so the figures are not directly comparable. ChatGPT also takes its business data from Google Business Profiles, as our sister study shows.36
2. How many businesses Google names
| Question | Answers | Businesses total | Avg per answer | in 8+ of 10 | only once | Deep Search | Follow-up |
|---|---|---|---|---|---|---|---|
| 1 Restaurants | 10 | 21 | 6.5 | 3 | 7 | 9 | 5 |
| 2 Dentist | 10 | 10 | 4.7 | 3 | 4 | 0 | 0 |
| 3 Vet practice | 10 | 13 | 3.8 | 0 | 4 | 0 | 6 |
| 4 Tax advisor | 10 | 10 | 5.5 | 4 | 2 | 0 | 9 |
| 5 Dental implant | 10 | 15 | 4.7 | 3 | 9 | 10 | 0 |
| 6 Estate agents | 9 | 11 | 6.9 | 5 | 3 | 0 | 4 |
| 7 Emergency vet | 10 | 10 | 6.8 | 6 | 2 | 0 | 0 |
| 8 Eating out | 10 | 20 | 4.3 | 2 | 13 | 0 | 0 |
| 9 Plumber | 10 | 10 | 4.5 | 1 | 3 | 0 | 6 |
| 10 Dental crown | 10 | 0 | 0 | 0 | 0 | 0 | 0 |
| 11 Agent commission | 10 | 0 | 0 | 0 | 0 | 0 | 0 |
| 12 Registering a trade | 10 | 0 | 0 | 0 | 0 | 0 | 0 |
Businesses = different businesses named in the answer text. For question 6, “8+” means at least 8 of 9. Deep Search and follow-up: number of answers.
Depending on the question, Google named 3.8 to 6.9 businesses per answer on average. Many answers ended with a follow-up question, for example whether you are looking for a tax advisor as a private person, a founder or a company, which pet it is about, or whether the water is still draining slowly. For tax advisors it came in 9 of 10 answers.
3. Who gets named
The tables show up to five of the most frequently named businesses per question. A mention reflects what Google answered. It is not a recommendation or rating by us. All names and runs are in the CSV.
Question 1 · Restaurants 10 answers, 21 businesses
| Business | named in | 95% range |
|---|---|---|
| Steirereck | 9 of 10 | 60% to 98% |
| Pramerl and the Wolf | 8 of 10 | 49% to 94% |
| SIXTA | 8 of 10 | 49% to 94% |
| Mraz & Sohn | 5 of 10 | 24% to 76% |
| Apron | 5 of 10 | 24% to 76% |
Question 2 · Dentist 10 answers, 10 businesses
| Business | named in | 95% range |
|---|---|---|
| Dr. Azzawi | 10 of 10 | 72% to 100% |
| DDr. Stefanie Weber | 9 of 10 | 60% to 98% |
| Dr. Lukas Hallmann | 8 of 10 | 49% to 94% |
| Dental Performance | 7 of 10 | 40% to 89% |
| Dr. Mathias Lidauer | 5 of 10 | 24% to 76% |
Question 3 · Vet practice 10 answers, 13 businesses
| Business | named in | 95% range |
|---|---|---|
| Dr. Florian Teufel | 5 of 10 | 24% to 76% |
| Fachtierärzte Althangrund | 5 of 10 | 24% to 76% |
| Vetklinikum | 5 of 10 | 24% to 76% |
| Tierarztpraxis am Floridsdorfer Markt | 4 of 10 | 17% to 69% |
| Tierarztpraxis Schwarzlackenau | 4 of 10 | 17% to 69% |
Question 4 · Tax advisor 1010 10 answers, 10 businesses
| Business | named in | 95% range |
|---|---|---|
| Steuerberatung Weiß | 10 of 10 | 72% to 100% |
| Michaela Wiesner Steuerberatung | 10 of 10 | 72% to 100% |
| G & W Steuerberatung | 9 of 10 | 60% to 98% |
| RSB Austrian Tax Partner | 8 of 10 | 49% to 94% |
| steuerexperten.at | 7 of 10 | 40% to 89% |
Question 5 · Dental implant 10 answers, 15 businesses
| Business | named in | 95% range |
|---|---|---|
| Dr. Dr. Felix Wick | 9 of 10 | 60% to 98% |
| Dr. Robert Deak | 8 of 10 | 49% to 94% |
| Dr. Azzawi | 8 of 10 | 49% to 94% |
| DDr. Gabor Tepper | 7 of 10 | 40% to 89% |
| DDr. Stefanie Weber | 4 of 10 | 17% to 69% |
Question 6 · Estate agents 9 answers, 11 businesses
| Business | named in | 95% range |
|---|---|---|
| Kaltenegger Realitäten | 9 of 9 | 70% to 100% |
| OPTIN Immobilien | 9 of 9 | 70% to 100% |
| BREZINA Immobilien | 9 of 9 | 70% to 100% |
| 4M Immobilien | 9 of 9 | 70% to 100% |
| ZELLMANN Immobilien | 8 of 9 | 56% to 98% |
Question 7 · Emergency vet 10 answers, 10 businesses
| Business | named in | 95% range |
|---|---|---|
| Universitätsklinik für Kleintiere | 10 of 10 | 72% to 100% |
| AniCura Erdberg | 10 of 10 | 72% to 100% |
| Tiernotarzt Wien | 10 of 10 | 72% to 100% |
| Tierklinik Hutter | 10 of 10 | 72% to 100% |
| Vetklinikum | 10 of 10 | 72% to 100% |
Question 8 · Eating out 10 answers, 20 businesses
| Business | named in | 95% range |
|---|---|---|
| EF16 | 8 of 10 | 49% to 94% |
| SIXTA | 8 of 10 | 49% to 94% |
| DAS LOFT | 5 of 10 | 24% to 76% |
| Verde | 3 of 10 | 11% to 60% |
Next come three venues with 2 of 10 each (Motto am Fluss, Pürstner, Maria und Josef), so only four rows here.
Question 9 · Plumber 10 answers, 10 businesses
| Business | named in | 95% range |
|---|---|---|
| G. Weiner | 8 of 10 | 49% to 94% |
| A-S-G Haustechnik | 7 of 10 | 40% to 89% |
| B-GAS Installateur | 7 of 10 | 40% to 89% |
| RohrMax | 6 of 10 | 31% to 83% |
| BM Installateur | 6 of 10 | 31% to 83% |
Questions 10 to 12, information questions: no business named. Here Google named only institutions: for the dental crown, for example, the health insurers ÖGK, SVS and BVAEB and the Chamber of Dentists; for registering a trade, WKO, GISA, the business service portal, the municipal district office, SVS and the tax office. On the commission question, 7 of 10 answers contained no name at all. The sources, on the other hand, were very stable: many sites appeared in every single answer, and for the dental crown and commission questions it was the same subpage each time.
| Question | Sources in 10 of 10 answers | Other frequent sources |
|---|---|---|
| 10 Dental crown | drkoban.at (dental practice, cost page for crowns), zahnarzt-kosten.at (information portal), zahnarzt-payerbach.at (dental practice in Lower Austria, advice page), ganz-wien.at (city portal) | ident.wien (9 of 10), ergo-versicherung.at and zahnregulierung-klagenfurt.at (7 of 10 each) |
| 11 Agent commission | immoverkauf24.at (referral portal), infina.at (mortgage broker, advice page), 4m-immo.at (estate agent, advice page), advokat-wien.at (law firm, page on purchase costs) | oesterreich.gv.at (9 of 10), vires.at and zp-immo.at (6 of 10 each) |
| 12 Registering a trade | wko.at, gisa.gv.at, usp.gv.at, ams.at, steuerexperten.at (tax firm), youtube.com | bmwet.gv.at (8 of 10), wien.gv.at (7 of 10), facebook.com (5 of 10) |
An observation, not an explanation: 4M Immobilien, named in 9 of 9 answers to the estate agent question, also appeared as a source with an advice page on agent costs in 10 of 10 answers to the commission question. Whether one helps the other is something our measurement cannot tell.
4. Location test: questions without a place
| Question | Answers related to Vienna |
|---|---|
| 7 „Tierarzt Notdienst heute“ | 10 of 10 |
| 8 „Wo kann ich heute Abend gut essen gehen?“ | 10 of 10 |
| 9 „Ich brauche dringend einen Installateur, mein Abfluss ist verstopft.“ | 10 of 10 |
Type just “Tierarzt Notdienst heute” (emergency vet today) in Vienna and you get Vienna emergency services. We cannot see how Google determines the location; the footer showed “1210 Wien”. For the emergency vet, besides Vienna clinics Google also named businesses in the surrounding area, such as Tierklinik Perchtoldsdorf Ost (6 of 10) or Tierklinik Wiener Neustadt (2 of 10). For businesses this means: questions without a place are local questions too.
5. How stable the answers are
| Comparison | Businesses | Pairs | Sources | Pairs |
|---|---|---|---|---|
| within day 1 | 47.1% | 90 | 55.2% | 93 |
| within day 2 | 48.3% | 86 | 59.8% | 79 |
| between the days | 47.7% | 220 | 55.6% | 214 |
Mean Jaccard value across all pairs of answers to the same question. Businesses: questions 1 to 9. Sources: only pairs in which both answers have captured sources.
Two answers to the same question share just under half of their businesses on average. Between the days the value is hardly lower than within one day. So the change is not an effect from one evening to the next; it happens from answer to answer. In our August measurement, when we asked AI Mode the same question twice a few minutes apart, the value was 47.4%.37
For sources, we count each website once and each Google profile separately, by name. Two restaurant answers with different profiles therefore share no source, even if both show only profiles. Sources are somewhat more stable than names: 55 to 60%.
6. AI Overviews
In normal Google Search, the AI Overview appeared for all 12 questions on both days, twice only in the replacement run. We were signed in, which according to Google is a requirement in Austria.17
| Question | Day 1 | Day 2 | Sources day 1 | Sources day 2 | of which Google profile |
|---|---|---|---|---|---|
| 1 Restaurants | yes | yes | 14 | 13 | 10 |
| 2 Dentist | yes | yes | 3 | 5 | 3 |
| 3 Vet practice | yes | yes | 3 | 4 | 3 |
| 4 Tax advisor | yes | yes | 8 | 9 | 8 |
| 5 Dental implant | yes | yes | 11 | 11 | 6 |
| 6 Estate agents | yes | yes | 10 | 8 | 7 |
| 7 Emergency vet | yes | yes | 11 | 8 | 7 |
| 8 Eating out | yes | yes | 7 | 6 | 4 |
| 9 Plumber | yes | yes | 6 | 6 | 0 |
| 10 Dental crown | yes | yes | 10 | 14 | 0 |
| 11 Agent commission | yes | yes | 5 | 7 | 0 |
| 12 Registering a trade | yes | yes | 10 | 10 | 0 |
“of which Google profile”: both days combined.
AI Overviews read more widely than AI Mode. Across all questions, 33% of their sources came from business websites, 24% from Google profiles, 15% from media and guides, 14% from directories, 11% from authorities and 4% from social media. For restaurants, the overview mixed Google profiles with Falstaff, Falter, Gault&Millau, the Michelin Guide and wien.info, among others. On the plumber question and the three information questions, it showed not a single Google profile.
7. The most frequent sources
| Source | in answers | mainly on |
|---|---|---|
| Google Business Profiles | 66 | questions 1 to 8, only once on question 9 |
| steuerexperten.at | 17 | registering a trade (10), tax advisor (7) |
| ganz-wien.at | 15 | dental crown (10), plumber (5) |
| youtube.com | 11 | registering a trade (10), plumber (1) |
| 11 sources with 10 each | 10 | Dental crown: zahnarzt-payerbach.at, zahnarzt-kosten.at, drkoban.at. Commission: immoverkauf24.at, infina.at, 4m-immo.at, advokat-wien.at. Trade: ams.at, gisa.gv.at, wko.at, usp.gv.at |
Number of AI Mode answers in which the source appeared at least once (out of 101 answers with captured sources). All Google profiles count as one row here. In 5th place, 11 sources share the value 10, hence one combined row.
8. Same questions, different AI
We asked ChatGPT two of our questions in the same week, word for word.36 For restaurants, Google named 21 businesses in 10 runs, ChatGPT 16 in 3 runs. 8 were shared: Amador, Figlmüller, Mraz & Sohn, Pramerl and the Wolf, Schachtelwirt, Skopik & Lohn, Steirereck and Wiener Wiazhaus. For tax advisors, both named 10 firms each, with only 2 in common: RSB Austrian Tax Partner and Steuerberatung Weiß. Because ChatGPT was asked only three times, this is a rough comparison. But it shows that appearing in one AI does not mean appearing in the other.
Before and after
We had asked Google five of the twelve questions word for word before, in June, July, August and September 2026. We check how often the businesses named back then are named today.
| Question | back then | named back then, named today in | today at least once | today 5+ of 10 |
|---|---|---|---|---|
| 1 Restaurants | 10 June 2026 | Steirereck 9 of 10, SIXTA 8, Silvio Nickol 2, Figlmüller 1, Glasswing 0 | 4 of 5 | 2 |
| 1 Restaurants | 4 September 2026 | Steirereck 9 of 10, Apron 5, Mraz & Sohn 5, Rote Bar 4, Griechenbeisl 3, JOLA 2, Silvio Nickol 2, Glasswing 0, Konstantin Filippou 0, Zum weißen Rauchfangkehrer 0 | 7 of 10 | 3 |
| 2 Dentist | 18 August 2026 | Dr. Azzawi 10 of 10, Dr. Mathias Lidauer 5, Dr. Zgkouri 1 | 3 of 3 | 2 |
| 3 Vet practice | 13 June 2026 | Tierarztpraxis am Floridsdorfer Markt 4 of 10, Dr. Tune Lazri 1, Tierarztpraxis am Stadtpark 1, Tierklinik Dr. Hutter 1 | 4 of 4 | 0 |
| 5 Dental implant | 18 August 2026 | Dr. Dr. Felix Wick 9 of 10, Dr. Robert Deak 8, Dr. Azzawi 8, Universitätszahnklinik Wien 1, myZahn 0 | 4 of 5 | 3 |
| 6 Estate agents | 21 July 2026 | EHL Immobilien 1 of 9; 3SI Immogroup, ANOBIS Immobilien, Dr. Funk Immobilien, Fermoso Immobilientreuhand, Immototal Immobilientreuhand, KLASAN & Partner and OTTO Immobilien 0 each | 1 of 8 | 0 |
| 6 Estate agents | 5 August 2026 (half captured) | 4M Immobilien 9 of 9, Kaltenegger Realitäten 9, OPTIN Immobilien 9, Börner Ihr Hausmakler 5; ANOBIS Immobilien, Dr. Funk Immobilien and 3SI Makler 0 each | 4 of 7 | 4 |
For question 6, “5+” means at least 5 of 9. Sources of the before data: see sources 38 to 41.38394041
For restaurants, dentist and dental implant, the top names from back then often appear today too, such as Steirereck, Dr. Azzawi or Dr. Dr. Felix Wick. For the vet practice, all four names from June still appear today, but none in more than 4 of 10 answers. The change is clearest for estate agents: of the July list of 8 firms, only one came back today, in 1 of 9 answers. From the August list, on the other hand, three firms now appear in every answer (4M Immobilien, Kaltenegger Realitäten, OPTIN Immobilien) and a fourth in 5 of 9.
What this means for businesses
Our measurement shows what Google displays, not why. No guarantees follow from it, but a few clear pointers:
- Local businesses: the Google Business Profile is the source AI Mode shows. On all questions with a place, it showed profiles only. What we measured is that the profile gets cited, not which details in it make the difference. According to Google, the profile is part of the data for AI answers, and whatever is outdated there can end up in the answer.7
- Buying and information questions: your own pages that answer exactly that question. On the dental crown question, Google cited the cost page of a dental practice (drkoban.at) and the crown page of the information portal zahnarzt-kosten.at in 10 of 10 answers. On the commission question, a law firm's page on purchase costs (advokat-wien.at) and a mortgage broker's advice page (infina.at) appeared in 10 of 10 answers. What they have in common is a subpage on exactly this question.
- No Reddit strategy needed, at least not for Vienna questions like these on Google. Reddit did not appear in our measurement.
- Never judge from a single query. About half of the names change from answer to answer, even on the same evening. A single test only shows a chance result.
- Check the Search Console switch. “Search generative AI control” should be set to “Include”; otherwise, according to Google, the website disappears from AI Mode and AI Overviews.27
How ChatGPT answers the same questions is shown in our sister study on ChatGPT and GPT-6. How often four AIs name Vienna businesses is what we measure in the KIGEO AI Visibility Index. This approach is called Generative Engine Optimization (GEO), and it is what KIGEO does.
What does Google's AI show about your business?
We check for free whether Google AI Mode, ChatGPT, Perplexity and Gemini name you, with which details, and who gets recommended instead.
get a free AI visibility check →Limits
- A small slice. 12 questions, 2 days in October 2026, Vienna, German, Chrome on desktop, signed in with a single Google account. Google can personalise answers through search history and the account. None of this can be transferred to other questions, places, devices or users.
- 10 runs is a minimum. A rate from 10 answers is uncertain, which is why the ranges are shown. “0 of 10” does not rule out a true rate of up to about 28%. So we never write that Google does not name a business.
- We only see what Google shows. The sub-questions in the background (fan-out) are invisible, and we cannot measure causes. That a page is cited does not mean it moves a business up.
- Model unknown. AI Mode does not display a model. Google may have switched it during the measurement.
- Error pages and replacement runs. On day two, Google gave no valid answer for 15 captures, so 11 runs come from replacement runs. One run is missing for the estate agent question, which has only 9.
- Gaps in the sources. For 18 answers no source cards were captured. They are missing from the source mix and from the source figures on stability.
- Time of day. Day one ran between 23:27 and 00:40. On “Wo kann ich heute Abend gut essen gehen?” (where to eat tonight), Google therefore mentioned opening hours and places still open late. On the second evening the question ran earlier.
- Manual coding. We standardised names and assigned source types by hand. Borderline cases, such as a mortgage broker with advice pages, could also be classified differently.
- Before data. Signed in, a single run each, partly incomplete (see above).
- Outside figures. Most figures in the research part come from US tool vendors and from English searches in the US.
Sources
- Google, blog: A new era for AI Search, Liz Reid, Google I/O (19 May 2026).
- Search Engine Roundtable: Gemini 3.8 Flash in AI Mode for AI Pro and Ultra (2 September 2026).
- Google, blog: Gemini 3 as the default model for AI Overviews (27 January 2026).
- Google, blog: Gemini 3 in Search and AI Mode (November 2025).
- Google Search Central: AI features and your website.
- Search Engine Land: Gemini 4 Argon rolling out in Google AI Mode & Gemini (13 October 2026); plus Google DeepMind: Model card Gemini 4 Argon.
- Google Search Central: AI optimization guide (last updated 10 July 2026).
- Google, blog: Ask a Techspert, Dounia Berrada on visual search in AI Mode (5 March 2026).
- United States v. Google, remedies ruling, Judge Amit Mehta (2 September 2025), findings 44 to 46.
- Search Engine Journal: Google Antitrust Case: AI Overviews Use FastSearch, Not Links.
- Ahrefs: AI Overview citations and the top 10 (March 2026).
- SE Ranking: AI Mode Research (data from June 2025).
- seoClarity: AI Mode and organic rankings overlap (data from September 2025).
- BrightLocal: AI directory sources (August 2026).
- SE Ranking: Google links in AI Mode answers (6 March 2026).
- Sistrix: Top-Quellen im AI Mode Deutschland: Google liebt sich selbst am meisten (top sources in AI Mode in Germany, October 2025, in German).
- Google, blog: We're bringing the helpfulness of AI Overviews to more countries in Europe (March 2025).
- Sistrix: AI Overviews in Deutschland: So stark sinken die Klickraten wirklich (AI Overviews in Germany and click rates, updated 13 October 2026, in German).
- Ahrefs: AI Overview triggers.
- Google, blog Germany: KI-Modus startet in DACH (AI Mode launches in Germany, Austria and Switzerland, 8 October 2025, in German).
- Google Ads Help: About ads and AI Overviews.
- 9to5Google: Google Ask Maps global (6 August 2026).
- European Commission: Press release IP/26/1670, DMA fines against Google (23 July 2026).
- NEMO Europe: Tracker Google DMA Search.
- CNBC: Google hit with EU antitrust probe over use of online content for AI (9 December 2025).
- Google, blog: New controls for website owners (3 June 2026, updated 31 August 2026).
- Google Search Console Help: Search generative AI control.
- Cloudflare, blog: Have it both ways, settings for mixed-use crawlers (15 September 2026).
- Google Search Central: Robots meta tag, data-nosnippet and X-Robots-Tag.
- Google Search Central: Spam policies for Google web search.
- Ahrefs: Schema and AI citations, before-and-after comparison with a control group (May 2026).
- Google Search Central: Updates (including 8 May 2026 on FAQ snippets and 15 June 2026 on llms.txt).
- Google: Common crawlers, section Google-Extended.
- PPC Land: Google reveals how AI decides which businesses get recommended (interview with Robby Stein).
- Data collection and raw data: KIGEO, own measurement of 13 and 14 October 2026, 119 AI Mode answers and 24 AI Overviews with all named businesses and sources, in full in the CSV below.
- Sister study: KIGEO, How does ChatGPT find its information? Before and after GPT-6 (13 October 2026), with the same questions 1 and 4.
- KIGEO, Dieselbe Frage, zehn Minuten später, andere Firmen (same question, ten minutes later, different firms; in German), repeat test of 6 August 2026 (Google AI Mode: 47.4%, rounded to 47% on that page).
- Before data, restaurants: KIGEO restaurant study (10 June 2026) and AI Visibility Index, edition 2 (4 September 2026).
- Before data, dentist and dental implant: KIGEO dentist study (collected 18 August 2026).
- Before data, vets: KIGEO vet study (13 June 2026).
- Before data, estate agents: KIGEO, own recordings of 21 July and 5 August 2026 (signed in, one run each, not published).
Download the raw data of this study as CSV ↓
Note: the CSV is in German.