How does ChatGPT find its information? Before and after GPT-6

Ask ChatGPT for a restaurant or a tax advisor and you get names. For businesses, what matters is where those names come from. So besides the answers, we also read the data ChatGPT sends along in the background: sources, business records and status messages.

The experiment

On the evening of 13 October 2026, we asked ChatGPT two questions from a Vienna browser, each three times in a new chat. Both questions are word for word the same as in our earlier measurements, so before and after can be compared.²³

Engine

ChatGPT, free access without login

Model

GPT-6, according to ChatGPT itself during the test (without login, ChatGPT shows no model name)

Date

13 October 2026, between 22:40 and 23:00 Vienna time

Location

Vienna, Austria (real IP, no VPN)

Runs

6, three per question, each in a freshly loaded chat

Recorded

Answer text, cited sources, business records and status messages from the answer's data stream

Why without login? In a logged-in account, ChatGPT may use its memory even in a “temporary chat”; it says so on screen. An account that has often asked about GEO and Vienna would colour the answers. Without login there is no memory and no history. Ahrefs measures the same way in its large analysis.⁴

Result 1: ChatGPT still searches on every question

All 6 answers went through a web search. ChatGPT reports this itself with a short status line such as “Searched 4 websites”. For restaurants it was 3 to 4 websites, for tax advisors 6 to 8.

That is notable, because overall ChatGPT searches fairly rarely. According to Ahrefs, only about one in seven free answers (14.4%) triggers a web search at all.⁴ When asked for a business in Vienna, ChatGPT searched every time. So the basic process has not changed: search, read, answer.

One thing we can no longer see is the wording of the searches. In June, ChatGPT still showed eight search terms for the tax advisor question, such as “beste Steuerberater 1010 Wien” and “tax advisor Vienna 1010”. Without login, that field in the data stream stays empty.

Result 2: Reddit has disappeared

None of the 6 answers cited Reddit. In June, Reddit was still one of the main sources for the restaurant question.

So the switch came before GPT-6. That fits the measurements of large tool vendors. According to Promptwatch, Reddit's share of ChatGPT citations fell from 3.83% to 0.52% after a change on 8 August 2026.⁵ For the US, Ahrefs counts a drop from about one in five answers with a Reddit link in March to one in 43 at the end of September.⁴ GPT-6 has not reversed this.

Result 3: The business data comes from Google

Behind every business ChatGPT names there is a record with address, phone number, opening hours, stars, number of reviews and website. These records carry clear traces of Google:

• Google Place IDs. Several records carry the ID Google Maps uses for a place. It always starts with ChIJ. That applied to 4 restaurants in one run and 2 tax firms in another.
• Links from the Google profile. The website links contain the tag the business added in its Google Business Profile. Plachutta's link contains utm_medium=gbp (gbp stands for Google Business Profile), and two tax firms' links even say mtm_campaign=Google+Unternehmensprofil (German for Google Business Profile). A business only puts this tag on the link in its Google profile.
• Review counts at Google scale. Figlmüller on Bäckerstraße has 33,689 reviews in its record, Plachutta Wollzeile 14,646.

Where a record names a provider, it says serp, the usual abbreviation for search engine results pages. No other provider appeared in our recordings. Google Maps itself is never cited as a source. Someone reading the answer sees Michelin or the firm's website, but not that address, stars and opening hours come from Google.

The Google Business Profile is ChatGPT's business card, too. What's in it ends up in the answer.

That leaves the question of whether ChatGPT now has its own search index. In September, Peec AI reported that ChatGPT uses its own index, internally called “Labrador”, and also buys Google results from service providers.⁶ OpenAI has not confirmed this. The settings the ChatGPT page sends when it loads do contain a switch called labrador_web_enabled, set to on. So the name exists. How often this index supplies an answer cannot be seen from outside.

Result 4: Own websites up, directories down

The change is clearest on the tax advisor question. In June, 7 of the 12 sources came from two business directories, steuerpedia.at (5) and bezirkstipp.at (2). In October, neither appeared.

ChatGPT named two firms in all three runs, Reschny Birox and Keber & Keber. ChatGPT had already read both websites as sources in June, but did not recommend either firm back then.

This also matches the large datasets. Ahrefs sees ChatGPT searching for official pages more often. The word “official” now appears in 13.1% of English background searches, up from 10.1% in August.⁴

Result 5: The names still wobble

One thing GPT-6 has not changed: who gets named varies from chat to chat.

• Restaurants: 16 different restaurants across three runs, only 3 of them every time, Steirereck, Plachutta Wollzeile and Mochi.
• Tax advisors: 10 different firms, only 2 of them every time, Reschny Birox and Keber & Keber.
• Compared with before: of the 11 restaurants on the September list, 6 came back. Of the 6 firms from June, 3 were there again, Steuerberatung Weiß, RSB Austrian Tax Partner and cSt causa.

The taste has shifted too. In September, ChatGPT named almost only fine dining. In October it mixed starred restaurants with Viennese classics such as Plachutta, Figlmüller, Meissl & Schadn and Huth.

This confirms our own measurement from August: the same AI repeats only 42.8% of its businesses just minutes later.⁷ A single query proves nothing.

New: answers with a map and follow-up questions

What is visibly new is the format. Every one of the 6 answers came with a map of all named businesses and a small clickable form. For restaurants, ChatGPT asks about budget per person, cuisine and occasion. For tax advisors it asks about your situation (employee, self-employed, limited company, property), what matters to you, and language. OpenAI calls this “Intelligent UI” and introduced it with GPT-6.⁸

For businesses this means the next answer depends on these clicks. If you offer advice in English or know property tax well, say so clearly on your website and in your Google profile. That is exactly what ChatGPT now asks about.

What OpenAI itself says about GPT-6

Not much. According to OpenAI, GPT-6 is better at deciding when it needs to search the web and more reliably finds information that actually supports the answer. On questions that need a web search, GPT-6 Instant starts answering 44% sooner on average than its predecessor.⁸ OpenAI does not say how source selection has changed. There are no independent figures on GPT-6 yet; the large Ahrefs analysis of 12 October still ran on GPT-5.6.⁴

We have summarised the full background on models, the search system and crawlers in our guide Where does ChatGPT get its information?

We also asked Google's AI Mode the same two questions, each ten times on two evenings. The results are in our sister study How does Google's AI find its information?

What this means for businesses in Vienna

• Maintain your Google Business Profile. Address, phone, opening hours, website and reviews go straight into ChatGPT's answer, even though Google is never shown as a source.
• Fill your own website with clear facts. Services, district, languages, target groups, prices. On the tax advisor question, it was the most important source.
• Stop counting on Reddit for ChatGPT, as of October 2026. Other AIs may behave differently.
• Do not block the OAI-SearchBot crawler. If you lock it out, ChatGPT cannot read you.
• Never judge by a single test. The names change from chat to chat.

This approach is called Generative Engine Optimization (GEO), and it is what KIGEO does.

Methodology & sources

ChatGPT was queried on 13 October 2026 between 22:40 and 23:00 Vienna time from a Vienna browser without login, with essential cookies only. We asked each of the two identical questions three times, each time in a freshly loaded chat. We recorded the data stream of every answer in the browser and read sources, business records and status messages from it. During the first restaurant run the recording was not yet running, so we took the answer and sources from the page. We counted every business named in the answer text.

Limits, stated openly:

• 6 chats is a small sample. That is why we compare patterns such as source types and search behaviour, not individual names.
• In June and September we measured logged in, this time without login. Logged in, ChatGPT may answer differently.
• We only know the model from ChatGPT's own statement. Asked, it replied “Ich bin GPT-6 von OpenAI” (I am GPT-6 by OpenAI). According to OpenAI, free access has run on GPT-6 Luna since 8 October; OpenAI says nothing separately about use without login.
• We did not record which model ran in June. So we write “ChatGPT in June”.
• Whether a difference comes from the new model or from the search system behind it cannot be separated.
• The figures from Ahrefs, Promptwatch and Peec AI come from companies that sell analytics tools. OpenAI has not confirmed the “Labrador” index.

1. User numbers: TechCrunch, ChatGPT reaches 900M weekly active users (27 February 2026), with 50 million paying consumer subscribers according to OpenAI. Our calculation: 50 million out of 900 million to about 1 billion weekly users pay, so 94 to 95% don't. Counting the 9 million paying business users as well, it is 93 to 94%. Models per plan (Free and Go: GPT-6 Luna, paid plans: GPT-6 Sol): Unite.AI (October 2026).
2. Before data, restaurants: KIGEO restaurant study (10 June 2026) and KIGEO AI Visibility Index, edition 2 (4 September 2026).
3. Before data, tax advisors 1010: KIGEO, own recording of 7 June 2026 (logged in, one run, 8 visible search terms, 12 sources).
4. Ahrefs, Louise Linehan, How ChatGPT Does Query Fan-Out (15 Million Queries Studied) (12 October 2026). 66.6 million free answers without login, August and September 2026, GPT-5.6 Luna.
5. Promptwatch, How ChatGPT Searches the Web (updated 25 September 2026).
6. Peec AI, Tomek Rudzki, ChatGPT built its own search index (4 September 2026).
7. KIGEO, Wie verlässlich sind KI-Empfehlungen? (How reliable are AI recommendations?, 6 August 2026, in German).
8. OpenAI, GPT-6 announcement of 7 October 2026, reposted in the OpenAI Developer Forum, and Search Engine Journal, ChatGPT Gets GPT-6 And Intelligent UI For Interactive Answers (7 October 2026).
9. Data collection and raw data: KIGEO, own measurement of 13 October 2026, in full in the CSV below.

Download the raw data of this study as CSV ↓
Note: the CSV is in German.