Does ChatGPT use Google results? What 1,200 answers show

The findings

The full study this publication is part of consists of 6,120 answers. This publication uses the baseline measurement: 400 questions, 3 runs, 1,200 answers per chatbot, plus the Google results for 400 search terms. Percentages are weighted across categories and question types unless marked as unweighted.

The core question

How often does ChatGPT cite a page from Google's top 3?

Rarely. There are two ways to count it, and they give different numbers:

So the 4% does not mean that only 4% of answers cite anything from the top 3. It is the probability for a single page. Because there are three top pages per question, the share of answers citing at least one of them is higher.

All 400 questions, 3 answers per question per chatbot. For each question we looked at which pages ranked at positions 1, 2, and 3 on Google for the linked search term. First column: unweighted count (supplementary). Second column: average over the three positions, weighted across 20 categories and both question types. Answers without a source count as not cited.

How we tested this is explained in the methodology.

For purchase questions, ChatGPT cites a top 3 page even less often

For commercial SEO, this is the most relevant difference. For purchase questions, a top 3 page was cited in 2.1% of ChatGPT answers on average; for informational questions, in 6.0%. For Gemini it is the other way around.

Averages per page, like the primary measure: the probability that one specific top 3 page is cited in an answer. Exploratory.

An example of a purchase question is “what is the best charging card for an electric car?” (asked in Dutch). The difference for Gemini is tied to how often Gemini lists sources: for purchase questions in 239 of 687 answers (35%, unweighted), for informational questions in 91 of 513 (18%, unweighted).

Google ranking

Does a higher Google ranking affect ChatGPT and Gemini citations?

Yes, but from a low base. Two things are true at once:

So a low percentage does not mean the Google ranking is irrelevant. We did not test how strong the relationship is; the table is descriptive.

Share of all 1,200 answers per chatbot that list the page at that position as a source. Weighted average.

Its own search queries

How does ChatGPT find its sources for an answer?

With many search queries of its own, which differ from the user's question. For purchase questions, the API logged an average of 9.5 search queries per answer. In 84% of the purchase answers in which ChatGPT searched, at least one query contained a brand name the user had not mentioned.

First two rows: average per answer (each question was asked three times), across all answers, including answers without a search (those count as 0), weighted by category and question type. Exploratory. Third row: share of the purchase answers in which the chatbot searched (ChatGPT 687, Gemini 239, Perplexity 553). Brand name: a brand that was not in the question. For Gemini, the Health and wellbeing category is missing because Gemini did not search on purchase questions there.

ChatGPT often uses the site: search operator, which searches within a single website, as in “site:moneybird.nl tarieven” (“tarieven” is Dutch for rates). A site: query on a brand's own website appeared in 67% of purchase answers (weighted by category; unweighted 473 of 687 purchase answers, 69%). ChatGPT searched in all 687 purchase answers.

Example: one purchase question, the Google results, and ChatGPT's search queries

Google's top 10 for that search term on September 26, 2026:

The search queries the API logged for this answer:

1. 1site:nespresso.com/nl/nl Essenza Mini 2 kopgroottes 19 bar prijs
2. 2site:delonghi.com/nl-nl Magnifica Start ECAM220 prijs onderhoud bonen koffiemachine
3. 3site:philips.nl Senseo Original Plus koffiemachine twee kopjes tegelijk prijs
4. 4site:moccamaster.eu cup one 300 ml koffiezetapparaat specificaties nederland
5. 5site:moccamaster.eu/nl cup one prijs 300 ml
6. 6site:philips.nl/c-p/CSA210_
7. 7site:nespresso.com/nl/nl/koffie original capsules vanaf prijs espresso lungo
8. 8site:philips.nl senseo original plus csa210 prijs koffiepad

The sources in the answer: moccamaster.eu/cup-one-matt-black-69221, philips.nl/c-p/CSA210_40R1/original-plus-coffee-pad-machine, and nespresso.com/nl/nl/koffiemachines/essenza-mini. None of those pages or websites appears in Google's top 10 for the linked search term.

This is a single example (answer P12C02, ChatGPT, first run in the data repository), chosen because the pattern is easy to see in it. All eight logged search queries target a brand and a model, on the brand's own website. None of them matches the generic search term we used to query Google (“beste koffiemachine woon alleen”).

Exploratory: in about 42% of purchase answers, a brand name already appears in the first logged search query. Whether ChatGPT actually ran the queries in that order cannot be proven. So we cannot establish whether ChatGPT picks brands before it reads sources. Nor did we measure how a brand ends up in those search queries.

Page or website

Does ChatGPT cite the same page Google shows, or only the same website?

The same website more often than the same page. Little overlap at the page level therefore does not mean ChatGPT uses entirely different websites. Even so, the overlap at the website level remains limited.

Comparison 1: how often does a result from Google's top 3 come back as a source?

Average share of all 1,200 answers per chatbot, per top 3 result. The first row is the primary measure. The website level was preregistered as a second measure and is reported as exploratory.

Comparison 2: how many of the chatbot's sources rank in Google's top 10?

Average share per answer with sources (ChatGPT 1,115 answers, Gemini 330). First row: secondary hypothesis. Second row: exploratory.

The primary measure counts only the exact page. A source on the same website but on a different page does not count as a match there. For ChatGPT, the website of a top 3 page is cited more than twice as often as the page itself. At the same time, on average two thirds of the websites ChatGPT cites in an answer do not appear in Google's top 10 for the linked search term.

Gemini

What changes for Gemini if you count only answers with sources?

Then Gemini looks much more like Google. Gemini listed sources in 330 of 1,200 answers (27.5%), for 138 of the 400 questions. The other answers count as “not cited” in the primary measure, which pulls the average down. In this measurement, those were the same 330 answers in which Gemini ran a search.

Average share of answers that cite one specific page at position 1, 2, or 3. The right-hand column is a supplementary measure that was not preregistered.

The 34.8% cannot be derived from 8.1% and 330 of 1,200. The calculation uses only the 138 questions for which Gemini listed sources at least once, and per question only the answers with sources. The same weighting by category and question type was then applied to the combinations in which such questions occur; 7 of the 40 combinations are missing. So the figure says something about those 138 questions, not about all 400.

In 207 of the 330 Gemini answers with sources (63%, unweighted, supplementary), at least one top 3 page is among the sources. Of all pages Gemini cites, 32% on average rank in Google's top 10, versus 11% for ChatGPT. Compared on the same 138 questions, Gemini is 14.7 percentage points higher (95% interval 11.1 to 18.4).

According to Google's documentation, Gemini searches through Google Search. OpenAI does not disclose which search index ChatGPT uses.

For SEO specialists

What does this mean for the work of an SEO specialist?

That you need to broaden your analysis. A Google ranking for a generic search term and a citation in a chatbot are two different things, and you need to measure them separately. This study does not deliver a proven method for getting cited more often. It does show where to look.

This applies to Dutch-language questions through the APIs, measured on a single day. Whether a better Google ranking leads to more citations, we did not measure: this study compares two lists on one day and does not track pages as they rise or fall.

Whether small websites stand a chance as a source next to big names is covered in our study on the Google traffic of AI sources. How much the brands mentioned shift when you repeat the same question is covered in our study on the consistency of brand mentions.

Deep dive

Deep dive: industries, retrieved sources, and ChatGPT's own searches on Google

The three parts below are exploratory. They round out the picture, but they do not carry the main conclusion.

In which industries do AI sources overlap most with Google?

For Gemini, the local service providers category stands out: there, a top 3 page is cited in 28% of answers on average. For ChatGPT, no category is above 11%.

Averages per page, like the primary measure. A selection of notable categories; the lowest for Gemini is Software and SaaS (0.0%, ChatGPT 2.5%). 20 questions per category, with purchase and informational questions weighted equally, so these are rough estimates.

How many top Google pages do ChatGPT and Perplexity find while searching?

Few here, too. Through the API, ChatGPT and Perplexity return a list of sources they found while searching (in the methodology: retrieved). These are URLs from their search results, not pages that are known to have been opened or read.

Which websites ChatGPT mostly finds, and why Reddit is often among them but almost never cited, is covered in our study on Reddit as an AI source.

Do ChatGPT's sources appear in the Google results for its own search queries?

Within the checked selection of search queries without site:, an average of 15% of the pages ChatGPT cited ranked in the matching Google top 10 (416 answers). For the generic search term, the figure for the same answers was 9.4%.

We retrieved the Google results for 1,735 search queries that the chatbots logged themselves. From all distinct search queries without site: in the baseline measurement, we drew a sample of 1,793 with a fixed seed, proportional to each chatbot's count and with no cap per answer; 1,735 returned results. For the 416 ChatGPT answers, this means an average of 1.8 of the 3.8 search queries without site: per answer were checked, out of an average of 9.1 search queries in total. Search queries with site: were not included.

Only answers for which we looked up the Google results of at least one of the chatbot's own search queries.

This is not a complete check. A cited page that does not match here may still appear in the Google results of a search query we did not check, such as the site: queries, which are especially common for ChatGPT. The rise from 9.4% to 15.0% fits the picture that ChatGPT searches with different search terms, but does not prove it. Whether the sources rank further down on Google, we did not measure.

Frequently asked questions

Questions about Google rankings and AI citations

Related research

This page is the English edition of a study first published in Dutch. Read the original Dutch version.

Glossary

Dutch terms on this page

The study used Dutch questions and Dutch websites. These are the Dutch names and words that come up.

AD

Dutch national daily newspaper (Algemeen Dagblad), owned by DPG Media.

Consumentenbond

Dutch consumer association that tests products and services, comparable to Consumer Reports.

Coolblue

Dutch online electronics and appliance retailer, comparable to Best Buy.

Fashionchick

Dutch fashion and lifestyle website with product reviews.

Fleur de Café

Dutch online retailer of coffee machines and coffee beans.

Koffie.nl

Dutch website with coffee machine guides; koffie means “coffee.”

koffiemachine

Dutch for “coffee machine.”

koffiepad

Dutch for “coffee pad,” a single-serve pod.

koffiezetapparaat

Dutch for “coffee maker.”

MediaMarkt

European consumer electronics chain with large Dutch stores, comparable to Best Buy.

Moccamaster

Dutch brand of premium filter coffee makers, sold in the US as Technivorm Moccamaster.

Moneybird

Dutch online invoicing and accounting software for small businesses, comparable to QuickBooks.

onderhoud

Dutch for “maintenance.”

prijs

Dutch for “price.”

tarieven

Dutch for “rates” or “prices.”

Tweakers

Dutch technology news and review site with a price comparison tool and a large forum.

Zwarte Roes

Dutch online shop for coffee equipment.