# ChatGPT, Gemini, Perplexity, Claude: why does one Turkish question get different brands? > Why does one Turkish question bring up different brands on ChatGPT, Gemini, Perplexity, Claude and Google? How engines see the web, what they read, retesting. Published: 2026-10-07 The same Turkish question brings up different brands in the AI answers of ChatGPT, Gemini, Perplexity, Claude and Google, because these engines reach the web in different ways, read different sites, and do not give the same answer to the same question every time. This difference is not a bug; it is a property that measurement has to account for from the start. The practical consequence is simple: measure each engine separately, repeat the questions, and report frequency of appearance rather than rank. ## Engines reach the web in different ways For an engine to recommend a brand, it first has to reach a page describing that brand, or the model has to know the brand from its training data. How the engines reach the web is described differently in their own documentation. ### ChatGPT According to [OpenAI's help page](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt), ChatGPT can search the web automatically when a question would benefit from up-to-date information. When it works with search partners, it usually turns the user's question into one or more targeted search queries and sends them to those providers. It can also use general location information inferred from the IP address to improve results. This means the same question may not turn into the same search query every time. Also, ChatGPT does not search for every question: in an answer with no search, the brands come from the model's own knowledge. ### Perplexity According to [Perplexity's help page](https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work), Perplexity searches the internet in real time for the question asked, and every answer includes numbered links to the original sources. The same page says Perplexity can use different language models to build the answer and that users can choose the model for some searches. ### Claude According to Anthropic's [web search tool documentation](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool), Claude decides whether a search is needed by looking at the content of the question. It searches for questions about current or changing matters or about specific organisations and products; for questions that rest on established knowledge it answers without searching. When it does search, the answer comes with citations to the sources. This documentation describes Claude's developer interface; the behaviour in the app may differ in detail. ### Google AI Overviews and AI Mode According to Google's [AI features documentation](https://developers.google.com/search/docs/appearance/ai-features), AI Overviews and AI Mode can run several related searches across subtopics and data sources while building an answer. For a page to be shown as a supporting link, it is enough for it to be indexed in Google and eligible to be shown with a search result snippet; there is no additional technical requirement. The same documentation also says AI Mode and AI Overviews may use different models and techniques, so the answers and links they show will differ. According to [the Google Turkey Blog announcement](https://blog.google/intl/tr-tr/urun-duyurulari/turkiyede-ai-modu-ve-ai-bakisi-donemi-basliyor/), AI Mode and AI Overviews began rolling out gradually in Turkey from 18 February 2026, in the Google app and on the search results page. So even within Google there are two separate AI answers, two separate surfaces that need to be measured separately. ## Citation habits vary by engine Engines do not only search in different ways; they also cite sources with different frequency. In our own measurement, between 16 and 25 August 2026, we asked only questions that do not name the brand for three corporate brands (a pharmaceutical company, a dairy brand and a technology group). Across the 1,646 measured answers, the share citing at least one source was as follows by engine: | Engine | Answers measured | Citing at least one source | Rate | |---|---|---|---| | ChatGPT | 362 | 180 | 49.7% | | Gemini | 341 | 321 | 94.1% | | Perplexity | 339 | 339 | 100% | | Claude | 322 | 269 | 83.5% | | Google AI Overviews | 282 | 228 | 80.9% | This table is not a ranking. Citing sources does not make an engine "better"; it only makes it easier to see what the answer rests on. What the table says is this: in roughly half of the measured answers on ChatGPT, brands cannot be tied to a cited source, whereas on Perplexity every answer has a source list. The limits of this measurement should also be stated. 189 of 1,835 planned answers (10.3%) could not be measured and were kept out of the rates: 166 returned an error, 16 came back without source information, and 7 had no AI Overview block to read. The brands measured are three corporate brands; the result cannot be generalised to all brands in Turkey. Google AI Mode was not part of this measurement. ## Engines read different sites Engines that cite sources do not read the same sites either. In 60 analyses we ran up to 3 August 2026 for our own client and comparison brands, we counted the number of analyses in which a site was cited at least once and which engines cited it: | Site | Analyses citing it (of 60) | Engines citing it | |---|---|---| | sikayetvar.com | 39 (65%) | All five engines | | eksisozluk.com | 37 (62%) | AI Overviews, Claude, Perplexity | | en.wikipedia.org | 27 (45%) | ChatGPT, Claude, Perplexity | | tr.wikipedia.org | 19 (32%) | AI Overviews, Claude, Perplexity | | trendyol.com | 17 (28%) | All five engines | | hepsiburada.com | 17 (28%) | All five engines | | akakce.com | 15 (25%) | All five engines | The pattern in the table is striking. The complaints site and the large marketplaces appeared on all five engines. Ekşi Sözlük, by contrast, came up as a source on only three engines. ChatGPT cited English Wikipedia but did not cite Turkish Wikipedia in this data; AI Overviews did the opposite, citing Turkish Wikipedia and not English Wikipedia. Gemini and AI Overviews never cited English Wikipedia in this data. A separate measurement showed a similar spread. In the measurement between 16 and 25 August 2026, Turkish Wikipedia was cited in 72 answers: 34 on Claude, 34 on Perplexity and 4 on AI Overviews. Of English Wikipedia's 26 answers, 23 came from Claude and 3 from ChatGPT. In this measurement Gemini's sources arrived as Google redirect links and were not resolved to sites, so this site count covers four engines, not five. Three limits should be kept in mind when reading these tables. First, the brands measured are our own client and comparison brands; they lean towards technology and consumer brands and are not a sample representing the market. Second, an engine not citing a site in this data does not prove it never reads that site; it only says this was not observed for these questions and in this period. Third, because the two measurements were made with different brands, questions and methods, their percentages should not be compared with each other; only whether a site appears or not can be read side by side. Still, the practical meaning is clear: if the pages describing a brand are on the sites one engine reads, they may not be on the sites another reads. This is one reason the same question brings up different brands. {{cta}} ## The same engine does not give the same answer every time either The difference between engines is not the only variable. The same engine can list different brands for the same question at different times. In [research](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/) published by SparkToro on 28 January 2026, 600 volunteers asked 12 different questions to ChatGPT, Claude and Google's AI (AI Overviews, or AI Mode when no Overview appeared) a total of 2,961 times. According to the research, when the same question is asked 100 times to ChatGPT or to Google's AI, the chance of the same brand list appearing in any two answers is under one percent. Claude came out somewhat more consistent but was still highly variable. The same ranking coming back is far rarer still. From this, the researchers conclude that a metric like "ranking in AI" is not meaningful, but that a percentage of appearance calculated by asking many questions several times is a reasonable metric. They also state the limits of their own work: they say they are not professional researchers and that it should be repeated with much larger samples. The research was done with English questions; we know of no data showing the same rates would hold for Turkish questions. ## Why we do not merge the engines into one number Adding the five engines into a single "AI visibility" number is tempting, but it erases the differences above. If half the answers cite a source on one engine and all of them do on another, the average of the two rates describes neither engine. Merging also requires weighting the engines. We know of no reliable data showing how much each engine is used for which question in Turkey; a number calculated with a weight that does not exist is an estimate that looks like a measurement. That is why our reports list each engine separately and give results as count pairs such as "12 of 40 questions". We do not produce a single composite score. Unmeasured answers also make merging harder. If the answers where an engine returned an error are counted as "your brand was not mentioned", that engine looks low only because it had an outage. Every unmeasured answer should stay on its own row. ## The practical upshot: measure separately, repeat, report frequency These are the measurement rules that follow from these differences: 1. **Measure each engine separately.** ChatGPT, Gemini, Perplexity, Claude, Google AI Overviews and AI Mode are separate surfaces; their results should sit on separate rows. 2. **Repeat the same questions.** A single answer shows not that engine's behaviour for that question but its answer at that moment. When the same question set is asked again on different days, you can see which brands come up regularly. 3. **Report frequency, not rank.** "We are third on ChatGPT" may no longer hold in the next answer; "we were mentioned in 12 of 40 questions asked without naming the brand" is a sounder sentence (example numbers). 4. **Read the sources engine by engine.** If you are absent from the sites one engine reads, the work to be done for that engine may differ from the other. 5. **Write the unmeasured separately.** An answer that returned an error or came back without a source list does not go on the same row as "absent". We explained the steps of measurement, how to write the questions and how to separate unmeasured answers in our guide on [how to measure brand visibility in AI](/en/blog/yapay-zeka-marka-gorunurlugu-nasil-olculur). For why and how to measure repeatedly, see our post on [tracking AI visibility every day](/en/blog/yapay-zeka-gorunurlugunu-her-gun-takip-etmek). All weights and thresholds are published on [our methodology page](/en/yontem). > [!IPUCU] > If you hear a competitor say "We're number one on ChatGPT", ask how many questions it was measured over, how many times they were asked, and on what date. A single screenshot may not repeat on the next question. ## Frequently asked questions ### Why do different brands come up when I ask the same question twice? AI answers change even for the same question. In SparkToro's January 2026 research, in trials where 600 volunteers asked 12 questions a total of 2,961 times, the chance of the same brand list coming back in two answers was under one percent on ChatGPT and on Google's AI. So a single answer does not give a reliable conclusion about a brand's visibility. ### Which engine knows Turkish brands better? We do not answer this question with a ranking. Engines reach the web in different ways, read different sites and cite sources with different frequency; ranking them on a single criterion hides these differences. How a brand appears on which engine is found by measuring each engine separately. ### Do Google AI Overviews and AI Mode give the same result? They may not. According to Google's own documentation, AI Mode and AI Overviews may use different models and techniques, and the answers and links they show differ. So the two should be measured and reported separately. ### What should I do if I appear on one engine and not on another? First look at which sites that engine cites. In our own measurement, Ekşi Sözlük appeared as a source on only three engines, and Turkish Wikipedia never appeared as a source on ChatGPT in this data; so the places where your story needs to be found may differ for each engine. Then ask the same questions again to check whether the difference persists. ## Sources - [Searching the web with ChatGPT](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt), OpenAI Help Center, accessed 7 October 2026 - [How does Perplexity work?](https://www.perplexity.ai/help-center/en/articles/10352895-how-does-perplexity-work), Perplexity Help Center, updated 3 September 2026 - [Web search tool](https://platform.claude.com/docs/en/agents-and-tools/tool-use/web-search-tool), Claude Platform Docs (Anthropic), accessed 7 October 2026 - [AI features and your website](https://developers.google.com/search/docs/appearance/ai-features), Google Search Central, updated 10 December 2025 - [Türkiye'de AI Modu ve AI Bakışı Dönemi Başlıyor](https://blog.google/intl/tr-tr/urun-duyurulari/turkiyede-ai-modu-ve-ai-bakisi-donemi-basliyor/), Google Turkey Blog, 18 February 2026 - [NEW Research: AIs are highly inconsistent when recommending brands or products](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/), SparkToro, 28 January 2026 - How often engines cite sources, questions that do not name the brand, 2026-08-25: 6 measurements between 16 and 25 August 2026, 3 corporate brands, ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews; 1,646 of 1,835 planned answers were measured, 189 (10.3%) could not be measured and were kept out of the rates; the site list does not cover Gemini - Which engine cited which site, 2026-08-03: 60 analyses up to 3 August 2026, 19 company records (our own client and comparison brands, not a market sample), 10,474 successful answers, 75,914 source links, five engines; recounted on 7 October 2026