Perplexity and Gemini citation sources: a manual analysis
How to analyse which sources Perplexity and Gemini show for Turkish-language queries without an aggregate dataset, using a repeatable prompt set.
By Roozbeh Nazari · CEO
Answering "which sources do Perplexity and Gemini cite when they talk about Turkish brands" at an industry level requires a citation dataset: a tool that collects the source domains behind thousands of answers and breaks them down by brand and competitor. We did not have access to such a dataset in the environment where this article was prepared, and we are not going to publish a "most-cited domains" list or a percentage share we cannot verify. Instead, we read how the two engines display sources from their official help pages, and we describe a repeatable manual analysis method you can run for your own brand and competitors. The method does not change when tool access is available; a tool only enlarges the sample and shortens the time.
Two engines, two different ways of showing sources
Perplexity states in its own help centre that every answer includes numbered citations linking to the original sources. That is the most convenient situation for analysis: which page each sentence rests on is marked with a number inside the answer, and the source list is part of the answer itself. Gemini is different. According to Google's help page, sources appear as "related links" in a side panel opened by a "Sources" button at the bottom of or inside the response, and that button appears "when available", not on every response. So a Gemini response without a sources panel does not mean the response had no sources; it only means they were not shown. AI Overviews and AI Mode inside Google Search are a third surface; Google documents that these features rely on its core ranking systems and need no additional optimisation. This article is limited to Perplexity and the Gemini app; AI features inside Search are a separate measurement topic.
Step 1: build the prompt set
Measurement starts with the query set. The 100-prompt benchmark setup we described for ChatGPT applies here too: write the questions your brand wants to answer, in the language of the patient or customer, in four types. Informational questions ("when can I wash my hair after a hair transplant"), comparison questions ("which clinics are recommended for dental implants in Istanbul"), local and logistics questions ("how much does laser eye surgery cost in Antalya") and brand questions ("is clinic X trustworthy"). At least twenty-five prompts per type, around a hundred in total; fewer than that does not separate random fluctuation from a pattern. If you write the prompts in Turkish, the target market is Türkiye; for international patients you need the same set in Arabic, English and Persian as well, because both engines choose different sources depending on language and location.
Step 2: the execution protocol
The most common mistake in manual analysis is not fixing the conditions. Record the following on every run: date and time, engine and model selection where there is one, whether you are signed in or not, browser language and location. Use a signed-out session or a clean profile to reduce the effect of personal history. Run each prompt three times within the same day; answers change every time and a single run misleads. The fields to record are fixed: prompt, engine, run number, the source domains shown in order, whether your brand appears in the answer text, whether competitors appear, and whether your own site is among the sources. Keep the record in a table and attach a screenshot or share link of the answer to every row; months later, that is the only way to answer "was it really like that on that day". Repeating the same protocol in the same week every month lets you see how the engines' source selection changes. If Gemini did not open a sources panel, mark that separately as "no sources shown"; do not leave it blank, because the share of those rows is a finding in itself for Gemini.
Step 3: classify the sources
Once the domains are collected, assign each one to a source type. The classes we use in practice are: the brand's own site, a competitor clinic or brand site, directories and marketplaces (clinic comparison platforms), news and media, official institutions (ministry, professional association, university), encyclopaedic sources, forums and user-generated content, and video platforms. Classification makes the answer to "why aren't we there" visible: if an engine mostly shows directories and forums for Turkish treatment questions, writing your own site better is not enough on its own; you need to appear accurately and up to date in those directories. Conversely, if official institutions and media dominate, being mentioned in institutional sources comes to the fore. Do this classification separately for each prompt type; encyclopaedic sources can dominate informational questions and directories can dominate comparison questions, and the work plan is different for the two types.
Step 4: competitor comparison and the gap table
Interpret the same prompt set for your competitors: in which prompts a competitor is mentioned, and which class the sources of that answer belong to. This sits in the same table as the citation gap method we described for competitor clinics: source domains in the rows, you and your competitors in the columns, and in the cells the number of answers in which you were mentioned through that source. The empty cells of the table become the work list. Examine the source pages themselves as well: cited pages mostly contain short, self-contained paragraphs that answer the question directly; the principles in our article on content format for AI search are visible on those pages.
What we are not publishing, and why
There is no domain list, no share percentage and no "this much of Turkish brands are cited from there" finding in this article. The only reason is that we do not have a verifiable dataset at the scale needed to carry that claim. With the protocol above, a set of a hundred prompts, three repetitions and two engines produce six hundred observations; that is enough to make decisions for your own brand, but not enough to draw an industry landscape. Commercial tools that measure at industry scale exist; when we have access to them we will share the findings in a separate article, naming the source and the date. Until then, the numbers in your own table are the only real numbers.
Conclusion
Perplexity shows numbered sources in every answer, and Gemini opens sources in a side panel when they are available; on both engines you do not need an aggregate dataset to see which sites are being cited. A fixed prompt set, a recorded execution protocol, source classification and a competitor comparison table show, in a verifiable way, which source types your brand is missing from. Repeating this measurement monthly and prioritising content and source work according to the table is the backbone of our AI search visibility service; you can also run the method with your own team.