No, ChatGPT does not give the same answer to everyone, and it rarely gives the same answer twice even to one person. We asked ChatGPT 10 buying questions, in English and Hinglish, 8 times each from inside India within one hour, which gave 160 answers. On average, two answers to the same question shared only 54.7% of the brands they named.
That was with no login, no chat history and one country. Real users add more variation: OpenAI says ChatGPT can use your memories and your approximate location, so two people asking the same thing start from different places. Below are our numbers for ChatGPT, Gemini and Perplexity, a fresh ChatGPT rerun from 15 Sep 2026, and what the variation means if you check your brand by hand.
What we measured
The data comes from Depra's English vs Hinglish study. We wrote 10 buying questions, 5 about skincare and 5 about fashion, such as "Which face wash should I buy for oily skin in India under ₹500?". Each question had an English version and a Hinglish version, which is Hindi-English typed in English letters. We asked each version 8 times to three engines on 14 Aug 2026, in one 57-minute window, located in India, with no account logged in. That gave 480 answers.
ChatGPT answers were captured from the chatgpt.com web app with web search on. Gemini answers came from the gemini.google.com web app. Perplexity answers came from Sonar, the answer model Perplexity sells through its API, with live web search.
To compare two answers, we listed the brands each one named and worked out the overlap: brands both answers named, divided by all brands either answer named. Here are two ChatGPT answers to the face wash question, asked 10 minutes apart:
- Answer 7: Minimalist, The Derma Co, Dot & Key
- Answer 8: Minimalist, The Derma Co, Cetaphil
They share 2 brands out of 4 different brands named, so the overlap is 50%. Identical lists score 100%. Lists with nothing in common score 0%. With 8 answers per question, language and engine, there are 28 pairs of repeat answers each time, or 560 pairs per engine across the study.
How different are ChatGPT, Gemini and Perplexity answers to the same question?
Every row compares repeat answers to the identical question, on the same engine, in the same language.
| Repeat answers to the same question | ChatGPT | Gemini | Perplexity |
|---|---|---|---|
| Average brand overlap | 54.7% | 51.2% | 68.3% |
| Exactly the same brands | 55 of 560 (9.8%) | 27 of 560 (4.8%) | 86 of 560 (15.4%) |
| Same brands in the same order | 29 of 560 (5.2%) | 5 of 560 (0.9%) | 34 of 560 (6.1%) |
| Same brand named first | 60.0% | 63.2% | 69.5% |
| Average overlap of cited websites | 33.4% | 33.1% | 85.2% |
What the table shows:
- ChatGPT repeated its exact brand list in about 1 pair of answers in 10, and its exact order in about 1 in 20.
- On average, two ChatGPT answers to the same question shared only a third of their cited websites. Each run read a different set of pages.
- Perplexity was the most repeatable engine, with 68.3% average brand overlap and 85.2% average cited-website overlap between reruns.
The face wash question shows the pattern up close. Across 8 English answers, ChatGPT named 10 different brands, 3 to 5 per answer. The Derma Co appeared in all 8. Minimalist appeared in 6, Simple in 5, Dot & Key and Chemist At Play in 3 each, Plum and Fixderma in 2, and Saslic, Deconstruct and Cetaphil once each.
Across the whole study, ChatGPT produced 157 brand and question pairings, meaning a brand named at least once for a given question in a given language. 37 of them appeared in all 8 answers. 29 appeared once and never again. The top of each list is fairly stable. The rest rotates.
Other research points the same way. SparkToro and Gumshoe ran 12 brand recommendation prompts 60 to 100 times per platform with volunteers. As Search Engine Journal reported in January 2026, ChatGPT and Google's AI answers returned the same brand list less than 1 time in 100, and the same list in the same order less than 1 time in 1,000. The original write-up is on SparkToro's blog. Our exact-list rates for ChatGPT (9.8%) and Gemini (4.8%) are higher. Our answers were collected within one hour, from one country, with no login.
Does ChatGPT give the same answer a month later?
The study covered one hour. On 15 Sep 2026, a month later, we asked ChatGPT the same 10 English questions 3 more times each: from India, logged out, with web search on. That gave 30 new answers.
- On average, two August answers to the same question shared 54.1% of their brands (280 pairs, English only).
- On average, a September answer and an August answer shared 50.1% (240 pairs).
- On average, two September answers shared 46.4% (30 pairs).
A month apart, answers differed only a little more than answers asked minutes apart. The core brands held: brands ChatGPT named in all 8 August answers to a question (19 English brand and question pairings) appeared in 53 of the 57 matching September answers. New names did turn up. One September answer to the vitamin C serum question recommended Cureskin, a brand that appeared in none of the 480 August answers. Three runs per question is a small sample, so read this as a spot check.
Method and limits
The August overlap, first-brand and cited-website figures come from the published study report, summarised on the English vs Hinglish AI shopping study page. The exact-list and same-order counts, the brand pairing counts and the face wash breakdown are our re-analysis of the same stored answers on 15 Sep 2026. Brands were counted with the study's fixed brand list and matching rules, so a brand missing from that list was not counted. For the September comparison we added 9 brand names that appeared in the new answers and were missing from the list, Cureskin and U.S. Polo Assn. among them, and re-counted both months with the longer list. That is why the August English figure there is 54.1%. Limits: 10 questions in two categories, one country, logged-out sessions only, and 3 September runs per question. Nothing here measures logged-in users with memory or chat history. Depra sells an AI visibility tracker, so weigh the numbers with that in mind. This post replaces an earlier Depra post on the same topic that misstated a confidence range; the corrected example is in the section on checking your brand once.
Why do AI answers change each time?
Three causes, each documented in public sources or visible in our data:
- Randomness in the writing. A language model writes an answer one word at a time and picks each next word with some randomness. OpenAI's API reference describes a temperature setting: higher values make output more random, lower values make it more focused. OpenAI's help pages do not state the setting the ChatGPT app uses.
- Live search. With web search on, ChatGPT rewrites your question into one or more search queries and sends them to search partners, according to OpenAI's help page on searching the web with ChatGPT. Search results shift, so the pages behind each answer shift too.
- Answer length. Some answers give one pick, others a long list. In August, ChatGPT's 8 answers to the running shoes question named anywhere from 2 to 7 brands.
A 2026 paper, Don't Measure Once by Schulte, Bleeker and Kaufmann, draws on published studies to reach the same point: AI search answers vary across runs, prompts and time, so one-off checks are unreliable. It calls for repeated measurement and for describing visibility as a spread of outcomes.
Does ChatGPT give different answers to different people?
Our data cannot fully answer this. Every request used the same question, the same country, the same hour and no login. Real people differ on all four, and OpenAI documents several ways that changes the answer:
- Memory. OpenAI's Memory FAQ says memory lets ChatGPT remember context from your chats, files and connected apps to personalise responses. Temporary Chats do not use existing memories.
- Search queries. OpenAI's search help page says that if memory is on, ChatGPT may use saved memories when it rewrites a search query.
- Location. The same page says ChatGPT may estimate your general location from your IP address, and a VPN can change that estimate.
- Wording. People rarely type the same words. In SparkToro's study, 142 volunteers wrote their own prompts about headphones and almost no two looked alike, yet Bose, Sony, Sennheiser and Apple still appeared in 55% to 77% of the 994 answers.
- Language. In our study, switching a question from English to Hinglish lowered brand overlap beyond normal rerun change by 2.2 points on ChatGPT, 7.8 on Gemini and 23.4 on Perplexity.
Two strangers asking the same thing can see different brands. In our data, one person asking twice usually would.
What this means if you check your brand once
One check is one sample. Minimalist appeared in 6 of ChatGPT's 8 English answers to the face wash question. The other 2 left it out, so 1 check in 4 would have suggested ChatGPT ignores the brand.
The fix is to count, then report a range, the way pollsters report an election: a share, a margin of error and the sample size. For AI answers, the share is how often an engine names you. The range comes from the Wilson score interval, a standard formula for the likely spread of a percentage measured on a small sample. A 95% range is the band the true share most likely sits in.
| Answers checked | Brand named in | Share | 95% range |
|---|---|---|---|
| 8 | 6 | 75.0% | 40.9% to 92.9% |
| 30 | 15 | 50.0% | 33.2% to 66.8% |
| 45 | 28 | 62.2% | 47.6% to 74.9% |
| 100 | 50 | 50.0% | 40.4% to 59.6% |
| 400 | 200 | 50.0% | 45.1% to 54.9% |
Our earlier post gave the 28 of 45 case as 54% to 70%. That range was too narrow. At 50%, 30 answers leave about 17 points either way, 45 answers about 14, and 100 answers about 10. If your share moves from 55% to 60% on 45 answers, the data cannot tell that apart from noise.
How to get a reliable read
- Ask each question several times. 5 runs per question per engine turns a yes or no into a share. More runs narrow the range.
- Test logged out or in a Temporary Chat, so your own memory and history do not tilt the result.
- Keep everything else fixed: the same questions, wording, country, language and engine each time.
- Leave your brand name out of the question. A question that names you will mention you.
- Track the share over days, and judge each change against its range.
Our guide to tracking brand mentions in AI answers covers a by-hand sheet, and the guide on how to measure AI visibility covers which numbers to report.
Depra runs this routine for you. It asks your buyers' questions to ChatGPT (the web app with search on), Gemini and Google AI Overviews every day, and to Perplexity every week, from inside your country with no logged-in history. Every full answer is stored, you see how often each engine names you over time, and questions that already contain your name are not counted. Your report shows a confidence range next to the number of answers behind it. See AI visibility tracking, the ChatGPT visibility tracker and how Depra measures AI answers.
See how often ChatGPT names your brand across repeat answers. Start 7 days free on any plan, no card.