Ask ChatGPT Twice and Its Top Pick Changes 56.3% of the Time vs 82.7% for Gemini

A marketing manager asks ChatGPT for the best CRM for a small business, sees their product named first and takes a screenshot for the board. Half an hour later a colleague asks the same question and gets a different product at the top. Neither answer is wrong; ChatGPT simply does not give the same answer twice.

For brands, the first product an AI assistant names to a "best" question is the most valuable spot in the answer. If that spot changes between two identical questions, a single check says almost nothing, and any test of how wording affects the answer has to be measured against that churn. Visibility in ChatGPT is not a rank you hold; it is a share of answers you win.

Asked the same question twice in a row, ChatGPT names a different top pick 56.3% of the time, and Gemini 82.7% of the time. Against that background, politeness, synonyms, typos and even the language of the question barely matter: each adds about 2 to 4 points. What does move the answer is meaning. A budget, a detailed list of priorities or the current year changes ChatGPT's top pick 65.4% to 87.4% of the time.

We ran three tests on September 23, 2026. In the repeat test we picked 300 "What is the best...?" questions, from vacuum cleaners and running shoes to accounting software and car insurance, and asked each one 10 times on ChatGPT and 10 times on Gemini, about 30 minutes apart: 3,000 answers from each assistant. In the wording test we asked ChatGPT, with web search switched on and from a US location, 500 "best product" questions in their plain form ("What is the best X?") and in six reworded versions: ending in "please", "top" instead of "best", a typo, adding "2026", adding "on a budget", and a detailed version listing the buyer's priorities. Every version was asked twice, giving 6,992 answers. The typo swapped two neighboring letters in the longest word of the question; 496 of the questions had a word long enough to misspell this way. In the language test we asked "What are the best [service] in [city]?" and its Spanish equivalent for 20 local services in 25 of the largest US cities, each twice, giving 995 English and Spanish pairs.

The top pick is the first product, or the first business, an answer recommends, treating small name variants as the same pick. In every test we compared each answer with another answer to the same question, and counted how often the top pick changed. The baseline is the change you get from asking the identical question again; the effect of a wording change is the gap above it. The wording and repeat tests use different question sets, so their baselines are close but not identical: 54.6% and 56.3%. The share of websites replaced is, on average, how much of two answers' combined source list appears in one answer but not the other.

Over 10 answers, ChatGPT named 3.7 different top picks per question on average. Its most common pick for a question won 61.2% of answers, so a typical leader holds the top spot about six times in ten. Only 8.3% of questions got the same top pick in all 10 answers, about one question in twelve.

ChatGPT runs a fresh search and writes a fresh answer every time, and for most "best" questions several products have a fair claim to the top. Small differences in which pages it reads, and in how it words the answer, are enough to move a different product to the front. The result is a rotation among a few contenders rather than a single fixed winner.

Gemini's top pick changed in 82.7% of back-to-back answers, about four times in five. It gave the same top pick in every answer for 1.3% of questions, against 8.3% for ChatGPT.

For a brand, that makes a single Gemini check even less meaningful than a single ChatGPT check. Being named first by Gemini once says little about the next answer; what counts is how often you are named first, and how often you are named at all, across many answers.

That baseline is the yardstick for every wording change. Ending a question with "please" changes the top pick 56.5% of the time, 1.9 points above it. Swapping "best" for "top" changes it 57.5% of the time, 2.9 points above. A typo changes it 58.8% of the time, 4.2 points above. All three look like the plain question asked again: most of the change would have happened anyway.

Politeness changes nothing that matters. Polite questions get a different set of cited websites 84.4% of the time against 82.2% for the plain question asked twice, ChatGPT cites 3.11 websites per answer against 3.10, and the answer runs 215 words against 216. "Please" replaces 53.4% of the cited websites, the fewest of any change. Where the polite version did change the pick, it was the usual reshuffle: for "best dog food brands", one answer led with Purina Pro Plan and the other with Hill's Science Diet, two brands ChatGPT treats as close alternatives.

Synonyms behave the same way. ChatGPT cites 3.06 websites per "top" answer against 3.10, and the set of cited websites changes 86.1% of the time against 82.2%. The one visible difference is length: "top" answers run about 189 words, the shortest of all seven versions and 27 words shorter than "best" answers, with no loss of sources. For double strollers, the Baby Jogger City Mini GT2 Double gave way to the UPPAbaby Vista V3 with RumbleSeat, the kind of reshuffle ChatGPT makes between two identical questions.

The seven versions fall into two groups. Budget, detailed and year wordings add information, and they move the top pick 11 to 33 points above baseline: 87.4% for a budget, 76.8% for a detailed question listing the buyer's priorities and 65.4% for adding the year. The sources follow the same split. A budget question replaces 78% of the websites ChatGPT cites and a detailed question 81.7%, against 57.1% for a "top" question.

When a buyer changes what they are asking for, ChatGPT searches for different evidence and reaches a different answer. When a buyer only changes how they phrase it, ChatGPT treats it as the same request. It responds to what the buyer needs, not to how the request is worded.

Where a typo does leave a mark is in the sources. 3.2% of misspelled product questions got an answer with no cited website, against 1.3% of correctly spelled ones, about two and a half times the rate. The set of cited websites changed in 90.6% of typo pairs, against 82.2% for identical questions. Once ChatGPT does search, it finds about as much: 3.04 websites per answer against 3.10, in an answer of about 203 words against 216.

The changed picks look like ordinary reshuffling among well-known options. For face moisturizer, the correct question led with Vanicream Daily Facial Moisturizer and the typo version with La Roche-Posay Toleriane Double Repair. ChatGPT works out the intended question and answers it, so pages built for misspellings of a category add little.

Most of that is ChatGPT's everyday variation: the English question asked twice changes the first business 46.8% of the time, so language adds about 4 points. The lists shift a little further too: English and Spanish answers share 3 to 6 points fewer businesses than two English answers do, so a Spanish speaker sees a somewhat different set of options, not a different market. ChatGPT writes 99.1% of its answers to Spanish questions in Spanish, and cites 3.12 websites per Spanish answer against 3.28 in English.

The research behind both languages is the same. The websites cited in Spanish answers are almost the same English-language directories as in English answers, in almost the same order, and none of the most-cited websites for Spanish answers is a Spanish-language site. ChatGPT draws on 428 different websites for Spanish answers and 447 for English ones, and in both languages the ten most-cited take about 45% of all citations.

A few questions produced the same answer every time. ChatGPT named QuickBooks Online the best accounting software for a small business in all 10 answers. It named Chase Sapphire Preferred for travel rewards credit cards, HubSpot for small business CRM, Spotify for music streaming apps and Breaking Bad for TV shows in every answer that named a top pick.

These are categories with a widely agreed leader, where reviews, rankings and comparison articles point to the same product. A fixed top pick goes with that kind of consensus. In crowded categories with many similar products, such as vacuum cleaners or running shoes, the top pick rotates.

Measure AI visibility as a share of answers. Ask your key questions many times, on different days, in ChatGPT and Gemini, and track how often you are the top pick and how often you are mentioned at all. One screenshot of ChatGPT is not a result, and a single before-and-after comparison of two wordings is mostly noise unless the gap clears the churn from asking twice.

Do not spend effort on surface variants. There is no need to test polite and casual phrasings, build separate pages for "best" and "top", or create pages for common misspellings: ChatGPT treats all of them as the same question. Build for the modifiers that change intent instead, such as price range, the buyer's priorities and the current year, because those decide which products appear and which websites get cited. If you serve Spanish speakers, your English-language directory listings feed those answers too, so keep them complete and say on them that you serve Spanish speakers.

The rotation is also an opening. If a category leader wins about six answers in ten, a challenger can take some of the rest, and the brands that hold the top spot every time are the ones independent reviews, comparison articles and buying guides agree on. Working with an ai seo company that tracks your top-pick share across repeated ChatGPT and Gemini answers shows which question variants in your category change the answer.

Embed this research

Cite this study