How ChatGPT’s writing style changes by model version

We tracked 22 stock AI words: words and phrases readers often flag as AI-sounding, namely robust, seamless, tailored, streamline, leverage, comprehensive, delve, elevate, versatile, user-friendly, intuitive, cutting-edge, ensure, crucial, landscape, navigate, holistic, game-changer, standout, top pick, note: and pro tip. We counted the share of answers that use at least one of them.

In January 2026, 78.4% of ChatGPT answers used a stock AI word. By September 2026 it was 16.7%. The drop happened inside each collection period: from 77.6% under gpt-5-2 to 42.9% under gpt-5-3, and from 32.1% under gpt-5-5 to 27.1% under gpt-5-6. Google's answers did not follow. Google AI Overviews used a stock AI word in 73.7% of answers from December 2025 to April 2026, and Google AI Mode in 75.2% from May to September 2026.

We checked this outside our fixed prompt set. On 3,000 random answers per group from all prompts, the share was 76.3% for gpt-5-2, 21.4% for gpt-5-6 and 75.8% for Google AI Mode.

Each version has a recognizable style. The table compares length, layout, punctuation and vocabulary across the same 1,500 prompts.

The most overused word is the word each version uses most often relative to the other ChatGPT versions, among words that appear in at least 2% of its answers. gpt-5-2 says tailor 18.5 times as often as the other versions. gpt-5-3 says actually 9.0 times as often. gpt-5-6 says particularly 23.1 times as often.

gpt-5-6 adds qualifiers to its recommendations. It uses distinction 61.0 times as often as earlier ChatGPT versions, particularly 23.1 times, caveat 20.7 times and roughly 17.3 times.

It also says currently 12.6 times as often and shortlist 9.3 times as often. These words place the answer in time and present the brands as a shortlist rather than a ranking.

ChatGPT now writes in the first person. 73.8% of gpt-5-6 answers contain I'd, as in I'd start with or I'd choose. Only 0.04% of gpt-5-2 answers did. Google's answers stay under 2% in every month.

First-person phrasing is not new. 34.5% of late-2025 ChatGPT answers used it, mostly as I'd lean toward or I'd go with. It vanished in December 2025 and came back with gpt-5-5 and gpt-5-6. A random check across all prompts agrees: 0.2% for gpt-5-2 and 74.9% for gpt-5-6.

The median ChatGPT answer had 510 words in October and November 2025. Under gpt-5-6 it has 265 words, 48% fewer. Part of the drop between gpt-5-3 and gpt-5-5 coincides with our switch from collecting ChatGPT to collecting ChatGPT Search. The gpt-5-5 to gpt-5-6 upgrade, on the same collection, cut the median from 311 to 265 words, 15% fewer.

From June to September 2026, ChatGPT Search and Google AI Mode wrote answers of nearly the same median length: 285 and 280 words. Length is the one measure on which the two engines matched.

In late 2025, 5.5% of ChatGPT answers contained a table. Under gpt-5-6, 55.5% do. Each named version used more tables than the one before it.

Lists shrank as tables grew. gpt-5-3 answers held 29.4 list items on average, often in short fragments of 7.4 words per sentence. gpt-5-6 answers hold 7.0 list items. Google AI Mode put a table in 12.6% of answers from June to September 2026.

The first two words of an answer are a strong version marker. 54.3% of gpt-5-2 answers began with Here are. 57.8% of gpt-5-6 answers begin with If, usually If you or If your, which sets a condition before the recommendation.

Late-2025 ChatGPT often opened with praise for the question: 17.3% began with Great question or Good question. Google's answers most often open with The best: 18.3% for Google AI Overviews and 14.7% for Google AI Mode.

When a prompt leaves out details, the two engines handle it differently. Google AI Mode asks the reader a question, such as What is your budget? or Do you need it for travel? From June to September 2026, 8.8% of its sentences were questions.

ChatGPT Search states the condition instead. 47.4% of its answers in the same months opened with If, against 1.0% for Google AI Mode. Only 1.2% of ChatGPT Search sentences were questions.

Google AI Mode's most overused words describe products the way a catalog does. Compared with ChatGPT Search on the same prompts from June to September 2026, it uses robust 15.4 times as often, specialized 6.4 times, offering 6.2 times and via 5.1 times.

Robust is also a stock AI word. ChatGPT Search almost stopped using it, at 0.07 uses per 1,000 words, while Google AI Mode used it 1.06 times per 1,000 words.

Answer overlap is the TF-IDF cosine similarity between the content words of two answers. It is 0 when the answers share no words and 1 when they use the same words in the same proportions.

Same prompt, same day, ChatGPT and Google AI Overviews peaked at an overlap of 0.55 in February 2026 and fell to 0.50 by April. ChatGPT Search and Google AI Mode went from 0.48 in June 2026 to 0.42 in September. Each engine overlaps more with itself four days later: 0.61 for ChatGPT Search and 0.56 for Google AI Mode in September. Within each period, the engines moved apart, not together.

We compared the last ChatGPT answer before an upgrade with the first answer after it, for the same prompt. As a control, we compared two answers from the same version that were the same number of days apart.

Every upgrade lowered overlap by 0.08 beyond the control. The gpt-5-5 to gpt-5-6 upgrade, four days apart, gave 0.61 against 0.68 for the control. On January 3, 2026, only the version label changed, not the writing style, and overlap matched the control within 0.01.

At the first mention of each brand, ChatGPT Search added a product or line word in 20.4% of mentions from May to September 2026, as in Ecovacs Deebot T50 or Kaweco Skyline Sport. Google AI Mode did so in 15.6% and used the bare brand name in 83.4%. Earlier, ChatGPT did so in 34.2% of first mentions and Google AI Overviews in 23.5%, from October 2025 to March 2026.

Neither engine uses legal names. Corporate suffixes such as Inc or Ltd appear in at most 0.13% of first mentions on any engine. A domain form such as brand.com appears in 1.0% or less.

We used one fixed set of 1,500 prompts that both engines answered in every month from October 2025 to September 2026. For each prompt and month we took the first answer of one day on each engine, and of the next day when available. We removed links, images, citation markers and shopping-card text before counting, and left out 1,169 answers with fewer than 20 words.

The engines changed during the year. Google answers are Google AI Overviews through April 2026 and Google AI Mode from May 2026. ChatGPT answers come from ChatGPT through April 2026 and from ChatGPT Search, with web search on, from May 2026. We compare engines only within the same period, and the ChatGPT version changes we highlight most (gpt-5-2 to gpt-5-3, and gpt-5-5 to gpt-5-6) each happened within one collection.

Our Google AI Overviews text from December 2025 to April 2026 carried embedded image data that made up 73% to 89% of the raw page text, which is why raw answers averaged 8,400 to 23,200 characters a month. After removing it, they averaged 1,700 to 1,900 characters. Before December 2025 that text arrived without list or table markup. ChatGPT did not report its version before January 3, 2026, or for part of September 2026; we grouped those answers by date and by the style breaks visible in the daily data. Our fixed prompt set has no ChatGPT answers from April 25 to May 28, 2026. In April 2026 our brand detection lapsed while the answer text stayed complete, so the brand-name figures leave April out.

The style of an AI answer is set by the vendor and changes with each model version, on the vendor's schedule. A page that copied gpt-5-2's phrasing in January would read like a different engine by September. Do not write to match a model's style.

Write what the current versions extract. ChatGPT now puts comparisons in tables in more than half of its answers, so publish comparison tables with consistent product names and plain attributes. Google AI Mode describes products with words like specialized and offering, so state plainly what your product offers and for whom. Name the product line with the brand, the way ChatGPT does. Track answers per model version, because a version change can move answer text more than re-asking does.

We measured vocabulary richness as distinct words per 100,000 words of answers, averaged over 20 random draws so that every group is measured at the same size. Late-2025 ChatGPT used 9,790 distinct words per 100,000. gpt-5-6 uses 11,260, about 15% more. Google AI Mode used 12,330, more than any ChatGPT version.

The word-frequency slope is the steepness of the line when word frequency is plotted against frequency rank on log scales, here for ranks 10 to 1,000. It stayed between -0.79 and -0.87 for every ChatGPT version. Upgrades changed which words appear, not how evenly words are used.