ChatGPT vs Gemini Blog Writing: Neither Sounds Like You Yet

Most ChatGPT vs Gemini blog writing comparisons ask which model produces the better post. We measured both and the more useful question is which one starts closer to the way you already write, because neither of them knows that. ChatGPT drafts like an enthusiastic columnist and Gemini drafts like a careful analyst. Your blog sounds like neither, and the fix is the same for both.

How I evaluated these

The rankings below come from our own stylometric benchmark rather than from reading a few drafts and forming an impression. We generated 320 writing samples across five models, eight prompt types and four languages, then scored each sample with the deterministic formulas we use on human writing for Writing DNA Snapshots. The full method is in how we measure average AI, and the three-way results including Claude are in our ChatGPT vs Claude vs Gemini comparison.

For this post I pulled out the two models the search term names, GPT-5.2 and Gemini 3 Pro, and read the six dimensions through the lens of a blog post rather than an email or a report:

1. Sentence complexity. How nested and clause-heavy the sentences are. Blogs want less of this than reports.
2. Vocabulary richness. How often the model reaches for a specific word instead of reusing a general one.
3. Expressiveness. Rhetorical questions, exclamation marks, attitude markers such as "importantly". Most of what readers call "personality".
4. Formality. Function-word density, hedging and semicolons, the markers that make prose read like a white paper.
5. Consistency. How uniform sentence lengths are. Low consistency means a bursty rhythm, which most good blog writing has.
6. Conciseness. Words per idea, where no model scores well; the overall average was 42 out of 100.

Two limits apply. The samples used plain prompts with no style instructions, because the point was each model's default. And a stylometric score describes the shape of prose, not whether the argument holds.

Where ChatGPT wins for blog drafts

ChatGPT's strength is energy. GPT-5.2 scored highest on expressiveness by a wide margin, with more rhetorical questions, more emphatic punctuation and more attitude markers than Gemini or Claude. It also scored lowest on consistency, which sounds like a flaw until you remember what a blog reads like: a short punch, then a longer explanatory sentence, then another punch. That variation is GPT-5.2's default, and it is why a ChatGPT draft of an opinion post often feels closer to finished than a Gemini draft of the same brief. It was also the most concise and the least formal of the three.

The weakness is the same trait pushed too far. The expressiveness that makes a hook land also produces three rhetorical questions in the first two paragraphs and a closing line that sounds like a poster. Readers learn the ChatGPT cadence within a few posts, and once they recognise it, everything on the blog reads as machine-assisted. Our post on why AI writing sounds generic lists the tells; several are simply GPT-5.2's expressiveness score made visible.

Where Gemini wins for blog drafts

Gemini's strength is structure. Gemini 3 Pro scored highest on formality and on sentence complexity, with heavy function-word use, careful hedging and a fondness for semicolons. For an explanatory post, a how-to with prerequisites, or a reference article people skim for one section, that discipline is an asset. Gemini drafts arrive with the sections in a sensible order and the caveats already attached, and Gems let you store a persona once and draft with it inside Docs.

The weakness is that Gemini was the least concise model we measured and the lowest on vocabulary richness. It reuses terms rather than varying them, and it pads sentences with qualifiers, so a 900-word brief comes back at 1,300 words with the energy of a compliance memo. It was also the least expressive of the three: accuracy without attitude, which is a hard combination to build an audience on. The setup guide for making Gemini write like you exists because that dryness is the first complaint bloggers have.

Which model sounds more like a human blogger?

Neither, in the sense that matters. Both models sit inside a narrow band that no individual writer occupies. The widest gap between any two models on any single dimension was roughly 16 points, while the typical gap between a model's default and a real person's writing was 20 to 40 points, on several dimensions at once.

Read that as a map rather than a scoreboard. If your published posts have short sentences, a bursty rhythm and visible opinion, ChatGPT starts closer on three dimensions. If your posts are long-sentenced, hedged and formal, Gemini starts closer. Either way, switching models trades one set of gaps for another:

The last column is the part a model choice cannot fix, and the rest of this post is about why.

The gap neither model closes

Model capability and user-data conditioning are separate layers, and a model comparison only ever looks at the first one. Capability is what the model can do with the instructions and text it is given: hold a register for 1,500 words, follow a long brief, keep a structure intact. OpenAI and Google improve that layer with every release. Conditioning is what the model is told about you in this session, and that layer is empty unless you fill it, because nothing you have ever published is in the weights.

This is why the two drafts feel different from each other and identical to everyone else's. GPT-5.2 has a house style and so does Gemini 3 Pro, and each applies it to every blogger who types the same brief. We covered the structural reason in why better models still do not know your voice: a lab can make its model sharper and cannot make it yours.

For a blog the cost is cumulative. One post in a house style passes. Twenty posts with the same hooks and the same closing cadence read as a feed, and readers who came for a person leave.

A tone label gets the same post from both models

Most bloggers try to fix this with an instruction: "Write in a conversational, expert tone, like a founder talking to peers." Both models follow it, and both produce a draft that could have come from any founder on the internet.

The instruction failed because AI does not know your tone; it knows words that describe tone. "Conversational" and "expert" point at a register each model learned from thousands of writers described the same way. Paste the identical sentence into ChatGPT and Gemini and you get two averages, one peppy and one formal, with no trace of you in either.

"Professional yet friendly" describes thousands of people, and so does every other adjective pair. That was the finding of our post on why tone adjectives are not your voice, and a blog is the sharpest case, because it is the one place readers expect to recognise a specific person.

What a blog voice is actually made of

A voice lives in sentence length, qualifier placement, transitions and punctuation. None of those are adjectives, and all of them are countable from posts you have already published. For a blog specifically, the countable features that carry recognition are:

• Opening move. A claim, an anecdote, a question or a number. Most writers use one move four times out of five.
• Sentence length distribution. Your typical sentence, your longest, and how often a one-line paragraph appears.
• Paragraph rhythm. Sentences per paragraph, and whether you break for emphasis or for topic.
• Qualifier placement. Hedging before a recommendation ("in most cases, ship it") or after ("ship it, with one caveat"), or not at all.
• Transitions. The two or three connectors you rely on, and the ones you never use.
• Punctuation habits. Colons versus parentheses, how you handle lists, whether semicolons appear at all.
• Closing move. A recommendation, a question back to the reader, a restatement, or nothing.

Two people can both be "conversational and expert" and differ on every item in that list. That is the difference between a label that describes thousands of people and a set of measurements that describes one. Our piece on how style extraction works walks through the full set of dimensions we measure.

Trained beats prompted for both tools

A prompt describes how you hope to sound. A profile built from your own writing is evidence of how you actually write, and that evidence is what turns either model's house style into yours. The process is the same whether you draft in ChatGPT or Gemini:

1. Collect ten to fifteen published posts from the blog in question, not from your email or your drafts folder.
2. Measure them for sentence length, paragraph rhythm, opening and closing moves, qualifier placement, transitions and punctuation. Do it by hand for a first pass, or let the free Writing DNA Snapshot score a sample for you.
3. Write the measurements as instructions. "Typical sentence 14 to 18 words, maximum 32. Open with a claim, never a question. One-sentence paragraphs for emphasis, at most two per post. Hedge before recommendations, never after. Never 'however' sentence-initial. Colons for lists, no semicolons."
4. Store them once per tool. The next section covers where.
5. Check output against the numbers. A draft either hits a sentence-length range or it does not, which is a faster check than asking whether it "feels right".

The measured version has a second advantage over the adjective version: it improves with the model. When a lab ships a release that follows nuanced instructions more faithfully, a profile written as numbers gets executed more precisely, while a tone paragraph stays exactly as vague as the day you wrote it.

Setting up the profile in ChatGPT and in Gemini

Both tools have a place to store standing instructions, and the profile belongs there rather than in the chat window each time.

In ChatGPT. Custom instructions apply to every conversation and suit a profile you use for all writing. A Custom GPT or a Project suits a profile specific to one blog. Put the measured instructions in the instructions field and keep sample posts as knowledge files for reference only; the model retrieves and paraphrases those rather than conditioning on them. Our guide to ChatGPT Projects for writers covers the layout.

In Gemini. A Gem is the equivalent of a Custom GPT: a saved persona whose instructions persist across sessions. Google's guidance for creating custom Gems recommends a specific role and detailed instructions, which is exactly what a measured profile is. Saved info in Gemini personalization holds facts about you, such as your audience and niche, and is not the place for the style rules themselves.

Because the profile is a document rather than a setting inside one product, the same block goes into both. Our post on deploying a Writing Twin across AI platforms shows the paste targets for ChatGPT, Gemini, Claude and Copilot. One asset, every tool.

Our own weakness belongs here too. A Style Profile needs real samples, at least three and ideally more, so a brand-new blog with two posts has nothing to measure yet. Setup takes an afternoon, not a minute. And a profile shapes how a post reads, not whether its argument holds; a weak idea in your exact voice is still a weak idea.

Can better models make this unnecessary?

No, and the direction of the effect is the opposite of what most people expect. Each new ChatGPT or Gemini release improves the capability layer: longer context, fewer drifts over a long draft, more faithful execution of detailed instructions. None of those changes add a sentence of yours to the model, because the conditioning layer is the one thing a lab cannot ship.

What a better model does change is how well it executes whatever you give it. Adjectives get you a more polished version of the same average. Measurements get you a draft that hits your sentence-length range on the first pass instead of the third. The better the model gets at following nuanced instructions, the better it executes your profile too, which is why a measured profile becomes more valuable with every release.

That also settles the ChatGPT vs Gemini question for anyone who plans to blog for more than a few months. The models will leapfrog each other. The profile built from your own posts works in whichever one is ahead this quarter.

A five-minute test before you pick a model

Before choosing between the two on feel, run this check on both.

1. Take a post you already published. Write the one-paragraph brief you would have given a colleague to produce it.
2. Give that brief to ChatGPT and to Gemini with your current tone instructions, and nothing else.
3. Put each draft beside your original and count five things: words per sentence, sentences per paragraph, the opening move, how many hedges and where they sit, and the number of rhetorical questions.
4. Whichever draft is closer on more counts is your better starting point. Expect it to still miss on at least two of the five.
5. Write the five measurements from your original as instructions, run the brief again on both models, and count again.

The second run is usually where the gap between the two models shrinks to noise and the gap to you closes for the first time. The tool comparison tells you where to start; the measurements get you home.

Frequently asked questions

Is ChatGPT or Gemini better for blog writing? Neither is better in general. ChatGPT scored highest on expressiveness and was slightly more concise, which suits opinion pieces and conversational posts. Gemini scored highest on formality and sentence complexity, which suits explanatory and reference posts. The widest gap between them on any dimension was about 16 points, while the gap between either model and an individual writer is typically 20 to 40 points across several dimensions at once.

Why does my post sound the same whether I use ChatGPT or Gemini? Because both models received the same instruction, and the instruction contained nothing about you. A tone label such as "conversational but authoritative" selects a register each model has already learned, and that register is an average of every writer described the same way. Your voice lives in sentence length, qualifier placement, transitions and punctuation, and neither model has those numbers unless you supply them.

Will a newer model finally write blog posts in my voice? No, because model capability and user-data conditioning are separate layers. A new release improves the first: longer instructions followed more faithfully, fewer drifts over a long draft. Nothing about you enters the weights, since your posts were never in the training data. The useful effect runs the other way: a stronger model executes a measured profile more precisely than the previous version did.

How do I make ChatGPT or Gemini write posts that sound like me? Replace adjectives with measurements. Take ten posts you actually published and count typical and maximum sentence length, sentences per paragraph, your opening move, where hedges sit, which transitions you use and never use, and your punctuation habits. Write those as instructions and store them once: in ChatGPT custom instructions, a Custom GPT or a Project, and in a Gemini Gem. A Style Profile built from your samples does the measuring for you.

Run the free Writing DNA Snapshot on one of your own posts to see the measurements ChatGPT and Gemini are both missing, then paste the profile into whichever model you draft in.