Last updated: October 2026
The short answer: most people need two or three tools, not one. A general assistant — ChatGPT, Claude or Gemini — for drafting and restructuring; a checker such as Grammarly for the final edit; and, if you publish for search, a research suite such as Semrush to decide what a piece must cover. Choose between them on four things that do not change with each model release: whether the tool can be held to your sources, whether it keeps your voice, how long its drafts take to make publishable, and what the vendor does with your text. On that last point the defaults differ more than most people realise, and the comparison below quotes each vendor’s own help pages, checked on 16 October 2026.
A ranked list of AI writing tools is close to the least useful thing anyone can hand you now, and it is worth saying why before spending your time. Most of these products are interfaces over a small number of underlying models, several of them over the same ones. Ranking their prose is largely ranking the model of the month, which changes without notice and takes the ranking with it.
What genuinely separates them is everything except the prose: whether the tool can be held to facts, whether it will hold a voice, how much editing the output needs before it is publishable, and what happens to the text you put in. Those are stable, testable, and almost never compared. So that is what this page does.
The Only Productivity Number Worth Measuring
Every vendor in this category quotes output speed. Words per minute, drafts per hour, "ten times faster". The number is real and it is measuring the wrong thing, because first drafts were never the bottleneck. The bottleneck is the distance between a draft and something you are willing to publish under your own name.
Measure time to publishable instead: start the clock when you begin the prompt and stop it when the piece is ready to go out, including fact-checking, restructuring and rewriting the parts that read like a machine. That is the number that decides whether a tool pays for itself, and it is the number that occasionally comes out worse than writing it yourself.
It also has an uncomfortable property that vendors have no incentive to mention: a fluent, confident, subtly wrong draft can take longer to repair than a blank page, because you must first work out which parts are wrong. Fluency is not correlated with correctness and it actively disguises the absence of it.
Grounding: Can the Tool Be Held to Facts?
This is the largest real difference between products in the category, and it is the one buried under prose comparisons.
A general assistant generates plausible specifics. Numbers, dates, names, statistics, quotations and citations are exactly the output most likely to be confidently invented and least likely to be caught by an editor reading for tone, because a fabricated statistic looks identical to a real one. Anyone who has fact-checked machine-written copy has found a study that does not exist, attributed to authors who did not write it, in a journal that does.
The products worth paying more for are the ones that ground their output in a source you supplied and cite back to it, so a claim can be traced to a page rather than to a model's memory. CustomGPT.ai is built around that shape — an agent over documents you already own, a sitemap, a set of PDFs, a Drive or Notion workspace, answering with citations to the source page. That does not eliminate error, but it converts an unverifiable claim into a checkable one, which is a categorical improvement rather than a marginal one.
Practical rule regardless of tool: treat every specific as unverified until you have seen the source. Names, numbers, dates, quotes, laws, prices, study findings. Everything else in a draft is cheap to fix; these are the things that damage you when they are wrong.
Voice, and Why Most Attempts at It Fail
Brand voice features are usually implemented as a settings panel of adjectives — professional, friendly, authoritative. Adjectives are a weak instruction, because every model already believes it is being professional, friendly and authoritative.
What actually works is examples. Three to five pieces of your own best writing, supplied as reference, will do more than any tone slider. When evaluating a tool, the question is therefore not "does it have voice settings" but "can I give it my own writing as the definition of the voice, and does it persist across sessions or must I paste it every time." The second half of that question is where most products fall down.
The failure to look for is regression to house style: a draft that opens with "In today's fast-paced digital landscape", uses three-item lists everywhere, and closes by restating the introduction. That is the model's default voice defeating your instructions, and it is the single most reliable tell that a piece was not edited by a person who cared.
Originality, and What Google Actually Cares About
Two questions get conflated here and they deserve separating.
The first is whether the text is duplicated from somewhere else. That is a real risk, it is checkable, and it is worth checking for anything commercial — particularly where a model has reproduced a distinctive phrasing from its training data.
The second is whether the text is detectable as machine-written. This one deserves much less of your attention than it gets. We went through the published evidence on detection tools in detail in our piece on whether AI detectors are accurate, and the short version is that the category is not reliable enough to be treated as proof in either direction. Optimising your writing to defeat a detector is a poor use of effort.
What actually matters for search is a different question again, and we cover it in AI content versus human content: the relevant standard is whether the page is useful, original and demonstrably informed, not which tool typed it. For the research and gap-analysis side of that work, a content and competitor suite such as Semrush answers a question a writing tool cannot: what the page needs to cover to be worth publishing at all.
What Happens to Your Text
Underexamined and occasionally decisive, particularly if you write about clients, under NDA, or ahead of an announcement.
- Is your input used to train the vendor's models, and can that be turned off? Look for the documented setting, not the reassurance in the sales conversation.
- How long is content retained after deletion, and after cancellation?
- Which model providers sit behind the product? A tool is a pipe; the sub-processor list tells you where the other end is.
- Can you get your own work out in bulk, in a normal format, if you leave?
These have real answers in the vendor's documentation. If the answers are hard to find for a product you would put confidential material into, that is itself an answer.
Here is what four of the most-used writing tools say about training on your text, from their own help pages:
- ChatGPT (OpenAI). OpenAI’s help article on how data improves its models says content from individual services such as ChatGPT may be used for training, and that you can opt out. It also says OpenAI does not train on inputs or outputs from its business products — ChatGPT Business, ChatGPT Enterprise and the API — by default.
- Claude (Anthropic). For the consumer Free, Pro and Max plans, Anthropic says chats are used to improve Claude when you choose to allow it, through a Model Improvement setting in Privacy Settings, and that incognito chats are not used even with that setting on. Chats flagged for safety review are an exception it names. Ratings you give with thumbs up or down are stored for up to five years.
- Gemini (Google). With the Keep Activity setting on, Google says it uses your activity to provide, develop and improve its services, “including training generative AI models”, and that a subset of chats is read by human reviewers. Activity auto-deletes after 18 months by default (you can choose 3 or 36 months, or never). With Keep Activity off, chats are still kept for 72 hours, and chats that human reviewers have already seen are kept for up to three years.
- Grammarly. Grammarly’s support pages describe a Product Improvement and Training control that is on by default for individual users and covers using your content to train and validate its AI models. Free and Premium users can switch it off in their account’s privacy settings; on Pro accounts with more than one user, only an administrator can change it.
The practical reading: on consumer plans, assume your drafts can be used for training until you have found and changed the setting; on business plans, check that the default you are paying for is the one written into the contract. Do this before the first confidential brief goes in, not after.
A Test That Takes an Hour
Pick three real pieces you have to write anyway. Then, for each candidate tool:
- Give it your voice samples and a normal brief — the brief you would give a freelancer, not a prompt engineered for the demo.
- Time yourself to publishable. Everything: prompting, restructuring, fact-checking, rewriting. This is the number.
- Fact-check every specific in the draft. Count how many were wrong. Do this even on a topic you know well — especially then, because you will catch what a reader would not.
- Count the house-style tells. How many sentences did you rewrite purely because they sounded generated?
- Try the second piece without re-pasting your voice samples. Does it remember, or are you starting from scratch every time?
- Compare against writing one yourself. This is the control condition, it is nearly always skipped, and occasionally it wins.
Turning the test into a decision. The sum is simple once you have the two times. Hours saved per month = (your unaided time per piece − your time to publishable with the tool) × pieces per month. The tool pays if those hours, at what an hour of your time is worth, exceed the subscription. Take an illustrative case, not a measurement: a piece that takes you three hours alone and two with the tool, four times a month, saves four hours. If the subscription costs less than four hours of your time, keep it; if the saving shrinks to zero once fact-checking is counted, as it sometimes does, cancel it. Use your own measured times, never the vendor’s.
Where a Writing Tool Is the Wrong Purchase
Three cases, stated plainly on a page that carries affiliate links.
You do not know what to say. No tool fixes an absent argument. It will produce competent paragraphs around an empty centre, and that is worse than nothing because it looks finished. The fix is reporting, thinking or asking someone who knows.
The value is in the specifics only you have. Original data, customer conversations, your own results, an opinion informed by having done the thing. A model cannot generate what it has never seen, and the parts of your writing that are worth reading are usually exactly these.
The volume is low. If you write four pieces a month, the setup, the voice tuning and the fact-checking overhead may never be repaid. This is a scale tool, and there is no shame in the honest conclusion that your scale does not justify one.
What We Removed From This Article
This page previously carried a per-tool pricing list. Every figure has been deleted rather than updated. Prices in this category change frequently, tiers are renamed and restructured, and a stale price is the single most actionable wrong thing an article can contain — it is the number a reader budgets against. Check the vendor's own pricing page on the day you decide.
The same reasoning applies to the model names and version numbers that comparisons like this usually lean on. Anything pinned to a specific version is wrong within a quarter. Judge on what the tool does with your brief this week.
Two products we hold and can describe honestly by what they do rather than by a score: PopAi for document-centred work, and AiPPT where the output needs to be a slide deck rather than prose. Run your own three pieces through anything before paying for it, ours included.
Disclosure: this article contains affiliate links. If you sign up through them we may earn a commission at no extra cost to you. It does not change what we recommend, and no vendor has paid for a position here.
Frequently Asked Questions
Does ChatGPT use what I write to train its models?
On individual plans it may, unless you opt out, according to OpenAI’s help centre. OpenAI says it does not train on content from ChatGPT Business, ChatGPT Enterprise or the API by default.
Can I stop Gemini from using my chats?
Yes: turn off Keep Activity in Gemini Apps Activity. Google says chats are then still kept for 72 hours, and any chat a human reviewer has already seen can be kept for up to three years.
Does Grammarly use my writing to train AI?
By default, yes, for individual users. Grammarly’s Product Improvement and Training control covers training its AI models, and Free and Premium users can switch it off in their privacy settings.
Which AI writing tool is best for factual content?
One that grounds its answers in sources you supply and cites back to them, so every claim can be traced to a page. Whatever you use, treat every name, number, date and quotation as unverified until you have seen the source.
Will Google penalise content written with AI?
Google judges whether a page is helpful, original and informed rather than which tool typed it; we go through its guidance in AI content versus human content. Detection tools are not reliable enough to treat as proof either way.
Related Articles
