Paste your ad copy, email sequence or sales page into the box below and get an AI likelihood score in a few seconds. Three checks a day at up to 1,000 words each, no account and no card. A free account takes that to sixty checks a day at 5,000 words.
The score is an estimate of how machine-written the text reads to our model. It is not proof of who wrote it, and this page will tell you where it gets things wrong.
Paste some text or load the example to estimate AI likelihood.
The gauge gives one figure: how much of the paste reads as machine-written to the model. The bands are fixed and the tool labels them for you.
The copy reads as human across the whole paste. Nothing to act on.
This is the band most conversion copy lands in, and it is the least useful result you can get. Direct response is full of slots that are formulaic by design: a mandated guarantee line, a legal disclosure, a compliance sign-off, a stock CTA the brand will not let you change. Those score the same whoever wrote them.
The fix is to stop scanning the page as one block. Run the hook on its own, then the body, then the close. Three numbers across three sections tell you something. One number across three thousand words tells you almost nothing, because a strong body will average out a weak opener and you will never see it.
Worth reading again before it goes to a client. Look for the flat stretch: a run of sentences that are all roughly the same length, category nouns where a specific one would do, benefit claims with no number in them. That is usually where the score is coming from, and it is usually a section written late.
The result also reports how many words it scored and how long the scan took, so you can tell at a glance whether it read the whole paste or hit the size cap. The counter above the box tracks that cap as you type. It reads 0 / 1000 words without an account and 0 / 5000 once you are signed in. Anonymous scans are not stored anywhere. Signing in saves them to a history.
Four things, stated plainly, because you are going to put this number in front of a client and you should know where it breaks.
A detector measures how text reads, not where it came from. A high score on copy you wrote yourself means the copy reads flat, not that you cheated. A low score on machine-written copy means it reads varied, not that a person wrote it. Treat every result as a prompt to reread, never as a verdict.
The tool needs at least 25 words and it is more stable well above that. A Meta headline, a 30 character Google RSA headline and an email subject line are all below the floor. Paste the whole ad set at once, or the whole seven email sequence, and read one number for the batch. Scoring a single headline gives you a number, but it is noise dressed up as a measurement.
The model is an English-tuned ensemble (DeBERTa, RoBERTa and ELECTRA). Running Spanish, French or German copy through it will return a number and that number means nothing. If you write in more than one language, only the English deliverables are worth scanning here.
This is a known weakness of English AI detectors as a class and ours has it too. ESL writing tends to use a smaller, more predictable vocabulary and more even sentence rhythm, which is exactly the pattern these models were trained to read as machine-written. If English is not your first language, expect your own unaided writing to score higher than a native speaker's comparable copy. We would rather say that here than have you find out in a client review.
Most people land on this page because someone else ran a check, not because they wanted to. Four moves, in order.
Detectors do not share a model or a threshold, so two tools can return different numbers on the same paragraph, and the same tool can read a 200 word excerpt differently from the full page it came from. Before you rewrite anything, find out which tool produced the number and whether it was run on the whole deliverable or one section a reviewer pasted in.
The strongest thing you can hand a client is not a second opinion from a second detector. It is evidence of how the copy was made: the document version history, the brief, the research file, the call recording or review mine the phrases came from, the dated drafts. A Google Docs version history is more persuasive than any score, because it shows the work rather than measuring the output.
Once you have the process evidence, a scan is useful supporting material. Send the number with one sentence of context: which tool, what it measures, and that it is an estimate. A copywriter who explains what a detector actually does reads like a professional. One who forwards a screenshot of a green number reads like someone hoping it settles the argument.
If a client asks for a minimum detector score as a delivery condition, push back. You would be tying payment to a third party model that can change without notice, that scores your formats unreliably, and that penalises some writers for how they learned English. Offer a process commitment instead: research led drafting, named sources, version history available on request. That is something you can actually guarantee.
The model is reading flatness. So is your reader, which is the useful part: the edits that lower the number are edits a good creative director would ask for anyway.
Machine-written prose settles into a uniform, medium length rhythm. Strong direct response never does. A one word line, a long benefit run, a fragment for the turn. If you read your copy aloud and it moves at one speed the whole way, that is the flatness the score is picking up.
"Drive growth", "maximise value" and "boost results" are the interchangeable middle of a thousand landing pages. "Cut onboarding from twelve days to three" is not. Specific vocabulary is the single most reliable difference between copy that reads written and copy that reads generated, and it is also what makes the claim believable.
Language pulled from review mining, sales call recordings and support tickets is so particular to one buyer that no model would produce it by default. The swipe file move that wins on conversion is the same one that lowers the score. Doing the research twice is not required.
Unlock, Discover, Transform, Elevate, Revolutionise, Master. These sit in the most predictable band of machine defaults, which is also why they have stopped working on readers. Write the opener last, after the offer and the proof are settled, so it has something specific to carry.
This is a measurement, not a guarantee. Editing for cadence and specificity usually lowers the score, because the score is reading those exact properties. It is not a switch, the result varies by format and length, and any tool that promises you a particular outcome on someone else's detector is selling you something it cannot deliver. Rewriting a tested headline purely to move a percentage is also how writers flatten the line that was carrying the click through. The score is a diagnostic. It is never the target.
How to scan each thing so the number means something.
Three to seven lines plus a headline is under the band where the model is stable. Paste the whole ad set, five to ten variants together, and read one number for the set.
Thirty character headlines and ninety character descriptions are far too short on their own. Scan all fifteen headline variants in one paste.
Same problem, same fix: paste the full five to seven email flow rather than one email. Subject lines and opening lines are the slots that drift into templated phrasing first, so read those two rows carefully once you have the number for the sequence.
Long enough to score reliably, long enough that one number hides the weak part. Scan the full page, then rescan the hook and the close on their own. Those two sections are usually written under the most deadline pressure. A full page runs past the 1,000 word anonymous cap, so either sign in free for 5,000 or scan it in sections.
Spoken word scripts score differently from written prose because the cadence is different: contractions, fragments, one word lines. A script transcribed from a real interview and a script written from an outline do not read the same way, to the model or to the viewer.
What you get without signing up, what a free account adds, and what the paid plans cost.
Three checks a day, up to 1,000 words each. No card, no email. The three are shared with the document detector on this site, so it is three checks in total across both, not three each. Anonymous scans are not stored.
Sixty checks a day, up to 5,000 words each, saved to a history you can go back to. That is the largest step on this page and it costs nothing: twenty times the checks and five times the paste size, which is the difference between scanning a sales page in sections and scanning it in one go.
Starter is $9.99 a month, or $7.49 on annual billing, and raises the daily limit to 200 checks. Pro $19.99 ($14.99 annual), Business $39.99 ($29.99) and Enterprise $79.99 ($59.99) remove the daily limit entirely and add per sentence highlighting in the app, which tells you which line is pulling the number down rather than just that something is. The rewriter runs 20,000 words a month on Starter, 50,000 on Pro, 100,000 on Business and 150,000 on Enterprise. Full detail on the pricing page.
More for copywriters.
The full content-writer workflow with delivery-attached scans and brand voice defence.
For writers →How agency teams running fifty-plus pieces a month build the scan into their QA workflow.
For agencies →Light, Balanced, and Maximum modes for fixing flagged passages without losing voice.
Read the guide →Free, Starter, Pro, Business. Yearly billing saves 25%. Solo to agency tiers.
See pricing →Headlines that match what the piece actually says, with no clickbait padding.
Generate titles →Scroll back up, paste the draft, read the number, then read the copy again. Three checks a day without an account. If a passage genuinely needs reworking, the rewriter is on the next page over.
How TextSight fits other teams and workflows.