How to Check AI Content for Plagiarism Before You Publish

Run the finished text through a plagiarism checker that searches the web, such as Copyscape Premium (3 cents for up to 200 words), Originality.ai or the checker in Grammarly Pro, then open every match and decide whether to quote and credit it, rewrite it, or cut it. An AI detector answers a different question and proves nothing about copying.

The checker's blind spots matter more than its score. It compares wording against pages it can find, so it catches copied sentences but not copied ideas, page structures, paywalled sources or print. And AI rarely lifts long passages on its own initiative. The usual culprit is close paraphrase of whatever someone pasted into the prompt as "inspiration", which is also the hardest case for a checker to flag.

Follow me on Instagram@sagnikteaches

What a checker compares, and what it can't see

A plagiarism checker breaks your text into fragments, searches for them in an index of web pages (Grammarly adds academic papers and published works), and lists each source that shares wording with yours, with the overlap highlighted. It's good at a narrow job:

Connect on LinkedInSagnik Bhattacharya

• Sentences copied word for word from a public page.
• Sentences copied with a few words swapped.
• A paragraph of yours that already appears on another site, whether you copied it or someone copied you.

It is weak or blind on everything else:

Subscribe on YouTube@codingliquids

• Copied ideas and structure. A services page that follows a competitor's sections, order and selling points in fresh words scores as original.
• Heavy paraphrase. Once most of the words change, matches drop below the tool's threshold.
• Sources it can't reach. Paywalled reports, PDFs behind log-ins, print books, a client's internal documents.
• Anything that isn't text. Images, layouts, code and design.
• Very recent pages. A page published yesterday may not be in the index yet.

You can learn where your checker's threshold sits in ten minutes. Take a paragraph from a public page, make three versions (unchanged, every fifth word swapped, and fully reworded with the same points in the same order) and run all three. Typically the first is flagged in full, the second in fragments, and the third not at all. That third result is the one to remember, because it's exactly what a "rewrite this in our style" prompt produces.

An AI detector is a separate tool doing a separate job. It estimates whether text was machine-written from statistical patterns in the wording. That tells you nothing about copying, and it isn't reliable evidence of AI authorship either. OpenAI withdrew its own AI-text classifier on 20 July 2023, citing its low rate of accuracy; at launch it had correctly flagged 26% of AI-written text while wrongly labelling human writing as AI 9% of the time. Newer detectors claim better figures, but a detector score is still a probability, not proof. Tools sold to push that score down are a different purchase again, weighed up in whether an AI humaniser tool is worth paying for.

Where AI copy picks up borrowed wording

In a web design studio that drafts client copy with AI, borrowed wording tends to arrive by one of five routes. Knowing them tells you where to look hardest.

1. Pasted competitor copy. "Write ours in the style of this" is the single biggest source of close paraphrase.
2. Supplier and manufacturer text. Product descriptions from a supplier's feed are identical on every retailer's site, and AI polishing often leaves whole sentences intact.
3. Well-known lines. Slogans, famous quotations and stock phrases can come back word for word, because they appear so often in training data.
4. Quotes and statistics carried with their original sentence. Ask for "a statistic about mobile browsing" and you may get someone else's sentence around it.
5. Your own earlier work. The same prompt template used for two clients produces two pages that read alike, which a web check won't catch unless one is already live.

The first route is worth seeing in action. A typical prompt:

Here's the services page of the best studio in our niche. Write our
services page in the same style: [pasted competitor page]

An illustrative line from the output, next to the competitor's original:

AI draft: "We craft pixel-perfect websites that turn visitors into loyal customers, built on a foundation of strategy, design and code."

Competitor's page: "We craft pixel-perfect websites that turn visitors into loyal customers, built on strategy, design and code."

Seventeen words in sequence are identical, and the AI would describe the draft as "inspired by" the example. A checker will flag this one, because the competitor's page is public. It won't flag the next paragraph, where the AI kept the competitor's three selling points in the same order and changed every word. The fix is upstream: don't paste competitor copy at all. Describe what you liked about it in your own words ("short sections, one proof point per service, prices shown") and give the AI your facts instead.

Write the services page for [studio]. Use only these facts:
- Services: brochure sites, online shops, care plans
- Typical build: 4 to 6 weeks; we show fixed prices on the page
- Proof points: 11 years trading; builds in WordPress and Shopify;
  one named client quote (supplied below, approved for use)
Structure: short sections, one proof point per service.
Don't use slogans. If a sentence sounds like something any studio
could say, replace it with a specific fact from the list.

The output from a prompt like this tends to be plainer, and it's yours: the only words it can lean on are the facts you gave it.

Three checkers, their prices and what each report shows

These are list prices checked in September 2026; confirm on each vendor's page before you buy, because pricing in this category changes often.

Each report lists matching sources and highlights the overlapping text, but the headline percentages aren't comparable between tools, because each counts overlap differently. Pick one checker, learn how its reports look on your own clean copy, and stick with it; the triage rules below matter more than the brand.

A quick sum shows how small the per-page cost is. A 650-word page on Copyscape costs 3 cents for the first 200 words and 5 cents for the next 450 (five blocks of 100 or part of 100), so 8 cents a check. On Originality.ai, the same page uses 13 of your 2,000 monthly credits for a combined scan. Cost is never the reason to skip the check.

The pre-publish routine, step by step

1. Check the final version, after every edit. A check run on the first draft says nothing about the paragraph a colleague pasted in later. (1 minute to confirm the version.)
2. Check the whole page. Include headings, FAQs, captions and any product text pulled from a supplier feed, not just the body. (2 minutes to assemble.)
3. Run the checker and save the report. Keep a PDF or screenshot named with the page and version number. (1 to 2 minutes.)
4. Triage every match using the table below. (3 to 8 minutes a page.)
5. Fix, then re-check only what changed. Rewrites can land on another source's wording too.
6. Record the result in the sign-off. One line is enough: "Copyscape, v4, no matches over 8 words outside credited quotes, report attached." If you run a formal sign-off, this line belongs next to the fact check; the draft, check and sign-off workflow shows where.

One shortcut to avoid: asking the chatbot that wrote the draft whether it's original. A prompt like "Check this text for plagiarism" typically gets a confident reply along the lines of "This text appears to be original and does not contain copied passages." The model hasn't compared anything against an index of pages; even with web search switched on, it runs a handful of searches, not a fragment-by-fragment comparison. In the competitor example above, that reassurance would have gone out alongside a 17-word copied sentence. Use the AI to rewrite a flagged passage from your facts, and use a checker to find the passage.

When the report comes back, sort each match with this table:

Reading a match report without panicking

Here is an illustrative report for a services page the studio drafted for an accountancy client, 780 words, with an overall similarity of 14%:

MATCH REPORT  services page v3  (illustrative)
1. 7 words  "cloud accounting software for small businesses"
                                                  [software maker's page]
2. 6 words  "fixed monthly fee with no surprises" [3 accountancy sites]
3. 31 words "Our cloud bookkeeping service means you always know where
            you stand, with real-time figures available whenever you
            need them, wherever you are."          [competitor's page]
4. 24 words client testimonial, in quotes, named   [client's review page]
5. 12 words "We take the stress out of the numbers so you can focus
            on running your business"              [14 sites]

Match 1 is a product category that software makers all use: ignore. Match 2 is a short stock phrase: ignore. Match 4 is the client's own customer, quoted and credited: keep. Matches 3 and 5 need work. Match 5 is a cliché found on 14 sites; nobody will sue, but it tells a reader the page could belong to anyone. Match 3 is a real problem: 31 consecutive words from one competitor, which came from a prompt that included that competitor's page.

Before and after for match 3:

Before: Our cloud bookkeeping service means you always know where you stand, with real-time figures available whenever you need them, wherever you are.

After: We reconcile your bank feeds every Monday, so the figures you see on Tuesday morning are no more than a few days old. Your dashboard shows cash in the bank, unpaid invoices and next month's bills on one screen.

The rewrite works because it's built from facts only this client can state: a weekly routine, a named day, three specific numbers on a screen. It also reads better, which is the usual side effect of fixing plagiarism properly rather than shuffling synonyms.

Two client sites, one prompt template

The route a web check misses is duplication across your own work. Say the studio builds sites for two accountancy practices a month apart, using the same well-tuned prompt for the About and Services pages. Neither page copies the web, so both pass a normal check. But when the second site goes live, the two pages share long runs of wording, and the second practice's owner eventually spots phrases from her page on a competitor's site that the studio also built. From her side, it looks like her copy was reused.

Copyscape's private index exists for this. Add every page you've written for clients to the index, at a cent per 1,000 words, and each new check compares against your own archive as well as the web. Thirty client sites averaging 5,000 words each is 150,000 words, or $1.50 to index. The prevention is cheaper still: put three client-specific facts into every prompt (how long they've traded, one thing they do differently, one real number) and templates stop producing twins.

Supplier text is the same problem at a larger scale. An online shop that lets AI "tidy up" 200 product descriptions from a supplier's feed may end up with pages that match dozens of other retailers. That's rarely a legal issue, since suppliers usually expect their text to be reused, but it's weak copy and gives search engines no reason to prefer your page. If you're building pages from scratch, writing your website copy with AI covers how to feed the AI your own material instead. If the pages are already live, auditing your website for AI content that needs fixing covers finding and prioritising them.

When a client's AI detector flags copy you delivered

A situation many studios and agencies will recognise: a client pastes your delivered copy into a free AI detector, gets a score of "87% AI-generated", and asks whether they've paid for machine writing. The answer rests on three things you should already have: a contract clause that says how you use AI, a plagiarism report showing the text is original, and your editing record showing the human work. A reply that uses all three:

Hi [first name],

Thanks for flagging this. As our agreement sets out, we use AI to
produce first drafts, and every page is then rewritten and checked
by our team against the facts you gave us. The services page went
through three rounds of editing (versions attached).

AI detectors estimate how "machine-like" wording is; they don't
show whether text is original or accurate, and they misfire on
plain, clear writing. OpenAI withdrew its own detector in 2023
because of its low accuracy. The originality check we ran before
delivery (report attached) found no copied passages.

If any part of the page doesn't sound like you, tell us which lines
and we'll rework them with you on a short call.

[first name]

If your contracts don't yet say how you use AI, that's the gap to close first; AI clauses for agency contracts has wording for disclosure and ownership. Avoid the temptation to run the copy through a humaniser until the score drops. It doesn't make the work any more original, and it tends to make it worse.

A 12-page launch, checked for about a dollar

Here's the routine on a real-sized job, an illustrative client launch: 12 pages averaging 650 words, 7,800 words in all, drafted with AI from a fact sheet and edited by the studio's copywriter.

• First pass. Twelve Copyscape searches at 8 cents each: 96 cents. Time to assemble pages and run the batch: about 15 minutes.
• What came back. Nine pages were clean apart from stock phrases. Two had a 20-to-30-word match each with the same competitor, traced to a single prompt that had included its About page. One product page matched the supplier's description on 11 other sites.
• Triage and rewriting. About 50 minutes: two passages rewritten from client facts, the product page rebuilt around the client's own photos and sizing notes.
• Re-check. Three pages at 8 cents: 24 cents. All clean.
• Record. Reports saved against version numbers; one line in each page's sign-off.

Total tool cost: $1.20. Total time: a little over an hour for a launch that took the studio several weeks to build. On Originality.ai the same launch would use 156 credits for combined scans of the first pass, well inside one month's Pro allowance, with a few more for the re-check. Either way, the cost is trivial next to rewriting a live page after a competitor's complaint, or explaining to a client why their homepage shares a sentence with a rival's.

Plagiarism checking questions

Is AI-generated text plagiarism in itself?

No. Plagiarism means presenting someone else's words or ideas as your own, and text an AI drafted for you isn't another person's work in that sense. What can still get you into trouble is a contract or a publisher that requires you to disclose AI use, or AI output that reproduces someone else's wording closely. Check your agreements as well as the text.

What match percentage is acceptable?

There isn't a safe number. A 3% match in a long page can be one whole copied paragraph, while a 20% match can be properly credited quotations and product names. Judge each match on its type and length, and ignore the headline percentage, which isn't comparable between tools anyway because each one counts overlaps differently.

Is it safe to put confidential client copy into an online checker?

Read the checker's terms on what it stores before you paste anything unpublished. Some services keep submitted text or offer to add it to an index, which may not suit embargoed or confidential material. For sensitive launches, check with a paid account whose terms you have read, and remove anything confidential that isn't needed for the check.

Do AI humaniser tools help with plagiarism checks?

No. Humanisers rewrite text so AI detectors score it lower. They don't make copied wording or copied ideas any more acceptable, and heavy automated rewording often introduces errors and awkward phrasing that a reader notices. Fix a plagiarism match by writing from your own facts, not by running the text through another rewriting tool.

Further reads

• How to Write an AI Content Policy for Your Marketing — Make the plagiarism check a written rule, with a word threshold.
• Grammarly vs ChatGPT for Checking Business Writing — Compare the two tools for the editing work around the check.
• How to Write Website Articles With AI That Still Rank on Google — Write AI-assisted articles that are original enough to rank.
• Should You Design Your Logo With AI? Copyright and Quality Risks — The copyright questions for images, which text checkers can't see.
• How to Check Sources and Citations in AI Research — Verify the sources behind any quotes you keep.
• How Much Does Human Editing of AI Content Cost? — Budget the editing time that plagiarism fixes add.
• AI Content Workflow for Agencies: From Brief to Approved Draft — Six gates from client brief to approved draft, with the AI's job and the human's job at each, plus a filled-in brief, outline prompt and claims log.
• How Tutors Should Check AI-Made Worksheets Before Using Them — The ten-minute, four-pass check that catches wrong answer keys, two-answer questions, off-level reading and garbled diagrams before a pupil sees them.
• How to Check AI-Generated Images for Errors Before Posting — An eight-point checklist and four viewing tricks for catching AI image mistakes, with a filled-in check and a campaign of eight images put through it.
• How to Choose a Marketing Agency That Uses AI Responsibly — Seven questions that separate careful agencies from careless ones, a test brief for the pitch, and how a wedding planner compared three agencies.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Copyscape Premium pages (pricing, batch search, private index); Originality.ai pricing page; Grammarly Support (Grammarly Pro pricing, Plagiarism Checker user guide); OpenAI's note withdrawing its AI text classifier (July 2023).