Pre-Publish Trust Checks

Summary#

Pre-Publish Trust Checks is two tools behind one sidebar entry: a Plagiarism Checker that searches the web for passages of your draft that already exist word for word, and a Fact Checker that pulls the factual claims out of your draft and verifies each one against live sources with citations. Their internal tool ids are originality_check and fact_check.

Purpose#

The two things most likely to damage a publication are text that is not yours and claims that are not true. Both are cheap to check before publishing and expensive to fix afterwards.

These checks exist so that the last step before publishing is evidence rather than confidence. The decision each supports is a go/no-go: is this passage original enough to publish, and is this claim safe to put in print?

Overview#

Both tools work the same way. You paste a draft into a large text box, the word counter updates as you type, and you click the run button. The result replaces the three information cards below with a full report.

Plagiarism Checker samples eight distinctive sentences from your text and searches each one on Google as an exact phrase. A quoted match means a verbatim copy exists somewhere. It reports the percentage of sampled passages that matched, ranks the sites those matches came from, and shows you the exact sentences that overlap. Separately, a language model estimates how likely the whole text was machine-written.

Fact Checker asks a language model to extract up to six specific, checkable claims — statistics, dates, named events, quantities, attributions — and skips opinions and generic advice. Each claim then goes to Perplexity, which answers with a one-word verdict, an explanation and citations. The report marks every claim Supported, Disputed or Unverified.

Switch between the two with the tab strip at the top: Plagiarism and Fact Check.

Benefits#

• Evidence you can act on: the exact sentences that matched and the exact sites they matched on.
• Every fact-check verdict links to the source it came from.
• An AI-likelihood read on the same run as the plagiarism scan.
• Clear thresholds rather than a vague score — the verdicts are defined, not impressionistic.
• Both accept pasted text, so a draft in any editor can be checked.

Use Cases#

Clearing a freelancer's first delivery. Run the plagiarism check before you pay for it, not after it is live.

Checking an AI-assisted draft. Generated copy is confidently wrong about numbers more often than about anything else, and Fact Checker finds exactly those.

Publishing a data-led post. Every statistic you quote gets a verdict and a link before the post goes out.

Auditing an old page you inherited. Paste a page you did not write and find out what it copied and what it claims.

Protecting a client relationship. A Disputed claim caught in review is a correction; the same claim published is an apology.

Requirements#

• A signed-in account. Signed out, the screen shows only Sign in to use this tool.
• A paid plan. Each check costs 3 credits, and any cost above zero is refused on Free.
• Available credits within your monthly quota.
• At least 25 words of text.
• Plagiarism Checker needs DataForSEO credentials configured on the deployment; Fact Checker needs a Perplexity key. Both fail with a clear message if not.

Permissions#

Any workspace member on a paid plan can run either check. Free accounts are refused with This tool needs a paid plan. Free includes the 10 technical SEO tools; upgrade to Pro to unlock the rest. Because these checks cost 3 credits, they are metered against your monthly quota and can also be refused with quota_exceeded when it is used up.

Cost#

3 credits per run, for each check. Running both on the same draft costs 6.

Three is the premium threshold, so unlike most tools these checks are counted against your monthly quota. There is no cache: every run is a fresh set of searches, and nothing is served from a stored result.

Neither run button shows a credit chip, so the cost is not visible on screen before you click. See How credits work and Quotas and rate limits.

Navigation Path#

Dashboard → Write & Optimize → Pre-Publish Trust Checks

Then use the tab strip: Plagiarism (the default) or Fact Check.

Inputs#

Step-by-Step Guide#

1. Open Write & Optimize and choose Pre-Publish Trust Checks.
2. Choose the Plagiarism or Fact Check tab.
3. Paste your draft into the text box. Watch the word counter clear 25.
4. Click Check originality or Run fact-check. The button disables while the run is in progress.
5. Wait for Searching the web for matching passages… or Extracting claims and verifying against sources….
6. Read the score ring and the verdict banner first, then work down the evidence.
7. Fix the flagged passages or claims in your draft, then re-run if the changes were substantial.

Reading the Results#

Plagiarism Checker#

The score ring. A large percentage with the caption duplicate. This is the share of the eight sampled passages that were found verbatim elsewhere — not the share of your document. Eight sampled sentences means the ring can only ever show 0, 13, 25, 38, 50, 63, 75, 88 or 100.

The verdict banner, keyed to that percentage:

The four stat tiles. unique content (100 minus the duplicate percentage), passages matched, passages checked (normally 8) and matching sites. If passages checked is lower than 8, your draft was short or had few distinct sentences to sample.

AI-likelihood. A badge, a meter and a line reading N% estimated AI-written, plus a short note. The label follows the score: Likely AI at 70 or above, Mixed signals from 40 to 69, Likely human below 40. Treat this as a signal, not a verdict — it is one model's estimate and it can be wrong in both directions. What it is genuinely useful for is triage: a Likely AI read on a draft that was supposed to be original is a reason to ask questions.

If the block is missing entirely, the AI read failed. The plagiarism result is still valid; the estimate is best-effort and never blocks the check.

Matching sources. Every site a match was found on, ranked by how many of your passages it matched, with a favicon, a proportional bar and a hit count. Up to 12 sites. Read the top row first: one site with several hits is usually the original your text came from, while many sites with one hit each is often a common phrase, a quotation or boilerplate.

Matched passages. The exact sentences that matched, each followed by up to four clickable source chips. This is the section you act on. Open the sources, confirm whether the overlap is genuine copying, a quotation you should attribute, or an unavoidable stock phrase — then rewrite what needs rewriting.

With no matches you get a clean line instead: No verbatim matches were found for any of the sampled passages.

Fact Checker#

The score ring. The share of checked claims that came back supported, with the caption verified.

The verdict banner. Accuracy issues found in red whenever any claim is disputed; Partly verified in amber when nothing is disputed but something is unverified; All claims verified in green when everything is supported. The guidance line names the count, for example 2 claims appear to be inaccurate or contradicted by sources. Review and fix before publishing.

The four stat tiles. supported, disputed, unverified, claims checked, followed by a legend spelling out what each verdict means: Supported — backed by sources, Disputed — contradicted by sources, Unverified — no clear source found.

Claim by claim. One numbered block per claim. Each shows the claim as it was extracted from your draft, a one-sentence explanation of the verdict, up to five clickable citation chips, and the verdict badge.

Read them in this order:

• Disputed first. Something in your draft contradicts the sources. Fix or remove it before publishing.
• Unverified next. This is not the same as false. It means no clear source was found, which most often happens with very recent figures, internal data or claims phrased too loosely to verify. Either add your own source, or soften the claim so it does not assert more than you can support.
• Supported last, and only to collect the citations — they are exactly the links your article should be citing.

When there is nothing to check. If no specific factual claims were found you get No claims to verify and No specific factual claims were found to verify. That is a real result: opinion pieces and how-to guides often contain no checkable facts.

What to do with both. Run Plagiarism first — if the text is not yours, the facts do not matter yet. Then run Fact Check, fix every Disputed claim, resolve or soften every Unverified one, and add the supported citations to your page as outbound links. Then publish.

Examples#

Screenshots#

Tips#

• Check a section at a time on long pieces. Eight sampled sentences spread across 3,000 words is a thinner sample than eight across 600.
• Use Load an example once to see the shape of a full report before spending credits on your own text.
• Paste plain text. Markup and formatting characters add noise to the sampled sentences.
• Keep the citations from a fact check — they are the outbound links your page should carry.
• Re-run only after substantive edits. Each run is another 3 credits.

Best Practices#

• Run both checks as the last step before publishing, not the first step after.
• Treat Unverified as "needs a source", not as "false".
• Attribute rather than delete when a match is a legitimate quotation.
• Record the result in your editorial workflow so a piece cannot be published twice without a check.
• Pair with Live Writing Editor for readability and GEO Content Scorer for citability. Trust, clarity and quotability are three different checks.

Common Mistakes#

• Reading the duplicate percentage as a document score. It is the share of eight sampled sentences, not of your whole draft.
• Treating AI-likelihood as proof. It is an estimate from a single model.
• Assuming Unverified means false. It means no clear source was found.
• Checking after publishing. The value is entirely in catching it first.
• Running the same draft repeatedly while making small wording changes — that is 3 credits each time.

Limitations#

• Eight sampled sentences per plagiarism run, not a full-document comparison.
• Matching is verbatim, phrase-level. Paraphrased copying is not detected.
• Coverage is what Google's index returns for an exact-phrase search. Content behind logins, in PDFs that are not indexed, or in print is invisible.
• Up to six claims per fact-check run, so a long article needs several passes.
• Claim extraction is a model's judgement of what is checkable. It skips opinions and generic advice by design and can miss a claim you cared about.
• Up to five citations per claim.
• Results are not saved to the Library and there are no export buttons on these screens. Copy anything you need to keep.
• There is no Get recommendations panel on either report.
• Neither run button displays a credit chip, so the 3-credit cost is not shown before you click.

Troubleshooting#

FAQs#

How much of my document is actually checked? Eight distinctive sentences are sampled and searched as exact phrases. The duplicate percentage is the share of those eight that matched, which is why it moves in steps rather than smoothly.

Does this detect paraphrasing? No. It finds verbatim matches. A rewritten passage with the same meaning and different words will not be flagged.

How reliable is the AI-likelihood score? It is one language model's estimate of how likely the text was machine-written, returned as a 0–100 score with the label Likely AI, Mixed signals or Likely human. It is useful for triage and should never be used on its own to accuse anyone.

Why does the fact checker only look at six claims? Each claim is a separate live verification with citations, and six keeps a run inside a sensible time and cost. Split a long article and run it in sections.

Why 3 credits when most writing tools cost 1? Each run fans out into several search or verification calls. Anything at 3 or above is also metered against your monthly quota. See Credit cost table.

Can I export the report? Not from this screen — there are no export buttons and the result is not saved to the Library. Copy the matched passages or the citation links you need before navigating away.

Is my draft stored anywhere? The text is sent to the check endpoint to run the searches and verifications. The report is not written to your saved results. See What data we store.