How to check AI content quality before you publish

You have a finished draft and a publish button. The question is whether the article is safe to put your name, or your client's name, on. This guide gives you six checks in order. Each has a pass/fail test, a fix for when it fails, and a note on what the check cannot tell you.

Short answer: what "good enough to publish" means

An AI-written article is good enough to publish when five things are true. It answers the query and fits the brief. Every claim traces to a source you can open. It adds something the sources alone do not. It sounds like the publisher, not like a template. And it clears a written pass mark.

Each of those can be tested. "Reads well" cannot, which is why gut feel lets weak drafts through on a busy day.

One thing is not on the list: an AI-detector score. A 2023 study of several widely used detectors found that they consistently misclassified non-native English writing as AI-generated, and that simple prompting strategies could bypass them [1]. A detector guesses at how a text was produced. It says nothing about whether the text is accurate, useful or original. We cover this in Check 3.

Before you check: have the brief and the sources in hand

You cannot judge a draft without the documents it was written from. If you have only the draft, you can proofread it, but you cannot QA it. Gather four things first:

• The target query and the intent behind it. Is the reader trying to learn, compare or buy?
• The angle. The one thing this article says that the pages already ranking do not.
• The must-answer questions. The specific questions the reader typed or would ask next.
• The source list. Every source the writer was allowed to use, with the passages it drew on.

If the brief lacks any of these, fix the brief before you check the draft. A draft written to a vague brief will pass a vague review. For agencies reselling articles, this is the step most often skipped, because the brief lives in someone else's head.

Check 1: Does it answer the query and match the brief?

Intent and must-answer questions

Test: write the must-answer questions in a list. Next to each, put the heading and paragraph where the draft answers it. Any question with no answer is a fail. So is any question answered only by a vague sentence such as "it depends on your needs" with no statement of what it depends on.

Read the first 100 words as the reader would. They should know within those words that the article addresses their situation and what the specific answer is. If the opening defines the topic or describes how fast the world is changing, the draft is padding before it starts.

If it fails: send the draft back with the unanswered questions listed. Do not patch with a new paragraph at the end. An answer buried in the last section is easy to miss and shows the draft was not planned around the question.

Angle: what it says that top results do not

Test: open the angle line from the brief. Find the sentence in the draft that delivers it. If you can find it only by squinting, the draft has drifted to the generic version of the topic.

A useful check is to cover the title and ask whether the article could sit on any competitor's blog unchanged. If yes, the angle is missing. For this article, the angle is a sequence of checks with a testable pass/fail for each, plus the limits of each. A bare list of tips would not meet it.

Limit of this check: it confirms the draft matches the brief. It does not confirm the brief was right. If the brief aimed at the wrong query, a perfect match is still a miss.

Check 2: Can every claim be traced to a source?

This is the check that catches the most damage. A fluent draft can contain a wrong date or an invented figure and read no differently from a correct one. Google's guidance calls manual fact-checking and review of AI-written articles critical before publishing, and says the review should extend to metadata: title elements, meta descriptions, structured data and image alt text [2].

Numbers, dates and named entities first

Fact-checking quickly means checking in the order of risk. Start with anything that can be wrong in a way a reader would notice:

1. Numbers, percentages and prices.
2. Dates and "as of" statements.
3. Names of people, companies, products and laws.
4. Direct quotes.
5. Superlatives and absolutes: "always", "the only", "the most".

Test: for each item, open the cited source, find the passage, and compare. The number in the draft must match the number at the source, with the same unit, the same time period and the same population. A claim passes only if the source says it. "Close enough" is a fail. Also check that the draft does not attach a number to the wrong thing, for example quoting a statistic about one group as if it described another.

A worked example, labelled as an illustration. Suppose a draft says: "Google says AI content breaks its spam rules." The cited source is Google's guidance on generative AI content. Opening it, the passage says that using generative AI tools to produce many pages that add no value for users may violate the spam policy on scaled content abuse [2]. The draft has turned a conditional into a flat statement and dropped two qualifiers: the volume of pages and the lack of value. Fail. The corrected claim keeps all of them. A few words changed the meaning, and a fluent draft would never have flagged them.

What counts as a source

Rank sources by how close they are to the facts:

• Official documentation and policy pages from the organisation the claim is about.
• Pricing and product pages for claims about what a product costs or does, with the retrieval date.
• Studies and datasets, with the publication date and the sample they describe.

A blog post that repeats a statistic is not a source for it. It is a pointer to where you should look. Follow it to the original; if you cannot find the original, you do not have a source.

Dates matter more than most checks allow. Google's pages on this topic carry visible dates, and policy wording changes. Note the date next to any claim about what a company currently says or does.

What to do with a claim you cannot trace

Cut it. Do not soften it.

Softening feels like a compromise: "many experts believe" in place of the statistic you could not source. But it keeps an unsupported claim in the article and removes the only thing that would let a reader check it. If the claim matters to the argument, send it back for research. If it does not, delete it and see whether the paragraph still works. Usually it does.

Limit of this check: tracing proves the draft says what the source says. It does not prove the source is right, and it does not catch what the draft leaves out. A reviewer who knows the field will still catch things a tracer will not.

Check 3: Is it original, and what originality checks can and cannot tell you

"Original" covers three different problems, and one tool does not cover them all.

Copied text. The draft reproduces passages from another page. A plagiarism scanner finds this. Pass means no copied runs of text beyond short, quoted, attributed excerpts.

Paraphrase of one source. Nothing is copied word for word, but the draft follows a single source's structure and points in order. Scanners often miss this. Google's helpful-content questions ask whether content built on other sources goes beyond copying or rewording them and offers substantial additional value and originality [3]. To test it, line the draft's headings up against the source list. If one source's outline maps onto the draft section by section, you have a rewrite.

Thin synthesis. The draft combines several sources but adds nothing: no worked example, no decision rule, no trade-off, no view on which source to trust. The test is to ask what a reader gets here that they could not get in ten minutes with the sources. If the answer is "a tidier summary", the draft is thin.

Only the first problem is something software handles well. The other two need a person, or a reviewer working from the sources, to look at structure and value.

What about AI detectors?

Detectors do not belong in this check. The 2023 study mentioned above found that the detectors it tested misclassified non-native English writing as AI-written, and that simple prompting could get around them [1]. That is a study of a specific set of tools at a specific time, and detectors have changed since. We did not find a newer primary source we could read, so we make no claim about current accuracy. What the study does show is the type of failure: a detector flags a style, and style is not quality. A human-written article can be wrong and a well-sourced one can be flagged.

Check 4: Does it read like your voice, without filler?

A draft can be accurate and original and still be unpublishable because it sounds like everyone else. Keep the voice document open while you read. Check four things.

• Terms. Preferred terms are used and banned words are absent. This is a search, not a read. Run it first.
• Openings. The first paragraph names the reader's situation and promises the answer.
• Repetition. No paragraph restates the one before it.
• Vague claims. Every adjective or claim of benefit is backed by a specific.

Four tells of filler, with a before and after for each:

Tell	Before	After
Throat-clearing opening	"In today's fast-moving world, content quality matters more than ever."	"Before you publish, check that every number in the draft matches its source."
Restated heading	Heading: "Check 2: Can every claim be traced?" Next sentence: "In this section we look at tracing claims."	Delete the sentence and start with the method.
Empty intensifier	"A powerful, streamlined review process."	"A review where one person checks each number against its cited source."
Summary paragraph	A closing paragraph to each section that repeats it in different words.	Cut it. End on the last new point.

Pass test: read each paragraph and ask what it would lose if deleted. If the answer is nothing, delete it. Count how many paragraphs survive.

Limit of this check: voice is the most subjective of the six. A voice document with concrete rules (terms, point of view, sentence length, banned words) turns part of it into a search. The rest is judgement, which is why a second reader helps.

Check 5: Score it against a rubric with a pass mark

Gut feel says "this seems fine". A rubric says what was checked, how much each part counts, and where the line is. It has three advantages. Two reviewers reach similar results. A failing draft comes back with reasons. And you can see over time which criterion fails most often, so you can fix the process rather than the article.

For a rubric you build yourself, include:

• Criteria that map to your checks: brief fit, claim traceability, originality, voice, structure and readability.
• A weight for each, so a fabricated number costs more than a clumsy sentence.
• A fail line, set in advance and not negotiated per article.
• A rule for automatic fails, such as any untraceable number, regardless of total score.

On the pass mark: there is no industry standard that we can cite, and we found no source for one. We pick ours by asking how much of the score a draft can lose and still be something we would sign. Ours is 80 out of 100.

Our own rubric has 20 criteria. Sixteen are scored from 0 to 5 and weighted, and their total is normalised to 100; the other four are pass/fail checks. Accuracy carries the most weight at 15; Sourcing, Intent match and Specificity carry 10 each. To pass, an article needs at least 80, no hard fail, and a score of at least 3 on each of Accuracy, Sourcing and Intent match. The hard-fail criteria are banned phrases, originality, expert input where the brief asks for it, and AI self-reference or first person outside the voice's point of view. An article that misses any of these goes back to be rewritten. The full criteria are published on our site, so you can test them against your own drafts.

Read the rubric every article must pass

Limit of this check: a score is only as good as the reviewer applying it, and a high score does not predict rankings. It records that the draft met the standard we set.

Check 6: Who checks the checker? Independent review

The writer cannot be the final reviewer. A writer reads what they meant to say. A reviewer reads what is on the page.

Google's helpful-content guidance suggests asking people you trust, who have no tie to your site, for an honest assessment [3]. We apply the same principle inside our process. Our editor is a separate AI session that never sees the writer's reasoning. It sees the draft, the brief, the sources and the voice document. It checks each number against the source it cites. If it cannot trace a claim, the claim is cut, not softened.

Keeping the writer's reasoning from the reviewer matters because reasoning persuades. "I included this because..." makes an unsupported claim feel supported. Without it, the reviewer has only the page and the sources, and the claim either traces or it does not.

Limits. A separate reviewer reduces shared blind spots, but does not remove them. If the brief is wrong, both writer and reviewer will work to the wrong target. An independent review also does not guarantee rankings. It guarantees that a second pass against stated criteria happened before publication.

A pre-publish checklist you can copy

Check	Pass test	If it fails
1. Query and brief	Every must-answer question has a located answer; the angle sentence can be found; the opening names the reader's situation	Return to the writer with the missing questions listed
2. Claims	Every number, date, name and quote matches the cited source, including in title, meta description and alt text	Cut the claim. Send back for research only if the argument needs it
3. Originality	No copied runs of text; no single source mapped section by section; at least one thing the sources do not give	Add the example, decision rule or trade-off, or reject as thin
4. Voice	Preferred terms used; banned words absent; no paragraph that could be deleted without loss	Delete filler; rewrite generic openings
5. Rubric	Score at or above the pass mark; no automatic-fail items	Rewrite against the lowest-scoring criteria, then rescore
6. Independent review	A reviewer who has not seen the writer's reasoning has signed off	Do not publish until a second reader has seen it

Common questions

Does Google penalise AI-written content? Google's stated concern is scaled low-value pages, not the tool. Its guidance says that using generative AI tools to produce many pages that add no value for users may violate its spam policy on scaled content abuse [2]. The spam policy defines that abuse by purpose: producing many pages mainly to manipulate search rankings rather than help users, however the pages are made [4]. So the test is value and intent. We found no Google statement that penalises AI use as such, and none that exempts it. Passing these checks does not guarantee any ranking.

How do I fact-check AI-written articles quickly? Check in order of risk: numbers, dates, names, quotes, absolutes. Open the cited source for each and compare it with the draft. Do the metadata too, as Google's guidance says [2].

Can AI detectors tell me if the content is good quality? No. A detector estimates how a text was produced, not whether it is correct or useful. A 2023 study found that detectors misclassified non-native English writing and could be bypassed with simple prompting [1].

What should a quality rubric include, and what pass mark makes sense? Weighted criteria that match your checks, a fail line set in advance, and automatic fails for untraceable numbers. We use 80 out of 100. We cannot point to a standard that says 80 is right; it is where we decided a draft stops being something we would sign.

How many checks can be automated, and which need a person? This is our judgement, not a measured figure. Term searches, banned-word searches and copied-text scans automate well. Tracing a claim to its source can be partly automated, but a person or a reviewer working from the sources should confirm the meaning matches. Brief fit, paraphrase, thin synthesis and the decision to publish need a person.

How do I know if the content is original? Run a copied-text scan, then compare the draft's structure with its sources, then ask what a reader gains that the sources alone do not give. Google's own questions frame it the same way [3].

What to do now

Take the draft you are holding. Pull the brief and the source list. Run Check 2 first, because that is where a confident draft hides its mistakes. Then work through the rest in order and write down the result of each, so the decision to publish is on record.

If you want to see the standard we hold our own articles to, read the rubric. If you want us to look at the content you already have, we offer a free audit.

Read the rubric every article must pass

Get a free content audit of your site

Sources

1. GPT detectors are biased against non-native English writers — arXiv (published version in Patterns, Cell Press), 2023-04-06
2. Google Search's guidance on using generative AI content (fact-check passage) — Google Search Central, 2026-10-01
3. Creating helpful, reliable, people-first content — Google Search Central, 2026-10-05
4. Spam policies for Google web search — Google Search Central, 2026-08-28

This article was researched, written and edited by AI under human supervision, and checked against Rankwrite's public editorial rubric.