A report on "How to Check an AI-Written Article Before Publishing"
The graded article is 1,158 words, written by DURBUN from one keyword. Open it below to read what the models read. The report took 11.7 seconds and cost 468 credits.
How to Check an AI-Written Article Before Publishing
For a reliable prepublication review, use DURBUN’s Article Quality Report alongside a manual reading. First confirm that the opening answers the search, check every fact, figure, and date against its source, cut repetition and filler, make broad claims specific, and get a second reading before publishing. DURBUN can help locate weak passages, but its score does not replace reading the source yourself.
From Search Answer to Source Check
Start with the opening. It should answer the search directly instead of making the reader work through several paragraphs. Compare the title, introduction, and conclusion so they make the same promise. If the title offers a figure, date, or verdict that the article cannot support, change the title or revise the claim.
Next, make a claim-and-source pass. Check every number, date, name, quotation, study, and conclusion against the source itself. Make sure a cited paper, journal, or DOI exists and supports the exact statement beside it; a working link can still lead to evidence that does not support the claim. Read the source rather than relying on a citation that merely looks plausible.
Specific wording makes this check easier. Replace “studies show” with the finding and its limits, or remove the sentence. “Research shows that shorter introductions improve engagement” needs the result that supports it and the condition under which it applies. A broad statement such as “this improves performance” should name the measured change or disappear.
A second reader should then review the finished draft before publication. That reading can catch a promise the source check missed, a repeated point, or a sentence that sounds precise while saying very little.
DURBUN’s report reads one article: either an article saved in your DURBUN dashboard or text you paste into it. Its quoted passages and short verdicts can show where to begin the manual review. After you edit a saved article in DURBUN, the next report scores the edited version.
Evidence and Consistency
Follow each important claim through the article instead of checking isolated sentences. Mark repeated figures, dates, definitions, and conclusions, then compare them with the source and with one another. If two paragraphs give different weights for the same product, keep the supported value and remove the other one.
Technical terms need a context check as well. Fluent specialist language can still combine ideas that do not belong together. Read the paragraph around each term and confirm that the explanation fits both the subject and the source.
A draft also needs a repetition pass. Cut an explanation when it adds no new evidence or qualification, especially if the same advice appears in the opening, a middle section, and the conclusion. Remove filler that announces the article’s structure or repeats a point the reader already understands.
DURBUN scores evidence and credibility, but that score cannot confirm a citation for you. Its report is an inspection of the article, not a substitute for checking the material behind its claims.
Specific Claims and Readable Prose
Smooth grammar is not enough. Each paragraph should make a precise point, give the reader a reason to accept it, and contain enough support for the strength of its wording. Soften a claim when its condition is limited, or state the condition directly instead of using a vague hedge.
Uniform structure deserves inspection too. Several paragraphs that open alike or repeat the same claim-support-conclusion pattern may be saying the same thing in different words. Read those passages for duplicated advice, missing detail, and sentences that sound polished without adding information.
The final pass should also check whether the article speaks plainly. Turn abstract instructions into actions: check the source, compare the figures, cut the repeated sentence, and ask another reader to review the result. A reader should be able to tell what to do without translating formal editorial language first.
How the Article Quality Report Works
Run the report by hand after the first editorial pass. You pick two or three models, and each reads the article independently in the way a search quality rater would. Every model scores seven axes from 0 to 10: search intent, writing quality, topical coverage, information gain, evidence and credibility, structure, and spam safety.
The report shows each axis average, where the models disagree, and how far apart their overall scores are. The models quote passages that work and passages that need fixing, and each writes a short verdict. Use those quotations as an editing list, then check any factual change against the underlying source.
The overall score provides a clear review threshold:
| Overall score | Report verdict | What you do |
|---|---|---|
| 8.5 and above | Would outrank the current results | Check the article and publish when the second reading passes |
| 7 and above | Competitive, worth publishing | Fix the quoted weaknesses before publishing |
| 5.5 and above | Publishable after the fixes below | Make the listed edits and run the report again |
| Under 5.5 | Not ready to publish | Revise the article before another review |
The gap between 8.5 and 5.5 is a practical difference in how much work remains, not proof that a report has checked every source.
You need at least 1,500 credits free to start a report. DURBUN bills actual use, usually a few hundred to about 1,300 credits, and returns the rest. Reports are saved, so you can open them again later.
Each SEO-Friendly Article DURBUN writes is researched from live search results and up to 5 ranking pages, then checked for overstated claims and edited. That writing workflow is separate from the report, which evaluates one saved article or pasted text.
Run the report on a new article, fix what it quotes, then run it again before publishing.
The Final Draft Record
Once the wording is approved, publish the article and keep its final Article Quality Report saved in your DURBUN dashboard as the record of the version you released. Make that report the last file you attach to the article’s editorial record.
Frequently Asked Questions
Does Google automatically penalize an article because AI helped write it?
Google does not automatically penalize every AI-assisted article. Its guidance focuses on scaled content abuse and low-quality content, whether people or automation produced the material. Use the Article Quality Report in DURBUN to inspect the draft, then check its claims and sources before publishing.
Can a watermark prove that a person wrote an article?
No, the absence of a watermark or provenance signal does not prove that a person wrote the article. Paraphrasing can weaken or remove that signal, so judge the article by its evidence, accuracy, and usefulness rather than by whether a marker remains.
What should I recheck after changing an article?
Recheck the article after substantial edits because a new headline, summary, image, metadata field, or structured data can add claims that the original draft did not support. Run the report again on the edited article, then verify any new claim against its source before publishing.
Scores by axis
Each axis is the average of the three models. Where they disagreed the spread is shown, because a wide spread usually means the article is ambiguous rather than bad.
- Search intent
- 7
- Writing quality
- 6.3
- Topical coverage
- 6.3
- Information gain
- 5.3
- Evidence and credibility
- 5.3
- Structure
- 5.7
- Spam safety
- 4.3
Strongest: Search intent, Topical coverage. Weakest: Spam safety, Information gain, Evidence and credibility.
What each model said
GPT-5.6 Luna, 6/10. This is a useful, well-organized editorial checklist that answers the query quickly and includes practical checks for sources, repetition, consistency, and second-reader review. However, it becomes a product-led explanation of DURBUN, presents several precise product and scoring claims without supporting context, and adds an SEO-style FAQ and repeated calls to rerun the report. It would be stronger as an independent prepublication workflow with the tool reduced to an optional example.
GPT-5 Nano, 6.6/10. Meets the basic need of showing a process to check AI-written content, but leans heavily on a single vendor's workflow and promotional framing. It would benefit from more neutral guidance and broader verification steps beyond the DURBUN system.
Gemini 3.8 Flash, 4/10. This piece hijacks a universal, informational editorial workflow query ('how to check an ai written article before publishing') and turns it into an onboarding manual and product pitch for a proprietary tool ('DURBUN'). While the foundational advice on fact-checking hallucinated citations and smoothing repetitive cadences is sound, half the article is spent detailing DURBUN's credits, dashboards, and internal grading thresholds. An independent searcher looking for practical verification methods, AI hallucination checks, and pre-publish tooling will find this heavily biased and incomplete.
What it got right
- The opening gives a direct, actionable answer: "First confirm that the opening answers the search, check every fact, figure, and date against its source, cut repetition and filler, make broad claims specific, and get a second reading before publishing."
- It correctly distinguishes automated review from human verification: "DURBUN can help locate weak passages, but its score does not replace reading the source yourself."
- The source-checking advice is concrete and appropriately demanding: "Check every number, date, name, quotation, study, and conclusion against the source itself."
- It identifies a subtle but important failure mode beyond grammar: "Fluent specialist language can still combine ideas that do not belong together."
- "For a reliable prepublication review, use DURBUN’s Article Quality Report alongside a manual reading."
- "Check every number, date, name, quotation, study, and conclusion against the source itself."
- "The final pass should also check whether the article speaks plainly."
- "Google does not automatically penalize every AI-assisted article. Its guidance focuses on scaled content abuse and low-quality content, whether people or automation produced the material."
- Identifies a specific failure mode in AI citations: 'Make sure a cited paper, journal, or DOI exists and supports the exact statement beside it; a working link can still lead to evidence that does not support the claim.'
- Gives practical, actionable guidance on pruning synthetic filler: 'Replace “studies show” with the finding and its limits, or remove the sentence... A broad statement such as “this improves performance” should name the measured change or disappear.'
What it wants fixed
- The article is overly promotional for the query. It repeatedly directs readers to DURBUN and includes product-specific material such as "You need at least 1,500 credits free to start a report" and "Each SEO-Friendly Article DURBUN writes is researched from live search results and up to 5 ranking pages." Remove most of this or move it into a brief optional-tool note after the independent checklist.
- Several precise claims about the tool's operation and output are asserted without enough explanation, including "You pick two or three models," "Every model scores seven axes from 0 to 10," and "The report shows each axis average." Either explain that these are features of this particular report, or present them as product documentation rather than general editorial guidance.
- The score table creates false authority for publication decisions: "8.5 and above | Would outrank the current results" implies that a model-generated score predicts rankings. Replace it with a clearly subordinate heuristic and state that no score establishes search performance or publication readiness.
- The page repeats the same tool limitation and call to action in multiple places: "Use those quotations as an editing list," "Run the report on a new article," and "Run the report again before publishing." Consolidate this into one short workflow step and give more space to missing editorial checks such as permissions, disclosure, originality, accessibility, and image or quote verification.
- The FAQ is only loosely needed and reads as search-engine scaffolding. Questions such as "Can a watermark prove that a person wrote an article?" are peripheral to the target query; remove the FAQ or integrate only the directly relevant guidance into the main checklist.
- The instruction "Once the wording is approved, publish the article and keep its final Article Quality Report saved" is oddly sequenced and product-centered. Say instead to retain the final fact-check and editorial approval record, with a tool report as optional supporting documentation.
- "DURBUN can help locate weak passages, but its score does not replace reading the source yourself." — Promotes a vendor and uses promotional framing rather than neutral guidance.
- "DURBUN scores evidence and credibility, but that score cannot confirm a citation for you." — Overstates the tool's limitations, relies on a single-system workflow.
- Heavy emphasis on DURBUN-specific terminology and workflow may reduce perceived applicability to non-DURBUN readers.
- The article is long and highly structured around DURBUN, which can feel like a funnel to a product rather than a general best-practices guide.
- Some claims (e.g., the scoring thresholds and outputs) are presented as facts without independent justification beyond the product context, which could mislead readers about general practices.
- The article turns an informational how-to search into a product promo for a specific commercial platform: 'For a reliable prepublication review, use DURBUN’s Article Quality Report alongside a manual reading.' Replace the proprietary platform shilling with an agnostic, comprehensive workflow covering hallucination auditing, entity verification, style de-biasing, and multi-model fact-checking.
- Devotes excessive space to the mechanics and credit billing of a single tool: 'You need at least 1,500 credits free to start a report. DURBUN bills actual use, usually a few hundred to about 1,300 credits...' Delete this commercial documentation entirely; it is irrelevant to the search intent.
- Neglects core technical verification practices for AI copy, including checking for hallucinated URLs, synthetic logic flaws, math errors, timeline/recency confusions, and prompt-bleed artefacts, offering instead generic editorial platitudes.
How to read the score
The grader keeps nine and above for pages it would rank above the current results and publish itself, and it puts most published writing on the web at six or seven. The verdict follows the average: 8.5 and above reads Would outrank the current results, 7 Competitive, worth publishing, 5.5 Publishable after the fixes below, and anything lower Not ready to publish. Every fix above is quoted from a model, so each one points at a passage you can change.