Content scoring rubric: the 20 criteria our editor applies

If you approve AI-written articles in bulk, "this looks fine" stops working around the tenth draft. You need a content scoring rubric: a written list of what a draft must do, and a number that says whether it did. Below is the one we use, with all 20 criteria, the weights, the pass rule and a worked example of a draft failing and being fixed.

What a content scoring rubric is and why it beats a gut check

A content scoring rubric is a set of named criteria, a way to score each one, and a pass mark. The University of Nebraska–Lincoln's teaching centre recommends usually 3–10 criteria, with 3–6 performance levels, and details that are objective, clear and non-overlapping [1]. A gut check has none of that. Two reviewers disagree, and the same reviewer disagrees with themselves on a Friday afternoon.

At high volume the gap grows. A rubric lets you say why a draft failed, point to the criterion, and hold every article to the same standard.

We apply ours through an independent editor session, not the writer. The editor sees the draft, the brief, the sources and the voice doc. The score and per-criterion notes ship with the article.

The rubric at a glance

The rubric has 20 criteria. Sixteen are scored 0–5 and weighted, then normalised to a score out of 100. Four are pass/fail hard checks. The numbering below is the rubric's own [2].

#	Criterion	Weight	What it checks
1	Accuracy	15	Checked facts are verified; dated facts are phrased as dated
2	Sourcing	10	Claims are cited; sources are primary where numbers appear, and recent enough
3	Intent match	10	Answers the query as the readers of top results expect, in the right format
4	Angle	8	Says something the top results do not, from the customer's position
5	Completeness	8	Every must-answer item and the table-stakes topics are covered
6	Specificity	10	Concrete examples, numbers, named things; no generalities
7	Structure	5	Headings true to content; logical order; scannable
8	Opening	5	Situates the reader and promises the answer within 100 words
9	Voice	8	Pov, formality, tone, preferred terms, sample-paragraph feel
10	Readability	5	Sentence variety, paragraph length, reading level as targeted
11	Banned phrases	Hard check	Any hit is an automatic fail
12	Originality	Hard check	Passes the plagiarism and AI-pattern checks
13	Internal links	3	Links to the customer's own pages where the brief gives them
14	SEO hygiene	4	Title length; query in title, opening and one heading; no stuffing; meta fits
15	Caveats and safety	5	Required disclaimers and trade-offs present; nothing that reads as regulated advice beyond the sources
16	CTA	2	Present, from the voice doc, placed as briefed
17	Length	2	Within range or justified
18	Formatting	2	Clean Markdown, tables where they help
19	Expert input	Hard check (if briefed)	Expert answers are used and attributed correctly
20	No AI self-reference	Hard check	No mention of AI authorship, no first person outside the voice doc's point of view

Read the rubric every article must pass

The criteria, grouped

Facts and sourcing

Accuracy (15) is the heaviest weight. Every checked fact must be verified, and a fact with a date must say so. Sourcing (10) asks that claims are cited, that numbers come from primary sources, and that sources are recent enough [2].

A typical loss: a draft states a percentage with no citation. That costs points on Sourcing, and if the number cannot be traced it is also an Accuracy problem.

Fit to the query and brief

Intent match checks that the article answers the query as the readers of top results expect, in the right format. Angle checks that the piece says something the top results do not, from the customer's position. Completeness checks that every must-answer item and the table-stakes topics are covered. Specificity checks for concrete examples, numbers and named things, with no generalities.

A typical loss: a piece about scoring that never says what score passes. That is a Completeness miss. A paragraph of general advice where an example belongs costs Specificity.

Reading experience

Structure checks that headings are true to content, in a logical order, and scannable. Opening checks that the first 100 words place the reader's situation and promise the answer. Voice checks pov, formality, tone, preferred terms and the feel of the sample paragraphs. Readability checks sentence variety, paragraph length and the reading level as targeted.

A typical loss: three sentences of scene-setting before the answer appears. The reader has left by then, and so has the Opening score.

Polish and conversion

Internal links, SEO hygiene, Caveats and safety, CTA, Length and Formatting carry smaller weights, from 2 to 5 points each. SEO hygiene looks at title length, the query in the title, opening and one heading, no stuffing, and a meta description that fits. Caveats and safety looks for required disclaimers and trade-offs, and for nothing that reads as regulated advice beyond the sources. CTA must be present, from the voice doc, and placed as briefed. Length must fall within range or be justified.

A typical loss: a research pack lists a caveat, and the draft leaves it out. That costs points on Caveats and safety.

Hard checks

Hard checks are pass/fail. A single failure fails the draft, whatever the total.

• Banned phrases. Any hit is an automatic fail.
• Originality. A plagiarism score above 5%, or an exact match over 12 words, fails. Attributed quotations are excepted from the 12-word rule. An AI-pattern score above the owner-set threshold also fails, and the flagged passages are returned. The default threshold is 40 [3].
• Expert input. If expert answers exist, they are used and attributed. If none exist, nothing is attributed to an expert.
• No AI self-reference. The text never refers to its own authorship and uses no first person outside the voice doc's point of view.

How scoring and the pass mark work

The editor scores each of the 16 weighted criteria from 0 to 5. The scores are weighted and the total is normalised to 100. A draft passes if the total is 80 or more, no hard check fails, and Accuracy, Sourcing and Intent match each score at least 3. Otherwise it fails [2].

The weights add up to 102, not 100, which is why we say "normalised" rather than "out of 100 points". We do not publish the exact normalisation formula here, so treat the arithmetic below as approximate.

The floor on those three criteria exists so a high total cannot hide a bad fact. A draft scoring 5 on everything except Accuracy at 2 would total about 91 (465 of 510), yet still carry a wrong fact. Requiring a 3 on Accuracy, Sourcing and Intent match stops that.

Three calibration principles keep scores consistent [2]:

• Plain and correct beats polished and vague. A simple draft is not marked down for being simple.
• A borderline draft is never rounded up to pass.
• The same standard applies to every customer.

The editor is a separate session from the writer. It sees the draft, the brief, the sources and the voice doc, so it can check each claim against what the sources say.

A worked example: scoring one draft

This example is illustrative. It is not a customer result, and the numbers are simplified.

Take a draft where most criteria earn a 4, but two do not. A statistic has no citation, so Sourcing scores 2. The advice stays general, so Specificity scores 2.

Multiply each score by its weight and add them up. With every criterion at 4, the weighted sum would be 408 out of a possible 510, or 80%. Dropping Sourcing and Specificity by two points each, at weight 10, removes 40, leaving 368. That is about 72 out of 100. The draft fails on the total, and it fails the floor too, because Sourcing is below 3.

The fix is concrete. The writer finds the primary source for the statistic and cites it, or cuts the sentence. The vague advice becomes a worked number and a named example. Re-scored, Sourcing and Specificity both reach 4. Because the uncited number is now checked, Accuracy also rises from 4 to 5.

The new weighted sum is 408 + 15 = 423 out of 510, about 83 out of 100. Total above 80, no hard check failed, and Accuracy, Sourcing and Intent match all at 3 or more. The draft passes.

Note what would have happened with no fix to Sourcing: even a draft that reached 80 elsewhere would still fail with Sourcing at 2.

How to adapt the rubric for your own team

Start with criteria a reviewer can observe in the text. "Engaging" is a feeling. "Opening names the reader's situation and answers within 100 words" is something two people can check.

Define the score levels so that descriptions are objective and do not overlap, as the Nebraska guidance advises [1]. Then calibrate: have two reviewers score the same draft separately, compare, and rewrite any criterion where they diverge.

Set a pass mark, and make a few criteria hard checks, such as banned phrases, copied text and uncited numbers.

Size matters too. The Nebraska guidance suggests 3–10 criteria [1]. Our 20 is a deliberate choice for an automated editor that applies every criterion on every draft. A human team may do better starting with fewer and adding as the gaps show.

Limits of a rubric

A rubric checks a draft against the brief and the sources. It cannot tell you whether the article will rank or bring traffic, and a passing score is not a promise of either. It is also only as good as the brief: a draft that follows a weak brief faithfully can still score well.

Want to see the full rubric in one place? Read the rubric every article must pass. If you would like us to look at your existing content, you can also get a free content audit of your site.

Sources

1. How to Design Effective Rubrics — University of Nebraska–Lincoln Center for Transformative Teaching, 2026-01-01
2. Rankwrite editorial rubric (policies/rubric) — Rankwrite (internal, first-party), 2026-10-20
3. Originality policy (policies/originality_policy) — Rankwrite (internal, first-party), 2026-10-19

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