How to Evaluate AI-Written Drafts: A Twenty-Point Editorial Rubric

How do you evaluate an AI-written draft?

Score it against a fixed list of yes-or-no checks, read the total against a threshold, and let accuracy failures veto publication. The rubric on this page has twenty checks in five groups of four: accuracy, originality, intent, voice and sourcing.

One qualification matters. The checks come from published guidance, but the grouping, the blockers and the score bands are Kanso's design and have not been validated on real drafts, so treat the thresholds as a starting point to adjust.

• Twenty checks, one point each, in five groups: accuracy, originality, intent, voice and sourcing.
• Three checks are blockers. A draft that fails any of them is not published, whatever its total.
• A total of 17 or more with no blocker failed passes. 12 to 16 is revised and rescored, and anything lower is rewritten from the brief.
• Two reviewers scoring the same draft should agree on at least 80% of the checks. Below that, the wording of the rubric needs work.
• Every pass and fail example is invented for this page. None comes from a real draft or a test.

score the draft → check the three blockers → read the total → revise or publish

Why does an AI draft need an editorial review before publishing?

An AI draft needs review because it can look finished while containing errors that only a reader with the sources open will catch. Metaflow names hallucinations, bias, off-brand tone and generic writing as the main failure modes, and says models predict likely text rather than retrieve verified facts.

Search guidance points the same way. In its post on AI-generated content, Google says its ranking systems aim to reward original, high-quality content that shows E-E-A-T (expertise, experience, authoritativeness and trustworthiness), and that its focus is the quality of content rather than how it is produced.

Google is not hostile to automation. The same post says automation has long produced useful content such as sports scores, weather forecasts and transcripts, while using AI primarily to manipulate rankings violates its spam policies.

Reviewing is how a team shows that a draft sits on the right side of that line. For where in a workflow a person should sign off, Kanso's article on where approval gates belong in an AI marketing workflow covers the placement question, and this page covers what the person checks.

How was this twenty-point rubric built?

The rubric was built by taking the quality criteria that four published sources name, sorting them into the five groups the brief asked for, and writing four yes-or-no checks for each group. No draft was scored and no test was run to produce it, so it is a synthesis of the sources and nothing more.

The sources are Google's Search Central Blog post on AI-generated content, plus pages from Metaflow, Walturn and BrightEdge. Google is the primary source. The other three are practitioner pages, and each claim taken from them is attributed to them by name.

Two design rules shaped the wording. Each check must be answerable by reading the draft with its sources and brief open, and each must be a yes or no, because half-points are where two reviewers drift apart.

The intent checks assume a written brief to test against. Kanso's content brief template with a filled example shows one way to write it.

Accuracy: can every claim in the draft be traced to a source?

Every claim passes only if a reviewer can open the source and see the claim there. Walturn calls accuracy a critical measure because AI systems often fabricate plausible-sounding but incorrect data, and Metaflow asks whether information is both correct and up to date.

Checks 1, 2 and 4 are blockers because they decide whether the page states something false. Check 3 is not, because an undated fact is a defect to fix, not a falsehood. Fail examples in every table below were written for this page.

The cheapest way to run check 1 is to highlight every number, date and proper noun in the draft and tick each one against its source. Anything left unticked is a fail until a source is found or the claim is cut.

Originality: does the draft add anything the ranking pages lack?

A draft is original when it contains at least one element a reader cannot already get from the pages ranking for the same query. BrightEdge advises original research, case studies or expert analysis over rehashing what exists, and Walturn says a lack of originality shows up as repetitive or generic responses.

Check 5 is the one that most AI drafts fail, because a model trained on existing pages tends to average them. A passing draft names its extra element, such as a table built from sourced figures or a worked example, and says how it was put together.

Check 8 follows BrightEdge's advice that every sentence should serve the reader's query. The test is mechanical: delete a sentence, and ask whether the page lost a fact, a step or a decision.

Intent: does the draft answer the question it was written for?

A draft answers its question when the reader gets the answer in the first paragraph and the format fits what they came to do. Metaflow lists relevance as whether the content addresses the intended topic and audience, and Walturn warns that irrelevant responses reduce usability even when they are grammatically correct.

Checks 9 to 11 test structure: the answer comes first, the format fits the query, and each section answers its own heading. Check 12 borrows the Why question from Google, which asks whether content was created primarily for people or for search ranking.

Google's post says that evaluating content with the Who, How and Why questions will help whether or not AI was used. Check 12 is the Why. Check 20, in the sourcing group, covers the Who and the How.

Voice: does the draft sound like your brand and read cleanly?

A draft passes on voice when a reader who knows your other pages would not notice a different author, and when the text is coherent, natural and free of bias. Metaflow says AI output tends toward a neutral, somewhat formal tone, so a playful or personal brand will find raw output jarringly off-brand.

Walturn separates three nearby qualities: coherence (logical flow and consistent tone), naturalness (how human-like the text reads) and bias (reproducing stereotypes or offensive assumptions from training data). Checks 13 to 16 cover them in that order, plus the brand voice guide.

Check 14 is the one a script can run. Keep a short list of phrases your brand bans, search the draft for each, and count the hits before a human reads a word.

Sourcing: does the draft show where its claims come from and who stands behind it?

A draft passes on sourcing when each claim names its source in the same sentence and the page is honest about who wrote it and how. Google's post suggests accurate bylines where readers would expect them, and AI or automation disclosures where a reader might ask how the content was created.

Google also says that giving AI an author byline is probably not the best way to make AI involvement clear. That is why check 20 asks for a named human author and a plain note on process, not an AI credit.

Checks 18 and 19 apply a ranking of sources: the platform's own documentation first, named studies that state their method next, and practitioner pages only as pointers or clearly attributed opinion. When two sources disagree, the draft should say so and attribute each.

How do you turn twenty checks into a decision?

Give one point per pass, add them up, read the total against three bands, and apply the blockers before the bands. A blocker failure means the draft is not published, even if its total would pass.

The 17-point line is a proportion, not a finding. Metaflow gives 85 out of 100 as an example threshold for publishing with minimal oversight, and 0.85 × 20 = 17 applies the same proportion to twenty checks.

Blockers exist because a total can hide a serious failure. A draft that passes 19 checks but fails check 2, an invented quote, scores 18 under this scheme (19 − 1) and would clear the 17-point line on total alone.

What does a scored draft look like?

A scored draft looks like a five-line table with a subtotal per group and a total. The example below is hypothetical: a draft on choosing a time-tracking tool, built for this page to show the arithmetic.

The arithmetic is 3 + 2 + 4 + 2 + 2 = 13. No blocker failed, since checks 1, 2 and 4 all passed, and 13 sits in the 12 to 16 band, so the decision is to revise and have a second reviewer rescore.

Now suppose the same draft also contained an invented customer quote. Check 2 would fail, accuracy would drop to 2 and the total to 12, and the blocker would stop publication regardless of any later improvement in the other groups.

How do you make two reviewers score the same way?

Have both score the same drafts independently, compare the results, and discuss each disagreement until the wording of the check is clear. Metaflow recommends this: multiple reviewers score the same content independently, compare, aim for 80% or more agreement, hold regular calibration sessions on edge cases, and update the rubric when the same disagreements recur.

To compute agreement on this rubric, count the checks where both reviewers gave the same pass or fail and divide by twenty. If they match on 16 checks, agreement is 16 ÷ 20 = 0.80, or 80%, which meets Metaflow's aim, and 15 matches give 75%.

Metaflow does not say how it computes agreement, so simple percent agreement is this page's choice. It does not correct for agreement by chance, which is one more reason to read it as a rough signal.

1. Pick three drafts, ideally one strong, one weak and one in between.
2. Have each reviewer score all three alone, using only the written checks.
3. List every check where the two scores differ.
4. For each, agree what the pass example and the fail example for that check should say, and edit the wording.
5. Rescore a new draft and recompute agreement.

Should a person, a model or both score the draft?

For most teams the answer is both, with a person making the call on the blockers. Metaflow describes three methods and gives rough weekly volumes for each, which are its own recommendations and not measured results.

Walturn adds a caution about automatic metrics. BLEU and ROUGE rely on n-gram overlap with a reference text, and Walturn says such metrics often fail to assess factual accuracy, originality or ethical soundness, so they are a poor fit for an editorial rubric.

Which checks can a model screen? In Kanso's judgment, checks 9 (answer first) and 14 (banned phrases) are mechanical enough for a first pass, while checks 1, 2, 4 and 17 to 19 need a person with the sources open. The research supplies no data on how often a model judge agrees with an editor, so that split is a recommendation and not a finding.

For the wider question of which marketing tasks to keep away from a model entirely, see Kanso's article on tasks where AI costs more than it saves.

What is documented and what is Kanso's analysis?

The facts below come from the four sources, and the judgments below them are Kanso's own. Keeping them apart lets you disagree with the analysis without rejecting the sources.

Documented (attributed to the named source):

• Google says, in its Search Central Blog post on AI-generated content, that it judges content on quality rather than on how it was produced, and that using automation mainly to manipulate rankings breaks its spam policies.
• Google asks creators to apply the Who, How and Why questions, and says bylines and AI disclosures are worth adding where readers would reasonably expect them.
• Metaflow lists accuracy, relevance, clarity, grammar and safety as core criteria, and describes manual, automated and hybrid methods with the weekly volumes quoted above.
• Metaflow advises that several reviewers score the same content independently, and that a team aims for 80% or more agreement.
• Walturn lists relevance, accuracy, coherence, originality, bias, ethics, naturalness and explainability, and says common automatic metrics such as BLEU and ROUGE rely on word overlap and miss factual accuracy.
• BrightEdge advises against filler and in favor of original content with cited sources and clear author information.

Kanso's analysis (not documented anywhere):

• The grouping of twenty checks into five groups of four is Kanso's design. No source prescribes it.
• The three blockers (checks 1, 2 and 4) are Kanso's choice, based on the sources' emphasis on hallucination and accuracy.
• The score bands (17 to 20, 12 to 16, 0 to 11) are Kanso's suggestion. They have not been tested against real drafts or real outcomes.
• The pass and fail examples are invented to show the standard. They are not drawn from real drafts, and no test or survey sits behind them.
• The worked example in this article is hypothetical, built to show the arithmetic.
• The rule that a person decides blockers is Kanso's recommendation. The research supplies no measurement of how well a model judge agrees with an editor.

Facts here were read on 28 October 2026. Google's guidance and the practitioner pages can change, so confirm against the linked source before building a policy on them.

Where does Kanso's Writer Agent fit in this review?

The Writer Agent produces the draft and this rubric is the review that follows it. Per its agent page, it turns Search Console opportunities into long-form article drafts in your brand voice, for you to review and publish yourself.

Kanso does not score drafts against this rubric, and nothing in the agent decides that a draft is ready. The review, the blockers and the decision to publish stay with a person.

The 20 checks work with a draft from any source: a model, a freelancer or your own first pass. The sheet only needs the draft, its brief and its sources open side by side.

Evaluating AI-written drafts: frequently asked questions

Sources

• Google Search's guidance about AI-generated content, Google Search Central Blog
• The Complete Beginner's Guide to AI Content Evaluation (2026), Metaflow
• Evaluating AI-Generated Content, Walturn
• Content Quality Signals for AI Algorithms, BrightEdge