AI Content Quality Score: A 100-Point Rubric and What Tools Miss
An AI content quality score should grade what readers and Google’s raters judge: accuracy, effort, originality and a clear answer. Here is a 100-point rubric with gates and pass marks, a scored example, and what Surfer, Clearscope, MarketMuse and Originality.ai measure and miss.
An AI content quality score is worth having only if it measures what makes an article trustworthy, and most scores on the market measure something narrower. Surfer, Clearscope and MarketMuse compare your terms with pages that already rank. Originality.ai estimates how much of the text an AI wrote. None of them checks whether a price is current or a quote is real.
This guide gives a 100-point rubric built on what Google’s raters evaluate, with weights, partial-point rules, four gates and pass marks, then scores an example draft twice. It ends with what each tool score measures, what it misses, and how to give it the right number of points.
Summary
Score AI-assisted articles on a 100-point rubric: answer and intent 20, accuracy and sources 25, originality 20, depth 10, structure 10, links 5, trust and settings 10. Publish at 85 or more with all four gates passed (no fabricated facts, accuracy 20+, originality 10+, a subject reviewer on YMYL topics). Tool scores from Surfer, Clearscope and MarketMuse measure term coverage against ranking pages, so they feed 5 points at most.
What should an AI content quality score measure?
A quality score is only as good as the things it counts. Google’s guidance on helpful content lists four attributes its quality raters are trained to evaluate in the main content of a page:
- Effort: how much human work went into it. Using generative AI to produce large amounts of text without oversight counts as little to no effort.
- Originality: information or perspectives not already available on other sites.
- Talent or skill: clear writing and functional page elements.
- Accuracy: factually correct content, and highly accurate content consistent with expert consensus on topics that affect health, money or safety.
The same page says trust matters most among the parts of E-E-A-T (experience, expert knowledge, authoritativeness and trustworthiness). The rubric below follows that: accuracy and sources carry the most points, originality comes next, and the parts a term-coverage tool can measure carry the fewest.
The rater guidelines also settle the AI question. In the September 2025 edition, the use of generative AI tools alone does not determine a page’s effort or quality rating. So the score grades the article, never the tool that drafted it.
What is in the 100-point rubric?
Seven categories, 20 criteria, 100 points. Copy the table into a sheet and score each criterion with whole numbers.
| Category | Criterion | Points |
|---|---|---|
| Answer and intent (20) | The opening 100 words answer the query | 6 |
| Title, H1 and headings describe the same job | 6 | |
| The searcher’s sub-questions are answered; nothing drifts into another page’s topic | 8 | |
| Accuracy and sources (25) | Share of factual claims linked to a source that says the same thing, x 10 | 10 |
| Sources are original: the study, the vendor, the official documentation | 5 | |
| Prices, dates and versions checked on review day | 5 | |
| Quotes traced to the speaker’s own words | 5 | |
| Originality and added value (20) | At least one element missing from the top results: worked example, own data, template, arithmetic | 10 |
| No near-copy of another page on the same site | 5 | |
| Named first-hand input: the reviewer, the person who did the work, the process behind the numbers | 5 | |
| Depth (10) | Subtopics and terms covered (a tool score can feed this one) | 5 |
| Examples specific enough to act on | 5 | |
| Structure and readability (10) | Headings make sense alone; tables and lists where the content is a comparison or a list | 5 |
| No repeated ideas, no filler, one capitalisation style | 5 | |
| Links (5) | Internal links with descriptive anchors, including one to the next step and one pointing in | 3 |
| Outbound links go to sources | 2 | |
| Trust and page settings (10) | Real byline; subject reviewer named where the topic needs one | 3 |
| Disclosure where readers would ask how the article was made | 2 | |
| Title tag, meta description and alt text written for this page | 3 | |
| Structured data matches what readers see | 2 | |
| Total | 100 |
The weights are a judgement call, and the reasoning is simple. A wrong fact can cost a reader money; a weak heading costs a click. So accuracy gets 25 points and structure gets 10.
How do you award partial points?
Two scorers should land within a few points of each other. These rules get them there:
- Proportional criteria: the sourced-claims criterion is the share of claims with a matching source, times 10, rounded down. 9 of 12 claims sourced: 9 / 12 x 10 = 7.5, so 7 points.
- Yes/no criteria: full points or zero. Either a near-copy exists on the site or it does not.
- Graded criteria: start at full points and take off one point for each fix the scorer has to note, down to zero.
- Score the version that will go live, in the CMS preview, with metadata in place.
Calibrate on your first ten articles: two people score each one separately. Where their totals differ by more than 10 points, the criterion that split them needs clearer wording.
Which pass marks and gates decide publishing?
A total hides the one fatal error, so the rubric has four gates that block publishing whatever the total says:
- Any fabricated source, quote or statistic. One is enough.
- Accuracy and sources below 20 of 25.
- Originality and added value below 10 of 20. Pages without added value are what the rater guidelines rate Lowest in section 4.6.6.
- A health, money, legal or safety topic without a named subject reviewer.
| Result | Verdict | What happens next |
|---|---|---|
| 85 to 100, no gate failed | Publish | Log the score with the date; rescore at the next refresh |
| 70 to 84, no gate failed | Revise | Fix the criteria that lost the most points, then rescore |
| Below 70 | Rewrite or drop | A new brief with sources attached beats patching |
| Any gate failed | Blocked | The total does not matter until the gate passes |
The 85 mark is our recommendation, set high on purpose: an article that loses 15 points has several real problems, and readers notice them before any search engine does.
What does a scored example look like?
Example: an 1,800-word AI draft comparing two SEO content tools, scored, rewritten from a brief with sources attached, then scored again.
| Category | Max | Draft 1 | Draft 2 | What changed |
|---|---|---|---|---|
| Answer and intent | 20 | 14 | 19 | Answer moved from paragraph three to the opening (2 to 6 points) |
| Accuracy and sources | 25 | 13 | 23 | Draft 1: 6 of 14 claims sourced (4 of 10 points), two outdated prices. Draft 2: 11 claims, all sourced; one quote paraphrased |
| Originality and added value | 20 | 8 | 16 | Added an own comparison table with prices checked that day and a named reviewer |
| Depth | 10 | 9 | 9 | Term coverage was high from the start |
| Structure and readability | 10 | 6 | 9 | Two repeated sections merged |
| Links | 5 | 2 | 5 | “Click here” anchors rewritten; link to the next step added |
| Trust and page settings | 10 | 5 | 9 | Disclosure added, meta description rewritten |
| Total | 100 | 57 | 90 | Draft 1 failed two gates; draft 2 passes all four |
The arithmetic: 14 + 13 + 8 + 9 + 6 + 2 + 5 = 57 for draft 1, and 19 + 23 + 16 + 9 + 9 + 5 + 9 = 90 for draft 2. Draft 1 also failed the accuracy gate (13 is below 20) and the originality gate (8 is below 10).
Look at the depth row. Draft 1 scored 9 of 10 there, the category where a term-coverage tool would have shown a strong score, while the same draft carried two wrong prices and eight unsourced claims. That gap is the reason the rubric exists.
What do Surfer, Clearscope, MarketMuse and Originality.ai measure?
Four tools sell a number that sounds like quality. Here is what each one measures in its vendor’s own words, with prices from each pricing page, checked 5 Nov 2026.
| Tool | Score and scale | What it measures, per the vendor | Free plan / paid from |
|---|---|---|---|
| Surfer | Content Score, 0 to 100 | An SEO Score (keywords, NLP terms, title and headings, length, images, alignment with top-ranking pages) plus an AI Search Score (Facts Coverage and Upfront Intent Alignment) | “Start for Free” buttons, no free plan terms stated; Discovery from €49 a month, billed yearly (shown in euros when checked from the Netherlands) |
| Clearscope | Content Grade, letters with A++ at the top | Relevance and comprehensiveness: how many of the relevant terms from top-ranking content the draft includes; a separate readability score based on Flesch Reading Ease | 14-day free trial; Essentials plan $129 a month |
| MarketMuse | Content Score, against a Target Content Score | How many subtopics from its topic model the page includes, compared with competing pages (definition on the pricing page) | Free plan: 1 user, 10 queries a month; paid plans list no price (book a demo) |
| Originality.ai | AI score in percent, plus a Content Quality Score of High, Good or Low | The share of the text showing AI involvement: under 10% is Likely Original, 10 to 80% Mixed, 81 to 100% Likely AI (5.0 model); the Content Quality Score rates writing quality | 100 free credits once (1 credit = 100 words); Pro $12.95 a month billed yearly, $14.95 monthly |
Three of the four compare your draft with pages that already rank. Surfer’s editor benchmarks against the top-ranking pages for the query, Clearscope weights each term by how much it appears in competitors’ articles, and MarketMuse calculates its target from competing pages. Originality.ai measures something else entirely: how the text was produced and how well it reads.
What do tool scores miss?
Every tool above answers a narrower question than “is this article good?” The gaps are predictable:
- Truth. Surfer’s Facts Coverage measures how fully a draft addresses the facts and entities common in AI-generated answers for the topic. Coverage and correctness are different tests: a draft can mention every expected fact and get three of them wrong.
- Originality. Term tools reward resemblance to pages that already rank. Google’s self-assessment asks the opposite question: whether the content adds value compared with the other pages in search results.
- Sources. No score checks whether a link supports the sentence it sits in.
- Effort and experience. No tool can see that a person worked out the arithmetic, interviewed a customer or checked a price that morning.
- Fact-check reliability. Originality.ai’s own accuracy study reports 86.69% accuracy for its fact checker, so about 13 in every 100 verdicts are wrong by the vendor’s own measure.
The vendors say some of this themselves. Surfer’s help center advises against chasing 100, recommends aiming 10 to 20 points above your main competitors, and warns that turning guideline boxes green mechanically can make the score drop.
How do tool scores fit into the rubric?
Give tool scores the one criterion they measure well: subtopics and terms covered, 5 points in the depth category. Our mapping:
- Surfer: 67 or higher earns 5 points (Surfer’s docs describe 67 to 100 as optimum quality and relevance), 34 to 66 earns 3, below 34 earns 0.
- Clearscope: a grade at or above the median of the top five competitors in its Competitors tab earns 5; one grade step below earns 3.
- MarketMuse: at or above the Target Content Score earns 5; within 10% of it earns 3.
- Originality.ai AI score: 0 points either way. It measures how the text was made, and the rater guidelines say that alone does not determine quality. Track it only if a client contract asks for it.
- Automated fact checkers: a pre-screen that shows the scorer where to look first. The accuracy points still come from a person with the source open.
With this mapping, a perfect tool score moves the total by 5 points at most. That is the right size for a measure of term coverage.
Which quality standards should every article meet?
Quality standards are the rubric written as rules a writer can read before drafting. Put these eight at the top of every brief:
- The article answers its title in the opening 100 words.
- Every factual claim links to a source that says the same thing.
- Prices, dates and versions are checked on publishing day, and the article says when.
- The article contains at least one thing the top five results lack.
- No other page on the site targets the same query.
- A real person is named as author, and a subject reviewer is named on health, money, legal and safety topics.
- Title tag, meta description and alt text are written for this article.
- The article scores 85 or more with no gate failed.
Standards describe the finished article. The process rules behind them (what AI may draft, what gets disclosed, which records to keep) belong in your AI SEO governance policy.
Which Copycat guides feed the score?
The score grades the result of the stages before it.
AI SEO editing checklist: the nine editing passes to run before scoring.
Human review for AI SEO articles: the fact-check log that the accuracy points come from.
AI SEO content writer tools: the drafting tools, compared with prices.
AI SEO governance policy: the process rules behind the quality standards.
Refresh old content with AI: rescoring and updating published articles.
FAQ
What score should an AI article reach before it is published?
On this rubric, 85 or more with all four gates passed. Between 70 and 84, fix the criteria that lost the most points and score again. Below 70, a new draft from a sourced brief is faster than patching.
Does a high Surfer or Clearscope score mean the article will rank?
No tool promises that. Both scores measure how closely your draft’s terms and structure match pages that already rank, which is one input among many. Surfer’s own help center advises aiming 10 to 20 points above competitors instead of chasing a perfect score.
Should an AI detection score count toward quality?
Leave it out of the total. Google’s rater guidelines state that using generative AI tools alone does not determine a page’s quality rating, and a detector’s percentage says nothing about whether the facts are right. Record it separately if a client requires it.
Who should score the article: the writer or someone else?
Someone who did not write or generate it. The writer can pre-score structure and links to catch obvious gaps, but accuracy and originality need a second person, the same way a fact-check does.
Can an AI model score its own draft with this rubric?
It can pre-score structure, repetition and link anchors. Accuracy stays with a person, because checking its own claims means comparing them with the memory that produced them. Treat any model-generated score as a draft for the human scorer to correct.
When should a published article be scored again?
Whenever a fact in them can change: a price, a plan, a version, a rule. Tool comparisons need a rescore every time a vendor changes its pricing page. The refresh itself is covered in refreshing old content with AI.