4-Dimension Content Scorecard: E-E-A-T, Depth, Freshness, Structure (Claude)

Last month I pulled 47 drafts from my "ready to ship" folder — the ones my gut said were good — and ran them through a 0-100 scorecard. 19 came back under 75. Six were under 60. I would have published all 47. The scorecard caught 19 posts that would have slowly dragged the site down, and I had been calling my gut "good enough" for years.

That scorecard is four dimensions, each scored 0-25: E-E-A-T, Depth, Freshness, Structure. The math is fixed, the rubric is in the prompt, and the same rubric can audit existing posts in batches of 10. This is the rubric, the prompt, the threshold logic, and the reason a scorecard survives staff turnover better than a style guide.

The Four Dimensions

Each dimension scores 0-25. Anything under 16 in a single dimension is a structural problem, not a line-edit problem — you don't fix it, you rewrite that section. The total is informational, not a judgment; what matters is the floor on any one dimension.

1. E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) — 0-25

E-E-A-T is the part most "AI SEO" articles get wrong. They chase E-E-A-T by adding author bios and "medically reviewed" disclaimers. That misses the point. Google's quality systems are looking for evidence in the content itself, not in the byline. The four letters, decomposed for a scorecard:

• Experience (0-6) — Does the writer demonstrate first-hand use? Screenshots from their own account, a number that came out of their own spreadsheet, a workflow with their real files in it. A post that says "I ran this 3 times in 2 days" scores higher than one that says "you can run this in your workflow." If you can't tell whether the writer actually did the thing, the score is 0-2.
• Expertise (0-6) — Specific technical depth, not jargon density. Naming the exact API call, the actual config file, the real error message you hit. "Configure your DNS" is 0. "Add an A record for mail.example.com pointing to 192.0.2.45 with TTL 300, and a TXT record for SPF" is 5-6.
• Authoritativeness (0-6) — Linked sources, named tools, named people, linked studies. Not "according to experts" — who? Not "studies show" — which one? A post that links to Google's own docs, an Ahrefs study, and a named person on LinkedIn scores 5-6. A post that links to "various sources" scores 0-1.
• Trustworthiness (0-7) — No unsupported claims, dates stamped on time-sensitive content, no broken-link claims, hedging on the things that genuinely deserve hedging. The 7-point weight here is intentional: trust is the costliest dimension to lose. A single confidently-wrong stat tanks the post's authority for everything else in it.

A 22/25 E-E-A-T post has a real example, a real number, a real screenshot, a real link, and a date stamp. A 12/25 E-E-A-T post has an author bio at the top and that's it.

2. Depth (0-25)

Depth is not word count. A 3,000-word post can score 8. A 700-word post can score 22. Depth means did this go past the surface-level take?

Three signals I score on:

• Numbers and examples (0-10) — Not "many marketers" but "63% of B2B SaaS sites." Not "this works well" but "this added 410 clicks/month to a category page on a 200-URL site." The unit of evidence is the specific number. Two or more in the post → 6-8. Five or more, including one contrarian or counter-intuitive one → 9-10.
• Counterarguments addressed (0-8) — Every post makes a recommendation. A high-depth post names the case where the recommendation is wrong, the cost of the recommendation, and the alternative. "Use Claude for content scoring" is shallow. "Use Claude for content scoring — but skip it for legal/compliance posts because the cost of a hallucinated citation is too high" is depth.
• Mechanism, not just outcome (0-7) — Did the writer explain why this works, or just that it works? "Posts with screenshots rank higher" is outcome. "Posts with screenshots rank higher because the SERP click-through rate (SERP 即搜索引擎结果页) is 1.6-2.1x for image-rich results, and dwell time is 18% longer because the user confirms the page matches the intent before scrolling" is mechanism.

Depth is the dimension where AI-generated content fails most reliably. A model can write a 1,800-word post that says nothing the top 5 results don't already say. Score: 9/25.

3. Freshness (0-25)

Freshness is the dimension editors skip because it requires checking, not writing. Three checks:

• References in the last 18 months (0-8) — Stats, studies, product launches, API changes. "According to a 2023 HubSpot report" is stale for almost every marketing topic. 18 months is the floor. Posts that rely on pre-2024 SEO data are now actively misleading because the March 2024 core update changed the underlying assumptions.
• Current product UI (0-8) — If the post has a screenshot of a dashboard, an admin panel, or a settings screen, is that screen the current one? Klaviyo's campaign builder has had three significant UI changes since 2023. Mailchimp's automation builder moved twice. A screenshot from 2022 is a 2/8. A screenshot from the last 6 months is 7-8.
• No broken-link claims (0-9) — A "Klaviyo does X" claim is a broken-link claim if it doesn't link to a Klaviyo doc. A "Google says X" claim is broken if the citation is to a 2021 blog post that has since been updated. The fix isn't to remove the claim — it's to add the link. Posts that link to live, current sources score 8-9. Posts that say "according to recent research" score 0-2.

The reason freshness gets its own dimension: it's the one that's hardest to fix after publish. A shallow post can be deepened. A stale post has to be re-researched from scratch, and most teams never do it.

4. Structure (0-25)

Structure is the easiest dimension to score because the rules are mechanical. Five checks:

• Scannable H2s (0-6) — Each H2 is a claim, a question, or a sub-promise — not "Introduction," "Background," "Other Considerations." "Why 'Helpful Content' Is Harder to Audit Than You Think" is 6. "Overview" is 0.
• Useful sub-bullets (0-5) — Bullets that compress information that would take 2-3 sentences to explain. A 14-bullet list where every bullet is one word is 0. A 5-bullet list where each bullet replaces a paragraph is 5.
• Working TOC (0-4) — For posts over 1,200 words, an in-post table of contents that links to the H2s. Not a sidebar widget — actual in-body links.
• No wall-of-text (0-5) — No paragraph over 5 sentences. No section over 4 paragraphs without a subheading, table, or list. This is the dimension where AI fails most reliably in the other direction: walls of 3-bullet lists with no connective tissue.
• Front-loads the answer (0-5) — The first 200 words answer the H1 question, or at least commit to an answer. Not "in this article we'll explore" — "the answer is X, and here's why." Readers (and Google's extraction systems) reward posts that give the answer up.

The Prompt

The whole scorecard runs from a single prompt. I paste the draft and the rubric; Claude returns four sub-scores, a total, and a remediation list. The prompt, exactly as I run it:

You are a strict pre-publish content reviewer. Score the following draft on four dimensions, each 0-25. Use the rubric below. For each dimension, give the sub-score, one sentence of evidence citing a specific line, and a one-line fix if the score is below 18.

E-E-A-T (0-25): Experience — does the writer demonstrate first-hand use (screenshots, real numbers from their work, named workflows)? Expertise — specific technical depth, not jargon? Authoritativeness — linked sources, named tools, named people? Trustworthiness — no unsupported claims, dates on time-sensitive content, hedging where deserved?

Depth (0-25): Numbers and examples (specific, not vague)? Counterarguments addressed? Mechanism explained, not just outcome?

Freshness (0-25): References in the last 18 months? Current product UI in any screenshots? No broken-link claims (every "X says Y" claim is linked to a live source)?

Structure (0-25): Scannable H2s (claims, not labels)? Useful sub-bullets (compression, not fragmentation)? Working TOC if over 1,200 words? No wall-of-text paragraphs? Front-loaded answer in the first 200 words?

End with: (1) sub-score per dimension, (2) total /100, (3) a numbered remediation list — one line per issue, phrased as a fix ("Add a date stamp to lines mentioning Klaviyo's UI; that section is from 2023"). If the total is below 75, the post is not publishable. If any single dimension is below 16, the post is not publishable regardless of total.

Draft: [paste]

The "remediation list" is the part that turns the scorecard from a verdict into a workflow. "E-E-A-T: 14/25" is a number. "Add a date stamp to lines mentioning Klaviyo's UI; that section is from 2023" is something you can fix in 4 minutes.

The Threshold Logic

A 75/100 total is the floor. A 16/25 sub-score is the floor on each dimension. Both have to clear.

The reason for two thresholds, not one: a post can score 78 by being 22/25 on three dimensions and 12/25 on one. That post looks good in the total but has a structural hole — usually E-E-A-T, usually because the writer skipped the first-hand evidence. The 12/25 dimension is the post's load-bearing wall, and the next Google update is the earthquake. A single bad dimension drags everything else down on re-read.

In my last 60 posts scored with this rubric:

• 38 cleared 75/100 with no dimension below 18 on the first pass. Shipped.
• 14 cleared 75 but had one dimension at 16-17. Rewrote that section, re-scored, shipped.
• 6 came in under 70. Killed the post, started over with a tighter brief.
• 2 cleared everything but failed a fact-check on a stat. Sent back to research before re-scoring.

The 75 threshold isn't a quality judgment. It's a publish-or-revise gate. The number is just a number; what it does is turn a vibes call into a decision. I no longer argue with myself about whether a draft is ready — the scorecard tells me, and the remediation list tells me what to fix.

Auditing Existing Posts in Batches of 10

The scorecard isn't just for new drafts. It's also the cheapest way to triage a backlog. The workflow I run on a 200-post archive:

Step 1: Export the URL, title, last-modified date, and first 200 words of every post. Drop into a sheet. The first 200 words are the only thing the scorecard needs to flag a post — the full draft is for the posts that get flagged.

Step 2: Run the scorecard in batches of 10. Paste 10 posts into a single Claude turn, ask for a per-post dimension score and a verdict (ship / refresh / delete). The first-200-words sample is enough to flag 80% of low-quality posts; the other 20% need a deeper read, but you've already cut the work by 5x.

Step 3: Triage into three buckets:

• Ship as-is — score ≥ 75, no dimension below 18. Do nothing.
• Refresh — score 60-74, or any single dimension 14-17. Add a 1-2 hour refresh pass (new screenshots, new stats, new links, new date stamp). Re-score.
• Delete — score < 60, or any single dimension below 14. The cost of refreshing exceeds the traffic value. 301 (永久重定向) to the closest related post and move on.

In a typical 200-post catalog I've audited with this workflow, the breakdown lands at roughly 50% ship, 25% refresh, 25% delete. The 25% delete is where the value is. Every post you keep that doesn't earn its keep is a tax on every post you're trying to rank. Site-wide quality signals are real, and a 200-post site with 50 dead-weight posts underperforms a 150-post site with zero.

The 1-2 hour refresh budget is the part most teams underestimate. Refreshing isn't "add a new paragraph." It's a real pass: re-screenshot the UI, re-link the sources, re-stamp the date, re-check that the example still works. A refresh that takes 20 minutes is usually a paint job, not a fix.

Why a Rubric Beats Vibes-Based Editing

The case for the scorecard isn't accuracy — Claude will miss things a human editor catches. The case is consistency across writers and survival across turnover.

When I had a team of three writers and a part-time editor, the "looks good" call lived in the editor's head. She had a feel for what a good post was. Then she left. The next editor had a different feel. The third editor had no feel and was learning from the two previous editors' comments, which contradicted each other. The quality of what shipped was a function of who happened to be reviewing, not what the post was.

A rubric fixes this in two ways. First, it forces every reviewer — human or AI — to score the same dimensions on the same evidence. "This post is bad" is a vibe. "This post scores 12/25 on E-E-A-T because it has no first-hand evidence and the only source is a 2021 blog post" is a verdict. The second is a 5-minute fix.

Second, a rubric survives the editor leaving. The prompt is the spec. A new editor — or a new model — can pick it up and run the same scorecard the same way. The institutional knowledge lives in the prompt, not in someone's head.

The harder version of the case: a rubric also surfaces the things you didn't know to look for. I added "counterarguments addressed" to the depth rubric after noticing that my best-performing posts all had a "but this is wrong when…" section, and my worst-performing ones didn't. The model can do that analysis if the rubric asks for it. It can't, if the rubric doesn't.

What the Scorecard Can't Catch

Two things, and they're both real.

First, factual accuracy on the specific claims. Claude can check that a claim is supported, but it can't verify that the underlying stat is correct. A post can score 24/25 on E-E-A-T (linked source, named study, current data) and the source itself can be wrong or fabricated. I still do a 60-second human spot-check on the 3-5 most load-bearing stats in every post. The scorecard gets a draft to the door. The human decides if the door is real.

Second, whether the post matches the audience's actual problem. A post can score 78/100 by every mechanical measure and still be the wrong post — answering a question the reader didn't ask, or recommending a tool for a problem the reader doesn't have. This is the hardest thing to scorecard because it requires knowing the audience's specific situation. The closest proxy in the rubric is the "intent match" check in structure: does the post actually answer the H1 question. But that's a sentence-level check, not a fit-with-the-reader check.

A scorecard is a forcing function, not a quality system. It forces consistency. It doesn't force correctness or fit. Use it for the first, keep a human in the loop for the second.

The Single Rule

If I had to compress the whole framework into one rule: a 0-100 scorecard with sub-scores and a remediation list is the cheapest way to make a content operation survive its own growth. Three writers, ten writers, a hundred posts, two thousand — the rubric runs the same. The vibe check doesn't.

Run it on your next draft. Then run it on the ten drafts you already shipped this month. The number is less interesting than what it forces you to fix.