E-E-A-T Review Checklist for AI-Assisted Blog Posts

You have a draft in front of you. An AI wrote most of it. Now you need to decide: publish, revise, or kill.

Google's Search Quality Rater Guidelines use E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) to evaluate whether a piece of content deserves to rank. AI drafts fail these checks in predictable ways. Not randomly. Not occasionally. Predictably. That means you can review for them systematically, which is exactly what this checklist does.

This post is not theory about what E-E-A-T means. It's a working checklist for the editor or founder doing the review. Read it with a draft open.

Why AI Drafts Fail E-E-A-T (and Where to Look)

E-E-A-T is Google's framework for assessing content quality across four dimensions: Experience, Expertise, Authoritativeness, and Trustworthiness. "Experience" was added in December 2022, specifically to distinguish first-hand knowledge from aggregated claims. That addition was a direct response to the rise of AI-generated content that sounds authoritative but has never actually done the thing it's describing.

That's the core failure mode. AI models synthesize patterns from training data. They produce confident prose. But confident prose and genuine expertise are not the same thing, and quality raters are trained to spot the difference.

Two failure categories show up most often in AI drafts:

• Fabricated citations. The URL is invented, or it resolves to a page that doesn't contain the claimed fact.
• Generic claims with no supporting data. Sentences that sound informative but could have been written by anyone who spent ten minutes on Google.

Both are checkable. Both are fixable, if you catch them before publishing.

Check 1: Sources and Citations

Start here. Citation problems are the most damaging and the fastest to spot.

Open every linked URL. Confirm it resolves. Confirm the page actually contains the information the draft attributes to it. Northwestern's AI evaluation guide recommends "lateral reading" for this: before trusting a source, check what other sites say about that source's credibility, not just what the source says about itself (Northwestern University Libraries).

A real URL that resolves to a real page is not the same as a citation that supports the specific claim. AI frequently cites a real source that does not contain the stated fact. Open the source. Find the sentence. Confirm the match. If you can't find the sentence, the citation fails.

Flag every statistic that traces back to a secondary source ("according to a study..."). That phrasing is a signal that the original source was never checked. Your job as reviewer is to trace the claim to the original paper, report, or dataset. If you can't find it, the stat gets cut or flagged for replacement.

Cross-check claims with at least one additional credible source: a claim that appears in only one source should be treated as unverified until a second independent source confirms it (St. Kate's Library). Single-source claims aren't automatically wrong, but they're higher risk.

Check 2: Factual Accuracy and Claim Precision

After citations, check the numbers. Every percentage, date, dollar figure, and study result needs to be verified against the primary source, not assumed correct because the citation resolved.

AI models round, misremember, and occasionally invent numbers. A stat might be directionally right but numerically wrong. That's still a factual error. Check the specific figure, not just the general claim.

Watch for what LlamaIndex's glossary calls "content faithfulness" failures: the draft states something the source does not actually say (LlamaIndex). This is different from a fabricated citation. The URL is real. The page exists. But the paraphrase is wrong, or the quote is stripped of context that changes the meaning. This requires actually reading the source, not just confirming it loads.

Check for outdated data. AI training cutoffs mean statistics from 2021 or 2022 may be presented as current without any date marker. Flag any stat older than 18 months. Either replace it with current data or add the date explicitly so readers know what they're reading.

Missing context is also a factual accuracy problem. A stat like "70% of marketers use AI tools" is not a meaningful claim without the sample size, survey date, and organization that ran the study. If that context isn't in the draft, find it or cut the stat.

Check 3: Experience Signals

Google's December 2022 E-E-A-T update added "Experience" for a specific reason: to reward content that reflects first-hand knowledge. An AI draft has none. The reviewer's job here is to find the places where a real human perspective needs to be inserted.

Look for process claims that float free of any specific example, tool name, or outcome. "The best way to do X is..." followed by a general recommendation. These are the insertion points. Replace them with the author's actual experience: the specific tool they used, the result they got, the thing that didn't work the first time.

Ask a simple question while reading: could this post have been written by anyone? If the answer is yes, it won't rank for competitive terms. Specific details signal experience. Vague recommendations signal aggregation.

Concrete experience signals include:

• Named tools the author has actually used
• Real numbers from the author's own work (not cited from somewhere else)
• Named mistakes and what they cost
• Comparisons that come from doing both things, not reading about both things

If none of those exist in the draft, the post needs a revision pass before it goes anywhere near publish.

Check 4: Authoritativeness and Brand Voice

Authoritativeness is partly off-page (backlinks, mentions), but the on-page signal is consistency. Does this post sound like the same person who wrote everything else on this site?

Check the draft against your established voice markers. For a research-first site, that means: direct sentences, specific numbers over vague claims, no hype words. Any AI draft that uses "leverage," "seamlessly," or "unlock" hasn't been voice-matched. Those words are pattern-matched from generic business writing and will make a post sound like it came from a template, not a person.

Verify the author bio or byline. Does the person's credentials match the topic? A post about enterprise data security signed by a general content writer is an authoritativeness mismatch. Quality raters notice. So do readers who are experts in the field.

Flag any claim that overstates the author's or company's position. "We've helped thousands of companies" without a number or case study is a trust liability. Either add the evidence or rewrite to something you can actually stand behind.

Check 5: Trustworthiness and Editorial Risk

Trustworthiness is the T in E-E-A-T. For YMYL (Your Money or Your Life) topics, Google's quality raters treat it as especially important. For non-YMYL content, it still determines whether a reader would recommend the post.

Check for missing qualifications. AI drafts routinely omit the "it depends" that a human expert would include. If the post recommends a specific tool, approach, or decision, the conditions under which that recommendation is true should be in the draft, along with the conditions under which it isn't.

Assess editorial risk directly. Ask: would publishing this create legal, reputational, or factual liability? Specific risk categories to check:

• Unverified health or financial claims
• Named competitor comparisons without supporting evidence
• Screenshots or quotes that can't be independently verified

The University of Pretoria's AI evaluation guide recommends a final gate for checking Currency, Relevance, Authority, Faithfulness, and Trustworthiness (University of Pretoria Library). Run each dimension before approving. Currency means the data is current. Relevance means the source actually applies to the claim. Authority means the source is credible. Faithfulness means the draft accurately represents what the source says. Trustworthiness means the overall post could be recommended to someone who needs accurate information.

Macquarie University's generative AI evaluation guide adds a practical note: when assessing AI-generated content, assume errors exist until you've checked, not after you've published (Macquarie University Library). That framing shifts the default. You're not looking for problems to fix. You're confirming the draft has no problems before approving it.

The Publish Decision: Pass, Revise, or Kill

Every reviewed draft ends in one of three outcomes:

Pass: Publish with minor copy edits. Every citation resolves and supports its claim. At least one section contains a specific first-hand detail. The voice matches the site's established tone. No claim creates editorial risk.

Revise: Return to the writer with specific flags. The structure is sound, but specific sections need experience signals, a citation needs replacing, or the voice needs a pass. The draft is fixable without rebuilding.

Kill: The draft's core premise is wrong or unverifiable, and revision would cost more than starting over. Kill when:

• More than 20% of citations are fabricated or unverifiable
• Experience signals can't be added without a complete rewrite
• Factual errors are structural (the argument depends on a false premise)

Run the checklist in order: sources first, then facts, then experience, then voice, then trust. Stop and flag as soon as you find a kill-level issue. There's no reason to finish the checklist on a draft you're going to kill.

One workflow note: if you find a kill-level issue in Check 1, close the doc and write the brief for a fresh draft. The time you'd spend revising around fabricated citations is time you could spend writing something publishable.

FAQ

How do I know if a citation is fabricated vs. just outdated?

A fabricated citation either returns a 404 or resolves to a page that doesn't contain anything resembling the claimed fact. An outdated citation resolves to a real page with real content, but the data is from an earlier year. Both are problems, but they require different fixes. Fabricated citations get cut. Outdated ones get replaced with current data or dated explicitly in the text.

Can I run this checklist on every post, or is it only for high-stakes content?

Run it on everything that goes under a real byline. The time cost is roughly 15-30 minutes per post depending on citation count. That's cheaper than publishing a post with fabricated stats that gets flagged by a reader with 40,000 followers. For lower-stakes content, you can run a lighter version: check all citations, check the experience signals, check the voice. Skipping the full trust audit only makes sense for posts where editorial risk is genuinely low.

What counts as a strong experience signal in a post I didn't personally write?

A strong experience signal names something specific: a tool, a result, a date, a decision and why it was made. "We use Ahrefs for keyword research" is an experience signal. "Keyword research tools can help you find opportunities" is not. If the author has real experience with the topic, they can add a paragraph. If they don't, the byline may need to change or a subject matter expert needs to review the draft.

Does E-E-A-T apply to every page, or just blog posts?

Every indexable page is evaluated against E-E-A-T in some form, but the weight varies by topic. YMYL topics (health, finance, legal) get the most scrutiny. Blog posts in competitive B2B niches get meaningful scrutiny. Product pages and landing pages matter less for E-E-A-T specifically but still benefit from accurate, credible content. Focus your review effort on anything where you're trying to rank for competitive terms.

If I add experience signals after an AI draft, does the post count as AI-generated?

That's a question about disclosure policy, and the answer varies by publication. From a quality perspective, a post that has been substantively edited by a human who added real experience, verified all citations, and matched the brand voice is a different artifact from a raw AI output. The edit matters. Whether you disclose AI involvement in the drafting process is a separate editorial decision that your publication should have a policy on.

Sources

• Northwestern University Libraries - Evaluating AI-Generated Content
• St. Catherine University Library - Evaluating AI
• LlamaIndex - Content Faithfulness Glossary
• University of Pretoria Library - AI Evaluation
• Macquarie University Library - Evaluating Generative AI

Copy this checklist into your editorial workflow doc before your next AI-assisted post goes live. If you're using Ryterr, the quality audit runs five of these checks automatically before the draft reaches you, with citation counts, stat counts, and a quality score visible in the pipeline. The checklist above covers what the human still needs to verify: experience signals, voice match, and editorial risk. Those three won't be automated away anytime soon, which is exactly why they need a human gate.