Most content audits are about technical SEO — broken links, duplicate meta tags, missing schema. This guide is about what has to happen before any of that matters: whether the text itself is worth publishing at all.
I'll walk through how to check an article in about three minutes, using a few signals that genuinely correlate with ranking and reader trust. Some of this you can do by hand. Some of it needs an online text audit tool, because humans are slow at counting patterns across two thousand words. But the thinking is the same either way, and the whole point is that you don't need a linguist on retainer to know if a piece is publish-ready.
Why this became a real problem
Up until 2023 or so, content quality was mostly a writer-versus-writer conversation. Good writers produced better text, bad ones produced worse, and editors existed in the middle to catch the difference. Then generative AI showed up, got cheap, got everywhere, and the ground shifted.
Here's the thing that's not discussed enough: the problem isn't that AI writes badly. In 2026, a halfway-decent prompt to Claude or GPT produces prose that beats the median human copywriter on pure fluency. The problem is that fluency isn't what readers want, and it's definitely not what Google rewards.
The March 2024 Helpful Content Update wasn't really about AI, even though everyone called it the "AI update." If you read Google's own guidance carefully, it was about content made for search engines rather than people. The confusion happened because AI-generated content happened to match that description more often than human content did. Sites that got hit were often running pure-AI content factories. Sites that used AI but layered in genuine expertise on top came through fine, or even grew.
This matters for how we think about auditing. If the goal were just "detect AI, remove AI," you could buy an off-the-shelf detector for ten dollars a month and be done. The actual goal is harder: detect low-quality content regardless of who wrote it. A human copywriter with no domain knowledge produces the same template-y, generic, authority-free text as a cheap AI prompt. Both underperform in the same ways, and both should fail the same audit.
There's also the trust side. A Reuters Digital News study in 2024 found that around 80% of readers report lower trust in articles they suspect are AI-generated, even when they can't always articulate why. They just bounce. Your content can rank and still fail if readers don't believe it. Which means a good audit has to predict not just ranking risk, but reader experience.
The method below is what I've been using — refined through building the Text Audit tool at Algorithm, but the logic works whether you use the tool or do it manually.
The four things worth measuring
You can spend months chasing metrics. Flesch-Kincaid score, sentence length variance, adverb density, passive voice ratio — the internet is full of these, and most of them are noise. I've been through this exercise myself more than once. The list below is what's left after throwing out the things that didn't correlate with anything that mattered.
1. How much the text sounds like a template
This isn't "is it AI?" — it's "does it follow a predictable shape?" AI content and bad human content fail here for the same reason: both default to the safest, blandest structural path.
The giveaways are things you can spot once you know to look: paragraphs that all cluster in a tight length range, with no one-line beats and no eight-line digressions. Transitions that recycle the same handful of phrases — "Additionally," "Furthermore," "Moreover," and the deathly "It's important to note that..." A tone that never shifts, even on a topic where you'd expect frustration or dry humor or at least a change of pace. Every section wrapping up with a mini-conclusion, even when it doesn't need one.
How to check by hand: read the draft aloud. Not in your head. Out loud, at normal speed. If your voice settles into a monotone and stays there for two thousand words, the text is too flat. Real writing forces your voice to shift — short sentence, long sentence, question, aside, back on topic.
How to check at scale: a tool can measure variance in paragraph length, count unique transition types, and flag when tonal variety drops below baseline. Our detector runs about ten such signals in parallel and combines them. But honestly, you'll get most of the value just from reading aloud.
2. Information density
This one is simple to describe and brutal in practice. What percentage of sentences in your article actually contain information?
A sentence contains information if it has at least one of four things: a specific number, date, price, or measurement; a named thing — tool, product, place, person, study; a cause-and-effect claim that could be argued with; or a comparison with a specific alternative. Sentences that fail on all four counts are generic. They sound reasonable, they fill space, they tell the reader nothing they didn't already suspect.
"Good content is important for SEO" is a generic sentence. "Headings with question format outperform declarative ones on mobile by around 12% in CTR, but only for informational queries" is not. The first one could be written by anyone, including someone who knows nothing about SEO. The second requires experience.
The audit question for your own draft: what's your generic ratio? If more than 40-50% of your sentences are generic, the piece is padding. Trim or replace. I regularly see articles of 2,000 words that would be stronger at 1,200 once the empty sentences are cut.
Hard to check manually on a long piece, because you have to classify every sentence. It's where automation helps most. The manual test, if you want one: pick a random paragraph, go sentence by sentence, ask "could this have been written by someone who knows nothing about the topic?" If yes — generic.
3. Experience and authority markers
This is what Google's quality raters are trained to look for under the E-E-A-T framework, and it's what readers use — mostly unconsciously — to decide whether to trust you.
Experience markers are details that could only come from having actually done the thing:
- Specific tools used ("I set up three LinkedIn campaigns with a $500 daily budget in November 2024")
- Real outcomes with numbers ("conversion dropped from 3.2% to 1.8% after the redesign")
- Limitations honestly stated ("this worked for B2B SaaS under 50 employees; it failed completely for enterprise")
- Mistakes and corrections ("my first attempt used the wrong attribution window and misread the results for two weeks")
Content without these reads like someone repeating what they've read elsewhere. Content with them reads like someone who's been in the room.
You cannot fake this without lying, which is a feature. A generic AI output will avoid specific claims because hallucination is a liability. A copywriter who's never done the thing they're writing about will avoid them for the same reason. The absence of concrete experience markers is itself a strong signal.
The audit question: count the sentences in your article that contain experience-specific content — a number from actual work, a named case, an explicit limitation, a correction. If there are fewer than five in a 2,000-word piece, the authority isn't there, regardless of who the author's bio claims to be.
4. Cognitive load versus coherence
Readability scores like Flesch-Kincaid try to measure this and mostly fail. They flag long sentences and reward short ones, as if short sentences were always clearer. They're not. "The cat sat on the mat. It was a red mat. The cat was orange." has a great Flesch score and is unbearable to read.
What actually matters is whether the reader can follow the thread. Signs that they can't:
- Paragraphs that switch topic mid-way without warning
- Pronouns whose antecedents are unclear
- Lists that mix things at different levels of abstraction
- Technical terms introduced without definition and never explained
Signs that they can:
- Each paragraph has a single point
- Transitions explain why the next paragraph follows (not just "Furthermore")
- Abstract claims are immediately followed by concrete examples
- The reader could summarize each section in one sentence without going back
This is the hardest signal to automate well, because it's really a semantic question. You can get partial signal from sentence length variance, paragraph-to-paragraph topical similarity, and technical term density — which is what tools do — but the final check is always "read it and see if you can follow it."
The three-minute audit in practice
Here's the actual workflow, which I do before publishing anything on Algorithm's blog — this article included.
Minute 1 — the read-aloud pass. I read the full draft out loud. Not in my head. Out loud, at normal speech speed. Things I listen for: monotone drift (am I still engaged by paragraph five?), awkward run-on sentences, places where I stumble because a pronoun is ambiguous, paragraphs that feel like filler. I mark them as I go. This catches maybe 40% of the problems.
Minute 2 — the detector pass. I paste the draft into the Text Audit tool. What I care about, in order: the AI likelihood score (is it under 30%?), the information density breakdown (what percentage of sentences are generic?), the authority score (are experience markers showing up?), and the top three suggestions from the tool.
The tool will tell me things I'd miss on a close read. The most useful output is usually the list of generic sentences — it's humbling, every time, to see how many sentences I wrote that genuinely say nothing.
Minute 3 — the revision decision. Based on the two passes above, I pick exactly one thing to fix first. Not everything. If information density is low, that's usually the highest-leverage fix, because replacing three generic sentences with three specific ones lifts multiple metrics at once. If the AI likelihood is high but density is fine, it's usually a structural problem — paragraphs too uniform, transitions too formulaic — and the fix is breaking the shape rather than adding content.
Then I revise, and either re-run the detector or call it done, depending on how bad it was.
Three minutes isn't enough to fix a bad article. It's enough to decide whether the article is worth fixing, or whether it needs to go back to the writer.
When the audit is wrong
No honest guide to this should skip the failure modes. A few I've hit repeatedly, worth naming:
Specialized domains confuse detection. A well-written medical article on drug interactions will use clinical terminology, repetitive structure (because the safety information has to be structured), and a clipped tone. A generic detector will flag it as templated. The fix is human judgment: if you're in YMYL territory, a high template score doesn't automatically mean low quality.
Short texts are unreliable. Anything under about 800 words doesn't give the statistical signals enough to work with. Our tool rejects text under 300 words outright for this reason — below that, the confidence intervals are wider than the score itself. If you're auditing product descriptions or short landing page copy, the audit methodology above is mostly useless. You need a different approach for short-form.
Translated content pattern-matches oddly. A high-quality article translated from Russian or Japanese into English often reads as AI-generated to detectors, because the translation process smooths out the native-speaker rhythm without introducing the statistical weirdness a human writer would. If you're auditing multilingual content, calibrate for this — the detection is telling you something real about the text (it doesn't sound native) but not what it thinks it's telling you.
Very good human writing can score high. Experienced writers who've spent years optimizing for clarity can produce prose so clean that detectors flag it. The irony is that the same cleanness is what makes the writing good. If you're running audits on top-tier writers and getting high AI scores, check whether the issue is real (templated patterns, no experience markers) or whether the writer is just unusually disciplined. The fix for that case is to relax thresholds for known strong writers, not to blame them for being good.
I mention these because when your audit and your instinct disagree, you should figure out which is wrong before acting. Sometimes the tool is. Sometimes you are.
What to do in the next hour
If you made it this far, here's what I'd suggest as practical next steps:
Open the last article you published. Do the read-aloud pass. Just that. Note what you hear.
Then run it through any content auditor — ours, or any other — and look specifically at the generic-sentences list. Pick the five worst and ask yourself whether they'd survive if you deleted them. Usually three or four would.
That exercise alone, done once on something you've already shipped, is worth more than reading another guide. The gap between "I think my content is fine" and "I can see the specific places it isn't" is where real improvement starts. Everything else is mechanics.
The hard part isn't the audit. The hard part is being willing to find out.