Why your good content still does not rank

You publish good stuff. It is clean, it is correct, it covers the keyword three different ways. And it sits on page four, where nobody will ever find it.

I get asked about this a lot in advisory work. A founder ships twenty posts, ticks every box on the old SEO checklist, and the traffic line stays flat. The checklist was not wrong in 2021. It is wrong now.

Here is the short version. On December 11, 2025, Google began rolling out its third core update of the year, and it finished eighteen days later. That update folded the helpful-content system into core ranking and pushed E-E-A-T past the old health-and-money boundary into nearly every competitive search, as Search Engine Journal documented. The bar for what counts as good moved up. Clean and correct is the floor now, not the ceiling.

What ranks is proof that a real person who actually knows the topic wrote the thing. You cannot fake that with a template. So I am going to show you what real proof looks like, using this blog as the worked example, including the script that scrubs my own drafts before they are allowed to publish.

Why good content stopped being enough

For years you could rank a single strong page on its own merits. That era is over. The helpful-content signal used to sit apart, a separate filter bolted onto the side of search. Now it lives inside the core ranking system, and it grades your whole site at once. One thin page, one section that reads like a robot wrote it, and the score for everything else you publish takes the hit.

This is the trap for anyone who publishes to a schedule. Forty solid posts and a dozen thin ones, knocked out to hit a weekly quota, and the thin dozen weigh the whole site down. On a signal that grades the domain as a whole, weak posts are not neutral. They are a tax on everything else you ship.

Google says this in plain language in its own docs. Its guide on creating helpful, reliable, people-first content asks whether your work shows “first-hand expertise and a depth of knowledge (for example, expertise that comes from having actually used a product or service)“. It then draws a hard line on motive. Content made “primarily to attract search engine visits,” Google writes, is “not aligned with what our systems seek to reward.”

Read that twice. The search engine is telling you, out loud, that it does not want content built for search engines. It wants content built for the person reading it. After December 2025, it got much better at telling the two apart.

The four authority signals you cannot fake

E-E-A-T is short for Experience, Expertise, Authoritativeness, and Trust. Google’s quality-rater guidelines have leaned on it for years, but until recently it mostly mattered for medical and financial pages. The December update stretched it across almost everything you might write.

Each of the four has a cheap surface and an expensive core. You can buy the surface in an afternoon. The core takes years, and it cannot be bought.

Experience is the easiest to test. Have you actually done the thing? A byline that reads “10 years in logistics” is surface. A paragraph describing the exact way a warehouse slotting system fell over on you, and what you changed afterward, is core. Google’s people-first guidance hangs on three questions: who made this, how was it made, and why. Faking the who is trivial. Faking a believable how, full of the specific texture of real work, is close to impossible at scale.

The other three behave the same way. Expertise is fakeable as a confident tone, and real as technical depth a layperson could not invent. Authority is fakeable as a long author bio, and real as a track record outsiders can check against public talks and products. Trust you earn last, by linking to primary sources and admitting plainly what you do not know. Notice the pattern. The fakeable half is always a claim you make about yourself. The real half is always evidence someone else can go and verify.

This is the part founders underrate. They treat authority as a credential to display. Real authority is something you demonstrate instead, sentence by sentence, with details only a practitioner would know.

What does ranking-grade proof look like

Let me make this concrete, with the site you are reading right now.

I draft with AI. There, the quiet part out loud. Most consultants who write about ranking will not admit that, which is odd, because readers can usually tell anyway. The writer Tiago Forte said in public that Claude drafts roughly 90 percent of his content, and his readers revolted. One of them wrote, “if a writer isn’t dedicated enough to write its own articles, I’m definitely not going to be dedicated to read.” I get the reaction. It also misses the point.

The tool was never the problem. The problem is shipping the tool’s raw output, which reads like everyone else’s raw output, and slapping your name on it. What makes the writing yours is the work that happens after the draft: the real opinions and sources you bring to it, plus the rules that scrub out the machine tells.

So here is mine. This blog runs a word blacklist. These are words AI reaches for and I do not:

delve, navigate, landscape, robust, comprehensive, leverage, facilitate, seamless, tapestry, multifaceted

And that blacklist is not a sticky note on my monitor. It runs as code before every commit. A short script greps every post in the blog and fails the build if any one of those word families creeps back above its allowed floor:

check_family() {
  local name="$1"; local regex="$2"; local baseline="$3"
  local count=$(grep -rniE "$regex" $POST_DIR/*.md $POST_DIR/*.mdx | wc -l)
  if [ "$count" -gt "$baseline" ]; then
    echo "FAIL: $name has $count instances (baseline $baseline)"
    FAILED=$((FAILED + 1))
  fi
}

If a banned word sneaks in, the build breaks and the post does not go live. I cannot forget to run the check, because the check is no longer up to me. That is the gap between a guideline and a system. A guideline is a good intention you keep on a wiki. A system is a thing that fails loudly the moment you fall short of it.

Write like a person who actually knows

Authority shows up in voice, and voice is the first thing AI flattens. Generic-but-correct prose is the default setting of every model, and Google has gotten good at spotting it. So have readers.

Two measures give the machine away. Perplexity and burstiness sound technical, but they are simple. Perplexity is how surprising your word choices are. Burstiness is how much your sentence length varies. AI writing scores low on both. It reaches for the statistically safe word and holds a flat, even rhythm. Human writing lurches around. When one study analysed roughly 12,000 AI texts, a Carnegie Mellon and NJIT team found GPT-4o leaning on some words at more than 150 times the human rate. The smoothness is the tell.

Here is my own writing, from an old travel piece, years before any of this debate existed:

Night fell across the Ligurian coast and was atrociously beautiful.

Venice didn’t seem romantic, the gondola rides are rubbish, and nobody seems friendly over there.

No model produces “atrociously beautiful.” No model calls gondola rides rubbish. That texture, the blunt opinion and the odd pairing of words, is exactly what a perplexity-based detector reads as human, and it is exactly what a reader remembers an hour later. You cannot prompt your way to it from a standing start. It comes from having a view and a way of putting it. If you want the method for capturing yours, I wrote a full piece on building a voice profile, and another on why general AI beats template tools for this kind of work.

Build the system, not the one-off post

If you take one thing from all this, take this. Ranking is now a property of your whole site and your whole process, not any single post you sweat over. So stop polishing posts one at a time. Build the system that makes every post clear the bar by default.

Mine has three moving parts, and none of them are fancy. First, a short voice profile I feed the model, well under 400 words, so the draft starts closer to my register. Second, the blacklist and the build-time check you saw above, so the machine tells cannot ship. Third, a rule that every factual claim links to a primary source, the regulator or the original study, never a third-hand write-up. I run all three on this blog, and I teach the same loop in my AI courses at WashU.

The discipline that holds it together is the same iteration habit I use for prompt engineering: write the rule down once, then let an automated check catch your slips so you do not have to remember them. It also squares with a thing I keep repeating about where these tools earn their keep. AI does tasks, not jobs. Drafting a paragraph is a task. Being the authority behind it is the job, and the job is still yours.

None of this is a trick, and that is the whole point. After December 2025 the tricks stopped paying out. What is left is the un-fakeable stuff: real experience, a real voice, real sources, checked by a system that will not let you cut the corner. If you have actually done the work, this is the best news you have had in years. The bar went up, and you can clear it. If you have been faking it, the bar went up, and you cannot.