How We Use AI for SEO Content (And Why We Stopped Following the Standard Advice)

We publish content on PromptVault. We care deeply about whether that content ranks. And we have spent the past six months running experiments on AI-assisted content workflows that were sometimes productive, sometimes humbling, and occasionally embarrassing. This post documents what we actually do now, and the specific reasons we abandoned the approaches that everyone else seems to still be recommending.

We will start with the disaster, because that is where the real learning is.

Phase 1: Full AI Drafts (The Disaster)

Like most content teams that started experimenting with AI in late 2023, we began by doing what felt obvious: using AI to draft entire articles from a brief. We gave it the title, the target keyword, the desired length, a rough outline. It produced complete, grammatically perfect, topically coherent articles in minutes.

Those articles performed terribly. Not "slightly below expectations" — genuinely terribly. Pages that went live with minimal human editing consistently underperformed in search, generated almost no engagement, and attracted no backlinks. We pulled most of them within 90 days.

The reason, which took us embarrassingly long to articulate, is this: Google's core ranking signals in 2025 heavily favour content that demonstrates genuine experience and perspective. Full AI drafts are by definition competent generalisations — they synthesise what is already known without adding any new perspective, experience, or specific insight. They give Google nothing to reward.

The Google Signals We Believe Penalise Pure AI Content

We are not Google engineers, and we are not privy to algorithm details. But from our testing and observation of what ranks in our space, here is what we believe is happening:

• Engagement rate signals: Full AI articles have lower scroll depth and lower time-on-page. Readers can sense the hollowness even if they cannot articulate it. Google's engagement signals pick this up.
• Backlink generation: Original opinions, counter-intuitive findings, and specific data get shared and cited. Generic AI summaries of common knowledge do not. Our AI-heavy content generated essentially no organic backlinks.
• Entity clarity: AI content tends toward vague, hedge-everything language. Google's understanding of entity relationships appears to reward content that takes clear positions about specific claims — the kind of clarity AI is trained to avoid.
• Topical authority signals: Sites with AI-generated content across many topics may struggle to build topical authority signals because the content lacks the depth and specificity that signals genuine expertise.

Our Controversial Finding: Shorter, More Opinionated Wins

Across our own testing and observation of content that performs in our niche, we have reached a conclusion that runs directly against standard SEO content advice: shorter, more opinionated AI-assisted articles consistently outperform longer AI-generated ones.

A 600-word article that makes three specific, arguable claims and supports them with real experience outranks a 2,000-word AI-generated "complete guide" that covers the same topic comprehensively but generically. The word count is not the driver — the specificity of perspective is.

This matters for how you use AI in your workflow. If you are using AI to generate length, you are using it wrong. Use it to generate structure and research, then add human length through opinion, experience, and specific examples.

The Actual 5-Step Workflow We Use Now

Step 1: Human-Led Topic and Angle Definition

No AI involvement here. We decide the topic, the unique angle, and the specific claims we want to make based on our actual experience and opinions. This is the work AI cannot do — it has no experience, so it cannot generate genuine angles. Everything downstream depends on getting this right.

Time: 15–20 minutes per article.

Step 2: AI-Assisted Research Synthesis

We gather sources manually, then use this prompt to process them:

"Here are 5–7 sources on [topic]. Identify: (1) the key claims that appear in most sources, (2) any significant disagreements between sources, (3) gaps — important questions these sources do not answer, (4) the most specific data points across all sources. Do not add your own opinion. Just synthesise what is here."

Time saved: 45–60 minutes per article.

Step 3: AI-Generated Structure Draft

We use AI to produce a detailed outline only — not content. The prompt:

"Based on this angle [angle] and these source findings [synthesis], create a detailed outline with H2 and H3 headings. Each heading should represent a distinct argument or insight, not just a topic label. Flag which sections are best supported by the source data and which require original opinion from us."

Why this works: We get structure benefits without content generation. The outline tells us exactly where we need to add human perspective and where the research can carry the section.

Step 4: Human Writing with AI Assistance for Transitions

All primary content — every opinion, every claim based on our experience, every counter-intuitive observation — is written by a human. We use AI only to help with transitions between sections and to suggest alternative phrasings when a sentence is awkward. Never to generate substantive content.

Step 5: AI-Assisted SEO Review

Final pass prompt:

"Review this article for: (1) keyword density — is the primary keyword [keyword] used naturally 4–6 times?, (2) heading structure — do H2s and H3s create a logical hierarchy?, (3) internal link opportunities — what related topics are mentioned that we should link to?, (4) meta description quality — does the draft meta description make a specific, compelling claim rather than a generic summary?"

The Results Since Switching

Articles published using this hybrid workflow have significantly outperformed our Phase 1 full-AI content on every metric we track: organic traffic, average position in search, time-on-page, and social shares. We will not publish specific numbers because a content site's traffic data is competitive information, but the directional difference is substantial and consistent.

The insight we want to leave you with: AI is a research and structure tool, not a content tool. The moment you treat it as a content tool, you have given away the one thing that makes your content worth reading — your actual experience and perspective.

For the prompts we use in our SEO research workflow, explore our SEO prompt library. If you want to understand the broader time-savings context for AI in business workflows, our post on AI time tracking across 5 workflows shows the honest numbers. The foundation for all of this is solid prompt engineering — our prompt engineering 101 guide covers the principles that underpin every prompt in this workflow. And if you have developed an effective content workflow of your own, search our library for what others have built.

The Honest Warning

This workflow is more work than full AI drafting. It requires real human expertise to execute Step 1 and Step 4. If your team lacks that expertise in a given topic area, this approach will not save you — you will need either to develop the expertise or to be honest that you should not be publishing on that topic.

The teams that build durable search traffic with AI assistance are the ones who use it to amplify genuine expertise, not to simulate it.