It’s a typical perception that AI-generated weblog posts are inherently low-quality and inferior to their human-made equivalents.
Corporations that scale AI-generated content material achieve this with the information that they’re making a trade-off, we imagine, selecting pace and scale on the expense of high quality. We agree that AI is quicker than any human, and it makes a passing first draft, however we all know we’re nonetheless sacrificing one thing vital through the use of it.
I now suppose this perception is outdated. I feel we’ve got reached the purpose the place generative AI can create content material indistinguishable from the huge corpus of human-written content material produced by content material entrepreneurs, like me, in earlier years.
AI has develop into a extra thorough researcher, a extra compliant adherent to model and voice tips, extra versatile in its response to suggestions, quicker, and extra environment friendly. There isn’t any longer a trade-off inherent in utilizing AI for content material creation.
This isn’t to say that every one AI content material is, by default, good; solely that the obstacles that prevented AI content material technology from being good have fallen. Entry to world-class writing via LLMs continues to be uneven, but it surely won’t keep this manner for lengthy. Functionally “excellent” AI content material is simply across the nook, for all of us, and it’s in our curiosity to acknowledge it.
Right here’s why.
Many individuals imagine there’s some inherent high quality inside human writing that AI can by no means attain, some artistic spark eternally unreachable for our silicon counterparts.
I received’t declare that AI will ever method the profundity of Shakespeare, however I’ll declare that “nice writing” is less complicated and extra mechanical than most individuals assume. A lot of the constituent elements of “nice writing” are issues that LLMs can do very, very effectively.
I’ve spent my total profession attempting to develop into a greater author by introspecting into the writing course of, asking why some issues work and others don’t. I’m not an knowledgeable on this discipline, however I’ve developed an efficient worldview of the mechanics of writing, and a set of writing rules that I comply with time and time once more.
A small snapshot of my modifying checklists for coaching new writers.
For instance, listed here are some random excerpts from my modifying guidelines:
- Have we addressed the obvious objections to this thought?
- Have we used dense phrases wherever potential? (“novel” as a substitute of “one thing new,” “worldwide” as a substitute of “on a world scale,” and so forth.)
- Have we changed weasel phrases with particular examples? (“enterprise outcomes”, “specialists imagine”, “analyzing knowledge”, “decide”, and so forth.)
- Can we keep away from making laborious issues sound straightforward?
- Can we open with an important info? (within the introduction, at the beginning of paragraphs)
- and so forth.
These rules are how I write, how I edit writing, and the way I train writing. They’re very simple, however executed in unison, they ladder as much as one thing that finally ends up fairly usually nearly as good, even nice, writing.
These rules are so easy, the truth is, that an LLM can carry out them completely—and often higher than I can. I usually apply these rules erratically, via fatigue or boredom or laziness. However for an LLM, these rules could be arrange as soon as and adopted indefinitely. They are often scaled evenly to a whole lot, 1000’s, tens of millions of outputs, codified in system prompts and SKILL information (extra on these within the subsequent part).
Should you acknowledge {that a} fundamental recipe exists for excellent writing (and I imagine it does), LLMs can comply with it very effectively. Chain many of those heuristics collectively in a dependable approach, and you’ll construct an distinctive AI writing course of.
And, ultimately, we’ve got the expertise to allow that.
For many individuals, their worldview of AI continues to be anchored within the chat expertise. However LLMs—and extra importantly, the infrastructure that surrounds them—have progressed massively in current months.
Even of their nascent days, giant language fashions confirmed sparks of superhuman brilliance in small areas. However like a precocious little one imitating its dad and mom with no actual understanding of its behaviour, it was laborious to think about these sparks progressing right into a blaze of real writing skill.
A handful of coherent sentences appears 1,000,000 miles away from reliably producing 1000’s of phrases of correct, useful, concise, on-brand writing; from figuring out and filling matter gaps, understanding the dominant search intent, differentiating from competitor articles, and so on.
Once I wrote about my earlier AI writing course of (utilizing customized GPTs based mostly on my editorial rules), I noticed many sparks of brilliance within the output, however the remaining product nonetheless relied on human intervention to create.
An early model of our AI content material pipeline, constructed utilizing ChatGPT initiatives and customized GPTs.
However that’s now not the case. Simply seven months later, the constraints of that course of have disappeared. At this time, my $20/m Claude subscription offers entry to talents that appear nearly science fiction. I can:
- Chain a number of LLM processes collectively into one steady workflow (Claude Code, OpenAI Codex, and different agentic fashions).
- Present guardrails to keep away from a lot of the probabilistic “wobble” we see when LLMs often attempt to comply with processes (SKILLs) and encourage them to recursively benchmark their efficiency and self-improve.
- Combine AI into present workflows throughout different instruments (MCP).
- Floor content material in our bodies of analysis, in samples of present writing, tone of voice, model tips (RAG, reminiscence, context).
(And this ignores the numerous enhancements flagship fashions themselves have proven lately.)
A number of the customized SKILL information we’ve constructed for Claude Code.
All of the vibe-coding infrastructure developed previously yr has had a transformative affect on the usefulness of LLMs typically. LLMs are nonetheless “simply” refined autocompletes—and we’ve got definitely not achieved AGI—however firms like Anthropic and OpenAI have succeeded in harnessing that behaviour in a approach that appears a lot, far more helpful than the sum of its elements.
And importantly, the duty set earlier than them—content material advertising—will not be notably difficult.
Most content material entrepreneurs spend most of their time creating informational, keyword-targeted content material: useful “how-to” articles or comparability lists. These are the time-tested archetypes of content material advertising, and they’re typically fairly easy to create.
Like earlier than, I imagine there’s a fundamental recipe for efficient search content material. Listed below are a few of the core rules that we attempt to comply with in our search content material:
- Tackle the first search intent
- Construct upon the consensus within the present search outcomes
- Fill any matter gaps apparent between your articles and your rivals’
- Add new, novel info above and past the present outcomes
- Reference any related present content material you might have already created on the matter
- Reference any related exterior content material which may assist the reader proceed their exploration
- Prioritise matters that help you naturally reference your product
- Make sure the article construction is mutually unique, collectively exhaustive
- Make sure the article construction really delivers on the title’s promise
- Hook the reader’s curiosity with the title and introduction
- Embody key phrases and key phrase variations naturally in vital elements of the article
- and so forth.
These are equally easy ideas that additionally ladder as much as efficient search content material. If an individual can comply with these processes, their search content material typically performs effectively. The identical is true of an LLM. If Opus 4.6 or GPT 5.4 can comply with these processes, their output will even carry out effectively.
Even essentially the most opaque of those processes is pretty trivial for an LLM to comply with, both by offering express steps to take (“use WebFetch to run a website: seek for ahrefs.com/weblog and return the primary three articles…”), examples of the specified output (like a reference file of your favorite article introductions), or entry to trusted knowledge sources (just like the Ahrefs MCP).
A preview of a SKILL file for retrieving present Ahrefs content material for a given key phrase.
As a lot as we would want it in any other case, efficient search content material is very formulaic (therefore the success of the skyscraper technique). There isn’t any want for excellent complexity or novelty, no want for poetry or disagreement with the SERP.
There may be some room for innovation and experimentation, however lower than you may assume: straying too far outdoors of the Overton window often degrades efficiency, as a substitute of enhancing it (I say that after many failed makes an attempt to create “intelligent” search content material).
If Claude can refactor a 100,000-line codebase, it appears smug to imagine that giant language fashions can’t write a wonderful piece of search-optimized content material. AI can’t write Shakespeare, but it surely doesn’t want to.
Remaining ideas
Whether or not I’ve persuaded you or not, on the time of writing, there are important elements of my position I’ve already outsourced to generative AI. I exploit Claude Code, the Ahrefs MCP, and a collection of ~15 customized SKILLs, chained collectively in sequence, to replace outdated articles and create useful, high-quality content material.
Claude embarks on its content material updating course of.
These articles sound the identical. They carry out the identical. They embody my expertise and perspective. They’re nearly as good as something I may have written; higher, as a result of I wouldn’t have had the time to create them in any other case. There isn’t any trade-off.
There may be nonetheless an unlimited gulf within the high quality potential between a talented author utilizing generative AI to its fullest potential, and the typical layperson prompting ChatGPT to “write a weblog submit”.
However this gulf is much smaller than it was; within the long-term, it is going to shut, as AI platforms proceed to democratize entry to all of this superb performance. The talents of the “content material engineer” will develop into simply one other workflow in each main LLM platform. Functionally “excellent” AI content material is simply across the nook, for all of us.
I’m comfy making this argument as a result of there are lots of elements of my job that I nonetheless can not outsource to AI; there are others I might not, even when I may (like this text).
The trail ahead will solely be discovered if we’re trustworthy about the place AI can, and may, be used. Till just lately, AI content material wasn’t adequate. Now, it’s. The earlier we will admit that, the extra time we’ve got to give attention to the elements of selling the place people could have an extended, happier tenure.
(And I don’t miss writing skyscraper content material.)
