Writing a blog with AI: from brief to measurement

AI can draft a blog quickly. Producing a good one takes a workflow. The difference isn’t the model; it’s what you do around it: check the search intent before you type, write a brief instead of a vague request, edit the draft yourself and measure its performance after publication.

ChatGPT and Claude are good drafting tools, and most people start there. If you publish every week, however, the same tasks keep coming back: reexplaining the brief and brand voice, drafting metadata and repeating measurements. AI agents with memory, reusable skills and scheduled jobs can handle some of that repetitive work. I cover that option along the way. Want to take it beyond blogging? AI automation for small businesses is the logical next step. Prefer to start from a generated design instead of written text? See how we go from AI design to real, maintainable web code.

Why most AI blogs don’t work

Open a chat, type “write a blog about X, 1500 words, SEO-optimized” and you get something that looks like a blog: an introduction, subheadings, a conclusion and an FAQ. But it contains nothing that only you could write. It has none of your own examples or opinions, and no reason for Google to rank this version above the ten that already exist.

A blog article competes with everything already online about the topic. AI has made drafting cheap, so merely publishing an article is no longer enough. You need a topic where your angle matters, a structure that answers the search intent and a publication process that checks whether the article works. One prompt will not make those decisions for you; they are the workflow.

Step by step: how to write a blog with AI

Step 1 deserves the most attention but often gets the least. Skip the search intent and you may write a polished article that nobody searches for, or one that competes with another page on your site. The guide to keyword research explains how to cluster search queries and match them to pages.

Writing metadata with AI: title, description and slug

AI is useful for metadata as long as you make the final decision.

For example, I requested title ideas for an article about getting an affordable website built. Here are four shortened suggestions:

• Getting an affordable website: what it really costs — factual, but dry.
• 5 ways to get an affordable website built (the number works, but promises a list).
• Getting an affordable website without surprises — this addresses the real concern behind the search query.
• What an affordable website costs in 2026 — the year makes it current, but ages quickly.

I chose the third variant. The competing results and search intent show that the question is not only about price, but also about hidden costs. A model will not reliably infer that on its own. You can spot it by reading the search results. AI supplied the options; I made the editorial decision. That is where AI helps most with metadata: it produces alternatives quickly, while you choose what fits.

• Title: have AI make five variants around your core term, each with a different angle (number, question, promise). Choose the version that earns the click and keep it under about 60 characters.
• Meta description: write a concise description that matches the article and gives the reader a reason to click.
• Slug: keep it short and readable, with the core term. Avoid dates and words the URL does not need.
• Alt text: describe what is actually in the image in plain language. Do not turn it into a keyword list.

A useful rule of thumb: use AI for options, not the final choice. The first title is rarely the best, and a description that must compete with four others in the search results deserves careful thought.

Image choice for an AI blog

One strong opening image that visually summarizes the topic works better than five stock photos scattered through the article. Choose or create an image that captures the article’s central idea, give it descriptive alt text and use it as the social preview too. Generative tools can produce good editorial images if you specify a clear style and visual metaphor. Keep text out of the prompt, because generated lettering is often unreadable.

Avoid the same generic laptop photo as every other blog, images with random text and screenshots of tools you do not have permission to publish.

From chat to agent: blogging with memory, skills and cron jobs

In an isolated chat without project memory, the session is the limitation rather than the model. You have to explain who you are, how you write, and what the brief was each time. An AI agent such as Hermes Agent can retain approved working context in three ways:

• Memory: approved brand guidance and editorial decisions can remain available across sessions.
• Skills: once you find a reliable process, such as your brief structure plus an editing checklist, you can store it as a reusable procedure. You build an editorial process, not a prompt history.
• Cron jobs: recurring tasks run on a schedule, such as fetching new measurements, preparing draft ideas and flagging seasonal articles in time. You decide what happens with the results.

Such a session looks like this:

One detail matters: the agent first checks what the site already covers and identifies what still needs a human decision. It prepares a brief and content plan; you decide the article’s angle, which claims are acceptable and whether to publish it.

Common mistakes

• Publishing whatever the chat produces: treat the first version as a draft. Without editing, your blog will be average and interchangeable.
• Taking AI facts at face value: models invent sources and figures with conviction. Check every factual claim yourself.
• Skipping search intent: writing what you want to say instead of what people search for. This is a common reason why a good article gets no visitors.
• Not measuring the result: without a baseline and follow-up measurement, you do not know whether the blog works or what to improve next time.
• Forcing everything into one session: trying to brief, draft, edit and write metadata in one prompt weakens every stage. Work in phases.

Conclusion

Blogging with AI works when you treat it as a workflow: search intent, brief, draft, editing, metadata, and measurement. ChatGPT and Claude handle the draft phase well. If you publish regularly, an agent such as Hermes adds memory, skills, and scheduled jobs that keep the process consistent. You still make every publishing decision. Want a blog that follows this process from day one, with a structure that matches how people search? See how Straffe Sites approaches copywriting and SEO, or start by learning how to do keyword research yourself.