Most small business owners who try AI content get the same result: a dozen posts that read perfectly well and rank for nothing. The writing is not the problem. The problem is that a generic prompt produces generic content, and generic content is exactly what search engines have spent the last two years learning to filter out. This guide shows the workflow I use to get genuinely useful posts out of Claude — where the AI does the execution and you supply the one thing it cannot fake. It is the hands-on companion to the JSB Media Plan – Content Marketing pillar.

Why Most AI Blog Posts Fail Before They Are Published

Ask any model to “write a blog post about roof repair” and you will get a competent, structureless article that says what every other roof repair article says. It is not wrong. It is just interchangeable — and interchangeable content has no reason to outrank the fifty pages that already cover the topic.

Google has been explicit about this direction. Its spam policies target content produced at scale primarily to manipulate rankings rather than to help people, and the June 2026 spam update pushed further in that direction. Note the actual test: it is not “was AI involved.” Google has said repeatedly that AI-assisted content is fine. The test is whether the result is useful and original. A post you drafted with Claude and then loaded with your own hard-won specifics passes that test easily. A post you generated in one prompt and published unread does not.

So the operating rule for everything below: use AI for execution, never for expertise. Claude is very good at structure, phrasing, rewriting, and the tedious parts. It does not know what went wrong on your job site last March. That knowledge is the entire reason your post deserves to rank, and getting it out of your head and onto the page is what this workflow is for.

Step 1: Start From Demand, Not From a Blank Page

Pick topics people are actually searching for. Cheap, reliable sources of real demand:

  • Your own inbox and call log. Every question a customer has asked you twice is a post. If one person asked, hundreds are searching.
  • Google autocomplete and “People Also Ask.” Type your service into Google and read what it suggests. Those are real queries.
  • Search Console. If your site has any history, the Performance report shows queries where you already appear on page two — the cheapest wins available to you.

Paste a raw list of twenty questions into Claude and ask it to group them by search intent and flag which ones a small local business can realistically compete for. That is a good use of AI: sorting and triage, not invention.

Step 2: Write the Brief Before You Write the Prompt

This is the step people skip, and it is the step that decides whether the post is any good. A brief takes ten minutes and turns a generic request into a specific one. Mine has six lines:

  • Target question — the exact phrasing someone would type or say.
  • Reader — who they are and what they already know.
  • The answer — your actual position, in one sentence.
  • Proof — the specifics only you have: numbers, timelines, brand names, mistakes.
  • What competitors miss — skim the top three results and note what is vague or wrong.
  • Action — what the reader should do when they finish.

Keep the brief in a Claude Project alongside a short description of your business and your writing style. Everything you draft in that Project then inherits that context, which saves you re-explaining who you are every single time.

Step 3: Let Claude Interview You

The highest-leverage prompt in this whole workflow is the one that extracts what is in your head. Before asking for any draft:

You are interviewing me for an article on [topic]. I run a [business type] in [location]. Ask me eight questions, one at a time, designed to surface specific details a competitor could not copy — real numbers, real failures, decisions I made that were not obvious. Do not write anything yet.

Answer in rough, unpolished sentences. Typos do not matter. Ten minutes of this produces the raw material that makes the finished post impossible to replicate — and it is far easier than staring at a blank document trying to remember what you know.

Step 4: Draft With Constraints

Now ask for the draft, and constrain it. Unconstrained requests are what produce filler:

Using the brief and my interview answers above, write the article. Requirements: lead with the direct answer in the first two sentences, no throat-clearing introduction. Use my specifics in at least four places. Short paragraphs. No phrases like “in today’s digital landscape,” “unlock,” or “dive into.” If a claim needs a statistic I did not give you, leave a [CHECK] marker instead of inventing one.

That last instruction matters. Models will produce confident, specific-sounding numbers that have no source. Asking for explicit markers means you can find and fix them rather than publishing something you cannot stand behind.

Step 5: The Specificity Pass

Read the draft once with a single question in mind: could a competitor two towns over have published this exact paragraph? If yes, that paragraph is filler. Either replace it with something only you would say or cut it entirely. Most drafts lose twenty percent of their length in this pass and get significantly better.

Then check the things AI reliably gets wrong: names, prices, dates, regulations, anything local, and any statistic. Verify each one yourself. This is the part of the process where your expertise is doing the work — and it is also the part that keeps you out of trouble.

Step 6: Handle the On-Page Basics

Once the writing is done, the technical layer is quick. Ask Claude for a title under sixty characters that contains the target question, a meta description around 155 characters written to earn a click rather than to repeat the title, and a suggested set of internal links to your existing pages. Then make sure the page has one clear <h1>, descriptive image alt text, and a URL that is short and readable.

If you want the structured data layer handled too — article schema, breadcrumbs, local business markup — that is a job for a coding agent rather than a chat window, and it is covered in using Codex to fix your own technical SEO.

Step 7: Publish, Then Actually Measure

Submit the URL in Google Search Console and, for Bing, ping it through IndexNow so it is picked up in minutes rather than days. Then leave it alone for a month. Organic results are slow; judging a post after four days tells you nothing.

After thirty days, pull the query data and look at which searches the post is appearing for. Those queries are frequently better than the topic you originally targeted — they are real demand, discovered rather than guessed. Feed them back into your topic list. Over a few cycles this compounds into a content plan grounded in evidence instead of intuition, which is precisely what an agency retainer used to buy you.

A Realistic Weekly Rhythm

None of this requires a full day. A workable cadence for one post a week is about two hours: twenty minutes choosing the topic and writing the brief, fifteen minutes on the interview, ten minutes generating the draft, forty-five minutes on the specificity pass and fact-checking, and twenty minutes on titles, links, and publishing. The bottleneck is your judgment, not the writing — which is exactly the right place for the bottleneck to be.

Once you have posts worth reading, the next problem is distribution. That is covered in turning one blog post into a month of social content, and the signals that tell search engines your work is worth surfacing are broken down in the signals that tell Google and Bing your content is good.