AI Blog Writing: From Brief to Published Post

The short answer

A good AI-assisted blog post is decided before any text is generated. The model is fast at turning a clear brief and a research pack into a draft; it cannot decide which question is worth answering, what you know that competitors do not, or which claims are true. So the workflow below puts the human effort at the two ends (the brief and the checks) and lets AI carry the middle.

Google's position makes the same split. Its guidance says using generative AI is not in itself a problem, but content produced mainly to manipulate rankings is, and generating many pages without adding value for users can fall under its scaled content abuse policy. The steps are built to keep you on the right side of that line.

Step 1: choose a question someone is actually asking

Start from evidence of demand, not from a keyword tool alone. Three places are usually enough:

• • Search Console queries where you already get impressions but the page ranking is weak or off-topic.
• • Sales and support questions that come up repeatedly in calls, emails and tickets.
• • Objections that stall deals ("is this worth it for a company our size?").

Write the question in the reader's words. "How much does a bookkeeper cost for a sole trader?" is a post; "bookkeeping costs" is a topic.

Step 2: write the brief

The brief is the single most important input. Fill every row; if you cannot, the post is not ready to be written.

Blog brief template (copy one per post)

Field	What to write	Example (illustrative)
Reader	Who arrives, and what they already know	Owner of a five-person trades business, no finance team
Question	The exact question, in their words	Do I need an accountant or just bookkeeping software?
Promised answer	The answer in two sentences, before drafting	Software covers records; an accountant earns their fee at year end and on tax decisions. Here is where the line sits.
What we know that others do not	Your own data, cases, process or opinion	The three mistakes we see most in year-end packs
Evidence	Sources you have read, with links	HMRC guidance page on record keeping
Must not say	Claims, promises and topics to avoid	No fee figures; no tax advice for specific cases
Format	Guide, comparison, checklist, calculator	Guide with a decision table
Internal links	Two or three pages that help the reader next	Pricing explainer; year-end checklist
Next step	One proportionate action	Download the year-end checklist

Step 3: assemble the research pack

Do not ask the model to research facts from memory. Gather the sources yourself (or with a search tool you can inspect), and paste the relevant passages into the conversation with their URLs. Add your own material: call notes, a worked example, the internal figures you are allowed to publish.

The model's job in this step is to organise, not to discover. A useful prompt:

Here is a research pack of sources and notes. List every factual claim it supports, with the source for each. Then list the questions from the brief that the pack does not answer.

The second list tells you what is still missing before drafting starts.

Step 4: outline, then argue with the outline

Ask for an outline against the brief, then critique it before any prose exists. It is far cheaper to fix structure here.

Using the brief and research pack, propose an outline. Each H2 must answer part of the reader's question. Lead with the answer. Mark any section that would need a claim the pack does not support.

Delete any heading that exists for search volume rather than for the reader. If the outline could belong to any company in your market, go back to the "what we know that others do not" row.

Step 5: draft section by section

Draft one section at a time, giving the model the brief, the pack, your voice guide and one or two approved posts as examples. Section-level drafting keeps each part tied to its evidence and makes the edit easier.

Draft the section "[heading]" only. Use only facts from the pack; where you need a fact that is not there, write [NEEDS SOURCE]. Follow the voice guide. Use one concrete example.

The [NEEDS SOURCE] marker matters: it turns a hidden fabrication into a visible gap.

Step 6: fact check every claim

A person does this step, every time. Go through the draft and, for each number, name, date, quotation and "research shows", find the line in the pack that supports it. Remove what you cannot trace. Check that each linked source says what the sentence claims it says.

Step 7: edit for voice and value

Now read it as the reader. Cut the introduction until the answer appears in the first two paragraphs (the inverted pyramid). Replace general statements with the specific example, number or opinion from your own material. Remove sentences that would survive unchanged on a competitor's site. The articles on robotic-sounding AI copy and the context gap cover this edit in depth.

Step 8: on-page and publishing checks

Pre-publish checklist

Check	Pass when
Title	States the question or answer plainly; no promise the post cannot keep
Meta description	A complete sentence that tells the searcher what they will get
Headings	Each H2 answers part of the question; no heading exists only for keywords
Internal links	Descriptive anchor text to two or three genuinely useful pages, and at least one existing page links to the new post
Authorship and dates	Visible byline or publisher and a visible published or updated date
How it was made	Where it helps readers, a short note on how the content was produced
Claims	Every factual claim traced to a source in the pack

Google recommends descriptive anchor text and that every page you care about has a link from at least one other page on your site. Its generative AI guidance also suggests considering telling readers how content was created when that is useful to them.

Step 9: measure at 28 and 90 days

Judge each post against the job in its brief, not against traffic alone. At 28 days, check in Search Console that it is indexed and which queries it appears for; if the queries do not match the question, the title or opening is probably misaligned. At 90 days, look at clicks, the next page people visit, and whether sales or support now link to it. A post that answers a question your team keeps being asked has value even with modest traffic.

Where AfM fits

We run this workflow as part of a managed service: clients supply context (their questions, notes and proof), we produce and check the posts, the client approves each one before it is published, and we measure the result. If you want an outside view first, the free written breakdown looks at your public content and any figures you choose to share.

FAQ

Should I tell readers a post was written with AI?

Google suggests considering it where readers would reasonably want to know how content was made. A short note on your process (for example, drafted with AI from our notes, then fact checked and edited by our team) is honest and costs nothing.

How long should an AI-assisted blog post be?

As long as the answer needs. Set the length by the reader's question in the brief, not by a target word count; padding to a number is exactly the kind of low-value text that makes AI drafts worse.

Can I skip the outline and just ask for a full draft?

You can, but structural problems are then buried inside finished prose and take longer to fix. Critiquing a ten-line outline takes minutes and catches a missing answer or an off-topic section before any drafting time is spent.

Which part of the workflow should never be automated?

The fact check and the final approval. A model can flag claims for checking, but only a person with the sources in front of them can confirm that a number, name or quotation is real and correctly attributed.