AI Marketing & AEO.

Also posted as

AI content VA, AI marketing assistant, GEO/AEO specialist, marketing automation VA

Core loop

Brief → prompt → draft → verify → publish → measure → report

Pays more than general VA by

Roughly 30–60% at the same experience level, more with automation setup work

Hardest part

Telling a client their 20-posts-a-week plan will hurt them, and having the workflow to prove it

Easiest way in

A documented content system with a fact-check log — module 8 builds exactly that

Progress is saved in this browser only. Nothing is sent anywhere.

1. 01

   What AI actually changed (and what it did not)

   You can explain to a client which of their marketing tasks AI genuinely speeds up, which it quietly ruins, and where you fit.FoundationsClient conversations25 min

   The AI-literate VA is not competing with ChatGPT. You are competing with VAs who either refuse to use it or paste its output unread. Both lose. The job clients are actually hiring for in 2026 is judgment: knowing what to ask, what to keep, and what to catch before it publishes.

   Large language models — ChatGPT, Claude, Gemini — predict text. They are extraordinary at structure, tone, summarising, reformatting and first drafts. They are unreliable at facts, numbers, names and anything that happened recently, and they will state a wrong answer with the same confidence as a right one. That failure mode has a name, hallucination, and managing it is most of the professional skill.

   Swipe the table sideways →

   Task	AI alone	VA + AI	Why
   First draft of a blog post	Generic, confident, sometimes wrong	Strong	You supply the brief, the facts and the edit. AI supplies speed.
   Client’s monthly report numbers	Dangerous	AI drafts the commentary only	Numbers come from the source. AI never touches them unverified.
   Social caption variations	Fine	Faster and on-brand	Low-stakes, high-volume. Perfect AI territory with a tone example.
   Legal, medical, financial claims	Never	Human sources, AI polishes wording	One made-up statistic in a YMYL niche can end the client’s business.
   Research on a new market	Starting point only	AI maps the terrain, you verify	Every claim needs a primary source before it reaches the client.

   • Say 'AI-assisted' honestly. Most clients now expect it; the ones who forbid it are telling you something useful about the engagement.
   • The market shifted, not shrank: fewer 'write 10 articles' jobs, more 'run our content system' jobs. Systems pay better than articles.
   • Free tiers of ChatGPT and Claude are enough for this whole course. Do not buy anything yet.

   Do this

   Take a real business you know and list ten of its recurring marketing tasks. Sort each into: AI-safe, AI-with-review, human-only. Write one sentence per row justifying the sort.

   Hand in: The ten-row sorted list with justifications.≈ 40 min

   Check before you call it done

   • Every AI-safe row is genuinely low-stakes if the output is mediocre
   • At least two rows are human-only, with the risk named
   • You could defend each sort to a sceptical client

   Mark module done
2. 02

   Prompting like a professional

   You can turn a vague client request into a prompt that produces a usable draft on the first or second try.Workflow1 template35 min

   A prompt is a brief. Everything you already know about briefing a writer applies: audience, purpose, format, tone, length, examples, and what to avoid. Vague briefs produce vague drafts, from humans and models alike.

   • Be precise about the deliverable. 'Write a blog intro' is a wish. 'Write a 120-word intro for Filipino freelancers who already know what SEO is, ending with a question' is a brief.
   • Name the audience and what they already know. The single biggest quality lever in any prompt.
   • Give it a role and a tone example. Two sentences of the client’s actual writing beats any adjective list.
   • Constrain the format: word count, structure, headings, what to exclude. Models follow structure instructions well.
   • Iterate instead of restarting. 'Keep the structure, make paragraph two more specific, remove the clichés' is the real workflow.
   • Save what works. A prompt library is a genuine client deliverable — agencies bill for them.

   The brief → prompt → edit loop

   Worked example — Baler Coffee blog intro, second pass

   v1 prompt → 'write an intro about storing coffee beans' → generic, could be any site

   v2 prompt → role + audience (QC coffee buyers) + tone sample + 120 words + end on a question

   v2 output → usable after a two-minute edit: one cliché cut, one local detail added

   saved to library as 'blog-intro.md' with a note on what the tone sample was

   Reusable content prompt (fill the brackets)

   Works in ChatGPT, Claude and Gemini. Keep one per content type in your library.

   Copy template to clipboard

   You are [role: e.g. a content writer for a specialty coffee roaster in Quezon City].
   
   Audience: [who they are and what they already know]
   Task: [exact deliverable, e.g. a 120-word blog introduction about X]
   Tone: match this sample of our writing: "[paste 2–3 sentences]"
   Must include: [facts, keywords, the one point that matters]
   Must avoid: [clichés, competitor names, claims we can’t back]
   Format: [structure, length, headings]
   
   If any information you need is missing, ask me instead of inventing it.

   Do this

   Pick one real piece of content you have seen a client publish. Reverse-engineer the brief, write the prompt using the template, run it, then iterate twice. Keep all three outputs.

   Hand in: The prompt, the three outputs, and two sentences on what each iteration fixed.≈ 45 min

   Check before you call it done

   • Your prompt names audience, tone, format and exclusions
   • The final output needed less than five minutes of hand editing
   • The last line of the template (ask, don’t invent) is still in your prompt

   Mark module done
3. 03

   A content workflow that keeps quality

   You can run an AI-assisted content pipeline a client can trust, with the fact-check built in rather than promised.WorkflowQuality control35 min

   Speed without a checking stage is just faster mistakes. The workflow below is what separates a VA who 'uses ChatGPT' from one who runs a content operation.

   Swipe the table sideways →

   Stage	Who does it	Time	What can go wrong here
   1 · Brief	You, with the client	15 min	Skipping it. The model fills every gap you leave with generic filler.
   2 · Outline	AI drafts, you approve	10 min	Approving an outline that answers the wrong search intent.
   3 · Draft	AI	5 min	Nothing — this is the safe part. Generate two versions.
   4 · Fact-check	You, against primary sources	20 min	The stage everyone skips. Every number, name, date and claim gets a source or gets cut.
   5 · Humanise	You	15 min	Leaving the tells: 'delve', 'in today’s fast-paced world', perfectly parallel paragraphs, no opinions.
   6 · Publish + log	You	10 min	Not logging the prompt and sources. The log is what makes the system saleable.

   • Fact-check against primary sources, not against the model. Asking the same model 'are you sure?' is not verification.
   • Keep a kill list of AI tells for the client’s niche and edit them out on every pass.
   • One honest disclosure line in your contract — 'drafts are AI-assisted, every fact is human-verified' — prevents the awkward conversation later.
   • A 75-minute pipeline producing a verified article is the honest pitch. Anyone promising 20 articles a day is selling the mistakes stage.

   Worked example — the fact-check log for one Baler Coffee post

   claim: 'beans stale within 2–4 weeks of roasting' → verified, two roaster sources linked

   claim: 'the Philippines is the world’s 4th largest coffee producer' → WRONG, cut (model invention)

   claim: 'freezing beans is fine if airtight' → verified with caveat, rewritten to include it

   log saved next to the draft → client can audit any line

   Do this

   Run the full six-stage pipeline on one 600-word post for any real local business. Time each stage honestly.

   Hand in: The published-ready post, the fact-check log, and your six timings.≈ 75 min

   Check before you call it done

   • Every retained claim in the log has a primary source link
   • At least one AI claim was caught and cut — if none, check harder
   • Total time is under two hours; note which stage you can compress next time

   Mark module done