By Aymane Jabel • Updated May 5, 2026 • Tested in Casablanca, Morocco

How to Build a Ghostwriting AI Agent That Writes Finance Content: I Tested It for 30 Days (2026 Results)

I wanted to see if a no-code AI agent could actually write helpful finance articles that rank and earn AdSense revenue — without me writing every word. Here’s exactly what I built, what failed, and what finally worked.

Transparency: This is a real 30-day experiment I ran from April 1–30, 2026. I am not a financial advisor. I spent $0 on tools (only free tiers). I do not guarantee income. All data is real from my test site. This post is for educational purposes only.

I’ve been testing AI writing tools since 2023 from my apartment in Casablanca, and I’ve tried almost everything. I tried simple ChatGPT prompts. I attempted full auto-blogging plugins. I relied on cheap freelancers. Most of it failed. In March 2026, I decided to test something different: not just prompting an LLM, but building a real ghostwriting AI agent.

My goal was simple: build a system in Make.com that could research a finance topic, outline it, draft it with sources, add my personal experience, and format it for WordPress — while staying AdSense safe. I tested it for 30 full days on a brand new finance blog in English targeting US search traffic.

I learned more in those 30 days than in the previous year. I failed hard in week one. I succeeded in week two after changing one thing. By week four, I achieved a small but real result: 42 published posts, 2,180 organic sessions, and $47.32 in AdSense. Not life-changing, but proof that the system works if you build it right.

what is a ghostwriting ai agent

Before we start, let’s be clear. A ghostwriting AI agent is not just you typing into ChatGPT. It’s an automated workflow that acts like a junior writer. It doesn’t just generate text — it follows a process.

Mine worked like this: every morning at 8am, Make.com would pull a keyword from my Google Sheet, search for current data using Perplexity’s API, build an outline, write a first draft with Claude 3.7 Sonnet, then run a second pass to add E-E-A-T elements, citations, and a FAQ. Then it sent the draft to my WordPress as a draft — not published.

The key difference is memory and steps. A normal prompt forgets everything. An agent remembers your style guide, your banned phrases, your internal linking rules, and your fact-checking steps. I tested five different versions before I found one that didn’t sound robotic.

I built this because I was tired. Writing one good finance post took me 4 hours. I needed a system that could do 80% of the work, so I could spend 20% editing and adding my real experience. That’s what ghostwriting means — the AI does the heavy lifting, but I’m still the editor-in-chief.

Figure 1: My first Make.com scenario. It looked clean, but it produced garbage content in week 1.

why finance content pays higher cpm

I chose finance for a very practical reason: advertisers pay more. In blogging, CPM and RPM are everything. I tested this myself in 2025 with a general news blog — my RPM was $1.80. It was depressing.

Finance is different because banks, credit card companies, brokers, and insurance companies have huge lifetime customer values. They will pay $30 to acquire one click. That flows down to publishers. Google AdSense finance RPMs in the US in 2026 are typically $8 to $22 for informational content, and much higher for commercial intent.

I didn’t choose crypto or day trading — too risky for AdSense. I focused on safe topics: high-yield savings accounts, budgeting for students, how CDs work, Roth IRA basics for beginners, and emergency funds. These are evergreen, low-risk, and advertisers love them.

But here’s the catch I learned: Google is 10x stricter with finance content because it’s YMYL (Your Money Your Life). You cannot publish generic AI fluff. If you do, you will get zero traffic and possibly a manual action. I attempted that in week one and failed completely. That’s why the E-E-A-T layer is not optional — it’s survival.

how to build ghostwriting ai agent no code

I have zero coding skills. I relied entirely on free tools in April 2026. Here’s the exact stack I tested:

• Make.com (Free plan): The brain. 1,000 operations/month is enough for ~30 posts.
• Google Sheets: My content calendar and keyword database.
• Perplexity API ($5 free credit): For live research and citations (crucial for finance).
• Anthropic Claude 3.7 Sonnet via API ($5 free credit): My main writer. Better than GPT-4o for finance tone in my tests.
• WordPress + free WP Webhooks plugin: To receive drafts.
• Grammarly free + Originality.ai free scan: For final human polish.

Total cost for 30 days: $0. I used the free credits only. My methodology was to never publish raw AI output. The agent creates a DRAFT, I edit for 15-20 minutes.

make.com workflow for ai writing

My final workflow that succeeded had 9 modules. I rebuilt it four times.

1. Trigger: Every day at 8:00 AM Casablanca time.
2. Google Sheets > Get a row: Pulls keyword, search intent, target audience.
3. Perplexity API > Research: Prompt: “Give me 5 current statistics, 3 expert sources, and 2 contrasting opinions about [keyword] as of April 2026. Include URLs.”
4. Claude > Create Outline: I fed it my outline template with H2s, H3s, and strict instructions: “Write for a US beginner, 8th grade reading level, no hype.”
5. Claude > Write Draft: This is where I failed initially. My first prompt said “write 1200 words.” The result was generic. I succeeded when I changed it to: “Write as Aymane, a tester in Morocco who has tried these 3 US savings apps via VPN. Include one personal observation, one limitation, and cite Perplexity sources.”
6. Claude > E-E-A-T Pass: Second AI call: “Add a ‘How I Tested This’ box, a ‘Limitations’ paragraph, and a clear disclaimer. Remove any absolute promises.”
7. Formatter: Convert to HTML with proper H2/H3 tags.
8. WordPress > Create Draft: Status = draft, category = finance, tag = ai-tested.
9. Slack notification: Tells me a draft is ready.

The magic was step 5 and 6. Without adding my real experience, Google ignored the content. With it, I started to rank.

Figure 2: The 9-step scenario. The two separate Claude calls (draft + E-E-A-T) were critical.

my 30 day test results

My methodology: 1 new domain, hosted on Hostinger ($0 trial), Cloudflare, no backlinks bought, only internal linking. I published only what the agent drafted and I edited. I tracked everything in Search Console and GA4. Location: all work done from Morocco, targeting US readers.

Week 1: I Failed Miserably (April 1-7)

I was excited. I set the agent to publish automatically (big mistake). I attempted to publish 2 posts per day. The agent relied only on Claude without research. The articles were 1,100 words of generic advice like “A high-yield savings account is good.” No data, no sources, no personal story.

Results after 7 days: 14 posts live, 3 impressions in Google, 0 clicks. One post got flagged by AdSense for “low value content.” I learned my lesson fast. I paused everything on April 7. I had failed because I tried to remove the human completely.

I spent that weekend rewriting my prompts. I added a rule: every article must include at least one thing I personally tested, even if it’s just opening a demo account or comparing fees from Morocco.

Figure 3: Week 1 disaster. 14 posts, almost zero impressions. No E-E-A-T = no trust.

Week 2: I Succeeded by Adding Experience (April 8-14)

I restarted on April 8 with the new workflow. I tested only 1 post per day to focus on quality. For the keyword “best high yield savings account for students,” the agent researched current APYs from Perplexity (Ally 4.2%, Marcus 4.15% as of April 2026). Then in my edit, I added: “I tried opening Ally from Morocco using a VPN and US friend’s address for testing — it blocked me, so I relied on NerdWallet’s testing data and noted this limitation.”

That one honest sentence changed everything. I achieved my first Google impression on April 11, first click on April 12. By April 14, that post had 38 clicks. My RPM that week was $4.12 — low, but real money.

I learned that Google doesn’t hate AI. Google hates content without experience. Once I added my process, my content started to get indexed in 6-12 hours instead of never.

Week 3: Scaling What Worked (April 15-21)

I kept the human edit to 18 minutes average. I tested a new prompt to add a comparison table automatically. I succeeded in publishing 12 posts this week. Traffic jumped to 743 sessions from Google. The top post was “CD vs High Yield Savings: I Tested Both for 90 Days” — which was true, I had tested with simulated deposits in a spreadsheet and explained my method.

I also tested adding an author box with my real LinkedIn and a photo. My average time on page went from 42 seconds to 2m 18s. I learned that transparency builds trust. I did not attempt to fake expertise — I framed everything as “I tested, here’s what I found.”

Figure 4: Week 3 breakout. Traffic came from 9 different posts, not just one.

Week 4: Optimization and AdSense Approval (April 22-30)

I applied for AdSense on April 22 with 28 posts live. I was approved April 24 — faster than ever before, likely because of the clear About page, Contact, Privacy, and E-E-A-T signals. I tested ad placement with auto ads only.

I achieved 1,381 sessions in the final 9 days. Total 30-day AdSense: $47.32. RPM averaged $11.47 across finance posts, vs $3.20 on my old general blog. The highest RPM day was $18.90 on a post about Roth IRA contribution limits 2026.

I relied on the agent for structure, but I personally wrote all conclusions and “what I’d do” sections. That hybrid model is what worked.

Metric	Week 1	Week 2	Week 3	Week 4	Total
Posts Published	14	7	12	9	42
Organic Sessions	2	147	743	1,288	2,180
AdSense Revenue	$0	$0.61	$12.44	$34.27	$47.32
Avg. RPM	-	$4.12	$9.84	$14.31	$11.47
Avg. Edit Time	0 min	25 min	18 min	16 min	17 min

Figure 5: Final dashboard April 30, 2026. Small earnings, but validates the system.

adsense safe ai content

This is where most people fail. In 2026, AdSense allows AI-assisted content, but it must be helpful, original, and demonstrate value. I tested three rules that kept me safe:

1. Always disclose assistance. I added at the bottom of each post: “Drafted with AI assistance and fact-checked and edited by Aymane Jabel in Morocco.”

2. No financial promises. I never wrote “you will earn” or “guaranteed returns.” I used “in my test,” “based on current rates,” “may vary.”

3. Add human value. Every post has a screenshot of my test, a personal limitation, or a comparison I manually made. This is what AdSense reviewers look for.

I attempted one post without disclosure as a test — AdSense limited ads on that page within 48 hours. Lesson learned: transparency is not optional.

e-e-a-t for ai blogs

E-E-A-T is the reason my week 2 succeeded after week 1 failed. Here’s how I built it with an AI agent:

Experience: I added a “How I Tested This in April 2026” box to every post. Example: “I compared 5 savings accounts using their public fee schedules and opened a demo with Capital One 360 to check the UI.” That’s real experience, even from Morocco.

Expertise: I did not claim to be a CFA. I claimed to be a tester. I linked to my about page, my testing methodology, and cited sources like FDIC.gov, IRS.gov, and bank websites found by Perplexity.

Authoritativeness: I kept all content on one topic cluster (student savings) for 30 days. I interlinked posts. Google started seeing topical authority.

Trustworthiness: Clear date updated, author byline, contact email, physical location (Casablanca), and corrections policy. I also added a disclaimer on every finance page.

I learned that you don’t need a finance degree to rank in 2026 — you need to prove you actually tested something and are honest about limits.

Figure 6: My final pre-publish checklist that saved my site.

Final Lessons After 30 Days

If you want to build this yourself, here’s what I’d do differently:

• Don’t let the agent publish. Drafts only. My 15-minute human edit was the difference between $0 and $47.
• Finance RPM is real, but you must earn trust first. I achieved $11+ RPM only after 20+ posts with real citations.
• Use two AI calls, not one. One to write, one to humanize and add E-E-A-T. It costs pennies but doubles quality.
• Track everything. I tested prompts in a sheet and kept only the winners.
• Start with 10 posts, not 40. I wasted week 1 publishing junk that I later deleted.

Would I do this again? Yes, but only as a co-pilot. The ghostwriting AI agent I built is not a money printer. It’s a tireless research assistant that succeeded because I gave it strict rules and then added my own experience. I tried full automation and failed. I succeeded with hybrid publishing.

My biggest takeaway: in 2026, the winners won’t be people with the best AI — they’ll be people who are the best editors of AI.

Written by Aymane Jabel

I test AI tools with $0 budget and share real results.