---
title: Writing advertorials with Claude — Stefan Georgi's extract-then-model process
slug: writing-advertorials-with-ai
summary: What an advertorial is for (pre-selling between an ad and a short PDP or long sales letter), the two templates (first-person testimonial story vs third-person news story), and Stefan Georgi's three-step Claude process — extract inputs from the sales page, analyse a proven advertorial's structure, write a new one per avatar — producing 3+ drafts in ~45 minutes, with the editing, hallucination and health-claim checks it needs, plus his stock-up upsell template.
domain: marketing
tags: [advertorial, presell, ai, claude, prompts, native-ads, direct-response, upsell, compliance]
related: [presell-funnel-paths-advertorial-vsl, persona-landing-pages-and-listicles, ai-landing-pages-with-claude-code, name-the-unique-mechanism, voice-of-customer-copy-mining, proof-and-believable-claims]
sources: [yt:BoD63RyEO1I, yt:TDT7Bu2cwhU, yt:MH_fU3JOP3M]
updated: 2026-10-24
---

# Writing advertorials with Claude

Source: Stefan Georgi ([[people/stefan-georgi]]), demos from 2024 (Claude 3 Opus, then Claude 3.5) and his RMBC2 course intro (2025).

## TL;DR
- **An advertorial is an ad dressed as an article**, labelled "advertisement" or "sponsored". It pre-sells when the destination alone can't, for example a short e-com page or cold email traffic sent to a product page[^1][^2].
- **Two templates:** a *first-person testimonial story* ("I was drowning in mom guilt… then I discovered…") and a *third-person news story*[^3].
- **Process:** (1) turn the sales page or PDP into a text PDF with FAQs and testimonials; (2) have Claude extract the inputs; (3) give it a proven advertorial and have it *analyse the structure*, then write a new one for one avatar; repeat per avatar or template[^4][^5][^6].
- **Speed:** three full advertorials in ~47 minutes on video. In 2025 he claims you can do "a hundred in a day" once the process is set up[^7][^8].
- **Fact-check everything.** Claude invented a real doctor as spokesperson, studies and testimonials. Replace them with real ones or remove them[^9][^10].

## Details
**Where advertorials sit.** On native ad networks (sponsored "from around the web" links on sites like weather.com), the click almost always goes to an advertorial, because it keeps the experience congruent with the article the person was reading. From there the paths diverge (see [[presell-funnel-paths-advertorial-vsl]])[^11]. In his April 2024 walk-through, a native ad for a doctor-led supplement went to an advertorial that linked to a VSL, while a Performance Golf ad skipped the advertorial and went straight to a VSL[^20]. To find models, browse native-ad placements or use spy tools. Stefan models off advertorials that are running, including one from a pet-supplement brand he has equity in[^12].

**Step 1. Source document.** Paste the landing page into a Google Doc, expand and copy the FAQs, add the testimonials, then download it as a PDF without images[^4].

**Step 2. Extraction prompts.** Run these one at a time in a single Claude chat, and save every output to a separate "inputs" doc so you can reuse them in new chats[^5][^13]:
1. Target market, problems and pain points, primary promises (with an example of the output format).
2. Credible proof (universities, studies, experts) and social proof (customer counts).
3. Common solutions the market has tried and why they aren't the real answer.
4. The "real cause" (mechanism), if the page has one.
5. Product name, how to use it, price and anchor price.
6. Ingredients or components, plus studies.
7. Avatar archetypes, for example "busy working mom", "post-pregnancy weight struggler", "perimenopausal woman".
Add what you know that the page doesn't say. For the keto product he told Claude to assume the market skews female, because weight loss does, before asking for archetypes[^21]. Short e-com pages yield thin inputs. That's fine, but it raises the hallucination risk[^14]. The prep took him ~17 minutes and is done once per product[^21].

**Step 3. Model a proven advertorial.** Upload an advertorial for an *unrelated* product and ask Claude to "analyse the structure" first. Stefan finds this step improves the output. Then prompt along these lines[^6][^15]:
- Write a new advertorial for [product], following the example's structure, style and tone.
- Use a spokesperson matching [avatar]'s demographics and psychographics.
- Blend storytelling with a news-report feel.
- Go section by section and ask me to continue when you run out of space.
- Don't condense: effective DR copy is often long.

The output had headline options ("From exhausted mom to keto queen…") and followed the model's beats: pain story, "the discovery that changed everything", mechanism explainer, trying the product, results, testimonials, offer and urgency[^16][^17]. The spokesperson is a made-up persona matching the avatar, which he calls common because the page is labelled an advertisement. The prompt itself says to swap in a real customer before going live[^6].

**Editing passes.** Always ask for more headlines, and push for specificity ("this delicious chocolate drink" rather than "this drink"). In his view AI copy beats 99% of copywriters, and the missing 1% is judgement and specificity[^22][^23]. Other follow-ups he uses: keep the product name blind until the last section, and keep sensory details like flavour, which matter for consumables[^24]. If the model advertorial is doctor-led and your product has no doctor, tell Claude to lean on "leading health experts" or brand credibility instead of the model's doctor[^25].

**Step 4. Multiply.** Re-run Step 3 for each avatar, and with different model advertorials (doctor-led, testimonial-led), to get 10–20 variants to test[^7].

**Tool version (July 2024).** freeadcopy.com (Stefan's tool) has both templates and scrapes inputs from a URL. He found that pasting key sections himself, such as the challenge's day-by-day content and the founders' names, worked better than the scraper. He also raised the anchor price in the inputs and gave the testimonial persona a named, specific story, which made the copy more on-point[^18][^19][^26]. Outputs vary run to run, so re-run a bad draft rather than fix it. GPT-4o once wrote from the wrong saved example (a dog product) and Claude invented product names, so check names on every draft[^27][^28].

**Health claims.** On a deliberately white-hat blood-sugar supplement, Claude wrote "reverse diabetes" and "cure". Stefan calls the copy strong but "a little bit black hat". He would add a compliance note to the inputs, such as: don't say cure, reverse or treat; say "erratic blood sugar"[^29][^30].

**Stock-up upsell script.** The same tool has a "more of the same" upsell template, modelled on a structure Jay Debolt taught at Stefan's mastermind in 2020. It first reaffirms the buyer's decision, then raises a threat to their results (running out and losing consistency), then offers extra bottles at a discount[^31][^32].

## Disagreements / caveats
- **Compliance:** invented spokespeople and "fake articles" are standard in supplement DR, but they risk FTC, ASA and EU consumer-law problems and Meta ad rejections. Stefan himself says not to name a real doctor who isn't associated with the product[^9], and Claude defaults to aggressive health claims[^29]. Use real customers' stories.
- The demos are from 2024 (Claude 3, message caps he worked around via Poe). Model quality and limits have changed, but the extract-then-model process doesn't depend on them.
- His speed claims assume an expert editor. Thin source pages produce generic, partly invented copy.
- Dose found keeping the buy box on the advertorial beat redirecting ([[persona-landing-pages-and-listicles]]). Stefan's examples link out to a PDP or VSL instead. Test both.

## See also
- [[presell-funnel-paths-advertorial-vsl]] — what comes after the advertorial
- [[ai-landing-pages-with-claude-code]] — turning the copy into a page
- [[voice-of-customer-copy-mining]] — better inputs than the sales page alone

[^1]: yt:BoD63RyEO1I @ 00:00 — Stefan Georgi
[^2]: yt:BoD63RyEO1I @ 00:00 — Stefan Georgi
[^3]: yt:TDT7Bu2cwhU @ 00:00 — Stefan Georgi
[^4]: yt:BoD63RyEO1I @ 01:31 — Stefan Georgi
[^5]: yt:BoD63RyEO1I @ 04:50 — Stefan Georgi
[^6]: yt:BoD63RyEO1I @ 21:59 — Stefan Georgi
[^7]: yt:BoD63RyEO1I @ 46:53 — Stefan Georgi
[^8]: yt:MH_fU3JOP3M @ 29:15 — Stefan Georgi
[^9]: yt:BoD63RyEO1I @ 27:50 — Stefan Georgi
[^10]: yt:BoD63RyEO1I @ 30:36 — Stefan Georgi
[^11]: yt:MH_fU3JOP3M @ 15:50 — Stefan Georgi
[^12]: yt:BoD63RyEO1I @ 18:46 — Stefan Georgi
[^13]: yt:BoD63RyEO1I @ 22:50 — Stefan Georgi
[^14]: yt:BoD63RyEO1I @ 31:30 — Stefan Georgi
[^15]: yt:BoD63RyEO1I @ 24:39 — Stefan Georgi
[^16]: yt:BoD63RyEO1I @ 26:37 — Stefan Georgi
[^17]: yt:BoD63RyEO1I @ 18:46 — Stefan Georgi
[^18]: yt:TDT7Bu2cwhU @ 05:02 — Stefan Georgi
[^19]: yt:TDT7Bu2cwhU @ 02:18 — Stefan Georgi
[^20]: yt:BoD63RyEO1I @ 20:23 — Stefan Georgi
[^21]: yt:BoD63RyEO1I @ 16:42 — Stefan Georgi
[^22]: yt:BoD63RyEO1I @ 34:49 — Stefan Georgi
[^23]: yt:BoD63RyEO1I @ 36:18 — Stefan Georgi
[^24]: yt:BoD63RyEO1I @ 38:39 — Stefan Georgi
[^25]: yt:BoD63RyEO1I @ 42:41 — Stefan Georgi
[^26]: yt:TDT7Bu2cwhU @ 06:35 — Stefan Georgi
[^27]: yt:TDT7Bu2cwhU @ 14:32 — Stefan Georgi
[^28]: yt:TDT7Bu2cwhU @ 16:12 — Stefan Georgi
[^29]: yt:TDT7Bu2cwhU @ 10:50 — Stefan Georgi
[^30]: yt:TDT7Bu2cwhU @ 12:36 — Stefan Georgi
[^31]: yt:TDT7Bu2cwhU @ 17:44 — Stefan Georgi
[^32]: yt:TDT7Bu2cwhU @ 19:15 — Stefan Georgi