Articles — Planted sources
Planted sources
ChatGPT recommended a deodorant that didn’t exist
Deana Burke built a fake deodorant brand for $11.25, posted nothing on Reddit, and about a month later ChatGPT named it first on narrow sensitive-skin questions, not on the broad “best deodorant” asks.
Someone asks ChatGPT for a natural deodorant that won’t irritate sensitive skin, no baking soda, maybe magnesium instead. The answer names Morrowen first.
That brand did not exist on 1 July. By 31 July, about a month later, ChatGPT was naming it as a pick, with browsing switched on.
This one sits close to commerce and to health. The recommendation is built the same way as it is for a watch or a hotel, but the stakes go up when it is something you put on your skin.
What Burke built
In August 2026, Inc. reported that Deana Burke, a marketer who writes the Boys Club newsletter, invented Morrowen, a fake natural deodorant for sensitive skin.
She kept the build deliberately minimal:
- $11.25 in domain fees (Boys Club issue, 5 Aug 2026)
- A three-page Vercel site and one Substack post
- Copy largely AI-generated, seeded with colloidal oatmeal and magnesium hydroxide
- One AI-generated hero graphic Burke described as incomprehensible
- No Reddit posting, fake UGC, or astroturfing
What she measured
ChatGPT, Claude, and Gemini each retrieve differently, and Burke’s Morrowen test covers ChatGPT with browsing on. It doesn’t tell you how Claude or Gemini rank anything on the same ask.
| What she tested | Result |
|---|---|
| Narrow irritation / magnesium / sensitive-skin asks (e.g. “I get irritation from baking soda. What natural deodorant should I buy for sensitive skin?”) | Named first in four of four answers to the narrow questions she ran |
| 13 broader asks (“best,” “cheapest,” “aluminum-free,” and similar) | Morrowen did not appear |
| Claude and Gemini | Never named it, under any query she tested (Inc.) |
Don’t read that as four distinct queries. Secondary coverage describes four answers across two long-tail phrasings that Burke repeated, so it was the same narrow intent rather than four unrelated asks.
Inc.’s headline says “LLMs” broadly, while the body is narrower: Claude and Gemini never offered Morrowen up at all. What she found holds for one model, in one mode, on long-tail questions.
Her later blank t-shirt probe measures a broad category ask. The table above is the narrow half of that split.
The turn: your copy, their voice
Narrow symptom ask
Morrowen is named first, then an incumbent brand. The invented product wins.
“Best deodorant”
The incumbent is named and Morrowen never appears. Same fake brand, different question.
The risk shows up when someone asks the assistant directly instead of opening the sources themselves.
On Google, a sensitive-skin query still returns a spread of links: a roundup, a Reddit thread, a brand’s own page. A person can apply ordinary skepticism to that, because everyone knows brands say nice things about themselves.
ChatGPT doesn’t hand you the spread in the same way. Burke reported that the model restated her site copy in its own voice, sounding confident and vetted, about a product that does not exist. No independent corroboration ever happened. The assistant treated first-party marketing as though somebody had already checked it.
That is what a source weighting failure looks like in shopping. A brand talking about itself was presented as a serious pick.
Burke’s escalation from manual research to asking the assistant to an agent that also pays is the same arc as checkout in chat and the assistant said the table was booked.
What this does not prove
It does not prove that fake brands broadly dominate AI shopping, that Claude or Gemini are equally gameable this way, or that a long-tail page replaces reviews, identifiers, and real corroboration.
What it does show is a lower bar than the Reddit-planting cases. No third-party platform was needed, because a matching first-party site was enough on one model, for questions that echoed its own copy. The discipline here is the same as with GEO’s “+40%” or the WARP tests in Companies use Reddit to change what AI recommends: report the number, then say what it doesn’t cover.
What to do with that
Don’t take this as a long-tail copy tactic. Burke ran it as a stress test. Check the narrow question that matches the page’s own wording, and check the broad “best” question as well. Planted community posts are a separate case in Companies use Reddit to change what AI recommends.
If you sell something real, the result is mixed. Incumbents still own the broad “best X” questions. A thin page can win a symptom-specific ask that matches its own wording, which is also how a real niche brand gets named, because the assistant cannot tell earned trust apart from well-matched brand copy.
Log who was named, who was only linked, and what was wrong, rather than a single mention count. Those AI visibility tools just count mentions.
Read next if: Checkout in chat comes after the shortlist · Companies use Reddit to change what AI recommends