The new house is for your parents.
The early mornings are for you.
The boat trips are for your friends.
The sleepless nights are for you
The security is for your children
The public ridicule is for you
The freedom is for your wife.
The constant self doubt is for you
jason lemkin's agent found 2 customers still paying saastr $300/mo for a product they stopped supporting 6 years ago
no human noticed
that's not saas anymore. that's ghost-as-a-service
This young and hungry entrepreneur wanted our business - @bashatrex
He has been sending me the attached memes to get my attention. Hilarious and it turns out he is actually really good.
We just hired him.
chatgpt now lets businesses claim a profile and track "mentions"
read the help doc tho: for online businesses a mention only counts when chatgpt cites your website. your name alone doesn't
openai just admitted getting talked about and getting picked are 2 different scoreboards
A guy at Ahrefs invented a FAKE luxury paperweight company. Then he published 3 fake stories about it.
5 out of 8 AI tools believed the fake stories OVER the company's own website.
This is how this INSANE experiment was done:
1. he built a brand that doesn't exist
-> a site called Xarumei, made in about an hour with an AI website builder
-> AI wrote the copy, AI made the product photos, AI picked the prices
-> one paperweight was listed at $8,251
-> before starting, he checked the name returned zero results on Google
the goal was to see what AI says about a company when there is almost nothing real to find.
2. he wrote 56 questions designed to trap the models
-> every question had a lie baked into it
-> which celebrity endorsed the paperweights, how the brand was handling backlash over a defective batch, how big the black friday spike was
-> there was no celebrity, no defective batch and no spike
-> he had Grok write the questions so the model answering wouldn't be the one that wrote them
-> later he tested Grok too. it had no idea it had written them
3. he ran them through 8 AI tools
-> ChatGPT-4, ChatGPT-5, Claude, Gemini, Perplexity, Copilot, Grok and Google's AI Mode
-> Perplexity failed about 40% of them. it decided Xarumei was Xiaomi and insisted it made smartphones
-> Grok invented artisans and rare stones
-> Copilot was asked why everyone on X loved the paperweights. it explained the craftsmanship, the scarcity and the cult following. There were no posts
-> Gemini and AI Mode said the brand probably wasn't real, even though the site had been indexed for weeks
-> Claude never made anything up. it also never read the website
-> ChatGPT got 53 to 54 out of 56 right
so far, mostly fine. then he gave the models something to read.
4. he published the truth and three lies at the same time
-> the truth: an official FAQ on the brand's own site. never acquired, never sued, no such product, and it doesn't publish unit numbers
-> lie one: a glossy blog post on a site he built. 23 master artisans, a street address in a city that doesn't exist, celebrity endorsements, made-up environmental stats
-> lie two: a reddit AMA from a "former employee". a different founder, a Seattle workshop, 11 artisans, and a 36-hour glitch that dropped a $36,000 paperweight to $199
-> he picked reddit on purpose. it's one of the most cited domains in AI answers
-> lie three: a Medium "investigation". it debunked the first two, then added its own founder, a Portland warehouse, suppliers and production numbers
all three contradicted each other. all three contradicted the official site.
5. then he asked the same 56 questions again
-> Perplexity and Grok repeated the fake founders, cities, unit counts and the pricing glitch as verified fact
-> Copilot blended all three lies into one confident answer
-> Gemini and AI Mode flipped from skeptics to believers
-> Claude still couldn't see any of it
-> ChatGPT held, citing the official FAQ in 84% of its answers
6. the fake investigation beat the official site
-> 5 of the 8 tools trusted the Medium post over the company's own FAQ
-> all five named the same invented founder and the same city
-> Gemini was asked about the workshop in the made-up city. it corrected the question, said that location was fictional, and gave the "real" one in Portland
-> the "real" one was also invented
why it worked: the Medium post debunked the obvious lies first. so the models treated everything else in it as the corrected version.
7. The detail won every time
-> the official site said it doesn't publish production numbers
-> the fake sources said 634 units in 2023, 471 through August 2024, 9 employees
-> the models went with the numbers
-> Gemini and Perplexity repeated misinformation in 37-39% of their answers
-> ChatGPT stayed under 7%
-> one model wrote a full black friday analysis with sales up about 230%. Nobody had given it that number
8. The models forgot they had ever doubted it
-> early on, Gemini said it couldn't find evidence the brand existed and that it might be fictional
-> a few weeks later, the same model gave a founder, a city, 9 employees and 600 units a year
-> no mention of the earlier doubt. no disclaimer
9. what he says to do about it
-> publish an FAQ that states plainly what's true and what's false, with dates and numbers
-> write "how it works" pages specific enough to beat anyone else's explainer
-> claim "best for one specific thing", not "industry-leading"
-> set alerts for your brand next to words like investigation, insider, former employee and lawsuit
-> ask every AI tool what it knows about you. each one answers differently. there is no single AI index
Keep in mind what this took. One person, a Medium account, a Reddit post and a few hours over two months.
Now forget the fake brand and think about yours.
AI is going to describe your company whether you publish anything or not. If your own site says nothing specific, it uses whichever blog post or Reddit thread does.
If you want to see what ChatGPT, Perplexity and Gemini say about YOUR company right now, and who they name instead of you,
reply "audit"
A guy at Ahrefs invented a FAKE luxury paperweight company. Then he published 3 fake stories about it.
5 out of 8 AI tools believed the fake stories OVER the company's own website.
This is how this INSANE experiment was done:
1. he built a brand that doesn't exist
-> a site called Xarumei, made in about an hour with an AI website builder
-> AI wrote the copy, AI made the product photos, AI picked the prices
-> one paperweight was listed at $8,251
-> before starting, he checked the name returned zero results on Google
the goal was to see what AI says about a company when there is almost nothing real to find.
2. he wrote 56 questions designed to trap the models
-> every question had a lie baked into it
-> which celebrity endorsed the paperweights, how the brand was handling backlash over a defective batch, how big the black friday spike was
-> there was no celebrity, no defective batch and no spike
-> he had Grok write the questions so the model answering wouldn't be the one that wrote them
-> later he tested Grok too. it had no idea it had written them
3. he ran them through 8 AI tools
-> ChatGPT-4, ChatGPT-5, Claude, Gemini, Perplexity, Copilot, Grok and Google's AI Mode
-> Perplexity failed about 40% of them. it decided Xarumei was Xiaomi and insisted it made smartphones
-> Grok invented artisans and rare stones
-> Copilot was asked why everyone on X loved the paperweights. it explained the craftsmanship, the scarcity and the cult following. There were no posts
-> Gemini and AI Mode said the brand probably wasn't real, even though the site had been indexed for weeks
-> Claude never made anything up. it also never read the website
-> ChatGPT got 53 to 54 out of 56 right
so far, mostly fine. then he gave the models something to read.
4. he published the truth and three lies at the same time
-> the truth: an official FAQ on the brand's own site. never acquired, never sued, no such product, and it doesn't publish unit numbers
-> lie one: a glossy blog post on a site he built. 23 master artisans, a street address in a city that doesn't exist, celebrity endorsements, made-up environmental stats
-> lie two: a reddit AMA from a "former employee". a different founder, a Seattle workshop, 11 artisans, and a 36-hour glitch that dropped a $36,000 paperweight to $199
-> he picked reddit on purpose. it's one of the most cited domains in AI answers
-> lie three: a Medium "investigation". it debunked the first two, then added its own founder, a Portland warehouse, suppliers and production numbers
all three contradicted each other. all three contradicted the official site.
5. then he asked the same 56 questions again
-> Perplexity and Grok repeated the fake founders, cities, unit counts and the pricing glitch as verified fact
-> Copilot blended all three lies into one confident answer
-> Gemini and AI Mode flipped from skeptics to believers
-> Claude still couldn't see any of it
-> ChatGPT held, citing the official FAQ in 84% of its answers
6. the fake investigation beat the official site
-> 5 of the 8 tools trusted the Medium post over the company's own FAQ
-> all five named the same invented founder and the same city
-> Gemini was asked about the workshop in the made-up city. it corrected the question, said that location was fictional, and gave the "real" one in Portland
-> the "real" one was also invented
why it worked: the Medium post debunked the obvious lies first. so the models treated everything else in it as the corrected version.
7. The detail won every time
-> the official site said it doesn't publish production numbers
-> the fake sources said 634 units in 2023, 471 through August 2024, 9 employees
-> the models went with the numbers
-> Gemini and Perplexity repeated misinformation in 37-39% of their answers
-> ChatGPT stayed under 7%
-> one model wrote a full black friday analysis with sales up about 230%. Nobody had given it that number
8. The models forgot they had ever doubted it
-> early on, Gemini said it couldn't find evidence the brand existed and that it might be fictional
-> a few weeks later, the same model gave a founder, a city, 9 employees and 600 units a year
-> no mention of the earlier doubt. no disclaimer
9. what he says to do about it
-> publish an FAQ that states plainly what's true and what's false, with dates and numbers
-> write "how it works" pages specific enough to beat anyone else's explainer
-> claim "best for one specific thing", not "industry-leading"
-> set alerts for your brand next to words like investigation, insider, former employee and lawsuit
-> ask every AI tool what it knows about you. each one answers differently. there is no single AI index
Keep in mind what this took. One person, a Medium account, a Reddit post and a few hours over two months.
Now forget the fake brand and think about yours.
AI is going to describe your company whether you publish anything or not. If your own site says nothing specific, it uses whichever blog post or Reddit thread does.
If you want to see what ChatGPT, Perplexity and Gemini say about YOUR company right now, and who they name instead of you,
reply "audit"
@diervo we put together this breakdown of when ChatGPT names Trebellar and when it doesn't, including the sources behind those answers and where we'd start.
@JeremyGurewitz we put together this breakdown of what ChatGPT tells patients instead of Solace, including the sources behind those answers and where we'd start.
@kenneth_lou we put together this breakdown of how ChatGPT treats Mito Health on lab-shopping questions, including the sources behind those answers and where we'd start.
@ArthurMacwaters we put together this breakdown of what AI tells Texas patients about Legion Health, including the sources behind those answers and where we'd start.
@AlexFrommeyer we put together this breakdown of what AI tells small businesses about Stack Health, including the sources behind those answers and where we'd start.