When you buy something through one of the links on our site, we may earn an affiliate commission.
Just before the holidays, I stumbled upon a LinkedIn post about ‘GPT-SEO’.
It’s an idea by Kristin Tynski, the co-founder of the content marketing agency Fractl.
She’s a brilliant resource for interesting insights and data in the industry and someone who frequently uses generative AI in her company’s workflow.
She wonders if companies could game GPT-AI-generated text to increase the likelihood of brand mentions in the output.
Kind of like what we do on Google with SEO.
Her theory is based on how the GPT model was trained.
GPT-3, in particular, is trained on:
- Common Crawl (filtered) – 410 billion – 60%
- WebText2 – 19 billion – 22%
- Books1 – 12 billion – 8%
- Books2 – 55 billion – 8%
- Wikipedia – 3 billion – 3%
So she considered ways to leverage into an actionable strategy possibly worth considering.
For instance:
WebText2 represents 22% of GPT3’s training.
This is the text of web pages from all outbound Reddit links from posts with 3+ upvotes.
In this way, she wonders if you could increase your chances of mentions on large AI language models by posting more on Reddit and linking to unique articles about your brand.
Of course, this would require a huge number of posts and votes from others, but perhaps companies could even use AI to help – at least with the posting.
Mentions in books or even self-publishing books that are indexed in future book corpora is another possible way.
She also mentions having a Wikipedia page.
The Content Marketing Institute has a helpful guide on how to do this in a way that’s likely to get approved:
- Research Wikipedia best practices and browse the Wikipedia Conflict of Interest guidelines
- Create your account with your real name and email, and start making minor changes on pages to build up your user profile until your user level reacher an “auto-confirmed user”
- Gather verifiable and objective third-party sources that are unrelated to you
- Write up a draft copy of your article following guidance and send it in for review
Considering GPT-3.5, the tech behind ChatGPT only read up to 2021; this theoretical strategy is more focused on gaming GPT-4 or so long as it’s not trained on a far greater dataset.
And perhaps that could be extrapolated and used to game other AI systems like Quora’s POE.
But it is an interesting idea and another opportunity for the symbiotic relationship between SEOs and AI.
Here’s her post below!
