What we learned building Goblyn, an AI that runs a brand blog
Goblyn is our autonomous AI platform for running a brand blog. The hard part was never the writing.
We build software for other people. We also build our own. Goblyn (goblyn.ai) is ours: an autonomous AI platform that plans, writes, publishes, and optimizes a full brand blog, including multi-tenant architecture, custom domains, subscriptions, and the publishing pipeline. Its feedback loop is tuned for two things at once, search rankings and citations inside AI assistants.
When we started, we assumed the interesting engineering was in generation. Get the drafts good enough and the rest is plumbing. That was wrong in both directions. Draft quality turned out to be the cheapest problem to solve. Deciding what deserves to be published, and then figuring out whether anything worked, turned out to be the whole product.
The writing was never the hard part
Google has been clear about where the line sits. AI-generated content is not banned. What violates its spam policies is "using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results." That distinction is not a technicality. It is the design constraint for any system that publishes on its own.
And the platform has acted on it. Google said its March 2024 core update, combined with prior work, produced 45% less low-quality, unoriginal content in search results. Scaled, undifferentiated output is not a gray area. It is an actively targeted failure mode.
So the question we had to answer in code was not "can the model write this post." It was "does this post deserve to exist." An autonomous publisher that cannot say no is just a spam generator with better prose.
Using automation, including AI, to generate content with the primary purpose of manipulating ranking in search results.
Clicks are not the scoreboard anymore
The second surprise was measurement. We built the first version assuming sessions and clicks would tell us whether the system was working. Then we looked at what is happening to clicks.
Pew Research found that Google users who encountered an AI summary clicked a traditional search result in 8% of all visits, compared with 15% of visits when no summary appeared. Clicks on a link inside the summary itself happened in just 1% of visits. If your blog exists to earn clicks from search, you are optimizing a number that is being squeezed from both sides.
Being cited and being named are two different problems
Once you accept that the answer is the destination, the obvious next move is to track citations. That is where we hit the third lesson.
A Semrush study with Kevin Indig looked at 3,981 domain appearances across ChatGPT, Gemini, AI Overviews, and AI Mode. It found that 62% of AI citations are "ghost citations," where the source is linked but the brand is never named in the answer text. Your page can be the thing the model read and the reader can still walk away with no idea who you are.
For a feedback loop, that means two separate signals. Was the page cited? Was the brand mentioned? A system that tracks only one of those is reporting on half the outcome.
An owned blog competes for a thin slice
There is a scale problem underneath all of this. Profound's analysis of a dataset of 680 million citations shows AI assistants concentrate their citations on a small set of aggregator and community domains. Wikipedia alone accounted for 7.8% of ChatGPT citations.
That is a useful thing to internalize before you fund a content program. An owned brand blog is not competing for the whole citation pool. It is competing for what is left after the aggregators take their cut. Which is another argument for publishing fewer, more specific, more genuinely useful things: broad summary content is exactly what the aggregators already win.
How that changed what we built
Three changes came out of this, and they are the reason Goblyn looks the way it does today.
- The publishing decision is a first-class part of the pipeline, not a formality. Usefulness is the gate, because volume is the thing Google says it targets.
- The feedback loop watches rankings and AI assistant citations together. One without the other gives you a distorted picture of where your readers actually are.
- Citations and brand mentions are measured as separate outcomes, because most citations never name the brand.
| Volume-first content engine | How we built Goblyn | |
|---|---|---|
| Success metric | Posts shipped per week | Search rankings plus citations inside AI assistants |
| Risk model | Treats scale as the strategy | Treats scaled, undifferentiated output as the failure mode Google targets |
| Visibility tracking | Clicks from search | Citations and brand mentions tracked as separate signals |
| Verdict | Same words on the page. A completely different system behind them. | |
What this means if you are building something similar
If you are putting an AI system into production that publishes, recommends, or answers on your behalf, the model is the commodity part. The judgment layer around it is the product. Decide what your system is allowed to refuse to do, and decide what you will measure before you turn it on.
That is the kind of work we do. to1 Labs is a software studio founded by Alec Kretch that takes software from first idea to launched product, with strategy, design, and full-stack engineering under one roof. AI-native builds, including LLM-powered products, agents, and content systems, are our specialty, and Goblyn is one of the reasons we can say that with a straight face: it runs in production. You work directly with Alec from start to finish, with no account managers and no hand-offs. Projects are mapped into fixed milestones with clear pricing before the build starts, so you know what ships and when. We read every inbound message and reply within two business days.
- Google does not ban AI content. It targets automation aimed mainly at manipulating rankings, and it reported 45% less low-quality, unoriginal content in results after its March 2024 core update and prior work.
- Clicks are a shrinking scoreboard. Pew found users who saw an AI summary clicked a traditional result in 8% of visits, versus 15% without one, and clicked inside the summary in 1%.
- Citation tracking and brand-mention tracking are different jobs. Semrush and Kevin Indig found 62% of AI citations never name the brand, and assistants concentrate citations on a few aggregator domains.
- The hard engineering in an autonomous publisher is the decision to publish and the feedback loop after, which is why Goblyn measures rankings and AI assistant citations together.
- Google's guidance is that AI content is not banned, but automation used mainly to manipulate rankings violates its spam policies.
- Google said its March 2024 core update plus prior work produced 45% less low-quality, unoriginal content in search results.
- Pew Research found users who saw an AI summary clicked a traditional search result in 8% of visits versus 15% without a summary, and clicked a link inside the summary in 1% of visits.
- A Semrush study with Kevin Indig analyzing 3,981 domain appearances across ChatGPT, Gemini, AI Overviews, and AI Mode found 62% of AI citations are ghost citations where the brand is never named.
- Profound's analysis of a dataset of 680 million citations shows AI assistants concentrate citations on aggregator and community domains, with Wikipedia at 7.8% of ChatGPT citations.
Is it against Google's rules to publish AI-generated blog posts?
No. Google's guidance is that AI-generated content is not banned. What violates its spam policies is using automation, including AI, to generate content with the primary purpose of manipulating rankings. An autonomous publishing system has to be built around usefulness rather than volume.
Why track brand mentions separately from citations in AI answers?
Because most citations do not carry your name. A Semrush study with Kevin Indig across 3,981 domain appearances in ChatGPT, Gemini, AI Overviews, and AI Mode found 62% were ghost citations, where the source was linked but the brand was never named in the answer. If you only count citations, you miss whether anyone learned who you are.