What we shipped
- An end to end research and briefing system that built in key SEO considerations like keyword research, internal-link research, heading structure, information-gain analysis, missing-entity detection and real E-E-A-T signals. All with the client's tone of voice and brand guidelines as well as expert commentary and real world anecdotes.
- A self-healing client profile that ingests feedback on Google Doc comments, synthesises them, diff-checks them and folds them into the client profile so each subsequent brief is better and more accurate than the last.
Why we built it this way
Automated content is rarely good. Like hen's teeth, golden goose egg and month of Sundays rare. We have all seen the Ahrefs graphs on X where a domain gets a massive traffic spike followed by a righteous slap from Google.
So giving AI a keyword and asking it to magically write content that ranks was out. We trialled little 'tricks' like using humanisers and AI detection software and found them mostly to be a big steaming pile of rubbish. We learned that humanisers work by changing the burstiness and perplexity of the AI content and changing some words for good measure. Whilst this allows the text to pass AI detection software, it makes the content materially worse.
We learned that there is a sort of poetic built in hard limit to using LLMs to write. The only way for a computer to deceive a computer with writing is to make the writing worse. Thus defeating the purpose. As an amateur writer, this was a delightful discovery.
How it benefits clients
- Content built on real research, in their voice, that does not read like everyone else's.
- True information gain in each article that ranks well (and stays ranked)
- Reducing the admin time associated with copy around research, feedback, edits, etc.
- Get better over time: every edit feeds a larger system so that one contextual update then goes online for all future articles we create
What's next
- Richer input capture via the phone so we can get our AI to talk to the experts in the clients office to build quick mini briefs for every article
- More indepth analysis of entities that exist/are missing to satisfy information gain
- Final content sweeper that scores each piece according a bespoke content quality rubric that we have for each client
Technical innovations
As mentioned earlier, we discovered a poetic in-built fault in using LLMs for writing content. The fault is to do with vector space. If the inputs are single keywords or traditional SEO research inputs then the brief creates content that is in a certain vector space and that vector creates a certain obvious machine-made shape that would be impossible for a human to replicate.
For example, when writing a brief for "pink gin" we wanted to reference the Barbie movie in the copy as a jovial throwaway line that made an otherwise basic article more interesting to read. However, we found unless you explicitly input things like "barbie" into the brief, it would be physically impossible for the LLM to ever write that because the 'next token' is so far away that it would never be able to reach the word "barbie".
We also realised that heavy briefing around style, word choice, etc was the only way to actually build out passable first drafts.
How it fits the agentic journey
In our AI manifesto, we talk about a restructuring of work, not a replacing of work. Our frameworks are built with a human in the loop at every stage and the AI is used to augment the human experience, not to replace it. So when we discovered the built-in tension between 'humanisers' and LLM text we delighted that it accidentally fit our thinking. AI is a great growth engine, metaphorically like a Japanese bullet train, but like the bullet train, it needs to be put on rails and driven by an experienced technician.