What we shipped
- An edit mode for the content agent. Instead of working from our brief, it reads a client's live page, cleans and chunks it, and vectorises it.
- It then pulls the top ten ranking pages for the same keyword, does the same to them, runs named-entity recognition across all of it, and shows exactly where the client's page is thin on information gain.
- Then it suggests internal links with the right intent, drawn from a clustered crawl of the site.
Why we built it this way
Most of the time you don't want to completely rewrite a client's landing pages as they have already invested in producing them and they are fulfilling some function in their marketing funnel already. Writing new content from scratch is straightforward but actually editing existing content to make it better is much harder as there is a defined flow and editorial voice that you need to sync with.
When understanding what to edit, we would typically do keyword and competitor research as well as some additional research on the SERP to get a feel for what the client's content was not answering for the user.
For one page this could take anything from 30 minutes to a few hours. So if we could use agents to automate away some of the research and analysis we would be onto a winner.
How it benefits clients
- Measured upgrades to existing pages that close genuine content gaps against the competition.
- Additional client context, voice, quotes and anecdotes that personalise the content to make it not only rank well but convert better
- Internal links that point at the right page to support rankings and customer information flow instead of blindly linking
What's next
- Finer control over exactly what changes on a page, so a rewrite can be as light or as deep as the situation needs.
- More research inputs and comparisons
- A finer grain rubric to decide the information gain score of each piece
- Scaled production that's linked to analytics so we can automatically attach a cash value to each page and go from highest value to lowest forecast
Technical innovations
The comparison is the clever part. By embedding both the client's page and the ranking set and running entity recognition across them, we can see what the competition covers that the client does not, in a way a human skim would miss. Linking is driven off a clustered crawl, so a suggested link reflects the site's real structure rather than a guess.
How it fits the agentic journey
Our BOOM framework is very hard to agentify. It's a messy judgement based framework where we decide what to implement based off of the data from the REST phase. By identifying the exact production bottlenecks and deciding which could be safely run by AI, we reduce the time our specialists spend combing through data and doing research and instead present them with rubrics to judge and build content plans from.