What we shipped
- An AI Visibility Index for the UK property sector: a study of which brands ChatGPT names when people ask it to recommend one, delivered as an interactive web report with a downloadable sixty page deck behind it.
- Built to the standard a client would expect when commissioning a report using real prompts, split by buyer persona and measured at depth. For transparency, we also published the full methodology and made the raw data downloadable.
Why we built it this way
Our research shows that "Freshness" is a core feature in AI answers so running the report on a regular basis gives valuable information about how the LLMs produce results. To help benchmark our clients' performance we needed to grab a picture of the entire industry to understand client performance relative to the rest of their cohort. Property went first because we have a lot of property clients from Foxtons to John D Wood to NMRK so our team had deep industry knowledge that they could easily ratify against live client data.
The worry with building any report to use as marketing and not as a client deliverable is that you go a mile wide and an inch deep as it's not 'billable' work. However, we made sure to spend the time and the money to build proper infrastructure to get meaningful results that businesses would actually use.
How it benefits clients
- It gives them something to benchmark against and compare to. AI performance is relative to the people you are competing against. If lots of news and media sources show up, it's not the same as a direct competitor owning that share of voice.
- It also provides raw backend reporting data to integrate into our live client reports so we can now show clients not only how they are individually performing but how well they are doing versus the top sites.
What's next
The plan is to build it across more countries and sectors and have it update quarterly so we can start to see trends over time.
Technical innovations
This was our first real test of our internal agentic stack that we are using to build out the agents across the business. Everything we built had DRY in mind (don't repeat yourself) and everything can be later reused: directed acyclic graphs orchestrating the run, and a medallion warehouse moving raw AI responses from bronze (ingest) to silver (clean, join, tag) to gold (modelled, ready).
We also used our internal methodology of shipping it as a vertical slice, zero to live on a focused MVP, ChatGPT only, instead of a three-month build.
How it fits the agentic journey
It runs on our FEED framework. So instead of just having some raw data turned into graphs and charts, we measure domains against our own methodology to see if our framework holds up to scrutiny with real world data. Does Freshness still matter as much? Which entities are being mentioned? Do client endpoints make that much of a difference? Are there obvious content distribution patterns the AI is using to cite our client.
As we have been shipping interactive data assets for clients as part of our digital PR data journalism offering for the last 8 years the front end build was the most trivial part for us. As always, getting, cleaning and making sense of the data was the hard part.