What we shipped
- An engine that produces a full AI-visibility report for an entire industry, around sixty pages, and auto-publishes it to our website. We run it across every sector our UK and US clients operate in.
- Reports are now data plus a library of composable modules (heatmaps, leaderboards, flow diagrams, prose), assembled by the CMS and mapped to a per-industry template, so a new report is mostly a new data file. Each one refreshes every quarter, ready for client business reviews.
Why we built it this way
We needed a live view of what is happening across whole industries, not one brand at a time, and we needed it fresh every quarter so there is always current data to talk through in a client's quarterly review. Building each report by hand would never scale to that. So we built the machine instead: define the data contract once, and let it produce, update and publish on a cadence.
How it benefits clients
- Current, defensible numbers at every review, refreshed on a predictable schedule rather than scrambled together.
- The same engine spins out marketing assets, a branded slide deck and press-release material for their sector, so a report becomes outreach, not just a document.
What's next
Automated quarter-over-quarter cutover, and per-client benchmarking against named competitors.
Technical innovations
Inputs come from our Keyword Universe agent plus a scrape of where real questions live: X, Reddit, Google's People Also Ask, industry FAQs. An LLM works out the personas that matter and the questions each would ask, then we run every prompt three times and average it to cut the model's noise, across each persona, so we can see the real players in each category. All of it moves through a three-tier warehouse, raw to modelled to published, so the figures are reproducible. The published pages are flat and script-free, so they load fast and render anywhere.
How it fits the agentic journey
This is 4Ds data journalism turned into a production line, and the manifesto's "it's all just databases" at full stretch: set the data flow up once, wire it with small primitives, and good reports stop being a scramble and start being the default.