An AI visibility tracker for brands.

Challenge

Buyers are getting their answers somewhere analytics cannot see

Their team could see the shift those clients were starting to feel: people asking an assistant, getting a usable answer inside it, and never landing on a page anyone could count. Rank tracking has nothing to say about that, and there was no shared unit for comparing a mention inside an AI answer against a third-place search result.

A zero-click answer leaves no session to count

When the answer arrives inside the assistant, no visit happens, so analytics goes quiet exactly where the brand most needs to defend itself. Measuring it means asking the platforms directly, on a schedule, and keeping what they say.

Constraints

The brief and the funding constraints

Approach

Inside the Total Search Visibility Platform

Citation monitoring across the AI platforms

A scheduled agent takes the target prompts from a table the agency controls, puts them to ChatGPT, Perplexity and Gemini, then parses each unstructured answer for mentions and links back to the brand. Rate limiting, retries and cost ceilings are built into the agent itself.

Every citation pinned to a date, a platform and a source

Per-prompt citation rate, average position and sentiment, with the run history underneath showing exactly when each platform cited the brand and which page it linked to. Two identical queries an hour apart can return two different answers, so the history is the only honest way to read the trend.

Search, social and AI through one ingestion path

Views, likes, comments and subscriber growth land in the same normalised insert path as search and AI, which is what makes a single index possible at all. Each source runs as its own function owning one connector and one validation pass, so a broken source stops ingesting and no bad numbers reach the score.

Scoring weights the agency edits in a screen

The scoring model runs from the database. Weights sit in a table with an admin interface over it, and a view recomputes the score from the latest weights on every read. When the agency decides a footnote in an AI answer is worth more than a position-eight result nobody clicks, they change a row and the next read reflects it.

Decisions

The design decisions behind the score

Every platform cites differently

One returns footnotes, another puts links inline, a third paraphrases the brand without linking anywhere. OpenKit wrote a parser for each platform, and that parsing logic is a meaningful part of what the agency now owns.

The weighting is the agency’s judgement

Deciding what a mention inside an AI answer is worth relative to a search position is a professional judgement, and it moves as the platforms do. That judgement belongs to the agency.

Result

What we handed over to Hallam

The platform runs in production, with Hallam’s own admin tier active, their beta clients on read-only dashboards, and the citation monitor working through prompts the agency controls. OpenKit wrote the methodology up as the agency’s own intellectual property, and the admin interfaces cover what a working team changes day to day: the prompts, the weights, and who can see what.

• The platform in production, hosted in the UK, with role-based access and per-tenant isolation enforced at the data layer.
• The citation monitor: the agent, the prompt management around it, and the parsing logic that reads the answers.
• The scoring engine, its documented methodology, the admin interfaces, and the repository itself.

The build

Supabase PostgreSQL (AWS London), Vite + React + TypeScript, Tailwind + Recharts, Supabase Edge Functions, LLM Citation Monitor (scheduled agent).

Controls on this project

UK GDPR, UK data residency (AWS London), Row-Level Security per tenant. OpenKit holds ISO 27001, ISO 9001 (UKAS-accredited) and Cyber Essentials.