How it works
Measure. Understand. Improve. Measure again.
The loop is the product. Each cycle re-runs an identical benchmark, so any movement is attributable to something that changed in the world rather than to something that changed in the test.
What we measure
Customer question
“Who would you recommend for tree removal near Gaithersburg, Maryland?”
Asked across AI platforms
- ChatGPT
- Gemini
- Perplexity
- Claude
- Google AI Mode
Businesses recommended
Which businesses were named, in what order, with which sources cited, and whether the facts stated were correct.
Recorded as evidence
Every answer stored verbatim. Every published number traces back to one of them.
The process
- 1
Measure
24 buyer-intent prompts built from your services and cities, run across 5 platforms, 2 times each. Every answer stored verbatim.
- 2
Analyse
Each answer is read for presence, position, citation, accuracy and recommendation strength. Judgements below a confidence threshold go to human review rather than being guessed.
- 3
Find the gaps
Which prompts you lost, which competitors won them, and which sources those winning answers relied on.
- 4
Fix the source of truth
Entity data, service and location structure, structured data, crawler access — the things that determine whether a system can establish what you are.
- 5
Build authority
Get properly represented on the third-party sources that demonstrably influence answers in your category.
- 6
Publish useful evidence
Original expertise and real proof, aimed at the specific questions the measurement showed you losing.
- 7
Measure again
The identical benchmark, re-run. Movement is attributable. No movement is reported as no movement.
Why the prompts never change
This is the single most important design decision in the whole system, and it is the one that most obviously works against our commercial interest.
If an agency can rewrite the questions between cycles, it can manufacture improvement by asking easier ones. Freezing the benchmark removes that option. In our system it is enforced by a database trigger rather than by policy — an update to a benchmark prompt’s text raises an exception.
Exploratory prompts live in a separate experimental set and never contribute to a published score.
Why every prompt runs twice
AI answers are non-deterministic. The same question can return different businesses minutes apart, so a single answer is an anecdote. Repetition is how a measurement stops being one.
Why geography is stated explicitly
Every prompt carries a stated location recorded on the observation. Location is never inherited from the device of whoever ran the test — a benchmark that moved because an operator was working from a different city would be worthless.
Why humans review the edge cases
Ambiguous readings — a parent brand named instead of a division, a hedged recommendation, a source that could support either business — are held for human judgement instead of being resolved by a confident guess.
The monthly rhythm
Measurement runs on a schedule so results stay comparable and drift is caught early.
| When | What happens |
|---|---|
| Start of cycle | The frozen benchmark runs. Raw responses are stored, write-once. |
| Scoring | Score, platform scores, share of voice and competitor frequency are computed from validated observations only. |
| Source analysis | Cited domains and pages extracted and compared against the prior cycle. Source gap report produced. |
| Planning | A prioritised 30-day plan built from score changes, missed prompts, source gaps and entity issues. Human approval required. |
| Between cycles | Weekly checks on reviews and reputation, entity accuracy, and crawler and indexing health. |
| End of cycle | Client report: current score, baseline, 90-day trend, platform differences, competitor movement, work completed, next-month plan. |
See it run on your business
The baseline is cycle one. Everything after it is measured against what it finds.