The constraint: model output varies
Every answer is sampled. Ask the same question twice a day and you can get different sources, which means a short window measures noise as much as signal. That single fact determines the whole schedule: the interval has to be long enough for real change to exceed run-to-run variation.
This is why weekly comparisons are usually misleading, even when they are run carefully.
Weekly: useful for faults, not for progress
A short interval still earns its place, just not for measuring improvement. It catches regressions: crawl access breaking, a template deploy removing a schema block, a page becoming unreachable, an answer becoming factually wrong.
- Small fixed question set, five to ten questions
- Pass or fail checks only: blocked, unreachable, wrong answer
- No trend lines, no percentages, no conclusions about progress
Monthly: the working interval
This is where most programmes should measure. A month is long enough for crawling, retrieval and model updates to have happened, and short enough that a course correction still fits inside a quarter.
Run the full question set, keep the raw responses, and compare against the previous round. Two data points is enough to see direction; three starts to show whether it holds.
- Full question set across all agreed platforms
- Store every raw response, not just the summary number
- Compare mention rate, cited pages and competitor presence
- Note which specific actions were taken in the period, to attribute movement
Quarterly: business outcomes
Monthly visibility numbers tell you whether the work is landing. Only a quarterly view tells you whether it mattered, because referral traffic and pipeline need time to respond.
Keep this separate from the monthly series. Mixing them produces a metric that moves for reasons nobody can identify.
- Sessions and enquiries referred from AI engines
- Conversion on AI-referred traffic against other sources
- Changes in the competitive picture: who is being cited now
Making each round comparable
Comparability is the part teams skip and later wish they had not. Fix the question set, the platform list and the phrasing for the duration of the programme, and record the timestamp of every run.
If you must change the question set, report on both the original and the new one. Never replace it silently — the series stops being a series.
What not to do with the numbers
Two habits produce conclusions that cannot be acted on. The first is comparing across platforms as though they were one aggregate: Perplexity citing you and ChatGPT not citing you says something about the two systems, not about you.
The second is drawing a conclusion from a single movement. One round showing a jump is a data point; three rounds in the same direction is a trend. Anything else is a story you tell in a review meeting.
Weekly for faults, monthly for progress, quarterly for business impact. Keeping those three apart is what makes the numbers worth arguing about. Sources reviewed on 2026-10-17. This article records a method with verifiable boundaries; it does not claim that any search or AI platform will index, cite or recommend anything.