The short answer, by situation
| Situation | Cadence | Why |
|---|---|---|
| Establishing a baseline | Once, properly | Multi-pass across every engine. Everything later is measured against this. |
| Steady state | Monthly | Long enough for published work to be found; short enough to catch a competitor moving. |
| After publishing a fix | 30–60 days later | Below thirty days you are usually measuring indexing lag, not effect. |
| Around a launch or pricing change | Before and 30 days after | The before-run is the only chance to have a comparison at all. |
| Agencies, per client | Monthly or quarterly | Match it to your reporting cycle rather than inventing a separate one. |
Why weekly is usually wasted money
AI answers vary between runs. Some of what looks like movement week to week is ordinary variation rather than a change in your position — which means a weekly cadence produces a chart that moves constantly and tells you very little.
It also does not match how the underlying thing changes. Publishing a page, having it found, and having an engine start using it as a source is a process measured in weeks. Checking every seven days mostly measures the noise floor.
There is a cost argument too. On pay-per-use, a weekly cadence is roughly four times the credits of a monthly one for information you cannot act on any faster.
What makes two runs comparable
This matters more than frequency. A comparison only means something if the runs are alike.
- Same questions. Change the prompt set and you have changed the measurement. Keep a stable core and add new questions as a separate group.
- Same engines. Dropping an engine changes the totals even if nothing about your visibility changed.
- Same number of passes. Three passes and one pass are not comparable, because one of them is far more exposed to variation.
- Note anything excluded. If an engine could not be retrieved, that run is missing data and should say so rather than counting a zero.
Where runs are not comparable, the honest thing is to say so rather than draw a line between them. A flattering trend built on mismatched runs is worse than no trend.
What to actually record each time
A cadence is only useful if you capture the same things every run. Four numbers are enough for most people.
- How many questions cite you at all. The blunt headline number.
- How many come back with no owner. Unclaimed answers are your cheapest opportunities, and the count moves as competitors publish.
- Which domains are being cited instead. This is your outreach list, and it changes faster than you would expect in a young category.
- Per-engine breakdown. Absent everywhere points at your page. Absent in one place points at that engine.
Everything else is detail you can pull from the report when you need it. These four are what tell you whether to change course.
A cadence that works
Something like this for most businesses:
- Month 0 — baseline. Your full question set, every engine, multiple passes.
- Months 1–2 — publish against what the baseline found, especially anything that came back with no owner.
- Month 2 or 3 — re-run the identical set. Record what moved, per engine, per question.
- Then monthly, with a bigger review each quarter.
On pay-per-use that is a predictable, small line item rather than a subscription you pay through the months you do not run anything.
Common questions
Is monthly really often enough?
For most businesses, yes. The underlying process - publish, get found, get used as a source - moves in weeks, not days. Monthly catches real movement without paying for variation.
What does a check cost?
It depends on size. A ten-prompt run across six engines is 15 credits; thirty prompts is 35, or 45 with the LLM Mentions add-on. $1 = 1 credit, with no subscription underneath.
Should I check more often after publishing something?
Wait thirty days at minimum. Below that you are usually measuring how long indexing took rather than whether the page works.
Can I compare a ten-prompt run to a thirty-prompt run?
Not meaningfully. Keep the question set stable if you want the comparison to mean anything, and treat added questions as a separate group.