x
Ten fixed questions, run across two assistants, logged in three columns, once a month. Ten minutes.
The reason it exists is simple: there is no rank tracker for AI answers. Nothing tells you when a model stops mentioning you. No dashboard, no email, no weekly report. If you don't go and ask on a schedule, you will never find out — and you'll be optimising in the dark while feeling productive, which is worse than not optimising at all.
Write them once, then never change the wording. That second part is the whole method.
They should be the questions a buyer types, not questions with your name in them. "What do you know about [my business]" proves nothing month to month — of course it can find you when you've named yourself. The useful questions are the ones where you might not appear:
Five like that, five more specific to what you actually sell. Save them in a note. Freeze the wording.
Frozen wording is not fussiness. If you reword the questions each month you are measuring your own edits rather than the world, and you'll produce a chart that looks like progress and means nothing. This is the same discipline as pre-declaring the metric before an experiment: decide what you're measuring before you can see the result, or you'll unconsciously choose the measure that flatters.
Three columns and a date. Resist anything more elaborate — elaborate logs die in month two.
| Column | What it means |
|---|---|
| Named | You appeared, described roughly correctly |
| Not named | You didn't appear at all |
| Named but wrong | You appeared with a material error — wrong service, wrong scale, wrong location |
Ten questions × two assistants = twenty data points a month. Tally them. Add one free-text line: what changed since last time.
Two practical notes. Run each assistant in a fresh session with no memory of previous questions, or you're measuring the conversation rather than the model. And run the whole thing in one sitting — spreading it across a week introduces changes you can't see.
Because a rough instrument you run beats a precise one that doesn't exist, and because the unreliability is bounded in a way that still leaves signal.
Be clear about the limitation first. Generated answers are non-deterministic: ask the same question twice and you can get different names. So any single month's reading is noise. Twenty data points is a small sample. Nobody should present this to anyone as a metric.
What it gives you is a trend across three months, which is a different and more defensible claim. If you were named twice in twenty in June and eleven in twenty in September, something moved. If it's two, three, two, nothing has, and that is also useful — it tells you the approach needs changing rather than repeating.
Three months of zero is data, not failure. It's the trigger to re-examine the method rather than to run it a fourth time hoping. Deciding that in advance is what stops a habit becoming a ritual — write the trigger down next to the questions, so the future version of you who's invested in this working doesn't get to renegotiate it.
The same logic applies to the whole theme: this newsletter's own publishing strategy is measured by citation checks rather than sessions, precisely because click-through data no longer measures what it used to. A strategy with no instrument can't be evaluated — but an instrument nobody looks at is the same thing with extra steps.
First working day of the month, ten minutes, before anything else opens.
Attach it to something that already happens. If you already do a Saturday close, make it the first Saturday's extra ten minutes. A habit with no anchor is a habit you'll do twice.
And keep the log where you'd find it in a year, not in a chat window. The point of a fixed measure is the comparison, and the comparison only exists if the old readings survive — same argument as the decision log, applied to your own visibility.
Ten questions. Two assistants. Three columns. Ten minutes a month. It is not a good instrument. It is the only one there is, and it's better than the alternative of finding out in a year that nothing ever mentioned you.
For the audit that finds where your week actually leaks — free, 5-min read: The 10-Hour Audit.
Related Decodes
AI the Brief (@aithebrief) — AI advice you can act on and trust, for time-starved founders.
TL;DR: There is no rank tracker for AI answers — nothing tells you when a model stops mentioning you — so the only option is to ask on a schedule. Write ten questions a buyer would actually type, not ones containing your name, then freeze the wording permanently: rewording each month means measuring your own edits rather than the world. Run them across two assistants in fresh sessions, in one sitting, and log three columns — named, not named, named but wrong — for twenty data points a month. Any single month is noise, because generated answers are non-deterministic and the same query returns different sources on different runs; what you're reading is the three-month trend. Three months of zero is data rather than failure, and it should trigger changing the method — write that trigger down in advance so you can't renegotiate it later. Anchor the ten minutes to an existing monthly habit, and keep the log somewhere it survives a year, because a fixed measure is only worth anything if the old readings still exist.
