When the answers move

Ask an assistant the same question on Monday and again on Tuesday. Most of the time you get much the same answer in slightly different words. Sometimes you get a different answer entirely — a different set of sources, a competitor named where you were named, a recommendation that has quietly changed its mind.

Nothing tells you which of those two days you are on. There is no changelog for an answer. The model may have been updated, its index may have been refreshed, or a news cycle may have moved underneath the question — and from the outside all three look identical, which is to say all three look like nothing at all.

One number, once a day

So we started publishing the one thing that can be measured from the outside: how far the answers moved from one day to the next, across every question we ask, on a scale of 0 to 100. It is free, it updates daily, and it lives at getlooploop.com/ai-volatility with no account.

It is a reading about the field. It is not a reading about you.

What is being compared

The same question, asked of the same assistant, on two consecutive days. That pairing is the whole method: comparing two different questions, or the same question across two different assistants, would measure the questions and the assistants rather than the movement.

For every such pair three things are compared, and the day's reading is the mean of them across every pair.

Reading the scale

The scale is the movement itself, not a grade. Three bands are worth knowing, and the middle one is where most days sit.

• 0 to 5 — settled. The assistants answered yesterday's questions much as they had the day before. A change in your own visibility on a day like this is about your pages, not about the field.
• 6 to 12 — ordinary movement. Where most days sit. Individual answers shift — a citation swapped, a brand dropped from one list of three — without the field moving as a whole.
• 13 and above — the field moved. Answers changed widely enough that a single day's reading of your own brand should not be treated as a verdict. Wait for the next one before rewriting anything.

What a normal week looks like

Since the reading began it has spent most of its days in single figures, with the occasional day in the mid-teens. The chart's ceiling is fixed at 40 rather than rescaling to whatever is in view, and that is deliberate: an axis that rescales itself makes a settled fortnight look identical to a turbulent one, which is the opposite of what the page is for.

Some days are drawn as hollow points and the line skips them. Those are days measured on too few comparisons to say anything — a reading of 14 from thirty comparisons is a sample, not an event. They stay on the chart, because hiding a thin day would be a quieter lie than drawing one.

What it cannot tell you

Four limits, and they are printed on the page itself rather than in a footnote nobody reads.

• It is not about your brand. It is the field, measured across every question in the corpus. Your own visibility is a separate reading.
• It cannot say why. A model update, an index refresh and a news cycle all look the same from here.
• It is not a live curve. The reading is computed once a day, and a day with too few comparisons is published as thin rather than drawn as a fact.
• The source-citation term is the thinnest of the three. Most answers do not carry citations on both days, and the page prints what share did.

How to use it

Two ways, and both are about restraint.

Before you act on a drop, look at the day. If the field moved and you dropped, the honest reading is that you may not have dropped at all — the answers were moving around everybody, and one day is not a trend. If the field was settled and you dropped, that is about your pages, and it is worth the work.

Before you celebrate, the same in reverse. A rise on a turbulent day is the weather. A rise on a settled day is yours.