How small a delta is too small to act on?

Any delta below the system's own variation floor should be treated as noise. That floor is found by re-running the measurement over an interval too short for real market change and observing how much the readings differ anyway.

Does a delta following a launch mean the launch caused it?

No. A delta establishes that a change occurred within an interval, which is sequence rather than cause. Confirming cause requires evidence that rules out the other candidate explanations.

Why is absence treated differently from a low score?

Something not found and something not looked for are distinct states with different meanings. A system that records both as zero produces deltas that mix genuine disappearance with gaps in coverage.

Reading market movement from benchmark deltas

Guide · Continuous Intelligence · 4 min read · last verified 2026-07-21

Reviewed before publication Editorial board Independent commercial review

In shortA delta between two benchmark measurements is evidence that something changed only if everything about the measurement except the market was held constant. It shows movement, never cause.

A delta between two benchmark measurements is evidence of market movement only if everything about the measurement except the market was held constant. Where that condition holds, a delta is one of the most useful objects an intelligence program produces. Where it does not, the delta is measuring the method, and reading it as market movement produces confident errors.

What a delta actually is

A delta is the difference between the same benchmark question set run at two points in time. It is not a metric in its own right; it is a comparison, and comparisons inherit every weakness of the two things being compared.

Four conditions have to hold for the comparison to be legitimate:

Same questions, phrased identically, in the same order
Same sources or surfaces consulted, with the same selection rule
Same measurement window length and position, so that both readings cover comparable spans
Same scoring rules, applied by the same procedure

Break any one of these and the difference between readings contains an unknown mixture of market change and method change, with no way to separate them after the fact.

What a delta proves

A clean delta supports exactly three claims:

Something in the observed surface differed between the two readings
The difference has a direction — more, less, present, absent
The difference occurred somewhere inside the interval between readings

That is a narrower set of claims than most reporting makes, and it is genuinely valuable. Direction and timing are the two things a single snapshot can never supply.

What a delta does not prove

Cause. A change coinciding with a product launch is not caused by the launch. Deltas establish sequence at best, and sequence is weak evidence for cause. This matters most immediately after a release, which is the case examined in when to rescan after a launch.

Business impact. A movement in observed position is not a movement in revenue, pipeline, or preference. The translation between the two is an assumption, and it should be written down as one rather than absorbed silently.

Permanence. A single delta describes an interval, not a trajectory. Two points define a line only in geometry.

Precision of onset. The change happened somewhere between the readings. A shorter interval narrows the window; it does not pinpoint the moment.

Establishing the movement floor

Every measurement system has a level of variation produced by the system itself. Until that level is known, a delta cannot be classified as meaningful.

The way to find it is to re-run the measurement over an interval short enough that real market movement is implausible, and observe how much the readings differ anyway. Whatever variation appears is the floor. Deltas smaller than the floor are not findings, however inconvenient that is when a chart looks compelling. Deltas well above it deserve attention.

Establishing this once, and re-establishing it whenever the method changes, is the difference between a trend line and a decoration.

Reading deltas as a set

Individual deltas are ambiguous. Groups of them are considerably less so.

Correlated deltas — several related questions moving together in the same direction suggest a change in the underlying surface rather than a quirk of one question
Isolated deltas — one question moving while its neighbors hold still more often indicates something specific to that question's wording or source
Opposing deltas — movement in both directions across a set usually indicates noise, a method drift, or a genuine redistribution, and the three are distinguishable only by looking at the captured evidence
Absence deltas — something that was present and is now missing is a finding in its own right, and is frequently more informative than a change in degree

Absence deserves particular care because it is easy to record as zero. A thing that was not found and a thing that was not looked for are different states, and a system that collapses them produces deltas that cannot be trusted.

A reading procedure

Confirm the method was identical across both readings before looking at any number
Compare the delta against the known movement floor
Check whether related questions moved together or independently
Open the evidence trail for both readings and look at what actually changed in the captured material
Write the candidate causes as a list, ranked, with what would confirm or eliminate each
Decide whether the delta warrants action now or another reading first

The step people skip is the fourth. A number tells you that something moved; the captured evidence tells you what moved, and the two are frequently different in ways that change the conclusion entirely.

What to watch

Watch for deltas that appear immediately after any change to the measurement itself — a reworded question, an added source, a revised scoring rule. Those are the highest-risk readings in any program, because the method change and the apparent market movement arrive together and are indistinguishable in the output. When the method must change, run both versions in parallel for at least one interval so the two series can be bridged. Watch also for a delta being cited a second time as though it were an established fact; a change observed once is a candidate, and it earns the status of finding only when a subsequent reading confirms it.