Freeze the margin, preserve the pair, narrow the claim
Use this calculator only for one fixed binary endpoint such as confirmed brand mention, confirmed brand/domain citation, or confirmed active recommendation. It does not fit rankings, visibility scores, sentiment, counts, unmatched samples, or two independent groups.
Ask whether the plausible change stays inside your margin
Name the decision and margin, affirm only supported assumptions, then add one row per exact matched prompt-and-repetition unit.
Formula and eligibility rules
- Eligible matched pair = both periods have reviewed binary evidence—confirmed presence or confirmed absence—for the same row key under confirmed matching conditions.
- Stable absence = absent before and after. Gain = absent before, present after. Loss = present before, absent after. Stable presence = present before and after.
- Paired change = (gains − losses) ÷ eligible pairs, in percentage points.
- Paired score interval = Newcombe method 10: Wilson score intervals for both marginal proportions, combined with the continuity-corrected within-pair correlation documented for correlated binary data.
- Equivalence rule = the complete
1 − 2αinterval must be strictly above the negative margin and strictly below the positive margin. Touching either bound fails strict inclusion.
What the result can and cannot support
A supported result means the score-based interval for this fixed binary endpoint, matched panel, conditions, and evidence window lies inside the supplied practical band. It does not mean the periods are equal, unchanged in every respect, interchangeable, permanently stable, adequately powered by design, commercially acceptable for every stakeholder, or caused by—or protected from—any intervention.
An inconclusive result means equivalence was not established. The interval may touch or cross a margin because the point change is large, the evidence is sparse, the panel is variable, or the interval remains broad. It is not automatically evidence of a meaningful difference.
Exploratory or multiple-endpoint use receives the descriptive interval but no unadjusted equivalence decision. This calculator does not select a multiplicity correction, choose a margin from observed outcomes, compute sample size, or certify a measurement program.
Method sources
- NCSS/PASS: Confidence Intervals for the Difference Between Two Correlated Proportions reproduces Newcombe method 10, the marginal Wilson construction, continuity-corrected correlation, and a published numeric validation example.
- NCSS: Two Correlated Proportions—Non-Inferiority, Superiority, and Equivalence Tests documents the paired-proportion setting and equivalence framing.
- NIST: McNemar Test defines the matched binary transition table and independence of paired analysis units.
These sources support the statistical construction. They do not choose a defensible margin, validate a supplied prompt panel, confirm matching or independence, or replace statistical review for a consequential decision.
Frequently asked questions
How do you test whether matched AI visibility results are equivalent?
Choose a negligible-change margin before reviewing outcomes, preserve the same binary prompt-and-repetition units in both periods, calculate a paired-proportion confidence interval, and support equivalence only when the complete interval lies strictly inside both margins and every design gate is satisfied.
Does a non-significant before-and-after test prove no meaningful change?
No. Failing to detect a difference is not evidence that the plausible difference is small enough. Equivalence requires a prespecified margin and an interval that fits completely inside it.
What if the interval touches one margin?
The result is inconclusive under the strict rule. Boundary contact is not rounded into equivalence, and it is not converted into proof of a meaningful difference.
Should missing or failed answers count as brand absence?
No. Missing, failed, unavailable, unreviewed, and ineligible observations remain excluded evidence. The memo discloses coverage and every supplied row disposition.
Does this calculator upload observation data?
No. Validation, calculation, memo generation, copying, and Markdown download happen locally in the browser.