Block what you can, rotate the engine order, freeze the plan
This tool uses a balanced Latin-square sequence family for engine order, then seed-shuffles engine labels, prompt-repetition blocks, and the sequence rows used in each cycle. A complete cycle has n blocks for an even engine count and 2n blocks for an odd count.
Generate a frozen prompt-by-engine run order
Use one line per exact prompt and one line per engine or surface. The seed is part of the plan identity; the same normalized inputs and seed reproduce the same rows and digest.
What the counterbalancing algorithm does
For n engines, the first sequence alternates from the start and end of the engine list; each next sequence rotates the labels. With an even n, the n sequences balance each engine once in every position and each distinct immediate predecessor once. With an odd n, reversed companion sequences create a 2n-row cycle in which those counts occur twice.
The page first seed-shuffles the supplied engine labels. It separately seed-shuffles prompt-repetition blocks and the sequence-row order inside each cycle. It never splits a prompt block across collection windows. A final incomplete cycle is retained and measured; the page reports its actual maximum position-count and predecessor-count spreads.
Keep failures visible and changes outside the frozen plan
- Download the as-run CSV before collection. Its planned columns are populated; timestamp, status, evidence reference, and note fields are blank.
- Execute the exact row order. Open a fresh session where the protocol requires one, preserve the exact prompt, and capture the named engine or surface.
- Record the observed state, including failure. A refused, failed, unavailable, or skipped row remains part of the planned denominator record; it does not become confirmed absence.
- Do not reshuffle after a result. If a recovery run is justified, add it to a separate deviation or recovery log with its own identity. Do not overwrite the planned row.
- Analyze with the correct unit. Counterbalancing controls planned order only. Choose uncertainty, missing-data, repeated-measure, and multiplicity methods separately for the eventual endpoint and decision.
Method sources and boundary
- NIST/SEMATECH: randomized block designs — block important controllable nuisance factors, then randomize remaining trial order.
- NIST/SEMATECH: Latin square and related designs — balance a treatment factor across two blocking dimensions and randomize as the design allows.
- Bradley (1958), Complete Counterbalancing of Immediate Sequential Effects in a Latin Square Design — the sequence construction used for the complete even-count cycle.
- IAB: Measuring Visibility in the AI Era — current AI-visibility stability, reproducibility, vocabulary, and disclosure context.
These sources support experimental-design structure and transparent measurement. They do not establish that an AI-visibility schedule is representative, adequately powered, unbiased, causal, or decision-grade for a particular business question.
Frequently asked questions
How should I order prompt runs across AI search engines?
Keep every engine inside each prompt-repetition block, rotate the engine order across blocks, and freeze that full plan before collecting or scoring outcomes.
Why not run one engine all the way through before the next?
That assigns early collection time, operator attention, and any short-run drift to the first engine and later conditions to the last. A counterbalanced block plan distributes planned positions instead of building that confound into the schedule.
Why do odd engine counts use twice as many sequence rows?
A single immediate-predecessor-balanced Latin square requires an even condition count. Reversed companion sequences create a complete 2n-row cycle for odd counts.
Can I use a plan with an incomplete cycle?
Yes, but call it incomplete. The tool reports the exact position and predecessor spreads. It does not fabricate extra observations or label a partial cycle perfectly balanced.
Does the seed make AI answers reproducible?
No. It reproduces this page's planned order. Model sampling, retrieval, personalization, platform changes, and time can still change the answers.
Does this generator call AI models or upload prompts?
No. All scheduling, validation, hashing, copying, and export work happens locally in the browser.