Methodology

Prompt Set Construction

How an ideal-customer profile becomes a statistically usable prompt corpus — stratification, phrasing discipline, localisation and versioning.

CiteRank AI Research·18 January 2026·11 min readMethod note

Method note. Method note. This entry documents how CiteRank measures. It reports no findings; any numbers shown are worked examples.

MethodologyPromptsCorpus Design

Key takeaways

—Keywords describe queries. Prompts describe situations. The translation is not mechanical.
—Unbranded prompts are where competitiveness is measured; branded prompts measure recall.
—Every corpus change is a version bump, not an edit.

From ICP to intent strata

Start with who is deciding, what they are deciding between, and what would disqualify you. Each of those becomes a stratum, and each stratum gets a target proportion of the corpus before a single prompt is written.

Phrasing discipline

Prompts should read the way a person types when they are not performing for a search engine: full sentences, constraints included, occasional imprecision.

›Include the constraint — team size, budget, region, regulation, timeline.
›Avoid stuffing the category term in unnaturally; engines are not matching strings.
›Write a mix of confident and uncertain phrasings; they retrieve differently.
›Keep one idea per prompt so the parse is unambiguous.

Localisation

Language, region and even device context change which brands are named. A corpus intended to guide a multi-market programme must be replicated per market, not translated once and reused.

Versioning and drift

Corpora age as categories change. The discipline is to add a new version rather than quietly edit the live one, and to report which version produced any given number. Without that, a trend line silently becomes fiction.

Evidence basis

How this entry was produced

Method note. This entry documents how CiteRank measures. It reports no findings; any numbers shown are worked examples.

Research type: Method note — documentation of measurement approach
Classification: Method note
Data collected: Published 2026-01-18. No data collection: this entry documents method rather than reporting a study.
Prompt sample size: Not applicable — no corpus was sampled for this entry.
Replays per prompt: Not applicable — no prompts were replayed for this entry.
Engines covered: Method applies to all eight supported engines: ChatGPT, Gemini, Claude, Perplexity, Google AI Overviews, Microsoft Copilot, Grok, Meta AI.
Engine versions: Not stated. Engine vendors do not expose a stable build identifier for every model, so a version cannot be claimed accurately.
Methodology: Written by the CiteRank research team from the production measurement pipeline. Internal editorial review only; no external or academic peer review.

Limitations

This entry reports no findings. Any number shown is a worked example chosen for clarity, not a measurement.
AI engines are non-deterministic: an identical prompt can return a different answer on replay, so any figure derived from them is an estimate, never a fixed value.
Engine vendors change retrieval and ranking behaviour without notice, so anything stated here describes the stated window only.

Methodology, limitations and disclosure

Every CiteRank study states who produced it, what it measured and where it stops being reliable. The full scoring model is documented on the methodology page.

Author: CiteRank AI Research
Author role: CiteRank AI Research team — measurement, prompt-corpus design and scoring
Review: Internal editorial review by the CiteRank AI Research team. No external or academic peer review was conducted.
Published: 18 Jan 2026
Last updated: 18 Jan 2026
AI engines: Not engine-specific
Model versions: Specific model build identifiers are not disclosed by every vendor and are therefore not claimed here.
Industry scope: Cross-industry
Geographic scope: Global
Language scope: English
Sample size: Not disclosed for this entry
Conflicts of interest: CiteRank AI publishes this research and sells an AI visibility platform. No third party funded, commissioned or reviewed this entry.
Data availability: Underlying raw data is not published. Method and scoring are documented on the Methodology page.

Limitations

AI engines are non-deterministic: an identical prompt can return a different answer on replay, so every figure is a sampled estimate rather than a fixed value.
Engine vendors change retrieval and ranking behaviour without notice. Findings describe the sampling window stated above, not a permanent state.
Results describe the prompt corpus that was designed for this study. A different corpus for the same brand can produce a materially different picture.
This entry does not disclose a sample size, so its figures should not be treated as statistically representative.

Corrections and revisions

No corrections have been issued for this entry since publication on 18 Jan 2026. If a figure or claim here is wrong, write to research@citerank.in. Substantive corrections are published inline with the date they were made, and the original wording is retained in the note.

Suggested citation

CiteRank AI Research (2026). Prompt Set Construction. CiteRank AI. https://www.citerank.in/research/prompt-set-construction

← Back to the AI Research Centre