How to Preserve Sample Size and Survey Wording in AI Research Notes

Summary

• Preserve sample size and survey wording by capturing the exact original text, the exact numbers, and where each came from.
• Separate four layers in your notes: original text, paraphrase, interpretation, and verification status.
• Use a reusable record format so every claim keeps its context (population, timeframe, question wording, response options, and denominator).
• Work from “smallest reliable unit” (one question + one result + one sample definition) to avoid accidental recombination.
• Include a worked example you can copy, plus a checklist to prevent common evidence-handling errors.

When you take AI research notes from surveys, two details get lost first: sample size (who and how many) and survey wording (the exact question and response options). That loss is not just cosmetic: it can change what a result means, make comparisons invalid, and create “clean” summaries that no longer match the underlying evidence.

This guide shows a practical way to preserve both sample size and wording in your notes, with a worked example and a reusable record format. The goal is simple: anyone (including future you) can trace a statement back to the exact question, the exact denominator, and the exact text without guessing.

Why sample size and wording are the “meaning” of a survey result

A survey statistic is not just a number. It is a number attached to:

• A population (who was surveyed)
• A timeframe (when the data was collected)
• A question (exact wording)
• Response options (what people could choose)
• A denominator (how many people were eligible to answer, and how many did)

If you drop any of those, you can accidentally change the claim. For example, “60% support X” can mean very different things depending on whether the question asked about “support,” “prefer,” “would accept,” or “would vote for,” and whether the denominator is all respondents, only those aware of the topic, or only those who answered that item.

The four-layer rule: Original text vs paraphrase vs interpretation vs verification

To preserve evidence cleanly, keep four layers separate in your notes. Do not blend them into one paragraph.

1) Original text (verbatim)

What it is: The exact survey question wording, response options, and the reported result text as written in the source.

Why it matters: This is the only layer that can be quoted or used to resolve disputes about what was asked.

2) Paraphrase (faithful restatement)

What it is: A plain-language restatement that keeps the same meaning, scope, and denominator.

Rule: If your paraphrase changes any of these, it is no longer a paraphrase; it becomes interpretation.

3) Interpretation (your analysis)

What it is: What you think the result implies, how it compares to other results, or what it suggests for a decision.

Rule: Interpretation must point back to the specific record(s) it is based on.

4) Verification status (what you checked)

What it is: A short status label describing whether you have confirmed the wording, denominator, and sample details directly from the source.

Why it matters: It prevents “draft notes” from quietly becoming “final facts.”

What to capture every time (minimum viable survey evidence)

For each survey finding you want to use later, capture the following fields. If you cannot capture a field, leave it blank and mark verification accordingly.

• Source pointer: URL, document name, page number, table/figure ID, or screenshot reference (whatever uniquely locates the evidence).
• Field dates: When the survey was conducted (or publication date if field dates are not provided).
• Population: Who was surveyed (country/region, age range, eligibility criteria).
• Sampling method (if stated): Panel, random sample, intercept, etc. (only if explicitly provided).
• Sample size (N): The base size for the overall survey and the base size for the specific question if different.
• Question ID: Q number or label if available.
• Exact question wording: Verbatim.
• Response options: Verbatim list, including “Other,” “Don’t know,” “Prefer not to say,” and whether multiple selections were allowed.
• Result values: Percentages and/or counts, including rounding notes if stated.
• Denominator definition: “All respondents,” “those aware,” “those who answered,” etc. (only what the source states).
• Notes on transformations: Any calculation you performed (e.g., combining categories) clearly labeled as your work.

A worked example (with a reusable record format)

The example below uses placeholder survey text to demonstrate the method. Replace it with your real source content. The point is the structure: it keeps wording and sample size attached to the claim.

Reusable record format: “Survey Evidence Card”

How this prevents common errors (using the example)

• Error: “60% are concerned” without a base.
  Fix: The paraphrase includes “Among respondents who answered Q3 (N=1,050).”
• Error: Dropping “Don’t know” and changing meaning.
  Fix: The response options are preserved verbatim, and the interpretation notes the presence of “Don’t know.”
• Error: Treating a combined category as if it was reported.
  Fix: The combined 60% is explicitly labeled as your calculation, while the original table values remain intact.

How to write AI research notes without “laundering” the evidence

If you use AI tools to help summarize, draft, or organize notes, the risk is that the output becomes a smooth narrative that hides denominators and wording. You can reduce that risk by forcing the workflow to keep the four layers separate.

Practical workflow: evidence-first, narrative-second

1. Create the Survey Evidence Card first (verbatim question, options, N, and results).
2. Only then generate a paraphrase that must include the denominator and population.
3. Write interpretation last and require it to reference the Record ID(s) it depends on.
4. Keep a verification line so “unknown” stays visible instead of being silently filled in.

Prompt pattern you can reuse (without inventing sources)

You can use a structured instruction like this when asking an AI assistant to help you rewrite your own notes. Only paste text you already have rights to use and that you have already extracted from the source.

• Input: “Here is the verbatim question wording, response options, base size, and table values. Produce (1) a faithful paraphrase that keeps the same denominator and (2) a separate interpretation section. Do not add any new numbers, dates, or claims not present in the input.”
• Output check: Confirm the paraphrase includes N/base and does not replace the question’s key verbs (e.g., “support” vs “prefer”).

Verification statuses that stay honest (and useful)

You do not need a complicated system. You need one that prevents accidental overconfidence. Here is a simple set you can use in your notes:

• Verified: You have directly checked the wording, response options, and the relevant base size in the source.
• Partially verified: You verified some elements (e.g., wording) but not others (e.g., base size or field dates).
• Unverified: You have a claim in your notes but have not checked it against the source yet.
• Not found: You looked for a detail (e.g., field dates) and could not locate it in the source you have.

Use these as labels in your record, not as commentary. The point is to keep your downstream writing aligned with what you actually confirmed.

Common pitfalls (and how to prevent them)

Pitfall 1: Mixing multiple questions into one “finding”

If you combine results from different questions, you can create a statement that no single question supports. Keep one record per question-result pair, then create a separate synthesis note that references multiple Record IDs.

Pitfall 2: Losing the question base when subgroups appear

Subgroup tables (age, region, role) may have different base sizes. If the source provides subgroup Ns, capture them. If it does not, do not infer them.

Pitfall 3: Turning “share who selected option” into “share who believe X”

Response options can be descriptive (“I have done X”), evaluative (“X is good”), predictive (“I will do X”), or normative (“X should happen”). Preserve the verb. Your paraphrase should not swap “would consider” into “will,” or “somewhat agree” into “agree.”

Pitfall 4: Rounding and totals that do not sum to 100%

Tables may not sum to 100% due to rounding or multi-select questions. Note what the source says about rounding or selection rules. If it says nothing, record the mismatch as an observation rather than “fixing” it.

Pitfall 5: Copying a secondary summary instead of the primary question

Press releases and blog posts may summarize a survey without giving full wording or base sizes. If you must use a secondary summary, mark it as such and treat the verification status as incomplete until you locate the primary instrument or table.

A compact template you can paste into your notes app

Copy and reuse this template for each survey item you want to preserve. Keep the “Original text” section verbatim.

Record ID:
Source pointer:
Field dates:
Population:
Sampling method (if stated):
Sample size (overall N):
Sample size (question base N):
Question ID/label:

ORIGINAL TEXT (verbatim)
Question:
Response options:
Selection rule (single/multi-select, if stated):
Reported result (table values / statement):

PARAPHRASE (faithful; keep denominator)
-

INTERPRETATION (your analysis; reference Record ID)
-

TRANSFORMATIONS (your math; label as your calculation)
-

VERIFICATION STATUS
- Wording:
- Base size:
- Field dates:
- Notes:

Frequently Asked Questions

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Back to FAQ Table of Contents

Related Guides

• What Is an AI Context Pack?
• How to Prepare Better Context for ChatGPT
• How to Organize Copied Text for AI Work
• Source-Labeled Context for AI: Why It Matters
• ChatGPT Context Pack for Consultants