Build Practical AI Systems
How to Build a Source Register for AI Marketing Research
Build a claim-level source register for AI-assisted marketing research, then manually check the source, passage, date, limitation and status before drafting.
A source register is a working table that connects each factual claim you may reuse to the original source, its date, the supporting passage, important limitations and a human checking status.
Build it one claim-to-source relationship per row. Then open the original source and confirm that it supports the claim as written before the claim enters a brief, article or campaign draft.
What is a source register?
A source register is a reusable record of what you may say, where it came from and what a person found when checking it. The unit of review is the claim: one page may support several claims with different limitations, while one claim may need several sources.
This practical marketing term is not a universal research standard. The same document may be called an evidence log, source log or claim-to-source table.
Why a bibliography is not enough
A bibliography lists sources. It may not show which source supports each claim, whether the wording exceeds the evidence or whether anyone opened the cited page.
OpenAI advises checking important information, quotations, data and references against reliable sources. A citation gives you somewhere to look; it does not complete the check. Read OpenAI's accuracy guidance.
Use eight practical fields
The following fields are a starter recommendation, not a prescribed standard:
- Claim ID: A stable label such as
C-01. - Proposed claim: The factual wording you may reuse.
- Source details: Publisher, title and direct URL or authorised file.
- Source date: Publication or meaningful update date, or
not stated. - Checked date: When a person reviewed the source.
- Supporting passage: A short extract, heading, page or paragraph.
- Limitation: What the source does not establish.
- Status and owner: The result and next-action owner.
The source date helps judge freshness. The checked date shows when somebody confirmed that it supported the recorded wording.
Give every row a clear status
Use a small set consistently:
- TO CHECK: No person has verified the row.
- SUPPORTED: The source supports the claim as written, within the recorded limitation.
- PARTIAL: Only part is supported. Narrow the claim or add evidence.
- NOT SUPPORTED: The material does not support the claim. Remove or replace it.
- RECHECK: The source, claim or context has changed since the last review.
SUPPORTED is dated and scoped. It does not mean universally true, approved for every channel or safe to reuse forever.
Build the register in six steps
1. Start with one research question
Define the question before collecting claims so irrelevant sources are easier to spot.
2. Turn findings into individual claims
Write one factual statement per ID. Keep interpretation and recommendation outside the row. Split wording that contains two testable ideas.
How to Use AI to Organise B2B Audience Research Without Losing Human Judgement explains the separate step of distinguishing evidence from interpretation and a content decision.
3. Record the original source
Use the page, report, dataset or authorised note containing the evidence. An AI answer or search snippet that merely points elsewhere is not the original source.
4. Capture enough context to test the wording
Record a short passage or precise location, then read around it. A matching phrase may concern a different audience, period, geography or condition. Check quotation and reuse requirements.
5. Write the limitation before the status
Ask what the source does not prove. It may describe one product or sample, present a provider's claim or lack a date. The answer may change SUPPORTED to PARTIAL.
6. Open the source and decide
The reviewer compares the wording with the source in context and assigns the status. If a claim needs two sources, use two rows with the same ID. Do not merge disagreement into one vague reference.
A completed three-claim example
The following register uses one real public source: Google's people-first content guidance, updated 10 December 2025 and manually checked on 9 September 2026.
| Claim and source | Dates | Supporting location | Limitation | Status |
|---|---|---|---|---|
C-01: Google advises creators to ask whether content has an intended audience that would find it useful. Source: Google Search Central. | Updated 10 Dec 2025. Checked 9 Sep 2026. | Focus on people-first content; “existing or intended audience”. | A self-assessment prompt, not evidence of demand or a ranking guarantee. | SUPPORTED |
C-02: Google says clear sourcing can help content appear trustworthy. Source: Google Search Central. | Updated 10 Dec 2025. Checked 9 Sep 2026. | Expertise questions; “clear sourcing”. | Does not prove that the source or resulting claim is correct. | SUPPORTED |
C-03: Google says it has no preferred word count. Source: Google Search Central. | Updated 10 Dec 2025. Checked 9 Sep 2026. | Avoid creating search engine-first content; preferred-word-count question and answer. | Does not mean length is irrelevant to a reader or that every short article is sufficient. | SUPPORTED |
Each claim was narrowed to what the page actually supports. The example does not show that these practices guarantee rankings, traffic or inclusion in an AI-generated answer.
Where AI can help
AI can extract candidate claims from authorised material, format rows and flag empty fields. Research tools may also return source links. OpenAI says its current Deep Research reports include citations or links so users can verify information. See the current product guidance.
A person must still confirm the source and date, read the context and assess the precise claim. Use an appropriate specialist for technical or consequential subjects.
A source can be traceable and still be weak
A complete row records a relationship. It does not establish that the source is accurate, independent, representative or suitable.
The House of Commons Library summarises source-evaluation questions around accuracy, authority, relevance, transparency and purpose. Read its source-evaluation guide. Apply them in proportion to the claim.
Do not put personal data, confidential notes, restricted material or licensed research into an AI tool or shared register unless the tool, access and purpose are authorised.
Common source-register mistakes
- Saving only the homepage: Link to the exact page or file containing the evidence.
- Writing a topic instead of a claim: Record the sentence you may actually reuse.
- Using an AI summary as evidence: Trace it to the original source.
- Skipping the limitation: State what the source does not establish.
- Treating checked as permanent: Recheck changing sources before publication or reuse.
- Overwriting failed rows: Keep
PARTIALorNOT SUPPORTEDvisible so the same claim is not reintroduced later.
Your next step: check three claims
Choose one public source and write three factual claims. Create columns for claim, source, dates, supporting passage, limitation and status.
Open the source and check every row. Narrow partially supported wording and record the change. Only move SUPPORTED claims into the next draft.
Further reading
You Might Still Be Wondering...
Frequently asked questions
No. A bibliography lists sources. A source register also connects a proposed claim to a specific passage, date, limitation and review status.
AI can suggest claims, format rows and identify missing fields. A person should still open each original source, inspect the context and assign the checking status.
Use two rows with the same claim ID and a different source in each. This preserves what each source contributes and where their limitations differ.
Write not stated rather than guessing. Record your checked date separately, consider whether the missing date affects suitability and find a more current source when the claim is time-sensitive.
It means the recorded source supports the wording within the stated limitation at the time of review. You must still judge the source's quality, independence, relevance and suitability for the intended use.