Base2026
Partner with Base2026
Menu

Build journal 002Alex Yarosh5 min read

Three Sources Are Not Always Three Independent Voices

A small, reproducible check before turning search results into advice.

Blue-and-white evidence card showing video, excerpt and source-link labels.
AI-generated editorial illustration, not a screenshot or real-source quote.

AI-assisted research and writing. This example was checked against Base2026’s public API on 30 August 2026. I’m building Base2026, so this is a test of my own project, not an independent product review.

A search for “AI citation tracking” returned 12 matching records and selected three sources. Those three sources came from two creators. That difference is easy to miss when the output looks like a tidy research brief.

It does not make the results useless. It changes what they can support. Three videos are three records, but two videos from the same person do not establish three independent opinions. Even different creators can repeat the same underlying claim.

Here is a check you can repeat without an account, a paid plan or an API key.

01 · Exact request

Start with the exact request

Open this public, read-only request:

https://base2026.dev/api/evidence-brief/v2?q=AI%20citation%20tracking

In the 30 August readback, its coverage object reported:

matched_records: 12
selected_sources: 3
distinct_creators: 2
published_date_min: 2025-11-28
published_date_max: 2026-07-14

These are observations about one query against one corpus. They are not the number of people discussing the topic online, a representative survey, or a benchmark of how any particular AI answer engine chooses citations. A later request can change as the index changes.

The response also says status: full. That is the API’s coverage classification. It does not mean “fact-checked,” “complete coverage,” or “safe to copy into a client recommendation.”

02 · Inspect claims

Count people, then inspect claims

The findings include creator_handle, published_date, base2026_url, original_source_url and evidence_excerpt. Keep those fields together. A note without its source link is hard to challenge; a source link without the specific claim is hard to review.

First, group the results by creator. Then ask whether the records actually address your question. In this example, a returned source also covers broader SEO and traffic themes. A search match can be adjacent to the topic without directly supporting a recommendation about citation tracking.

That is a useful limitation to notice before you summarize. Do not smooth an adjacent result into agreement just because the interface selected it.

Next, follow the original source link and check the context. Is the speaker reporting a measured result, proposing a tactic, promoting a service, or describing an opinion? A strong number in a source title is still that source’s claim until you have checked its method and evidence.

03 · Record the observation

Keep the observation separate from the recommendation

My suggested research note has six fields:

  1. The exact question and request URL.
  2. The readback date and corpus version.
  3. The claim being considered.
  4. The creator, source date and original link.
  5. What the excerpt supports—and what it leaves unanswered.
  6. The decision: use, investigate further, or exclude.

“Two creators discuss this tactic” is an observation. “This tactic will work for our business” is a recommendation. Moving from one to the other requires more context: the market, the prompt set, the time window and an explanation of what success would look like.

This also helps when records disagree. Keep the disagreement attached to the source instead of averaging it into a confident sentence. If every record repeats one person’s assertion, look for a different source or say that independent support is missing.

04 · Review the brief

Make the brief easy to challenge

Base2026 returns bounded public excerpts and links back to the original material. The example does not need a full transcript dump, and it does not imply that the creators endorse the project. It gives a reader a route back to the evidence—and enough metadata to notice gaps.

The same check works with a spreadsheet or a manually assembled reading list. Count records and creators separately. Check relevance before agreement. Save the exact observation before writing the recommendation.

Which check has caught more weak research in your work: repeated sources, missing context, or a claim the cited page does not actually support?