Study Design
Does AI cite the same sources when you repeat a test?
This study design uses the same three dietary-supplement questions to compare websites cited by two models in same-day repeats and tests 17 days apart. This page explains the schedule, citation-counting rules and how to compare website lists, rankings and citation volume separately. The published observations are presented in a separate comparison article.
The study examines citations visible in AI answers, not product efficacy, brand recommendations or revenue outcomes.
Read the separate article for the published citation-source comparisons.
Read the AI citation comparisonsHow was this study set up?
- Question set
- Three questions in one dietary supplement topic; original questions and brand data are not disclosed.
- Test dates
- 2026-08-10 and 2026-08-27; three test batches.
- Tested models
- Export version labels: gpt-5.6-sol and gemini-3.6-flash.
- Units of comparison
- For each model, count which websites were cited and how many citation records there were. Keep rankings pooled across all three questions separate from rankings for each individual question.
First test date
One batch provides the starting point for comparisons across dates.
Same-day batch A
Compare with August 10 separately for each model.
Same-day batch B
Compare with batch A to examine repeat-run differences, and also compare with August 10.
What questions does the study compare separately?
Same-day repeats
In a same-day repeat, which of the ten most frequently cited websites appear again?
Across test dates
After 17 days, which websites remain on the list, and which have entered or left?
Between models
For the same questions, do the two models cite the same websites? How do citation counts for those websites differ?
Do matching source lists mean matching answers?
No. Even if both top-10 lists contain the same websites, the ranks, citation counts and answer content can differ. First count the shared websites, then examine rankings, citation volume and answers separately.
- How many websites do the two lists share?
- Websites shared by two top-10 lists
This asks whether the same websites keep appearing. It does not compare their ranks or the content of the answers.
- Citation share of leading sources
- Citations to the top 10 websites ÷ all citations for that model
This asks whether citations are concentrated on a few websites. The total used in the calculation is all citation records—not the number of answers or brand mentions.
This shows the comparison method, not study results. Apply the same clearly stated counting rules to each model, then interpret each model’s results separately.
Why compare models separately?
Separate lists show which websites each model cites often and how concentrated those citations are. One combined ranking can hide these differences. Retain each model’s list and denominator before comparing shared sources. Citation concentration does not tell us which model is more accurate.
How can a PR team use this research?
It can help the team identify what needs a closer look. Open the cited pages and check whether they support what the AI said. Then decide, in light of the brand’s questions, whether to add information, retest or keep observing.
Keep the research record
Read the answers and citations alongside the questions, models and test conditions.
Check the original pages
People check whether the source supports the claim and where it applies.
Choose the next step
The team chooses what to improve, retest or monitor, rather than acting on rankings alone.
How are citation records counted?
Within one answer, a repeated URL counts once. Before counting, remove URL fragments and trailing slashes, then remove duplicates. Two different pages on the same website still count as two citations. Website rankings follow the study’s stated domain-grouping rules.
These are this study’s counting rules; other GEO tools may differ. Before comparing reports, check how each removes duplicate URLs, groups URLs into websites and defines valid answers. Figures with different denominators need interpretation before they can be compared.
Understand mentions, citations and outcomesHow should you read this study?
The study covers three questions within one dietary-supplement topic, two models and two test dates. Read the questions, models and dates together when comparing records. Other industries, AI interfaces and long-term trends need their own evidence. A change in a citation list alone cannot establish whether a model update, search index or content edit caused it.
Citation rankings can help identify sources worth checking further. Content accuracy, media-buying choices and brand outcomes require additional evidence and professional judgment.
What are the sources, and what has been published?
This guide draws on three August 2026 test exports, the source-stability research article and its supplement. It explains the test design, metric definitions and interpretation. To protect the research data, original questions, brand data and citation-level records are not available for public download.
Published by REALI. This page sets out the study design, counting rules and the scope of the publicly available information.
Read our research publication and correction policy