Industry Insights · Preliminary test observations
Do AI models keep citing the same websites?
In the available test records, the two models often cited different websites. Their top-10 lists were closer in same-day repeats, with more changes across a 17-day interval. This article explains how to compare those patterns by model and date, and how brand teams can use them to guide source checking.
Data scope: this analysis uses the available records from three batches—47, 45 and 48 rounds—rather than the planned 50 rounds in each. The reason for missing records is unknown. Differences in the available samples may affect comparisons, and complete records could change the figures.
Do the two models cite the same websites for the same questions?
The two top-10 lists shared just one website. This comes from the 27 August batch A records: for each model, we combined citations across the three questions and identified its ten most frequently cited websites. Combining both models in one ranking could hide these differences. Keep their records separate so the team can choose which sources to examine.
For each model, we combine citation records from the same three questions and select the ten most frequently cited websites. The figure counts websites appearing in both lists. It does not measure matching ranks, matching answers or factual accuracy.
Source: available records from three REALI test exports from August 2026, recalculated on 2026-10-04.
Model names follow the August 2026 test-export labels. This article compares cited websites in those records; it does not represent current consumer ChatGPT or Gemini performance. Detailed conditions are given in the methods below.
This is a result for the three questions combined. It does not mean that each question has only one shared website, or that other model versions and industries would show the same pattern.
What differs between a same-day repeat and a test 17 days later?
These comparisons answer different questions. A same-day repeat looks at differences over a short interval; a test 17 days later looks at changes between two dates. Both comparisons below use the top-10 lists pooled across three questions. Matching lists can still have different ranks, citation counts and answers. Individual questions need a separate look.
First batch
47 rounds available
Same-day batch A
45 rounds available
Same-day batch B
48 rounds available
For each model, we combine citation records from the same three questions and select the ten most frequently cited websites. The figure counts websites appearing in both lists. It does not measure matching ranks, matching answers or factual accuracy.
Source: available records from three REALI test exports from August 2026, recalculated on 2026-10-04.
*In Gemini’s 8/27-B records, two websites tie at 56 citations at the top-10 cutoff. Alphabetical domain order gives 10/10 shared websites; choosing the other tied site gives 9/10. This compares list membership, not matching ranks or answers.
For each model, we combine citation records from the same three questions and select the ten most frequently cited websites. The figure counts websites appearing in both lists. It does not measure matching ranks, matching answers or factual accuracy.
Source: available records from three REALI test exports from August 2026, recalculated on 2026-10-04.
This compares one same-day repeat and records from two dates. A team can retain the study as a reference, then choose which questions deserve further checking or observation.
Turn source differences into a focused PR review
In the 27 August A batch, pooling three questions for each model produced top-10 website lists with just one website in common. A PR team can then ask: do those different sources also lead to different descriptions of the brand?
See the comparison chart and its data scope
01 · Identify claims worth checking
Read the original answers by model and question. Find statements about the brand’s identity, product scope or other important information, then open the cited pages.
02 · Decide what needs attention
Check dates, context and whether each source supports the answer. Different sources do not necessarily mean an answer is wrong, and frequent citation does not establish accuracy.
03 · Choose an action and observe changes
If information is outdated, incomplete or confused with another brand, decide whether to update your website, clarify a public description or request a source correction. Record the date and retest relevant questions under comparable conditions.
This is a way to interpret research and plan work. The source differences do not establish model accuracy or demonstrate PR results after an intervention.
See how GPS connects research with routine monitoringHow were these figures calculated?
Show counting rules and model conditions
This analysis uses 839 non-empty answers and 8,908 citation records from the available data, after duplicate URLs within each answer are removed. The records cover three questions within one dietary-supplement topic, pooled separately for each model. These counts describe the available records, not verified citation content.
URL counting: use the exported normalized URL where available, remove the fragment after # and the trailing slash, then remove duplicate URLs within each answer. Different pages on one website count separately. Website grouping: lowercase domain names, remove www. and retain the original study’s three domain-alias rules; keep other subdomains separate. Rank by citation count, using alphabetical domain order to break ties.
Changing the counting rule changes the result. If a website counts only once per answer, regardless of how many of its pages appear, the two models share 2 websites in their 8/27-A top-10 lists, and Gemini shares 8 across the 17-day comparison. Read the figures on this page using the citation-record rules given above.
The exports label the models gpt-5.6-sol and gemini-3.6-flash, shortened here to GPT and Gemini. We have not independently verified the API endpoints, full prompts, search settings or caching used at the time. These records therefore cannot represent the performance of today’s consumer ChatGPT or Gemini interfaces.
This is a descriptive comparison. No statistical significance test was performed, and the contents of the cited pages were not individually verified.
How do website coverage and citation-volume coverage differ?
- Website coverage
Website coverage: the share of distinct websites in the reference data found in fewer rounds.
- Citation-volume coverage
Citation-volume coverage: the share of all citations in the full reference data accounted for by those websites.
How to read these observations
This article describes available citation records for one dietary-supplement question set, two models and two test dates. The observations can help identify source differences worth checking or following. Other questions, industries and model versions may produce different results. Citation changes alone do not establish their cause, source accuracy, or brand and business outcomes.
What are the sources, and who publishes this article?
Source: available records from three REALI test exports from August 2026, recalculated on 2026-10-04.
Written and published by REALI. To protect the research data, brands, original questions, full answers and identifiable source lists are not public. This page provides the methods, data scope and content-update information.
Where can you start with your own brand?
Start with the question you want to answer about your brand. Then discuss the research scope, the materials to be delivered and how your team will use the findings. GPS (GEO Positioning System) is REALI’s workflow for organizing that brand research.