A hypothetical quarter, shown four ways.
The example assumes a panel of 1,200 prompts — 300 per engine, spanning definitional, comparative and recommendation intents. These are illustrative inputs, not a completed Q3 or Q2 run. Citation rate measures how often a tracked brand's own domain appears as a cited source inside the answer.
| Engine | Sample rate | Sample change | Hypothesis to test |
|---|---|---|---|
| Perplexity | 13.6% | ▲2.1 | Whether recently updated pages differ in citation rate |
| ChatGPT | 12.8% | ▲0.9 | Whether brand-search demand relates to citations |
| Claude | 11.2% | ▲1.9 | Whether dated, sourced claims differ in quotation rate |
| Gemini | 8.1% | ▼0.4 | Whether structured data relates to citation rate |
A fictional follow-up could assign 58% of citations to fresher pages in a sample comparison. That number is invented for this layout. No matched-pair prompts were run, and the example establishes neither a freshness preference nor a change in Perplexity's behavior.
What a narrower source pool might look like.
Another fictional scenario could show 3.1 unique cited domains per answer in one period and 2.6 in another. A sample table might also show a two-to-one citation ratio for pages with structured data versus pages without it, or a higher quotation rate for dated claims. These are invented illustrations, not observations about ChatGPT, Gemini, Claude or any other engine.
Sample question: are citations spread across more sources or concentrated in fewer?
The invented four-point change in top-decile share demonstrates one way to display concentration. A real interpretation would require a defined brand set, comparable readings and preserved source responses. This sample does not establish a market shift, identify winning brands or show that a content change improves visibility.
Four characteristics a study could examine.
A future study could examine definitional pages, dated statistics in quotable passages, corroboration on other sites and differences across engines. These topics connect to the proposed method. They are candidate questions, not characteristics observed among a measured top decile or a demonstrated ranking of what improves visibility.
How to read these numbers.
The proposed panel would preserve prompts, categories and collection windows across comparison periods. Actual sampling and scoring rules must be specified and validated before a collected benchmark is published.
Illustrative setup: 1,200 prompts · 4 engines · example period JUL 1–14, 2026. No collection took place for this article. All figures are hand-authored fictional sample values used to demonstrate the Index format. They do not become evidence when a monitoring integration is added; a real report would require newly collected data.
- Numerical exhibits: hand-authored sample values. No Q3 or Q2 panel collection underlies them.
- Engine interpretations: hypothetical scenarios and untested questions; no external study is offered as evidence for these sample claims.
- Future measured report: would need preserved prompts, dated engine responses and sources supporting each interpretation.