ChatGPT vs Perplexity vs Gemini: Citation Overlap Study Protocol
A reproducible protocol for measuring citation overlap across ChatGPT, Perplexity and Gemini without mistaking one-off outputs for evidence.
ChatGPT, Perplexity and Gemini can return different sources for the same buyer question. A credible citation-overlap study must use matched prompts, repeated runs, canonicalized URLs and a published source taxonomy. This page defines that protocol and explains what current research establishes; it does not invent a TotalAuthority overlap percentage before collection.
Why Citation Overlap Matters
Overlap answers a practical portfolio question: does one source strategy create visibility across multiple engines, or does each platform require separate monitoring?
Low domain overlap could mean platforms retrieve from different indexes, interpret the task differently, use different product modes or simply vary between runs. High overlap could indicate genuinely shared authoritative sources—or repeated syndication of the same underlying material. The metric alone cannot explain the cause.
The Research Questions
The study should answer:
- How often do the three surfaces cite the same domains for matched prompts?
- How often do they cite the same canonical pages?
- Which source categories recur across engines?
- Does overlap change by buyer stage, industry or query freshness?
- How stable is overlap across repeated runs?
- When a page is cited, does its evidence appear to be absorbed into the response?
Define the Product Surfaces
Record the exact consumer products and modes. “Gemini” can refer to an assistant, model family or Google Search experience. “ChatGPT” may or may not use web search. Perplexity offers different modes and model options. Do not compare an API model without retrieval to a consumer search product and call the result an engine study.
For every observation, store:
- platform and product surface;
- model or mode where visible;
- date and local time;
- country and language;
- account and personalization state;
- conversation state;
- whether search was triggered;
- exact answer and source list.
Build a Matched Prompt Set
Use identical core tasks across platforms. A balanced 120-prompt design might include:
| Prompt class | Count | Example task |
|---|---|---|
| Problem education | 20 | Explain causes, options or risks |
| Solution exploration | 20 | Compare methods or approaches |
| Provider discovery | 20 | Identify suitable providers by criteria |
| Product comparison | 20 | Compare named alternatives |
| Validation | 20 | Check suitability, evidence or limitations |
| Fresh/current | 20 | Answer a time-sensitive factual question |
Publish the prompts or generation template. Avoid silently changing wording between platforms. If a platform requires a different interaction, document the adaptation.
Repeat the Runs
One answer per prompt cannot distinguish engine differences from random variation. Start with at least five independent runs per prompt and increase repetition for unstable categories. Reset conversation context between independent runs.
With 120 prompts, three platforms and five runs, the design produces 1,800 answer observations before exclusions. Report the achieved sample and failure rate.
Normalize Citations
Raw URLs require cleaning before overlap analysis:
- Remove tracking parameters.
- Resolve redirects.
- normalize scheme, host case and trailing slash policy.
- identify declared and search-selected canonicals where feasible.
- distinguish syndication from the original source.
- preserve both the displayed URL and canonicalized analysis URL.
This prevents mobile URLs, fragments and campaign parameters from appearing as separate sources. The canonical URL guide covers the technical edge cases.
Classify the Sources
Use a mutually exclusive primary taxonomy:
- owned brand or vendor content;
- independent editorial media;
- professional association or regulator;
- academic or primary research;
- government or public body;
- review or marketplace platform;
- community or social content;
- retailer or distributor;
- reference/database;
- other or unclassifiable.
Add secondary flags for first-party data, named expert, current date, paywall, structured comparison and commercial relationship. Double-code a sample and report reviewer agreement.
Calculate Overlap
Pairwise Jaccard overlap
For source sets A and B:
J(A,B) = sources in both sets / sources in either set
Calculate it at domain and canonical-page level for each matched prompt-run group. Report the distribution, not only the mean.
Shared-source rate
observations with at least one source shared across platforms / matched observations
Unique-source contribution
Measure what proportion of each platform's cited domains did not appear on the other two platforms in the same matched group.
Rank-weighted overlap
If source order is meaningful and consistently available, use a documented rank-weighted measure. Do not assume source order equals influence.
Add Citation Absorption
The 2026 geo-citation-lab research distinguishes source selection from answer absorption. A follow-on sample can compare answer claims with cited-page evidence and classify whether a source supplied a definition, fact, comparison, procedure or other support.
This is labor-intensive and should use blinded human review on a sample. A link in the source panel is not proof that every sentence derives from that page.
Predefine Exclusions
Exclude only according to published rules:
- platform error or unavailable answer;
- no search in a prompt class requiring current retrieval;
- wrong language;
- refusal unrelated to the intended task;
- source panel not capturable;
- duplicate answer caused by collection error.
Preserve valid surprising results. Do not remove outliers simply because they weaken the story.
How Results Should Be Reported
The completed report should contain:
- collection dates and surfaces;
- valid observations and exclusions;
- domain and page overlap distributions;
- source-category shares;
- overlap by prompt class;
- run-to-run stability;
- confidence intervals;
- worked examples;
- negative and mixed results;
- privacy-safe dataset and code.
No proprietary findings appear here yet. Current published research already shows that citation breadth and absorption patterns differ across the studied platforms, but it does not provide a permanent universal overlap rate for every market.
What Brands Can Do Before the Study Runs
Use a multi-source strategy. Maintain accurate owned pages, earn relevant third-party corroboration, preserve primary data, and monitor each important platform separately. Do not infer that a Google citation guarantees visibility in ChatGPT or Perplexity.
The earned media versus owned content framework explains how to build and measure that portfolio.
Frequently Asked Questions
Why compare both domains and pages?
Domain overlap can hide important differences. Two platforms may cite the same publisher but different articles, dates or authors.
Should social and community citations be included?
Yes when they are visible sources. Classify them separately and preserve context rather than excluding them because they are not traditional publications.
Can overlap reveal an engine's ranking factors?
No. It describes observed source sets under defined conditions. It cannot expose the full retrieval pipeline or prove why a source was selected.
Will this page publish findings later?
Yes, only after the prompt set, raw evidence, analysis rules and reproducible dataset exist. The protocol will remain visible alongside results.
