---
title: "AI Citation Sources in 2026: Which Domains ChatGPT, Perplexity and Google AI Cite, and How Often They Rotate"
canonical: https://snezzi.com/blog/ai-citation-sources-2026/
source: https://snezzi.com/blog/ai-citation-sources-2026/
published: 2026-11-15
modified: 2026-11-19
author: "Team Snezzi"
category: "Industry Insights & Research"
---

> Canonical page: https://snezzi.com/blog/ai-citation-sources-2026/

# AI Citation Sources in 2026: Which Domains ChatGPT, Perplexity and Google AI Cite, and How Often They Rotate

AI platforms can rotate their displayed source sets daily or weekly, so one cited answer is a snapshot rather than a fixed ranking. In specific 2026 studies, ChatGPT Search showed 74% weekly domain churn-in and Google AI Mode 56%, while a separate daily study measured total source churn of 79.2% for ChatGPT, 75.9% for Google AI Mode, and 44.4% for Perplexity. [SISTRIX](https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/) [GetMentions](https://www.getmentions.ai/blog/ai-citation-volatility-study)

The leading domains also depend on the dataset. In Profound’s historical August 2024 to June 2025 dataset, Wikipedia was ChatGPT’s leading cited domain, while Reddit led both Google AI Overviews and Perplexity. These are observed dataset patterns, not a permanent cross-platform leaderboard.

The useful rule is simple: measure citations by engine, surface, prompt, location, and date before deciding that a domain has a durable position. A displayed citation is an observed link in one answer. It is not proof of a permanent preference, model training data, or a search ranking.

## A citation is not the same as training data or a permanent preference

When an AI answer displays a source, it does not reveal every page the system may have used or considered. Separate the process into three layers:

1. **Training data:** Material used to develop or improve a model. It is not a visible citation log.
2. **Search or grounding activity:** The process used when an engine looks for current information. The full candidate set is usually not visible.
3. **Displayed citations:** The source links shown with a specific answer. This is the observable layer that publishers can track.

OpenAI says ChatGPT responses using web search may include citations. Its Sources view can show cited sources and other relevant links. OpenAI also warns that search results and citations can be incomplete, outdated, or incorrect. [OpenAI’s ChatGPT Search documentation](https://help.openai.com/en/articles/9237897-searching-the-web-with-chatgpt)

The same prompt can produce different source sets because the search process is not always a fixed lookup. ChatGPT may search automatically when a question needs current information. It may also rewrite a prompt into targeted searches and use approximate location to improve local results.

Keep Google AI Overviews and Google AI Mode separate. Google says they may use different models and techniques, so their response links can vary. Both may use query fan-out, which means issuing related searches across subtopics and data sources. AI Overviews also do not appear for every query. [Google Search Central](https://developers.google.com/search/docs/appearance/ai-features?content_language=English)

## Which domains does each platform cite most?

The table below summarizes **historical platform patterns from Profound’s August 2024 to June 2025 dataset, not current universal rules**. [Profound’s citation dataset](https://www.tryprofound.com/blog/ai-platform-citation-patterns)

| Platform or surface | Leading domain in the dataset | Overall citation share | Share within that platform’s top 10 |
|---|---|---:|---:|
| ChatGPT | Wikipedia | 7.8% | 47.9% |
| Google AI Overviews | Reddit | 2.2% | 21.0% |
| Google AI Overviews | YouTube | 1.9% | 18.8% |
| Google AI Overviews | LinkedIn | 1.3% | 13.0% |
| Perplexity | Reddit | 6.6% | 46.7% |

The two percentages answer different questions:

- **Overall citation share** is the domain’s share of all citations recorded for that platform.
- **Top-10 share** is the domain’s proportion after limiting the calculation to the ten most-cited domains.

A domain can have a modest share of all citations but a large share among the top ten. These measures should not be compared as if they were the same metric.

In this dataset, Perplexity cited Reddit more often than Google AI Overviews, at 6.6% versus 2.2% of all citations. That comparison applies to Perplexity and Google AI Overviews in this dataset. It does not establish how Reddit performs in Google AI Mode, Gemini, or every future sample.

### Newer data shows different citation behavior by engine

Tonecast’s September and October 2026 API study also found sharp differences between engines. Perplexity returned citations in 100% of sampled answers. ChatGPT returned citations in 54.8%, and only when its web-search tool ran.

The sample included 1,730 answers, 81 questions, seven subjects, and specified API models. It did not cover every consumer interface. Tonecast’s results are therefore a directional comparison of those settings, not a universal rate for each product.

This helps explain why citation counts from different engines are difficult to compare. An engine that cites in almost every answer will naturally produce more observed links than one that often answers without web search.

## Which sources does AI cite for software recommendations?

Software recommendation prompts call for different evidence types. No universal list of sites should be treated as the target.

- **Vendor product, pricing, and documentation pages** can support capabilities, technical requirements, integrations, and current terms.
- **Independent reviews and category comparisons** can support alternatives and third-party evaluation.
- **Review and marketplace profiles** can show buyer feedback and category fit.
- **Community discussions** can provide practical use cases and user experience.
- **Reference sources** can support definitions and neutral factual context.

Tonecast’s multi-topic API sample illustrates this mix. Perplexity cited a subject’s own site in 35.6% of answers and a review platform in 17.1%. Gemini did so in 38.9% and 11.7%. ChatGPT’s equivalent figures were 13.5% and 0.7%. These percentages describe the share of sampled answers that cited each source family, not a software-only study.

The decision rule is to inspect the cited source mix for the exact category, comparison, pricing, and alternatives prompts your buyers ask. A documentation page, review profile, and community thread may each support a different part of the answer.

For more context, see this guide to [how AI chatbots pick sources](https://snezzi.com/blog/how-ai-chatbots-pick-sources-an-inside-look-marketers/). Video can also vary by engine and prompt type. The [YouTube AI search citations by engine](https://snezzi.com/blog/youtube-ai-search-citations-by-engine/) analysis explains why a video citation in one surface does not establish a broader YouTube preference.

## How often do AI citation sources rotate?

Define the measurement terms before using the figures:

- **Churn-in:** The share of domains in the current run that were absent in the prior run.
- **Retention:** The share of domains from an earlier run that remain later.
- **Total source churn:** Sources that disappeared plus sources that appeared. This is not the same as churn-in.

### Weekly changes in the SISTRIX study

SISTRIX analyzed 82,619 prompts across 17 weeks and six countries. It found:

- ChatGPT Search had 74% weekly churn-in.
- Google AI Mode had 56% weekly churn-in.
- For AI Overviews, 53% of prompts showed no source changes during the study.
- Another 19% of AI Overview prompts had 46% churn-in.
- ChatGPT showed 85% weekly drift at the URL level, compared with 74% at the domain level.

These results suggest that a response can contain a persistent domain while the specific page from that domain changes. That is why domain persistence and URL persistence should be measured separately. [SISTRIX’s citation drift study](https://www.sistrix.com/blog/ai-citation-drift-how-stable-are-sources-in-ai-search-results/)

### Daily changes in the GetMentions study

GetMentions tracked 2,398 B2B and consumer queries once daily for seven days in June 2026. It measured total daily source churn of:

- 79.2% for ChatGPT
- 75.9% for Google AI Mode
- 44.4% for Perplexity

Perplexity retained 66.8% of day-one sources on day two and 48.6% on day seven. The publisher describes the single seven-day window as directional.

The two studies use different definitions and collection windows. Their figures should not be placed in one ranking. They point in the same broad direction: source rotation is common, and its rate depends on the engine, surface, prompt, and test conditions. [GetMentions’ volatility study](https://www.getmentions.ai/blog/ai-citation-volatility-study)

Citation rotation does not establish a model-retraining schedule. It may reflect live search, query rewriting, query fan-out, changing web results, location, interface behavior, or answer generation. The observed citation change is real within the study. Its hidden cause is not necessarily identifiable from the displayed links.

## Compare citation rates within one engine first

A 2026 preprint based on 602 controlled prompts reported mean citations per prompt of 6.88 for ChatGPT, 12.06 for Google, and 16.35 for Perplexity. The study covered specific platforms, prompts, and collection conditions. These numbers are useful for comparing citation volume within that experiment, not for creating a universal visibility score. [2026 preprint](https://arxiv.org/html/2604.25707v2)

| Measurement rule | Why it matters |
|---|---|
| Compare citation prevalence within the same engine first | Engines attach different numbers of links to answers. |
| Keep AI Overviews and AI Mode separate | Google documents that their models and techniques may differ. |
| Compare domains and URLs separately | A domain can remain while the specific cited page changes. |
| Report citations per eligible answer | A system that often returns no citations cannot be compared directly with one that cites every answer. |

**Raw citation totals are not a cross-engine visibility score.**

A higher total can mean that an engine cites more pages per answer. It does not necessarily mean that a domain is more visible, more trusted, or more likely to be named in the response.

## A weekly tracking template for buyer prompts

The following is a hypothetical example.

A B2B software team tracks 25 buyer prompts across four surfaces:

- ChatGPT Search
- Perplexity
- Google AI Overviews
- Google AI Mode

The team runs each prompt once per surface each week for four weeks. That creates:

**25 prompts × 4 surfaces × 4 weeks = 400 prompt-surface-run observations**

Extra reruns can be added when a result looks unusual. They should be recorded as separate observations rather than replacing the original result.

For every run, record:

- Exact prompt and intent group
- Engine and surface
- Date and time
- Country and language
- Account or session state
- Whether web search or grounding was enabled
- Whether an AI Overview appeared
- Each displayed citation URL and parent domain
- Source type, such as vendor, documentation, review, editorial, community, or reference
- Brand mention status separately from own-site citation status
- First-seen and last-seen dates

A no-result condition should remain separate from citation loss. If an AI Overview does not appear, record that the surface returned no Overview. Do not automatically classify it as a lost citation.

A [citation tracking workflow](https://snezzi.com/features/ai-citation-tracking/) can save cited links over time when the collection conditions and definitions remain consistent. Snezzi describes its tracking process this way:

> Snezzi records every cited URL per prompt, engine and run across ChatGPT, Claude, Perplexity, Google AI Overviews and Google AI Mode, so you can see which sources each engine uses for your buyers’ questions and how often they change.

### Calculate domain citation prevalence

For each prompt and engine, calculate:

> **Domain citation prevalence = runs citing the domain ÷ valid runs for that prompt and engine**

A valid run is a completed observation that was eligible for the measurement. A run with no citation can remain a valid no-citation observation, depending on the question being measured. Failed or incomplete collections should be recorded separately.

For example:

> Domain X appeared in 7 of 10 valid Perplexity runs for this prompt during the four-week period.

That is a prevalence statement. It does not claim that Domain X ranks seventh, owns a position, or will remain visible next week.

### Calculate consecutive-run retention

For each pair of consecutive runs, calculate:

> **Consecutive-run retention = domains shared with the prior run ÷ domains cited in the prior run**

Keep the denominator visible. If a response has eight domains in one run and twelve in the next, churn-in and retention will not add up to 100% because they use different reference sets.

### Review domains and URLs separately

Start with the parent domain, then examine the exact URLs.

A domain can remain present while the cited page changes. That differs from the entire domain disappearing. Report both:

- **Domain persistence:** How often the parent domain remains in the source set.
- **URL persistence:** How often the same page remains in the source set.

A useful report would say:

- Domain X appeared in 7 of 10 valid Perplexity runs.
- The same URL appeared in 4 of those runs.
- Other URLs from Domain X appeared in the remaining runs.

This is more informative than saying that Domain X “ranks” in AI based on one answer.

## Common mistakes in AI citation analysis

### Calling a citation a brand mention

A brand can be named without its website being cited. A website can be cited without the brand being named in the answer. Track citations and mentions as separate fields.

### Combining Google surfaces into one metric

AI Overviews and AI Mode are separate measurements. Gemini is also a separate product and data source. A blended Google number can hide meaningful differences between surfaces.

### Comparing unlike runs

Runs from different locations, languages, dates, account states, or prompt wording can measure different retrieval conditions. If those variables change, label the comparison rather than presenting it as source rotation.

### Publishing a list of domains AI “trusts”

Any domain list needs the prompt set, time window, engines and surfaces, sample size, geography, language, and domain-counting rules. Without those details, it can create a false impression of a permanent hierarchy.

### Reporting URL volatility before domain volatility

Check whether the parent domain remained before concluding that a source disappeared. Page-level and domain-level changes answer different questions.

## The right way to interpret AI citation sources in 2026

The strongest conclusion from the available evidence is not that one domain wins across AI search. It is that each engine and surface produces a conditional source set.

A useful analysis should tell you:

1. Which prompts were tested.
2. Which engine and interface produced each answer.
3. When and where the run occurred.
4. Whether the surface returned citations or no result.
5. Which domains appeared repeatedly.
6. Whether the same URLs persisted within those domains.
7. How much of the source set rotated between runs.

That turns AI citation analysis from a one-answer screenshot into a repeatable measurement process. A domain that appears once is an observation. A domain that appears across many valid runs, prompts, and dates is evidence of a stronger pattern, but still not a permanent platform preference.

## FAQs

### How often do AI engines update the sources they cite?

The observed studies show both daily and weekly changes. GetMentions measured total daily source churn of 79.2% for ChatGPT, 75.9% for Google AI Mode, and 44.4% for Perplexity during a seven-day June 2026 study. SISTRIX measured weekly churn-in of 74% for ChatGPT Search and 56% for Google AI Mode across 17 weeks.

These figures describe displayed citation changes under specific study conditions. They do not establish a universal refresh or retraining schedule for any engine.

### Does Perplexity cite Reddit more than Google?

In Profound’s August 2024 to June 2025 dataset, Perplexity cited Reddit in 6.6% of all recorded citations, compared with 2.2% for Google AI Overviews. That is a historical comparison between Perplexity and one Google surface.

It should not be extended automatically to Google AI Mode, Gemini, or every prompt set. Measure each surface separately.

### Which sources does AI cite when recommending software?

The mix can include vendor product and documentation pages, independent reviews, marketplace profiles, editorial comparisons, community discussions, and reference sources. The right source type depends on the buyer’s question.

A pricing prompt may need current vendor terms. An alternatives prompt may draw on review or comparison pages. An implementation question may favor documentation or community experience. Track the exact prompts and engines instead of relying on a universal target list.
