What AI Engines Cite: A Source Map From 27,007 Citations

Key findings

0.9%of citations on category questions pointed to the brand’s own website. 99.1% pointed somewhere else.

40.1%of visible citations went to decision pages: best-of lists, alternatives, comparisons and reviews.

17.5%of Google AI Mode’s citations came from community and video platforms. In ChatGPT’s answers, that share was 0.0%.

35.1%of Gemini’s citations were Google links that hide the original source.

56.3%of answers carried at least one citation, with 9.7 citations per cited answer on average.

15.9%of citations were linked by the engine to a specific brand. Most support the answer as a whole.

Context

Why this matters

When a buyer asks an AI engine which product to choose, the engine answers with a shortlist and, often, the pages it read to build it. Those pages decide which names reach the shortlist. If a brand’s own site supplied most of them, marketing teams would already control the input. This study shows that it does not: 99.1% of citations on category questions point somewhere else.

Knowing which pages, on which sites, each engine reads is therefore the starting point for any plan to appear in AI answers. This report maps those sources across ten engines.

Research question

AI engines increasingly act as the first shortlist in a software purchase. For marketers, the practical question is not only whether a brand is mentioned, but which sources the engine relies on when it decides what to say. If engines cite a brand’s own pages, owned content is the lever. If they cite third parties, the lever is coverage elsewhere.

How often do AI engines cite sources on category questions, and how many?

What kinds of sources do they cite?

Do engines differ in where they look?

What kinds of pages get cited?

Data and method

Category questions only. A prompt is excluded when its text names the scanned brand. Questions that name a brand are answered differently: engines cite the brand’s own pages heavily, which would distort a study of how buyers discover options. After exclusion, 4,938 answers and 27,007 citations remain, from ChatGPT, Gemini, Google AI Mode, Google AI Overviews, Perplexity, Microsoft Copilot, Claude, Grok, DeepSeek and Meta AI.

Results are reported for all answers and for two groups: SaaS categories (such as CRM, workflow software and real estate technology) and AI visibility tools.

Source types

Each cited domain is normalised and assigned to a source type from fixed domain lists. Domains on no list count as “other”.

Page types

Each visible citation URL is assigned a page type from patterns in its address, such as best-of list, comparison or pricing.

Visible citations

All citations except hidden Google links. Shares by engine, page type and domain use visible citations; source types use all.

How often engines cite

Of 4,938 answers, 56.3% carried at least one citation. Answers that cited anything cited a lot: 9.7 sources on average.

EngineAnswers with a citationPer cited answer

Grok100.0%8.0

Google AI Mode83.4%5.0

ChatGPT75.0%10.3

AI Overviews69.0%6.2

Perplexity66.7%10.4

Copilot47.2%4.3

Gemini42.5%11.5

Meta AI31.6%46.9

Claude23.9%11.4

DeepSeek0.3%n/a

Engines differ in style as much as in frequency. Google AI Mode and Copilot cite in most answers but list few sources. Meta AI cites less often but, when it does, lists dozens. DeepSeek returned almost no citations, so it is left out of the source analysis below. Grok, Meta AI and Claude contribute fewer answers than the larger engines, so their rates are less precise.

04 · Finding

Where citations come from

0.9%

On category questions, about one citation in every 108 pointed to the scanned brand’s own domain. In the SaaS group, where brands are more established, the share was 1.7%. In the AI visibility tools group, it was 0.2%.

All 27,007 citations

Other domains · 81.7%Hidden Google links · 8.8%Named source types · 8.6%Brand’s own site · 0.9%

Named source types, up close

Tech media2.3%

Social and publishing1.3%

Review sites1.3%

Video1.2%

Community1.2%

Brand’s own site0.9%

Analysts0.5%

Reference (Wikipedia)0.4%

Press wires0.3%

Share of all citations. The bars below the line are scaled to the largest named type.

Hidden sources are not trivial. 8.8% of all citations were Google grounding links that redirect to the real source without naming it. All of them came from Gemini, where they made up 35.1% of citations. Any analysis of Gemini’s sources that reads only the answer will miss a third of them.

The long tail dominates. The well-known source classes, from tech media to press wires, together account for fewer than one in ten citations. More than four in five go to domains no standard source list covers: vendor sites, specialist publishers and category blogs. This is a finding, not a gap in the method. The material AI engines draw on is far wider than the review sites and analysts most marketers track.

05 · Finding

Engines look in different places

The source mix varies sharply by engine. Shares below are of each engine’s visible citations.

Community and videoTech media, reviews and analystsBrand’s own site

Google AI Mode17.5%2.9%1.3%

AI Overviews15.4%0.1%0.0%

Meta AI3.7%2.6%0.0%

Gemini3.2%0.6%1.2%

Perplexity0.5%0.8%0.1%

ChatGPT0.0%8.2%1.5%

Copilot0.0%2.8%0.0%

Google’s answer surfaces lean on Reddit and YouTube. ChatGPT does not: across every answer we collected, including those to questions that name a brand, ChatGPT cited Reddit zero times in 14,590 citations, and YouTube accounted for under 0.1%. Instead, ChatGPT gave tech media, review sites and analysts a larger share than any other engine.

There is no single source strategy for “AI search”. A brand discussed on Reddit and YouTube but absent from tech media and review sites may be visible in Google’s AI answers and invisible in ChatGPT’s, and the reverse.

06 · Finding

What kind of page gets cited

40.1%

of visible citations went to decision pages: best-of lists, alternatives, comparisons and reviews.

Best-of lists24.9%

Blog posts and articles13.8%

Homepages8.4%

Alternatives pages7.5%

Comparisons4.4%

Reviews3.3%

Documentation and guides2.9%

Pricing pages2.0%

Other pages32.8%

Share of visible citations by page type. Decision pages in red.

The share was higher in the AI visibility tools group (47.5%) than in the SaaS group (31.2%), driven by best-of lists (29.1% against 19.9%) and alternatives pages (10.6% against 3.7%).

This matches how the questions are asked. A buyer asking “which tools are best for…” gets an answer built from pages that already rank and compare tools. Being included on those pages, accurately, matters more than any single page a brand publishes about itself.

A long tail, with a few dominant voices

Citations are spread across thousands of domains, and most domains appear rarely.

SaaS categoriesAI visibility tools

Top 10 domains hold11.6%16.8%

Top 100 domains hold36.1%49.5%

Domains cited only once58.5%48.1%

Decision pages31.2%47.5%

Most-cited domains, SaaS categories

TechRadar3.5%

google.com1.9%

Gartner0.9%

Semrush0.9%

Wikipedia0.8%

YouTube0.8%

G20.7%

Reddit0.7%

SoftwareConnect0.6%

HubSpot0.6%

AI visibility tools

7 of 10

of the most-cited domains belong to vendors in the category. In a young category with little independent coverage, vendors’ own comparison and best-of content becomes the material engines cite.

Domains belonging to the scanned brands are excluded from the list.

Citations support answers more than brands

15.9%

Engines link some citations to a specific brand named in the answer. Only 15.9% were linked in this way to a tracked brand, meaning the studied brand or one of the competitors its scan follows. The rest support the answer in general: market overviews, buying guides and lists. Visibility in AI answers depends heavily on being present in the general material engines retrieve about a category, not only in pages about the brand itself.

09 · Example

One question, five engines

To show what these figures look like in a single answer, we took one category question from the dataset, asked about product walkthrough software: “What tools help sales teams share a self-serve product walkthrough after a demo call?” Five engines answered it with citations. Vendor names are withheld.

EngineCitationsWhat the engine cited

Gemini14Six best-of lists and five other vendor blog posts about demo software, plus a Reddit thread, a YouTube video and an analyst review page.

Perplexity10All ten from blogs of vendors in the category, four from a single vendor. Seven were best-of lists.

ChatGPT8Two best-of lists, vendor and CRM blogs, a homepage, a Japanese blog post and one page unrelated to the question.

Copilot5Vendor blog posts, an alternatives page, a product directory and a software catalogue.

Google AI Mode2One best-of list and one YouTube video.

Of 39 citations, only one article was cited by more than one engine. The brand whose scan asked the question was not cited by any of them. Most cited pages were written by competitors in the same category, and most were titled as lists of the best tools. This is the pattern of sections 4 to 7 in miniature: each engine built its answer from a different set of pages, mostly decision pages, mostly from the long tail.

Our view: why this happens

The figures above are measured. This section is our interpretation.

Engines borrow other people’s judgement.On a “which tool” question, an engine prefers pages that have already compared the options. A product page describes one option; a best-of list describes the market. That is why decision pages earn 40.1% of visible citations and the brand’s own site 0.9%.

In many categories, competitors write the lists.Where independent coverage is thin, vendors’ comparison and best-of posts fill the gap: seven of the ten most-cited domains in the AI visibility tools group belong to vendors, and in the example above most cited pages were written by rivals. The vendor that publishes the most useful list gets a voice in its competitors’ answers.

Each engine has its own reading habit.Google’s answer surfaces read the community web; ChatGPT reads editorial and review sites. These differences most likely come from different search indexes and retrieval choices, so they are unlikely to disappear on their own.

The long tail is an opening.Because 81.7% of citations go to domains outside the usual lists, a specialist publisher or a well-made category guide can be cited as often as a major review site.

What marketers can do

Map sources per engine, for your own questions

Run the category questions your buyers ask on each engine, collect the citations, and sort them by source type and page type. A single list of “AI sources” hides the differences between engines.

Get onto the decision pages already being cited

List the best-of, alternatives, comparison and review pages cited for your category. Check whether you appear on each one, and whether your description, pricing and use cases are correct. Ask authors to add or correct entries, with evidence.

Publish the comparison content your category lacks

Where rivals’ lists are the main source, publish fair, specific comparison and alternatives pages with clear criteria and dated facts. In categories with little independent coverage, these pages become source material.

Match each engine’s reading habit

For Google AI Mode and AI Overviews, take part in the Reddit and YouTube conversations about your category. For ChatGPT, invest in coverage from tech media, review sites and analysts.

Keep owned pages as the reference

Owned pages matter when buyers ask about the brand by name. Make pricing, features and integrations easy to read, so an engine has a correct source when it gets there.

Account for hidden sources

Resolve Gemini’s grounding links before drawing conclusions about its sources. Otherwise, about a third of them stay invisible.

Re-check over time, and plan for the long tail

Answers and citations change between runs, and most citations go to domains outside any well-known list. Track the source map for your category over time, not once.

Limitations

SampleThe dataset covers the categories MentionX has scanned. It is not a census of the web or of all software categories.

Unequal groupsThe AI visibility tools group is smaller and more concentrated than the SaaS group; its figures should be read with that in mind.

Answers varyAI answers can differ by user, location and time. Results describe these answers, not every person’s experience.

Generated promptsPrompts are generated from each brand’s website and market, not sampled from real buyer searches.

Rule-based classificationSource types use fixed domain lists and page types use URL patterns: transparent, but imperfect. A best-of article with a plain URL counts as “other page”.

Citation is not influenceA citation shows what an engine linked, not how much a source shaped the answer.

Engine coverageDeepSeek returned almost no citations, and some engines contribute fewer answers than others.

Appendix: classification rules

Source types (by domain). Community: reddit.com, quora.com, news.ycombinator.com. Video: youtube.com. Social and publishing: linkedin.com, x.com, twitter.com, facebook.com, medium.com. Review sites: g2.com, capterra.com, softwareadvice.com, getapp.com, trustradius.com, sourceforge.net, softwarereviews.com, softwareconnect.com, peerspot.com. Analysts: gartner.com, forrester.com, idc.com. Reference: wikipedia.org. Press wires: prlog.org, openpr.com, prnewswire.com, businesswire.com, globenewswire.com. Tech media: techradar.com, pcmag.com, zdnet.com, forbes.com, techcrunch.com, cnet.com, tomsguide.com, businessinsider.com. Hidden Google links: vertexaisearch.cloud.google.com. Brand’s own site: the scanned brand’s domain and its subdomains.

Page types (by URL, first match wins). Comparison: “vs”, “versus” or “compar”. Alternatives: “alternative”. Best-of list: “best” or “top” as a path word. Review: “review”. Pricing: “pricing” or “/price”. Documentation or guide: docs, help, support, knowledge, learn or guide paths. Blog or article: blog, insights, resources, article or news paths. Homepage: a domain with no path.

How a MentionX scan works, and every KPI, is in our first report, How MentionX turns AI answers into evidence. Standards for every statistic on Learn are in the editorial policy.

How to cite and data requests

Please cite this report as: Jones, D. (2026). What AI engines cite: a source map from 27,007 citations. MentionX Research. mentionx.ai/learn/research/what-ai-engines-cite

Data requests

Need the underlying tables?

Journalists and researchers can request the aggregated tables behind this report. Figures may be quoted with attribution to MentionX Research.

press@mentionx.ai

Questions

Does a brand's own website matter for AI answers?

Yes, but mainly when buyers ask about the brand by name. On category questions, the brand's own site made up 0.9% of citations.

Which sources do AI engines cite most?

Mostly vendor sites, specialist publishers and blogs. Among named source types, tech media, social platforms, review sites, video and community sites each held between 1% and 2.5% of citations.

Do ChatGPT and Google cite the same sources?

No. Google AI Mode took 17.5% of its citations from community and video platforms; ChatGPT cited Reddit zero times across all 14,590 of its citations in the dataset, including answers to questions that name a brand, and gave tech media, review sites and analysts a larger share.

What type of page is cited most?

Best-of lists, at 24.9% of visible citations. Together with alternatives pages, comparisons and reviews, decision pages made up 40.1%.