Source citations in AI search: how assistants pick them and how to measure them

The links attached to an AI answer are now the part of AI search that site owners can count. Search Console reports how many times links to your site were shown in AI Overviews and AI Mode,¹ and Bing Webmaster Tools reports how often Copilot and Bing's AI summaries cite you.² This post covers what a citation is, how each major product shows and chooses its sources, what the research says about which pages get cited and how accurate the citations are, and how to measure citation rate, share of citations and volatility. The wider mechanics of retrieval are in the AI search guide.

A citation is a link that an AI answer presents as a source for what it says. The first-party reports define it that way: Bing's AI Performance report counts citations displayed as sources in AI-generated answers.² Underneath, a citation is usually tied to one part of the answer. In the Gemini API each citation links a span of text, marked by its start and end positions, to a source URL,³ and each citation from Claude's web search tool carries up to 150 characters of the text it cites.⁴

A mention is a name in the text of the answer, linked or not. Ahrefs measured the two separately when it tracked AI Overviews, counting the people, organisations, places and brands named in the text apart from the cited URLs.⁵ They often come apart. To take two invented examples, an assistant can name a local accountant because a directory page praises the firm, and cite the directory. It can cite a manufacturer's guide for a specification and recommend a competitor's product. A mention shows how the assistant describes you; a citation shows which page it relied on. How mentions build up across the web is covered in brand signals in AI search.

The interfaces also show links that are not citations. ChatGPT's Sources panel lists cited sources and other relevant links,⁶ and the links Gemini shows may include content related to parts of its response.⁷ A link attached to a sentence and a link in a side panel are different events, and they are worth counting separately.

Every major product can attach sources when it answers from the web. How much each says about choosing them varies a great deal, so the table keeps to what the providers have published.

The pattern is the same in every product: the system searches, reads what comes back and cites from that set. A page that is never retrieved cannot be cited, so crawler access comes first; robots.txt for AI crawlers covers which bots to allow. Google and Microsoft tie retrieval to their search rankings. OpenAI says only that it ranks results on multiple factors, the others describe their systems in general terms, and none publishes how it weighs one source against another.

Ranking still matters, across more searches than the one typed. Ahrefs studied 1.9 million citations from 1 million AI Overviews in July 2025, taking the top three citations in each, and found that 76.1% of cited pages ranked in Google's top 10 for the same query, with a median position of 3.¹⁵ Its March 2026 update covered 863,000 keyword results pages and 4 million AI Overview URLs. Counting standard organic links only, 37.1% of cited URLs ranked in the top 10, 26.2% in positions 11 to 100 and 36.7% outside the top 100.¹⁶

Ahrefs suggests AI Overviews now draw more on the results of fan-out searches, possibly since they moved to Gemini 3 in January 2026. It had also improved its parser to see more citations, so I would put part of the change down to method.¹⁶ The practical point holds either way: many cited pages do not rank for the query typed, and the likeliest explanation is that they rank for the narrower searches behind it.

Being cited in one Google feature says little about the other. Across 540,000 query pairs from September 2025, Ahrefs found AI Mode and AI Overviews cited the same URLs only 13.7% of the time, although their answers had 86% semantic similarity. AI Overviews gave no citation at all in 11% of responses, against 3% for AI Mode.¹⁷

Assistants also lean towards newer pages. In Ahrefs' analysis of almost 17 million cited URLs from July 2025, the pages cited by ChatGPT, Perplexity, Copilot and Gemini averaged about 2.9 years since publication, against about 3.9 years for Google's organic results, and AI Overviews were closest to organic search.¹⁸

Third-party pages take a large share. In a September 2025 preprint comparing AI search with Google across several verticals, languages and paraphrased queries, Chen and colleagues found a systematic and overwhelming bias towards earned media, meaning third-party authoritative sources, over brand-owned and social content.¹⁹ Semrush tracked more than 230,000 prompts weekly from July to October 2025 and found Reddit and LinkedIn among the five most-cited domains on ChatGPT search, AI Mode and Perplexity alike,²⁰ and in March 2026 Ahrefs' own tracking data showed YouTube as the most-cited domain in AI Overviews.¹⁶

What people mean by source signals

No platform publishes a list called source signals. The phrase is used loosely for whatever makes a page likely to be retrieved and cited. On the documentation and studies above, these have evidence behind them:

• The page can be crawled, is indexed and is allowed a snippet. Google requires indexing and snippet eligibility for its AI features,¹⁰ and ChatGPT search requires access for OAI-SearchBot.⁶
• It ranks for the searches behind the prompt. Google retrieves with its core ranking systems across fan-out queries,⁹ and Copilot centres its answers on high-ranking content.¹³
• It has a passage that answers part of the prompt on its own. Perplexity, for one, scores parts of documents against the query.¹²
• It is reasonably current. Most assistants cite younger pages than Google ranks organically.¹⁸
• Third-party sites cover you, because AI search draws heavily on earned media.¹⁹

Anything beyond that, such as a weighting of these signals, is a guess. None of the providers publishes one, and I would distrust anyone who sells a ranked list.

Researchers measure two things: whether every claim has a citation that supports it, and whether every citation supports its claim. Nelson Liu, Tianyi Zhang and Percy Liang of Stanford University audited Bing Chat, NeevaAI, Perplexity and YouChat with human reviewers in 2023. On average only 51.5% of generated sentences were fully supported by their citations, and only 74.5% of citations supported the sentence they were attached to.²¹ Those products have changed since, but the two measures are still the right questions to ask of any citation.

In a study published in March 2025, the Tow Center for Digital Journalism tested eight AI search tools, giving each excerpts from 10 articles by each of 20 publishers and asking for the headline, publisher, date and URL: 1,600 queries in all. Together the tools answered more than 60% of queries incorrectly; Perplexity got 37% wrong and Grok 3 94%. More than half of the responses from Gemini and Grok 3 cited fabricated or broken URLs. Several tools cited syndicated copies on Yahoo News or AOL in place of the originals, and ChatGPT cited a Yahoo copy of a USA Today article although USA Today blocks its crawler.²² The excerpts were chosen because Google found the original in its top three results, so the test measures attribution, a task ordinary search handles well.

A study coordinated by the European Broadcasting Union and led by the BBC, published in October 2025, had journalists from 22 public service media organisations in 18 countries, working in 14 languages, review more than 3,000 answers about news from ChatGPT, Copilot, Gemini and Perplexity. Of those, 45% had at least one significant issue, and 31% had serious sourcing problems: missing, misleading or incorrect attributions. Gemini had significant issues in 76% of its responses, mainly because of poor sourcing.²³

The providers say much the same. OpenAI's help page warns that search results and citations can be incomplete, outdated or incorrect,⁶ and Microsoft's transparency note says Copilot's responses might misrepresent even high-authority web content.¹³ For a site owner this has two consequences. A citation to your page may sit beside a claim your page never made, so check the sentence as well as the link. And a copy of your content can take your citation, which is worth weighing before you agree to syndication.

Building the prompt set, controlling the conditions and recording each answer are covered in how to measure visibility in AI search. This section covers the citation numbers themselves. A March 2026 preprint by Ronald Sielinski, revised in August, gives precise definitions for the metrics most trackers report.²⁴

Citation rate      = responses citing your domain / all responses
Share of citations = citations to your domain / all citations in those responses
Run overlap        = URLs cited in both runs / URLs cited in either run
Margin of error    = 1.96 x sqrt(rate x (1 - rate) / number of responses)

A worked example with a margin of error

Take an invented accountancy firm that runs 20 prompts five times each in ChatGPT: 100 responses. Its domain is cited in 18 of them, 24 times in all, and the 100 responses carry 600 citations between them. Its citation rate is 18 ÷ 100, or 18%, and its share of citations is 24 ÷ 600, or 4%.

Sielinski gives the width of a 95% confidence interval for a proportion as 3.92 × √(p(1 − p) ÷ n), which is a margin of 1.96 × √(p(1 − p) ÷ n) either side.²⁴ For 18% over 100 responses that margin is about 7.5 points, so the true rate lies somewhere between about 10.5% and 25.5%. A rise to 22% next month tells you nothing. Halving the margin takes four times as many responses. That is the best case, because the formula assumes every response is independent, and in Sielinski's data the intervals for Perplexity and SearchGPT did not shrink as smoothly as it predicts. Getting the interval on citation share down to 5 points took roughly 30 to 50 queries on Gemini and 90 to 100 on Perplexity, and SearchGPT converged slowest and least evenly.²⁴

Why citations change between runs

Answers are generated afresh each time, and the sources change with them. Ahrefs followed more than 43,000 keywords, each with at least 16 recorded AI Overviews, for a month. Between consecutive checks only 54.5% of cited URLs overlapped on average, so 45.5% were new, and the text changed 70% of the time while its meaning stayed almost the same, at 0.95 cosine similarity.⁵ Profound, which sells AI visibility tracking, compared about 80,000 prompts per platform in June and July 2025. On average, 59.3% of the domains AI Overviews cited for a prompt in July had not been cited for it in June, against 54.1% for ChatGPT, 53.4% for Copilot and 40.5% for Perplexity.²⁶

Some changes come from the platform. In Semrush's tracking, ChatGPT cited Reddit in close to 60% of responses in early August 2025 and around 10% by mid-September, and Wikipedia fell from roughly 55% to under 20% over the same weeks.²⁰ Reddit was more stable on AI Mode and Perplexity over the same period, which points to a change on ChatGPT's side. Bing gives the causes of shifting citation patterns as changes in user behaviour, evolving models, freshness signals, partner refresh cycles and changes across the web.²⁵ Across Perplexity, SearchGPT and Gemini, Sielinski found that domains whose citation share differed by less than 5 to 7 points usually had overlapping confidence intervals, so they could not be told apart.²⁴

To measure volatility, keep the prompts fixed and run each one several times at every check. Overlap between runs within one check gives you the noise floor. Drift between checks gives you change over time. I treat a movement as real only when the drift between checks is clearly larger than the churn within a check, and when it shows up across many prompts. That is my own working rule; I know of no published standard.

How accurate AI search tracking tools are

Every tracker samples answers from the outside. Google says no third-party tool has access to its internal ranking or AI systems.⁹ How a tool collects answers matters a great deal. Surfer, which sells a tracker that reads the consumer interfaces, ran 1,000 prompts through ChatGPT, Perplexity, Gemini, AI Mode and AI Overviews both ways in 2026, through each product's interface and through an API, collecting 13,779 answers. The domains cited by the API and by the product overlapped by as little as 4.8% on ChatGPT and at most 26.7% on Perplexity.²⁷ Surfer has an interest in that conclusion, but the size of the gap is hard to explain away.

Method changes move a tool's numbers too: Ahrefs' share of AI Overview citations from the top 10 halved between two studies, after a change of parser as well as a change of model.¹⁶ Before trusting a citation report, ask the vendor whether it reads the consumer interface or calls an API, whether it tests signed out and from which location, how many runs it makes per prompt, whether it counts links in the answer text separately from the sources panel, and whether it shows a margin of error. Search Console and Bing Webmaster Tools are first-party counts, but each covers only its own products.

• Be retrievable. Allow the search crawlers, including at the CDN: ChatGPT search needs OAI-SearchBot,⁶ and Google needs the page indexed and eligible for a snippet.¹⁰
• Rank for the narrower searches. Ahrefs’ data suggests cited pages increasingly come from fan-out results.¹⁶ Bing’s grounding queries show a sample of the phrases its AI used when it retrieved pages it went on to cite.²
• Write passages that can stand alone. Put each answer in a clear section with the specific figures, dates and conditions. Bing recommends clear headings, tables and FAQ sections, and examples, data and cited sources.² In the GEO study, which tested rewritten pages in a generative engine built for the purpose, adding citations, quotations and statistics raised a source’s visibility in the answers by 30% to 40% on its main measure, while keyword stuffing did little or nothing.²⁸
• Keep pages current. Bing recommends regular updates and IndexNow so AI systems use the latest version of a page,² and most assistants cite younger pages than organic search.¹⁸
• Get cited by others. In Chen and colleagues’ tests, AI search cited third-party sources far more than brand-owned pages.¹⁹ Earn it honestly: Google says seeking inauthentic mentions is not as helpful as it might seem.⁹
• Protect your originals. Syndicated copies can take the citation for your work,²² so weigh that before agreeing to republication.
• Skip the shortcuts. Google says you do not need new machine-readable files, AI text files, markup or Markdown to appear in Search.⁹ Does schema markup help AI citations sets out the evidence on markup, and who llms.txt actually helps covers that file.

None of this guarantees a citation, and nobody can promise one. What it does is put your pages in the retrieved set more often, with passages an assistant can quote accurately. Measure the result as a rate over many runs, compare it with Search Console and Bing, and judge it over months.