AI Search Citation Sources Study: A Repeatable Method

An AI search citation sources study is only useful if its method is published. Pick one question, state the engines, dates and number of runs, classify every cited URL into a fixed set of source types, and chart each engine's share by type. Ahrefs and geoSurge publish their methods, which is why their numbers can be checked.

Key takeaways

• Publish the question, engines, run dates, runs per prompt and the classification rules next to the chart. A reader who cannot rerun your study cannot trust it.
• Report per engine. In geoSurge's Mexico study, ChatGPT searched the web in 97% of answers and Gemini in 62%, so a pooled "AI" number hides two different routes to an answer.
• Write a codebook before you collect data: about ten source types, one rule per type, and a second person to tag a sample.
• Show counts as well as percentages. Every bar needs its n, and "share of citations" must be kept apart from "share of answers that cite at least one".
• Treat results as description, not cause. Ahrefs says outright that it cannot show mentions cause AI visibility.
• Run the study again in a month. A single run is a snapshot of one day.

Why does a citation study need a published method?

A citation percentage with no method is a claim you cannot test. "Reddit is cited most" can mean most citations, most answers containing a Reddit link, or most brand mentions that correlate with Reddit threads. Those are three different measurements. Only the method tells you which one you are reading.

Three recent studies show how much or how little gets disclosed. The table lists what each states on its own page, as checked on 2026-11-09.

The geoSurge arithmetic closes, which is the point of publishing: 1,400 questions, times 5 runs, times 2 engines, times 2 search settings, gives 28,000. You can check it in your head. Ahrefs' table of what each mention type counts does the same job for a different design.

Neither of the two outside references classifies cited sources by type per engine. That gap is the reason for the method below. Conductor is the closest published example of the result you are after.

How do I pick the one question that frames the study?

Start with one buyer decision, not a topic. "Which project management tool should a 10-person agency choose?" is a decision. "Project management" is a topic. A decision gives engines something to recommend and cite, so every answer has comparable material to classify.

Then write the prompt set around that decision:

1. Core prompt. The question as a buyer would type it, with no brand names in it.
2. Phrasings. Five to ten rewordings of the core prompt, including a shorter keyword-style version and a longer one with context (team size, budget, country).
3. Intent variants (optional). One prompt each for research, comparison and purchase, labelled so you can split results later. Conductor groups results by intent, which is why it can show one engine behaving differently for support and education questions.

Keep brand names out of the prompts, as Ahrefs did by leaving out any prompt that named the brand. A prompt naming you will pull your own site into the citations and inflate the brand-owned share. Noma's guides on choosing prompts to track and running an AI answer benchmark by category give scoring sheets for this step.

How should an AI search study state sample size, dates and engines?

State six fields in a box above the first chart. If a field is missing, a sceptical reader will assume the worst about it.

Worked example: sizing a small study

This is a design illustration, not a result. Take one question with 8 phrasings, 4 engines and 5 runs each: 8 x 4 x 5 = 160 answers, 40 per engine. If an answer cites about five sources on average (an assumption you replace with your own count), that is roughly 200 citation rows per engine. A 200-row engine sample can show a gap between 40% forum and 10% forum. It cannot separate 22% from 25%.

The citations inside one answer are not independent, and the answers to one prompt are not independent either. Report counts, show the range across phrasings, and avoid claiming a ranking between types that sit within a few points of each other. Noma's AI Visibility Measurement Standard sets out the numerator, denominator and margin-of-error habit in more detail, and the post on measuring AI share of voice across engines works through sample sizes.

Plan at least two collection waves, a week or more apart. Noma's post on how long AI citations last covers why a single wave overstates how stable a cited URL is. Source types are steadier than URLs, but you should show that with your own two waves, not assume it.

How do I classify cited sources in AI answers by type?

Write the codebook first and freeze it before you look at any results. Classify at domain level by default, then override at page level where the page type changes the answer.

Add two flags rather than two more types. One marks competitor-owned (a vendor's own "best tools" page sits in brand-owned, but flagged so you can count it separately). The other marks suspected AI-generated. A May 2026 arXiv audit of four generative search engines found about 16% of cited sources showed evidence of being AI-generated (arXiv, May 2026), so the flag is worth a column even if you never chart it.

Check your own consistency

1. Tag every row yourself.
2. Ask a colleague to tag a random 10% blind, using only the codebook.
3. Report the percentage of agreement, or Cohen's kappa if you can calculate it, in the method box.
4. Fix the rule behind every disagreement, then retag the whole set.

Unlisted domains will be the slow part. Keep a lookup sheet of domain to type so a domain is tagged once and reused across engines and waves.

How do I chart which sources each engine cites most?

Use three charts and label the n on each.

• 100% stacked bar, one bar per engine. Segments are source types, share of citation rows. This answers the title question at a glance.
• Heatmap, engine by source type. Useful when you have more than five engines. Put counts in the cells, not just colour.
• Top ten domains per engine. Types hide the fact that one domain may be doing all the work. Conductor's finding that engines lean on a default source type is only actionable once you can name the domain.

Define two denominators and never mix them:

Answers with no citation at all belong in a separate line. They are an engine behaviour in their own right, and dropping them silently makes engines that cite less look more source-heavy.

What should the downloadable summary contain?

Put a one-page summary beside the chart so journalists, clients and AI engines can quote it with its context. A workable template:

1. Headline finding in one sentence, naming the engine, the source type and the measure (for example, share of citations).
2. Method box with the six fields from the table above.
3. Chart with n per engine and the wave dates in the caption.
4. Codebook as an appendix: types, rules, flags, agreement score.
5. Raw data link to a CSV with columns: prompt, engine, setting, run date, run number, cited URL, domain, source type, flags.
6. Limits in three bullets: what the study cannot say, what was excluded, what you would change next time.

The CSV matters most. Ahrefs and geoSurge publish charts, but a reader can rebuild a result only from the rows.

How do I run the study with Noma or the free tools?

You can run a small version by hand: ask each engine your prompts in fresh sessions, copy every cited URL into a sheet, and tag the types. Expect it to take hours per wave, and keep the same account state and location each time.

To see the first signal before you commit, the Free Competitor AI Share of Voice tool compares how often ChatGPT, Gemini, Claude and Perplexity name you against up to three competitors, and which sources they cite. Use its source list as the seed for your domain lookup sheet.

Noma by Nurdd asks 13 AI engines, including ChatGPT, Gemini, Perplexity and Claude, a fixed set of buyer prompts every day. It lists the sources each answer cites, so the rows for your CSV are already collected per engine, and it separates your own pages from earned sources in its owned and earned views. You still apply your own codebook.

See it on your own brand: book a demo or start the 3-day free trial. A card is taken at signup so billing can start on day 3, and cancelling before then costs nothing. Noma's help pages explain what sources are and why they matter and how the earned and owned views work.

What can the Ahrefs 260M-prompt study teach about running a citation study?

Four design choices are worth copying.

1. Exclude prompts that name the subject. Ahrefs compared how often AI names 10,000 US brands across 260M non-branded prompts, leaving out any prompt that named the brand. Without that rule, a brand's own name drives its own visibility.
2. Define every measure in a table. Ahrefs lists what each of its seven mention types counts and where the number came from. Your codebook should be as plain.
3. Handle entity ambiguity. Ahrefs says Brand Radar now counts Apple the company separately from apple the fruit. For source types, the equivalent is separating a brand's own domain from pages that merely share its name.
4. State what the result cannot show. The authors write: "We can't say mentions cause AI visibility" (Ahrefs, November 2026). Their top figure, a Spearman correlation of about 0.75 for Reddit mentions, describes how rankings line up, nothing more.

Ahrefs also reports that ChatGPT was the least tied to off-site mentions on six of seven factors, which is a reminder to split by engine. Noma's write-up of the study is at which mentions drive AI visibility.

What does the geoSurge Mexico study add?

geoSurge's 28,000-answer study shows why engine settings belong in the method box. According to geoSurge, ChatGPT searched the web in 97% of answers and Gemini in 62%. When they searched, ChatGPT ran 6.9 searches per answer against 2.3 for Gemini (geoSurge, November 2026).

For a source-type study this has a direct consequence. An engine that searches in two thirds of answers produces fewer citations per answer than one that searches in nearly all of them. If you pool engines, the engine that searches most dominates your citation rows. Chart per engine, and report answers with no citation as their own line.

geoSurge also ran every question with web search on and off. If you can control that setting, record it. If you cannot, say so.

Which sources do ChatGPT, Perplexity and Gemini cite most often?

The best published answer to that question is Conductor's seven-month analysis, and it is specific to the study period and intents it measured. Conductor reports that Perplexity had YouTube as the top source for Education and Recommendations queries every month, that Gemini had YouTube on top for Support intent every month, and that ChatGPT led Recommendations with Wikipedia every month while Reddit led Support in five of seven months (Conductor, study period September 2025 to March 2026).

Conductor also reports that across 16 rank-one citation slots tracked in February and March 2026, Claude never surfaced YouTube, Wikipedia or Reddit. Sixteen slots is a small sample, which is why the method box matters. Treat each of these as a hypothesis to test on your own question, not as a rule for your category. Your category, prompts and country will shift the mix.

What are the limits and common mistakes?

• Source type is not a cause. An arXiv study of 252,000 trials across six models found that topic match, price, recency and position dominated citation preference, with formatting having negligible impact (arXiv, May 2026). A chart of types describes where citations land. It does not tell you what to publish.
• One question is one market. A result for agency project management software says little about a consumer product. Say so in the limits.
• Pooling engines. Averaging ChatGPT and Gemini produces a number that describes neither.
• Changing the codebook mid-study. Retag everything after any rule change.
• Calling gaps significant without a test. With a few hundred rows per engine, report ranges and say which differences are large enough to act on.
• Citations are not mentions. An engine can name your brand without citing you, and the reverse. Noma's post on citation versus brand mention explains the split.

How we researched this

We reviewed primary study pages from Ahrefs and geoSurge, a seven-month engine comparison from Conductor, and two arXiv papers on citation behaviour and AI-generated sources, and read the methodology sections where they exist. Figures are quoted as each publisher states them. We did not run a study for this post, and it reports no Noma data. Facts, dates and figures were checked on 2026-11-09.

Next step

Pick one buyer question this week, write the codebook, and run one wave on four engines by hand or with the free share of voice tool. When you want per-engine cited-source lists without copying URLs, book a demo or start the 3-day free trial.