Meet an Expert

How AEO Tools Define Mentions and Citations

AEO tools disagree on what counts as a citation. Here is how the leading monitoring tools define it, and what to test before you buy.

Gilad Bechar Gilad Bechar
Nov 2, 2026 7 min read Tools

Run the same prompt through two AEO tools on the same morning and you can get two different citation numbers for the same brand. Neither tool is broken. One counted every time the brand was named in the answer. The other counted only the brand’s own URLs in the source list. So before you compare daily prompt-audit AEO tools on price or engine coverage, find out which of those two events each one is actually recording.

Key Takeaways

  • A mention is your brand named in the answer text, and a citation is your page used as a source. Tools that blur the two give you a number you cannot act on.
  • Most trackers count a citation only when your own domain appears, so a review site or forum thread that recommends you can score as zero.
  • Running a prompt daily adds repeats of the same question. Research on LLM brand answers finds that spreading across paraphrases, models and languages improves reliability far more than repeating one prompt.

What Monitoring Tools Say They Count

One monitoring tool’s documentation is unusually explicit. For each monitored answer it identifies brand mentions, cited sources, competitors and sentiment. Its Mention Rate is how often your brand appears among answers that mention you or a direct competitor. Its Citation Rate is how often your domain is cited among answers that cite you or a direct competitor’s domain. Both rates use your competitive set as the denominator, and the denominator does not include every answer the tool collected.

Another vendor draws the line in plain terms. Its product page defines a mention as your brand named in the answer and a citation as one of your pages used as a source, and says it reports the two separately because they need different fixes. Each run logs whether you were named, whether your pages were cited, which competitors appeared and which sources sat behind the answer. Runs can be daily, weekly or monthly.

Other tools in the category publish less. For any tool whose definitions you cannot find in writing, ask for the counting rules before the trial starts. A sales demo shows the dashboard. It does not show the rule that decided what counted.

Watch the labels as well. One tool uses Share of Voice to mean your share of mention occurrences against tracked competitors. At Answerburst, Share of Voice means how often a brand is mentioned across the web overall, and Share of Citation means how often AI models name the brand as a source. The same label can stand for very different measures in different tools.

Why the Same Answer Produces Different Numbers

Every AI answer has two layers: the text and the sources. A brand can turn up in either one, in both or in neither. Named and cited is the strong case. If the brand is named but not cited, the model is often working from training data, or it pulled your name from someone else’s page. A citation without a mention means your page supported a claim while the answer recommended someone else, or spoke in generic terms.

Those states map onto the two layers of AEO. Training-data presence comes slowly, through wide coverage, and tends to show up as mentions. Retrieval presence comes from structured, current content, and it shows up as citations. A tool that merges the two hides which layer is working.

Third-party sources make it harder. A Pew Research Center analysis of Google searches in March 2025 found that Wikipedia, YouTube and Reddit were the most cited sources in AI summaries, at 15% of all sources listed, and that 88% of summaries cited three or more sources. If an answer recommends you because a Reddit thread or review page recommended you, a tool that matches only your own domain records a mention and no citation. In practice that page was the citation that got you into the answer.

A citation count is only as meaningful as the rule that decided what counted, and most dashboards never show you that rule.

Page-level data is what makes citation tracking useful, because it tells you which assets models actually retrieve.

309
unique Moburst pages cited by AI models, alongside 42,435 total AI citations and a mention rate near 49%.See the case study.

Does Running Prompts Daily Make the Data More Reliable?

Only partly. A 2026 arXiv paper on non-determinism in LLM brand answers notes that common practice is to resample each prompt a small fixed number of times, usually five, and average the results. The authors analysed 12,933 responses about 20 Central and Eastern European brands across 8 languages and 3 models. Query language accounted for 26.5% of the variance in a single response, while brand identity accounted for 1.5%. Their conclusion is that one AI answer carries almost no signal that tells brands apart.

Adding languages and models cut error far more than adding repeats did. A repeat past the fifth reduced it by only 0.0003. One caveat: the outcome they modelled was sentiment toward the brand, and the authors list a refit on recommendation data as future work.

The direction still matters to a buyer. A daily schedule mostly buys more repeats of the same wording. What moves reliability is how many paraphrases of each question the tool runs, and how many engines and markets it covers. Ask how many variants sit behind each tracked prompt. Ask whether the trend line is smoothed across runs.

Six Tests to Run During the Trial

  1. Hand-label a sample. Run 20 of your prompts yourself in each engine on the same day, mark every answer as named, cited, both or neither, and compare your labels with the tool’s.
  2. Break the name matcher with abbreviations, product names, misspellings and a possessive form. If your brand name is a common word, check for false positives.
  3. Break the domain matcher. Subdomains, country domains, a blog hosted elsewhere, redirected URLs and links carrying query strings should all resolve to you.
  4. Find out whether you can tag third-party pages, such as a review listing or partner page that names you, and report them as earned citations next to your owned citations.
  5. Read the denominator. Remove one competitor from your tracked set and see whether your rate moves. If it does, the metric measures relative share.
  6. Export the raw record for one run: full answer text, source list, timestamp, model and location. If you cannot audit a number, you cannot defend it in a budget review.

Also ask how the tool queries each engine. Answers from an API and answers from the consumer interface can differ, and one vendor’s own documentation says its monitoring records what a model returned for a given prompt, country, persona and time, not every answer every user receives. A vendor who explains that limit is easier to trust than one who leaves it out.

What to Do Before You Sign

Write a one-page definitions document before any demo. It should define a mention, an owned citation, a third-party citation and the denominator you want, then go to each vendor with a request to map their metrics onto it. Treat vague answers as a result in themselves.

Next, build a prompt set of 30 to 50 questions split across brand, comparison and category intent, with two or three paraphrases each. Run two shortlisted tools in parallel on identical prompts for two weeks, score both against your hand-labelled sample, and pick the one whose errors you can live with. When you report upward, put mentions and Share of Citation on separate lines. They respond to different work, and citations compound once a model starts using you.

FAQs

What Is the Difference Between a Mention and a Citation in AI Answers?

A mention is your brand named in the answer text. A citation is one of your pages used as a source for the answer. A brand can have one without the other, and each points to a different fix.

Do AEO Tools Count Third-Party Pages as Citations?

Most match citations against your own domain, so a review site or forum thread that recommends you usually registers as a mention only. Ask whether the tool lets you tag third-party pages and report them separately.

Is Daily Prompt Tracking Better Than Weekly?

Daily runs add repeats of the same prompt, and research on LLM brand answers finds repeats past about five add little precision. Paraphrase variety and engine coverage matter more than cadence.

How Do I Test an AEO Tool Before Buying It?

Hand-label 20 prompts across each engine on the same day, compare your labels with the tool’s output, stress-test its name and domain matching, check its denominator and confirm you can export the raw answer and source list for any run.

Gilad Bechar

About the author

Gilad Bechar

Gilad Bechar is the Founder & CEO of Moburst. Gilad serves as a mentor to rising startups at Microsoft Accelerator, The Technion, Tel-Aviv University, Unit 8200 and for strategic Moburst clients, and is the Academic Director of the Mobile Marketing and New-Media course at Tel-Aviv University.

Meet the whole team

Ready to make your brand the answer?

Our growth strategists take brands from overlooked to recommended. In one call you will see how the AI engines read your brand today, and leave with your own roadmap to dominate the new era of search.