Most guides to tracking AI citations end with one number: the share of answers that cite the brand. That number says whether the brand was linked. It says nothing about which sources the engines actually draw on for the category, and those sources are where the work is.
In a 19-day Zumi study of one buyer question, Gemini, Google AI Overviews, and Google AI Mode cited 362 different domains across 77 answers. Most of them, 253 or 70%, appeared in one answer and never again. A brand-level rate taken on a single day cannot tell a source the engines rely on from one they used once.
This guide sets out a different record: a source-first citation log, built one row per cited source per answer, with the number of answers behind every figure.
Key takeaways
- A source-first log records every cited source, not only the brand's own pages, with its engine, prompt, date, page, and owner: the brand, a competitor, or a third party.
- One run is a sample. In the Zumi study, 70% of the cited domains (253 of 362) appeared in a single answer, so a source should be called recurring only after repeated runs.
- Engines need separate columns. Across the same question, Gemini cited 38 domains, AI Overviews 129, and AI Mode 276, and only 11 domains were shared by Gemini and AI Overviews.
- A link and a mention are different events. Gemini linked the study's most cited domain in 21 of 26 answers and mentioned it in the answer text in none.
- Every rate in the log is written with its count, such as "cited in 18 of 25 answers," because a percentage alone hides whether it rests on 4 answers or 400.
What is a source-first citation log?
It is a table where each row is one cited source in one answer. An answer that cites 12 pages produces 12 rows. The brand's own pages sit in the same table as competitor pages, review sites, forum threads, and news articles.
The usual starting point, checking a prompt list by hand in a spreadsheet and dividing cited answers by answers tested, still works as a first pass. The source-first log keeps that spreadsheet and changes what goes in it: instead of a yes or no for the brand, it records everything the answer cited.
That change matters because the brand's absence is rarely the useful finding. The useful finding is who is present instead. A category where engines lean on two review sites calls for different work from one where they cite competitor blogs.
What goes in each row?
| Field | What it records | Why it matters |
|---|---|---|
| Answer ID | A unique label for the answer the source came from | Lets a reader reopen the answer and check the fit |
| Engine | ChatGPT, Gemini, Perplexity, AI Overviews, AI Mode, and so on | Each engine cites differently, so totals are kept per engine |
| Prompt | The exact wording, unchanged across runs | A reworded prompt is a new prompt |
| Run date | The day the answer was collected | Separates change over time from noise |
| Domain and page | The site and the specific URL | A domain can be cited through one page or through 20 |
| Owner | Brand, competitor, or third party | Decides whether the fix is content on the brand's site or coverage elsewhere |
| Source type | Review site, forum, news, vendor page, video | Each type calls for a different response |
| Brand mentioned in text | Yes or no, for this answer | Keeps mentions apart from links |
The owner field carries the most weight. Without it, a log of 300 cited pages is a list. With it, the same log shows what share of the category's citations the brand controls, what share competitors control, and how much sits with third parties the brand can only influence.
How many runs before a source counts?
AI answers change between runs, even with identical wording. Zumi measured the same effect on brand mentions: across 7,280 consecutive-day answer pairs in an India beauty run, only 7.3% of pairs mentioned exactly the same brands.
Citations move too. In the 19-day study, 77% of the domains AI Overviews cited appeared in one answer only. A source seen once might be a regular, or it might be gone tomorrow.
The fix is a repeat-run rule, set before the first run and applied the same way every time:
- Repeat runs. Each prompt runs several times on each engine before anything is read into the result.
- A stated threshold. A source counts as recurring only when it appears in a set share of those runs, written as a count: "cited in 9 of 12 runs."
- Single appearances stay unconfirmed. A source seen once is held as unconfirmed until it returns.
The exact threshold is a team's choice. Stating it, and holding it fixed, is the part that makes two months of logs comparable.
Why does each engine need its own columns?
Engines asked the same question draw on different parts of the web, and they cite at different rates. The Zumi study measured three engines on one question:
| Measure | Gemini | AI Overviews | AI Mode |
|---|---|---|---|
| Answers collected | 26 | 25 | 26 |
| Distinct domains cited | 38 | 129 | 276 |
| Citations per answer | 7.7 | 12.4 | 25.7 |
| Most cited domain: linked, then mentioned in the text | 21, then 0 of 26 | 18, then 7 of 25 | 20, then 11 of 26 |
AI Overviews and AI Mode, both run by Google, shared 58 domains, while Gemini and AI Overviews shared 11. The citation rates matter too: a 5% share of AI Mode's citations works out to more than one citation per answer, while the same share on Gemini is less than half of one.
These figures come from one buyer question in one category, so they show how far engines can differ, not how far they differ in every category.
So the log keeps engines apart, and any combined figure states which engines went into it. How Perplexity cites sources and how Gemini, AI Overviews, and AI Mode differ show how far apart the patterns run.
Should mentions and citations be counted together?
No. A citation is a source the engine linked or quoted, and a mention is the brand appearing in the answer text. Each can happen without the other.
The study's most cited domain shows the gap. Gemini linked it in 21 of 26 answers and mentioned it in the answer text in 0 of 26. AI Overviews linked it in 18 of 25 and mentioned it in 7.
A team counting links alone would call that domain fully present. A buyer reading the answers would have seen its name in a minority of them, or not at all.
The log handles this with one extra field per answer: was the brand mentioned in the text. Citations and mentions are then reported as two counts from the same set of answers, and neither is folded into the other.
How should figures from the log be written?
Every figure carries its count. "Third parties hold most of the category's citations" becomes "third parties hold X of Y citations across Z answers on three engines," with the real numbers filled in. The second version can be checked, compared next month, and challenged.
Two habits keep the counts honest:
- A share says what it is counted out of. A share of citations and a share of distinct pages give different numbers from the same log, because one page cited ten times counts as ten citations but only one page. The figure names which one it uses, and the log keeps that choice from month to month.
- Prompts and repeat runs are disclosed alongside answers. Ten runs of eight prompts produce 80 answers, but they are not 80 independent observations. The figure reads in answers. The method note says how many prompts and runs produced them.
What can a citation log not show?
A citation log shows which sources an engine used. It does not show that a source caused the brand to be mentioned, or that winning a citation will change an answer. Google's guidance for site owners notes that AI Mode and AI Overviews "may use different models and techniques, so the set of responses and links they show will vary" (Google Search Central, 2026).
Two other signals are often offered as substitutes. AI referral traffic in analytics counts only the citations that produced a click, and server logs of AI crawler visits show that a page was fetched, not that it was cited.
Both are worth reading beside the log, and neither replaces it.
The log's job is direction: it shows where the engines look for a category, and therefore where content and coverage effort will be tested next.
Where does a team start?
A first log fits in a spreadsheet: a fixed list of buyer prompts, the engines the brand's buyers use, a weekly run, and the fields in the table above. The first month sets the baseline. The baseline guide covers how to set one, and GEO metrics beyond mention rate covers what to report from it.
A manual log gets slow as prompts and engines grow. At that point the same fields apply to a tool's citation view, and what to check in a citation analysis tool lists them.
Zumi, an AI Search Intelligence Platform, reports citation share as one of four AI Visibility signals for each tracked engine. Book a demo to see the sources engines cite for a brand's own category.
Sources
- Zumi, Gemini, AI Overviews, and AI Mode cite differently, one buyer question, 77 answers, July 20 to August 7, 2026.
- Zumi, how to evaluate AI visibility tools, India beauty prompt run, 7,280 consecutive-day answer pairs.
- Google Search Central, "AI features and your website".