How to Track AI Citations, Brand Mentions and Share of Answer

To track AI citations, record the exact question, engine, search context, answer and source URLs for each observation. Mark a citation only when the answer attributes information to a URL on your website. Record brand mentions separately, exclude failed requests from outcome rates and compare the same question cohort over time.

An AI visibility report is useful when it tells you what changed and where to act. A single percentage without its questions, denominator and source evidence cannot do that. This guide provides a measurement workflow and a blank CSV tracking template you can use before adopting a tool.

Define the outcomes before collecting answers

Outcome	Definition for this panel	What to retain
Cited	The answer attributes information to a URL on your website	Exact citation URL and answer
Named, not cited	The answer names your brand but does not cite your website	Brand text and answer
Retrieved, not cited	A visible search trace includes your page without an answer citation	Retrieved URL and trace, if available
Absent	An eligible answer neither names nor cites your website	Answer and competitor sources
No AI answer shown	The selected search surface does not return an AI answer	Surface, question and observation time
Failed	The measurement could not obtain a usable response	Error and retry status
Not asked	No measurement was made	Planned question and engine

A third-party citation about your company does not count as a citation to your website under this definition. Track it separately if third-party brand coverage is part of your objective. If a response both names and cites you, use “cited” as its exclusive outcome; an additional mention flag can preserve the detail.

Also define which hosts belong to you. Decide whether subdomains count, list approved aliases and retain the original URL before normalizing it. Otherwise a change in your matching rules can look like a visibility gain.

1. Build a panel around customer questions

Use actual sales questions, support requests, search queries and product decisions. Include several intent types:

• Branded: “What does [brand] do?” checks knowledge of your own product.
• Category discovery: “Which tools measure AI citations?” checks whether you enter the consideration set.
• Comparison: “[brand] versus [competitor] for a content team” tests a specific decision.
• Problem solving: “Why is my website missing from ChatGPT Search?” tests practical expertise.

Keep branded and non-branded results visible separately. A panel dominated by your name measures a different challenge from one dominated by category questions.

Give each question a stable ID, intent and target page. Preserve its exact wording. Mark it as a new cohort when you change wording materially, rather than silently replacing its historical identity. A hundred paraphrases of one question do not represent a hundred independent customer needs.

2. Document the measurement context

For each observation record the engine, surface or interface, model where disclosed, web-search setting, language, country and timestamp. Also record whether a conversation had prior context. Use a fresh context for baseline comparisons.

These controls matter because “ChatGPT visibility” might mean a consumer app answer, a search-enabled official interface or a manual screenshot. They are related measurements with different conditions. Avoid pooling them without labels.

Google AI Overviews, Google AI Mode and Bing Copilot Search are separate surfaces. A search with no AI answer is an observation of availability, not an answer that rejected your website. Likewise, a failed request provides no evidence of absence.

In Vercatus, the Answer Share documentation explains the question selection and outcome rules. The engine guide identifies supported engines and how their observations differ. The engines shown depend on configured availability.

3. Use metrics with explicit denominators

Choose one primary outcome metric and retain the underlying counts. These are useful definitions for a manual panel; a vendor’s similarly named metric may use a different unit.

Metric	Formula	Unit and use
Answer citation rate	Eligible answers citing your website ÷ eligible answers × 100	Per answer; split by engine and cohort
Named-only rate	Eligible answers naming but not citing you ÷ eligible answers × 100	Per answer; reveals mentions without owned-source attribution
Question citation coverage	Questions cited in at least one eligible answer ÷ questions with at least one eligible answer × 100	Per question; do not confuse it with answer rate
Citation share among tracked brands	Brand-answer citation occurrences for you ÷ occurrences across the defined brand set × 100	Competitive share; disclose the brands included
AI-answer availability	Searches showing an AI answer ÷ successful observed searches × 100	Per search surface; report beside citation rate

For competitive share, count each brand at most once per answer even if it receives several links. Multiple brands may occur in one answer. This share describes your selected competitor set; it is not a measure of the entire market.

When a denominator is zero, report the rate as unavailable with its count, rather than as zero visibility.

“Share of answer” is often used loosely. State whether it means cited answers, named brands, citation links or question coverage. Never switch the definition between reporting periods.

A worked example

Suppose you plan 30 observations for one engine. Two requests fail, eight successful searches show no AI answer and 20 produce eligible answers. Six eligible answers cite your website and four name it without citing it.

• Answer citation rate: 6 ÷ 20 = 30%.
• Named-only rate: 4 ÷ 20 = 20%.
• AI-answer availability: 20 ÷ 28 = 71.4%.
• Request failures: 2 of 30 planned observations, reported separately.

These are illustrative figures, not Vercatus results. Dividing six citations by all 30 planned observations would produce 20%, but that would be a different metric mixing answer outcomes, availability and collection failures.

Question citation coverage cannot be derived from these totals. It also needs the mapping from each observation to its question ID. Preserve that mapping in the tracking file.

4. Capture evidence, not just scores

Use the CSV template with one row per question-engine observation. It contains fields for the context, outcome, cited and retrieved URLs, answer text, competitors and the version of your target page. It is a manual evidence log, not a Vercatus prompt-import file.

Store an answer snapshot where permitted by the service you are using, or a reference to one in your evidence system. Preserve citations as returned, then apply documented URL normalization for analysis. Mark retrieval “unknown” when the interface does not expose it; an unavailable trace does not mean the engine retrieved nothing.

When a score moves, inspect the changed rows. Did the engine switch sources? Did it mention a competitor? Did the question stop producing an AI answer? A headline percentage can conceal all three.

5. Repeat the same panel and keep cohorts separate

Repeat observations on a consistent schedule and after material content changes. Answers vary, so show counts and the observation period alongside percentages. Small panels deserve cautious interpretation; one extra citation among five eligible answers changes the rate by 20 percentage points.

Compare engine-specific trends before an aggregate. An aggregate can move because one engine answered more questions this week. If you need a combined view, define a consistent weighting and report engine availability alongside it.

Keep new questions in a separate cohort initially. Archive retired ones without deleting their historical answers. When country, language or interface changes, mark the break in the series.

Vercatus’s poll asks each topic’s questions on a schedule and keeps the answers and their sources. Answer Share keeps cited and named outcomes apart. A manual panel should preserve the same distinctions even if it has fewer questions.

6. Connect measurement to a content decision

For questions where you are absent, check access, discovery and answer coverage. For questions where you are named without a citation, examine whether your own page supplies a clear, substantiated answer. Read cited competitor pages to identify the contribution missing from yours.

Assign one owner and one canonical target page per content task. Record the change date and a specific hypothesis, such as “adding the deployment checklist will answer the implementation question.” Compare the same cohort afterwards. A before-and-after movement alone cannot prove that your edit caused the change.

Keep traffic and business outcomes separate. A citation is source attribution; it is not a visit, qualified lead or sale. Track referrals and conversions through your analytics, then assess whether the pages receiving citations also help visitors complete a valuable task.

Common questions

Is a brand mention the same as a citation?

No. A mention names the brand; a citation attributes the answer to a source. Track mentions without owned-site citations as their own outcome.

Should “not asked” count as zero visibility?

No. It means there is no observation. Show measurement coverage separately so an untested question cannot look like a proven absence.

Can I compare scores from two tracking tools?

Only after checking their prompts, interfaces, locations, sampling periods and metric definitions. Identical percentage labels can hide different denominators.

How many questions do I need?

Enough to represent the decisions you are trying to improve, with a panel you can repeat consistently. Start with a manageable cohort across intent types and expand deliberately. The question mix and evidence quality matter more than a large unstructured total.

Evaluate a tracking tool

The evidence behind this

• Retrieved is not attributed. Engines expose two different things: the sources they retrieved or considered, and the sources a given passage of the answer is attributed to. A source retrieved but not attributed was not cited for that passage (grounding and attribution, grade A).
• Bing publishes its own record. Bing Webmaster Tools’ AI Performance report (public preview from February 2026) shows total citations, average cited pages per day, a sample of the grounding queries and page-level citation activity across Copilot and Bing summaries (Bing AI Performance, grade A). It is the engine-side counterpart to a panel you run yourself.
• Few teams measure at all. In a small survey of 100 practitioners, 14% said they track AI search citations (the attribution gap, grade D); treat it as a reason to write the method down, not as a benchmark.

Use our AI visibility tool comparison to check measurement scope, evidence access and pricing units before choosing a platform.