AI Citation
When an AI answer references or links your store as a source.
AI citation reporting becomes meaningless when teams count every visible brand appearance as the same event. A named source attached to an answer is evidence of attribution in one observation; a plain mention, shopping card, or referral visit is a different event with different measurement rules.
Treat citations as sampled interface behavior, not permanent rank. Name the system and mode, preserve the exact prompt and settings, capture the answer and link, record misses, and report a rate with its denominator. That discipline makes changes comparable without pretending a citation proves accuracy, endorsement, influence, clicks, or revenue.
What exactly counts as an AI citation?
An AI citation is visible source attribution or a link associated with a generated answer on a named surface during a recorded observation. The record must connect answer text to a source URL or domain. A brand mention without attribution may matter, but it is not an AI citation.
OpenAI explains that ChatGPT search responses can include linked sources, while Google documents links and supporting material in AI features. Those descriptions establish platform-specific presentation possibilities, not a universal UI definition. Always write what was visible, rather than inferring a hidden retrieval event.
How is a citation different from a mention, link, card, or backlink?
A citation attributes part of a generated answer to a source; the neighboring events do not necessarily do that. A mention names a brand, a product card presents commerce information, a link may sit outside the answer’s attribution, and a backlink is a conventional web-page link. Code each separately.
| Event | Minimum visible evidence | Safe interpretation |
|---|---|---|
| Citation | Source or URL tied to answer text | Attribution in this observation |
| Brand mention | Name without source attachment | Presence, not citation |
| Product card | Shopping module or merchant unit | Commerce display, not necessarily attribution |
| Standalone link | URL without answer-source association | Link exposure only |
| Traditional backlink | Link on a conventional web page | Web link, not AI citation |
A card or citation can also coexist, so use nonexclusive fields. Do not treat any event as endorsement, proof of factual accuracy, fixed rank, or guaranteed click.
Why does citation behavior change between observations?
Citation behavior can change because the system, mode, prompt wording, date, locale, account state, personalization, retrieval results, and interface all vary. The same URL may be cited once and absent later. A screenshot captures one outcome; it does not establish stable visibility or a universal ranking position.
Platform controls differ. OpenAI publishes its crawler and user-agent information, Perplexity documents its bots, and the Robots Exclusion Protocol defines standard crawler directives. Access can support eligibility, but permission does not compel crawling, indexing, retrieval, citation, or display.
What makes an ecommerce source eligible and worth citing?
Eligibility starts with accessible, accurate public content that answers a real question clearly and can be associated with a stable URL and entity. Citation worthiness improves when claims are specific, sourced, current, and distinguishable from sales copy. None of those inputs can force a system to retrieve or cite the page.
- Answer one buyer question directly and support it.
- Use stable URLs and consistent product identity.
- Cite primary evidence for specifications, tests, and policies.
- Disclose original-research sample, method, date, and limits.
- Keep price, availability, variants, and safety facts current.
- Publish genuine reviews and substantiated claims only.
- Separate observation from inference; correlation from causation.
Google’s Search Essentials provide a crawl and indexing baseline. Use Article and Product markup only where visible content supports each property. JSON-LD can describe facts; it cannot cause citation. The FTC’s advertising guidance still governs truthful claims.
How should teams build a reproducible citation sample?
A reproducible sample fixes the questions and observation conditions before results are collected. Build prompts from real buyer intents, cover discovery through comparison, and avoid changing wording after seeing favorable outputs. Include repeated runs and deliberate non-citation recording so the denominator represents the full test, not selected wins.
- Define surface, mode, locale, account, and device.
- Freeze exact prompts, categories, and personalization state.
- Set dates, cadence, repetitions, and unavailable-feature rules.
- Capture answer text, linked URL or domain, and screenshot.
- Code position, type, mention, card, link, and miss separately.
- Preserve raw records before summarizing outcomes.
- Repeat after a documented content or platform change.
Start with AI search visibility measurement and adapt the surface-specific steps in getting cited by Perplexity. The protocol measures observations; it does not reverse-engineer proprietary indexes.
What evidence should each monitoring record contain?
Each record should preserve inputs, the observed answer, attribution details, and downstream business evidence without mixing them. Input evidence makes the run reproducible; sample evidence supports a citation rate; business evidence shows visits or outcomes. No layer alone proves that a citation caused traffic, conversion, or revenue.
| Evidence layer | Record | Supports | Does not prove |
|---|---|---|---|
| Input | Surface, mode, exact prompt, settings, locale, account, date | Reproducibility | What the system indexed |
| Observation or sample | Answer, URL or domain, position, type, screenshot, miss | Counts and rates | Stable rank or endorsement |
| Business | Referral sessions, landing pages, conversions, revenue | Downstream association | Citation causality |
Report citations ÷ eligible observations, with both numerator and denominator, then segment by surface and prompt class. Use the Search Analytics API for search-performance context and Google Analytics traffic-source dimensions for referral evidence, while keeping those datasets distinct from citation observations.
How should citation rates be interpreted and reported?
Citation rates describe a defined sample, not the whole market. State the surface, observation window, prompt set, repetitions, locale, account conditions, and denominator beside every rate. Compare like with like, retain misses, and label changes as associations unless the design supports a stronger causal conclusion.
Separate three questions: Was the brand present? Was a source visibly cited? Did measurable business behavior follow? A rising AI visibility score may summarize a model, while AI search visibility defines broader presence. Neither substitutes for row-level evidence or proves influence.
Publish methodology with any original research: sample construction, exclusions, field dates, coding rules, limitations, and version changes. Avoid “most cited” or “best” claims when the sample cannot support them.
What can StoreCited diagnose, and what cannot it know?
StoreCited can run a point-in-time public storefront readiness diagnostic, highlighting accessible content, crawler signals, structured facts, and evidence gaps it can observe. It does not live-monitor prompts or citations, query proprietary indexes, identify the competitors actually selected by an AI system, or guarantee traffic, conversion, or revenue.
StoreCited is an Organization, not a fabricated Person expert. Run the free StoreCited readiness scan to find public eligibility issues, then use the monitoring protocol above for named surfaces. Treat fixes as controlled inputs and citations as observed outputs; do not promise that one causes the other.