AI Answer Citation and Brand Presence Auditor

What it does

Audit AI answers you already collected or are authorized to analyze. Compare literal brand mentions with the domains linked or separately supplied as citations. Each observation retains URL provenance and original text evidence. A report highlights mentions without a tracked first-party citation for human review. This is useful as a dataset QA and reporting step for GEO/AEO and SEO agencies; it does not generate AI answers or query AI providers.

Input

Supply answers and brands as JSON arrays. Each answer has a unique id and nonempty answer, plus optional query, provider, model, observedAt and citations. Missing metadata stays null or UNSPECIFIED and sourceMetadataVerified remains false. observedAt, when provided, requires a real calendar date and ISO timestamp with seconds and a timezone. Provider labels do not verify which model generated the answer.

Each brand has name, optional aliases and bare domains. Aliases shared across brands are rejected. With includeSubdomains enabled, cross-brand parent/child domain overlap is also rejected. Unicode domains use URL-standard ASCII normalization. www and trailing DNS dots are normalized for domain attribution, with exact suffix boundaries. A hostname such as brand.example.evil.test is not brand.example.

Output

One dataset report contains answer-level observations, brand summaries, separate supplied provider/model groups, source URL candidates, gap review items and limitations. AUDIT is complete JSON, CITATIONS_CSV is a spreadsheet-safe comparison table and REPORT is a readable Markdown report. Each brand observation is MENTION_AND_CITATION, MENTION_ONLY, CITATION_ONLY or ABSENT.

Mention matching uses literal NFKC-normalized aliases with Unicode letter/number boundaries. URL destinations are masked for prose matching; Markdown link labels remain visible. Overlapping aliases for the same brand count once. Up to three excerpts per answer and brand are copied from the original input with original offsets, while mentionCount includes all non-overlapping matches. Negative statements still contain literal mentions; no sentiment, endorsement or ranking is inferred.

Links in Markdown, reference lists and raw text are URL candidates, not verified supporting evidence. Supplied citations remain explicitly supplied. URL fragments are removed for deduplication, query strings retained, and both inline and supplied provenance survive deduplication. The same domain repeated within one answer does not inflate answer-level coverage. Malformed URLs produce review data; structural input errors fail before a billable report.

sameSuppliedQueryMultiset is true only for multiple groups with exactly identical supplied query strings and sample multiplicities. Missing or different queries keep it false. controlledExperimentVerified is always false because dates, settings and capture conditions are unverified. Rates describe the supplied sample only.

Pricing and cost control

The report price is USD 0.03 per completed citation-report, for up to200 supplied answers and20 tracked brands. Always verify the current Store price before running. The planning cost estimate is not a cloud measurement. Collection of upstream AI answers is outside this product and can have its own costs. This Actor performs no model inference, paid API calls, provider scraping, proxy use or outbound data requests.

The total input limit is3MB, each answer30000 characters, combined inline and supplied URL candidates100 per answer, up to10 aliases and10 domains per brand. maxAnswers can lower the limit. The serialized report has a7MB safety cap because repeated evidence can expand beyond input size; split unusually large reports into smaller batches. Overflow fails before exports and without partial report billing. Exports are saved before the single report emission. A valid sample with no tracked mentions still returns one report.

Limitations and fair use

This is an offline pilot, not a provider-wide visibility tracker. It cannot verify link availability, page content, citation entailment, source accuracy, model identity or domain ownership. Inline links may be examples or incidental URLs. A mention without a tracked citation is a review observation, not proof of a marketing failure or an SEO recommendation by itself. Short aliases can create incidental matches.

Input and reports use Apify run storage under its retention settings. Do not upload secrets, private chat exports or personal customer records without authorization. Nothing is sent to a third-party AI provider. No ads, budgets, accounts or customer websites are changed.

Example

examples/input.json contains two synthetic answers to the same query from two explicitly fictional engine labels and two tracked brands. It yields two unique answer-level URL candidates and three mention-only review items. Run npm ci --ignore-scripts, npm test, then the existing factory release.mjs --local-run --write-example. The source remains in this Actor directory; Apifi boty is a generated view.