How We Track
AI Visibility

The prompt text is appended with an instruction to name specific companies and include URLs where available. We use temperature 0.3 to minimise hallucination variance. Maximum 900 tokens — enough for a detailed recommendation list.

"{prompt}? Please mention specific companies, brands, or websites by name where relevant. Include URLs where available."

Three parallel checks determine whether the entity is mentioned: exact bare-domain match in the response text, brand-name substring match (minimum 3 characters to avoid false positives on acronyms), and a word-boundary regex for the domain base label.

mentioned =
  lower.includes(bareDomain) ||
  lower.includes(brandLower) ||
  /\b{domainBase}\b/i.test(lower)

The response is split by newline. Any line matching a numbered list pattern ("1.", "2)", etc.) increments a counter. The first numbered line that contains the brand name becomes the rank position. If mentioned but not in a numbered list, rank defaults to 1.

for line in response.split("\n"):
  if /^\s*\d+[.)]/.test(line): count++
  if line contains brand: rank = count; break
if mentioned && rank == null: rank = 1

A 450-character context window around the first brand mention is scanned against two word lists. Presence of any negative signal word takes priority over positive. If neither list matches, sentiment is neutral. Sentiment is only computed when the brand is mentioned.

window = text[idx-150 : idx+300]
sentiment =
  NEGATIVE.some(w => window.includes(w)) ? "negative"
  POSITIVE.some(w => window.includes(w)) ? "positive"
  : "neutral"

Two regex passes extract URLs. First pass captures markdown-formatted links [text](url) and preserves the anchor text as the citation name. Second pass captures bare https:// URLs not already captured. Output is capped at 6 citations per response to focus on the most prominent references.

// Pass 1 — markdown links
/\[([^\]]+)\]\((https?:\/\/[^\s)]+)\)/g
// Pass 2 — bare URLs (deduped)
/https?:\/\/[a-zA-Z0-9-.]+\.[a-z]{2,}[^\s)]*/g

Each known competitor domain is tested against the response using the same three-check pattern used for the entity itself. Competitors are stored as a JSON array in the prompt result and their count is written to the rank-history snapshot for trend tracking.

competitors = monitor.competitors.filter(comp =>
  lower.includes(bareDomain(comp)) ||
  /\b{base}\b/i.test(lower)
)

All results for a prompt are written in a single Promise.all: a PromptResult upsert (deduplicated by promptId + engine + cachedDate), a PromptRankHistory insert (immutable historical record), and the Citation Pipeline (see below). The three writes run in parallel — if one fails, the batch logs the error and continues to the next prompt.

await Promise.all([
  prisma.promptResult.upsert({ ... }),
  prisma.promptRankHistory.create({ ... }),
  processCitationPipeline({ ... }),
])