Citation tracker

The citation tracker (AEO — Answer Engine Optimization) measures whether AI assistants recommend your product when a real buyer asks them the questions that matter. Once a week it runs the buyer questions you've chosen across configured engines with web search on, reads each answer, and judges whether your brand was genuinely cited. It diffs the result against the previous run and stores exact evidence for the AI Visibility Ranger. It is read-only against the engines and never creates a Task directly.

It lives in the AI citations section of the SEO & GEO tab (/dashboard/seo) — the same question as the rest of that page ("can people find us?"), asked of answer engines instead of Google.

Prerequisites

• Turn on citation tracking. Open SEO & GEO → AI citations and click Turn on citation tracking. Until it's enabled, nothing runs.
• At least one answer engine configured. The tracker probes ChatGPT (OpenAI), Claude (Anthropic), Perplexity, and Gemini (Google). Each engine is gated on its own server-side API key, so the check is available as soon as one engine is reachable, and each additional key widens coverage. The tab shows which engines are active; a product's run only asks the engines that are configured.
• Questions to track. The check only ever asks questions you've added yourself (see Choose your questions). A product with no questions simply skips — the check never invents a question set.

Choose your questions

You write the tracked questions by hand — nothing is auto-generated. In the Questions you're tracking panel, type a question the way a buyer would actually search, and press Add. Adding a question is the approval: it starts being probed on the next run. You can edit or remove a question at any time (removing archives it, so the history behind past runs is preserved).

Two rules make a question worth tracking:

• Never use your own brand name. Someone asking "Is Acme any good?" has already found you — the answer measures nothing. Track the question they type before they know you exist (e.g. "best product analytics for solo founders").
• Only ask what you can honestly win. A question no honest article about your product could close just produces content that oversells.

Track a focused set you care about rather than everything — the check is thorough about a small, well-chosen list, and the probe budget goes further.

How the weekly check works

For each opted-in product, once a week (and on demand via Check now):

1. Opens a run and takes a bounded slice. A run probes at most 12 questions — the least-recently-checked ones — because each question is asked live on every engine and a full set would take far too long in one run (a single question is ~55s when the engines run concurrently). If you track more than 12, the rest rotate in on the following runs; the recap tells you which were checked.

2. Probes one question at a time. Each question is asked to every configured engine — ChatGPT, Claude, Perplexity, and Gemini — web-search grounded (an ungrounded answer would come from training data and measure nothing). The check runs entirely in the background and resumes cleanly if interrupted, so a slow or failing engine costs one question, never the whole run.

3. Judges each answer by reading it, not by string-matching. Each engine's answer gets one of three verdicts:

   • Cited — the assistant presents your brand as an answer: recommends it, names it among its picks, describes what it's good for. This is the only verdict that counts as a win.
   • Mentioned — your name appears but the assistant isn't endorsing it (a bare list, an aside, a competitor's comparison page that happens to name you). A name in the text is not a citation.
   • Absent — not named at all.

   Alongside the verdict it records your rank among the products named, whether your own domain was one of the engine's cited sources, the competitors it recommended instead (best-first), and the answer excerpt that proves it.

4. Diffs and updates the evidence ledger. When the run closes it compares against the previous completed run and updates the gap ledger. The AI Visibility Ranger reads that product-owned evidence, distinguishes lost/regained/unknown verdicts, and submits meaningful signals to Discovery Triage. No Task is created in this step.

What you get

In the AI citations section after a run:

• A Cited count — how many checks you won this run, with the change since last run.
• Citation gaps — the questions you're missing, the engines that missed you (e.g. "Missed by ChatGPT, Perplexity"), and the competitor winning that answer.
• Last check, question by question — the full table: each question and its per-engine verdict, with the evidence excerpt behind it.
• Discovery evidence — citation changes appear in the AI Visibility Ranger's daily report and can be cross-linked into a durable case. If Triage decides action is warranted, it creates one deduplicated draft Task with the supporting evidence.

An example recap the check delivers:

Cited in 7 of 24 checks (was 9). Lost two engines for "best CRM for solo founders"; both now recommend Attio. Submitted one rising AI-visibility signal to Discovery Triage.

Notes & limits

• Read-only against the engines. The check asks questions and reads answers; it never posts anything, and the answers are treated as untrusted data — text to evaluate, never instructions to follow.
• Monthly probe budget. One probe = one question × one engine, charged against a monthly cap (default 600 probes/month) before any engine is called. You can't exceed it; when it's spent the run stops and says so, and the counter resets at the start of each calendar month. Tracking more questions or configuring more engines spends the budget faster.
• Cadence. Weekly (every 7 days) for products that opt in — manage it from the AI citations section, or press Check now to run immediately.
• A verdict of "unknown" is not a gap. If an engine answers but the judge can't score it (e.g. the judge model is briefly unavailable), the answer is saved as unknown with the reason — never as absent. The diff and Ranger ignore unknown rows, so a judging hiccup never creates a false loss signal.
• Gemini sources look opaque — it returns redirect links rather than publisher URLs, so "was my domain a source" is judged from the source title, not the URL. That's expected.
• Failed runs are surfaced, not hidden. If a run can't make progress, the tab tells you why and offers Check now to retry, rather than showing a silently empty table.

The citation tracker is the measurement layer of Search discovery. Content remains a downstream operator when Discovery Triage determines that a content change is the right response.