How Cite actually tracks citations.

Cite reads four sources in two categories: answer engines it asks directly, and the search consoles you connect.

These are supported source types, not a promise that every source was checked for an account. Provider credentials and active console connections determine availability; inspect the latest successful check and any failures in /cite/status.

Two signals count as a citation:

1. Direct citation — the LLM cites a URL whose domain matches the founder's domain
2. Earned-media citation — the LLM cites a third-party URL whose page mentions the founder (brand-name substring or outbound link to founder's domain)

Per the 2026 5W study, 85.5% of AI citations come from earned media. Excluding the second signal would systematically underestimate citation share by ~6×. Cite tracks both and visualizes them separately on the Status surface.

Per-platform share of voice is computed over a 30-day window with normal-approximation 95% confidence intervals:

p   = cited / samples
SE  = √(p(1-p) / n)
CI  = [p - 1.96·SE, p + 1.96·SE], clamped to [0, 1]

When fewer than 10 samples exist for a platform, Cite flags the row as low confidence. We'd rather show you the uncertainty than read noise as signal.

This is the same methodology Profound, Otterly, and the broader GEO measurement community converged on in 2026 (Maximus Labs et al.). The math is published; the parameters (1.96, a 10-sample threshold, a 30-day window) are fixed constants in Cite's share-of-voice code, the same for every customer.

When a founder publishes an article, they pick 1–3 anchor queries the article is meant to win. ChatGPT and Claude take turns, one a day: each asks your tracked queries every other day, up to 30 queries a turn, the longest-waiting first.

The first time a ChatGPT or Claude answer cites a page on the founder's domain for an anchor query, Cite stamps the article with first_cited_at, llm_source, and position. The stamp is never overwritten on subsequent polls.

The first-citation moment fires a notification email (opt-out via /cite/settings/notifications) and surfaces on the Status panel with a time-to-first-citation calculation against the article's publish date.

For every cited URL that isn't the founder's domain, Cite fetches the page and scans for:

• Brand mentions (case-insensitive word-boundary match on the founder's SLD)
• Outbound links to the founder's domain (href patterns)

Hits are stamped on the cited source with earned_media: true, a mention count, and a short excerpt. Aggregated into earned_media_count on the citation row for the "Mentions you" UI badge.

Performance: parallel fetch with a per-URL 5s timeout, max 8 concurrent. Polite user-agent FeedSquadCite/1.0. No HTML parser dependency — regex strip + plain-text scan catches ~95% of mentions; the 5% buried in JS-rendered DOM aren't citation-worthy anyway.

Honest disclosure of current gaps:

• Daily or intra-day sampling per engine — ChatGPT and Claude take turns, so each engine asks a query at most once every other day, and with more than 30 tracked queries some wait longer. The 95% CI is computed over the samples in the 30-day window. Real share fluctuates faster than that; the CI captures most but not all of that variance.
• Per-query share-of-voice — current SoV is aggregated per platform. Per-query SoV (which query has the lowest founder share?) is on the roadmap.
• Competitor SoV alongside founder SoV — we track which competitors are cited (via the displacement surface) but don't render a side-by-side board-deck chart yet.
• Time-series trend lines — current state, not trends. Roadmap.

Methodology is code, not marketing. The parts that do the work:

• A citation tracker that schedules every check and keeps every result
• One reader per source: ChatGPT, Claude, and your Google Search Console and Bing Webmaster Tools data
• An earned-media detector that spots citations of other sites that mention you
• The share-of-voice calculation described in section 3

If a technical buyer wants to audit any of this before signing up, point them here. If they want to read the actual code, we'll share it.