Citation Patterns (Ultra)

See which content signals are getting cited by AI engines for each of your monitored prompts — and compare your own pages against the patterns.

Citation Patterns

Ultra & Enterprise only. Citation Patterns is a multimodal analysis layer that runs on top of prompt monitoring. It tells you why the AI engines cited the sources they did — and what's different about your own content.

What it does

Every time the prompt-monitor scrapes a prompt across the 7 AI engines (ChatGPT, Perplexity, Gemini, Claude, Grok, Google AIO, Copilot), the Citation Pattern agent picks up where the scrape leaves off:

  1. Reads every cited source — not just the URL, but the actual content. HTML pages, YouTube videos (audio + visual), images (vision-mode), and PDFs are all supported.
  2. Cross-references against the AI engine's response — to infer the specific signals that likely caused each citation.
  3. Compares against your own content — using either an explicit asset URL you provide, your latest GEO audit, or a sample of your top pages.
  4. Produces actionable recommendations — phrased in your brand voice and grounded in specific cited exemplars.

Where to find it

In Prompt Monitoring, click the ✨ sparkle icon next to any prompt's citation count. A modal opens with the latest snapshot.

What you'll see

Signal frequency

A horizontal bar chart of the signals that appear across the cited sources for this prompt. Red bars are gaps where the signal is frequent but your content lacks it. Green bars are signals you already have.

Signals come from a closed, curated vocabulary that includes:

  • Text signals: original statistics, freshness/recency, schema markup, author expertise, structured lists, expert quotes, editorial authority, outbound citations.
  • Visual signals: charts or graphs, infographics, diagrams, product screenshots, branded visuals.
  • Video signals: tutorial format, demonstrations, transcript match to the prompt, channel authority, view-count weight, recency.

Platform preferences

A small heatmap showing which signals each engine tends to favour for this specific prompt. ChatGPT might lean on editorial authority while Perplexity prefers original statistics. Use this to pick which signal to chase per engine.

Brand gaps

A ranked list of signals where your content trails the cited sources. Each gap has a severity badge (high / medium / low), an indicator showing whether competitors have it, and a one-line recommendation.

Recommendations

3–5 short, opinionated actions. Each one links to exemplar URLs from the cited sources so you can see what "good" looks like.

Top cited domains

The most-cited domains across this prompt's responses, with each domain's source type (UGC / Editorial / Corporate / Competitor / Other) and a media-type hint icon.

The asset-override workbook

The default brand baseline is automatic — it uses your GEO audit if available, otherwise it samples a few of your top pages. But for the highest-quality recommendations, you should tell the agent exactly which of your assets to compare against the cited sources.

In the modal, look for the "Not seeing precise gaps?" form. Paste 1–5 HTTPS URLs of your matching content (a product page, a whitepaper, a blog post — whatever you think should be getting cited for this prompt). Click Recalculate.

The agent will:

  1. Persist the URLs onto the prompt so future scrapes reuse them.
  2. Fetch those specific pages, extract signal presence, and feed it into the comparison.
  3. Reuse all the cited-source analyses already in cache (so the recalculate is fast — usually 5–15 seconds).
  4. Return asset-specific recommendations: "On your asset at <url> you lack X, while 78% of cited sources have it."

This converts Citation Patterns from a passive dashboard into an interactive content workbook.

How often it runs

Citation Patterns aggregates automatically after every monitored-prompt scrape, provided at least 2 of the 7 AI engines returned usable responses. Snapshots are idempotent by UTC date — running twice on the same day overwrites the morning's snapshot rather than duplicating it.

On-demand recalculates via the asset-override form bypass the daily cadence — they invoke the agent synchronously and produce a fresh snapshot in seconds.

What runs under the hood

  • Cited HTML, PDFs, and YouTube-with-metadata → Gemini 3 Flash (text mode)
  • Cited images → Gemini 3 Flash Image (vision mode, with the image bytes attached)
  • Cited YouTube videos → Gemini 3 Pro with native video understanding (the URL is passed via file_data.file_uri — Gemini processes audio + visual at 1 fps + 1 Kbps directly, no transcript scraping)
  • Per-prompt + topic-level synthesis → Gemini 3 Pro

All cited-source feature extraction is cached globally across the customer base (one fetch + extraction per URL per week, no matter how many tenants cite it), while the per-tenant reasoning is computed against your specific responses and brand context.