llm-mention-tracking

LLM Mention Tracking

Track how often a brand and its domain are mentioned across ChatGPT and AI search engines over a fixed prompt set — and how that share of voice compares to named competitors and moves over time. Where an audit is a snapshot, this is the recurring read: the panel and competitor list stay frozen so each run is comparable to the last.

This is an enhanced skill: it reads live public data through UnifAPI. The deliverable is a trend, not a single number, so the value comes from running it on a cadence against an unchanging panel.

Use UnifAPI for live evidence

This is an enhanced skill: it reads live public data through UnifAPI. Use the unifapi skill to connect (OAuth MCP), then discover these GEO operations. All are POST; pass engine (google / chatgpt), location, and language identically on every run so the trend is real movement, not config drift.

• Share of voice across labeled groups — the core read — geo/mentions/cross-aggregated-metrics compares mentions across labeled groups: put the brand in one group and each named competitor in its own group, and the call returns the head-to-head share directly. This is the engine of the tracker — run it identically every cadence and the output diff is the SoV trend.
• Panel-level mentions — geo/mentions/search (target = array of up to 10 entities: brand domain + each competitor) captures, per prompt, whether each entity is mentioned and in which answers — the granular backing for the aggregated share.
• Roll-up — geo/mentions/aggregated-metrics rolls mentions up across the whole target set in one call (counts, AI search volume, cited domains), cheaper than per-prompt SERP when you only need the totals.
• Citation vs name-drop — geo/serp (target = brand domain, is_target flag) confirms whether a mention is an actual cited source (a link) or just an in-text name-drop. Citations are the stronger signal; track them on a separate line.
• Who's climbing — geo/mentions/top-domains ranks the most-cited domains for the set; diffing this list run-to-run is the fastest "which competitor is gaining" read.
• Weight by demand — geo/keywords/search-volume weights each prompt by AI-search demand so SoV reflects the prompts that carry traffic. Pull once at panel creation and reuse across runs (re-pull quarterly).

UnifAPI reads public data only. Keep each billing block for the cost line.

Workflow

(Read .agents/product-marketing.md / .claude/product-marketing.md first if it exists — reuse the brand, domain, and competitor set so the panel matches everything else.)

1. Fix the panel — once, then freeze it. Lock a stable set of prompts (10–30), the brand + domain, and the named competitors (≤10 total entities — the mentions/search target cap), and assign each entity a fixed group label for cross-aggregated-metrics. Store this panel verbatim; it must not change between runs or the trend breaks. Record AI search volume per prompt now (geo/keywords/search-volume) and reuse it.
2. Run the share read. Call geo/mentions/cross-aggregated-metrics with the locked group labels to get brand-vs-each-competitor share in one shot. Back it with geo/mentions/search per prompt for the granular panel and geo/mentions/aggregated-metrics for the roll-up; for each (prompt × platform × entity) record: mentioned (yes/no) and cited-as-source (yes/no via geo/serp/is_target).
3. Compute Share-of-Voice. Apply the SoV rubric below, per platform and overall, separating cited slots from name-drops.
4. Diff against the prior run. Load the last run's stored figures and compute deltas: which prompts gained/lost the brand, which competitor climbed (diff geo/mentions/top-domains), net SoV change. Stamp the run date and note that AI answers vary by session.
5. Persist this run so the next run can diff against it. The trend table grows one column per run; the group labels and prompts never change.

Share-of-Voice scoring

SoV is computed over the fixed panel so it is comparable run to run. For each tracked entity:

mention SoV(entity)  = entity mentions / Σ mentions(all tracked entities)   [per platform + overall]
citation SoV(entity) = entity cited-slots / Σ cited-slots(all tracked entities)
weighted SoV(entity) = Σ_p [entity mentioned on p ? volume_p : 0] / Σ_p [any tracked entity mentioned on p ? volume_p : 0]

Counting rules (keep identical every run):

• Count an entity once per (prompt × platform) even if mentioned by multiple references — avoids verbose-answer inflation.
• A cited slot = is_target/citation link present. A name-drop = named in text, no link. Track both; citation SoV is the headline, mention SoV is the leading indicator.
• Drop (prompt × platform) cells where no AI answer rendered from both numerator and denominator.
• Weighted SoV (by AI search volume) is the primary headline; report unweighted alongside.

Output

Lead with the trend, because the point is movement.

Headline: "Brand citation SoV [this run] X% vs [last run] Y% (Δ +/−Z pts) across N panel prompts; [competitor] climbed from A% to B%." With run date and panel size.

Share-of-Voice — this run:

Entity	Mention SoV	Citation SoV	Google AI	ChatGPT	Weighted (by AI vol)
our brand	24%	18%	21%	14%	16%
competitorA	41%	47%	44%	51%	49%
competitorB	35%	35%	35%	35%	35%

Trend vs prior runs (one column per run; this is the deliverable):

Entity	2026-03-04	2026-04-04	2026-05-04	2026-06-04	Δ last run
our brand	12%	14%	16%	18%	+2 pts
competitorA	52%	50%	48%	47%	−1 pt
competitorB	36%	36%	36%	35%	−0 pts

Movement notes: prompts the brand gained/lost this run, which competitor is climbing (from top-domains diff), and the net SoV change.

Quick-win list: prompts where the brand is mentioned but not cited (name-drop only) — the model knows you, you just need extractable structure to earn the link.

Cost: UnifAPI records consumed, or best estimate. Each figure carries its run date.

Worked example

Panel: 20 prompts, brand acme.dev, competitors g2.com, rivalapp.com, US/English, monthly cadence. June run via geo/mentions/cross-aggregated-metrics (three frozen group labels) + geo/mentions/search + geo/serp: Acme cited in 9 of 50 tracked cited slots → citation SoV 18%, up from 16% in May (+2 pts) — driven by gaining "best [category] for startups" after a content fix. g2.com slipped 48%→47% on the top-domains diff. Three prompts show Acme name-dropped but uncited → quick-win extractability list.

Guardrails

• Read-only: it measures and reports a trend. It never posts, seeds mentions, edits content, or touches any account — the operator's own assistant runs any follow-up.
• AI answers are non-deterministic and vary by session, region, and time. Treat each run as a dated sample; trend across ≥3 runs before declaring a direction — don't over-read a single delta.
• Keep the panel and competitor list constant between runs. Changing prompts, adding a competitor, or switching engine/locale breaks comparability — if the panel must change, reset the trend baseline and say so.
• Separate citation SoV from mention SoV; reporting only one hides half the picture.
• AI search volume is a public-data estimate; ChatGPT engine is US/English only — mark it N/A for other markets rather than zero.

Related Skills

• ai-visibility-audit: the one-time, diagnostic version that explains why mentions are missing and seeds the panel. Its references/geo-methodology.md holds the shared cited-slot counting rules and the miss-cause (Structure / Authority / Presence) decision tree this tracker's quick-win column relies on.
• ai-answer-gap: turn the prompts where the brand is absent into a prioritized content backlist
• unifapi: the shared data skill (connect MCP, discover the GEO operations this tracker reads)