Tracking AI citations: the dashboard you actually need.

AI citation tracking is the gap in almost every analytics stack in 2026. Google Search Console tells you how many people clicked from the classical organic SERP. It tells you nothing about the queries where Google AI Overviews, ChatGPT, Perplexity, or Copilot generated an answer that named — or ignored — your brand. That blind spot is where the new top-of-funnel lives.

This article covers what to track, which tools cover which engine surfaces, and how to build a minimal dashboard that produces numbers worth reporting. The goal is a weekly tracking loop you can run in 90 minutes that replaces gut feel with structured data.

Why your current analytics stack is blind to AI citations

The short answer: the tools were built for a world where traffic came from a click.

Google Search Console reports impressions and clicks from the classical blue-link SERP. When a user reads a Google AI Overview and does not click, that interaction registers nowhere in GSC. When a user asks ChatGPT a question and ChatGPT cites your page, that session appears in GA4 as direct traffic (if the user then navigates to your site) or not at all (if they read the answer and move on). The referral header from ChatGPT’s browsing interface is inconsistently passed; Perplexity passes a referral header more reliably, but its traffic is often bucketed into a generic “other” channel.

The effect is structural, not a lag or a temporary gap. Classical analytics was designed to attribute clicks. AI citations increasingly produce no click at all — the user read the answer, formed an opinion about your brand, and moved on. That impression counts for something in any B2B sales motion. It measures nothing in your current stack.

This is the core problem that generative engine optimization introduces to the measurement layer: the distribution channel changed before the measurement tooling caught up.

The 5 metrics that matter for AI citation tracking

Before picking tools, define what you are measuring. Five metrics form the core of any GEO tracking dashboard.

1. Citation share by engine

Citation share is the percentage of your tracked prompt set that returns a response citing your domain, per engine. Tracked separately for Google AI Overviews, ChatGPT, Perplexity, Copilot, and Gemini.

Example: if you track 80 prompts relevant to your category on Perplexity and 22 of them return a citation from your domain, your Perplexity citation share is 27.5%.

Track this weekly per engine. The trend matters more than the absolute number, and the per-engine breakdown reveals which surfaces are growing or shrinking independently.

2. Cited URLs

Which pages on your site are being cited — and in what context. This is the diagnostic layer underneath citation share.

When Perplexity cites you 22 times, it might be citing 3 pages repeatedly (entity authority concentrated in a few pages) or 18 different pages (broad coverage). Both situations call for different interventions. Concentrated citation usually means the uncited pages lack direct-answer openings or structured extraction blocks. Broad coverage with low total share usually means competitor authority is stronger across the board.

The entity-authority concept underpinning this is covered in entity authority for LLMs.

3. Prompt coverage

Prompt coverage answers a different question: of the queries your buyers are asking, how many generate any response that references you?

If your citation share is 27.5% and your prompt coverage is 55%, you are cited in more than half the prompts but often only once per prompt. If your citation share is 27.5% and your prompt coverage is 27.5%, every cited prompt cites you exactly once and many prompts return nothing from your domain at all.

Low prompt coverage on high-commercial-intent queries is the highest-priority fix. The engine is answering your buyers without mentioning you. That is a content and authority gap, not a measurement problem.

4. Share of voice in answer engines

Share of voice (SOV) in AI search benchmarks your citation rate against named competitors across the same prompt set.

If your citation share is 27.5% on 80 prompts, and Competitor A’s share is 41% and Competitor B’s is 19%, your position in the competitive landscape is clear. SOV is the metric that survives a CFO review because it connects to competitive market position, not just internal content performance.

Tracking SOV requires running the same prompts against competitor domains, which most dedicated tools support natively.

5. Sentiment of cited context

When an engine cites your page, what does it say about you? Neutral factual citation (“according to [domain]…”), positive framing (“experts at [domain] recommend…”), or a correction or hedge (“while [domain] suggests X, other sources indicate…”).

Sentiment tracking is the least automated of the five metrics — most tools flag citation without parsing the surrounding context. Manual spot-checks on 10 to 15 cited responses per week are the practical approach for most programmes. Look for patterns: if your brand is cited frequently but almost always in a hedging context, the page content may be generating citation for the wrong reasons.

The tools that actually cover each engine

No single tool covers all five major answer engines with equal fidelity. Here is what each major platform covers as of mid-2026.

Profound

Profound is the most comprehensive dedicated GEO tracking platform currently available. It tracks citation data across Google AI Overviews, ChatGPT, Perplexity, and Copilot from a defined prompt set, surfaces share-of-voice comparisons against competitors, and logs which URLs are cited per prompt and per engine.

The prompt library is the core configuration: you define the queries that matter to your programme, Profound runs them on a scheduled cadence, and the dashboard shows citation trends over time. Prompt sets of 50 to 500 queries are typical for mid-market B2B programmes.

Pricing is not published as a fixed menu — plans are scoped to prompt volume and query cadence. Treat the budget as comparable to a mid-tier rank-tracking subscription.

Otterly.ai

Otterly.ai focuses on Perplexity and ChatGPT citation tracking with a clean interface optimised for marketing teams rather than SEO operators. Its differentiation is automated sentiment flagging — it surfaces responses where your brand is cited in a negative or hedging context, which saves the manual spot-check time.

Where Profound is more complete on engine coverage, Otterly is faster to set up for teams that primarily care about Perplexity and ChatGPT and want the sentiment layer without building it manually.

Peec AI

Peec AI tracks citation share and share of voice across ChatGPT, Perplexity, and Gemini. Its reporting is oriented toward regular competitive benchmarking — the default view shows you and up to four competitors across your tracked prompts.

The Gemini coverage is relevant for programmes targeting Google’s ecosystem beyond AI Overviews, where Profound’s coverage is stronger but Gemini’s conversational surface (as distinct from AI Overviews in Search) requires separate tracking.

Ahrefs Brand Radar

Ahrefs Brand Radar is part of the Ahrefs suite and tracks brand mentions and citations across Google AI Overviews and, in part, Copilot. For teams already inside Ahrefs, it is the lowest-friction entry point — no new tool, no new login, direct integration with the keyword and backlink data you already have.

The limitation is engine breadth: Brand Radar is strongest on the Google AI Overviews surface and does not cover Perplexity or ChatGPT with the same fidelity as Profound or Otterly.

The AI Overview optimization audit covers the content-side work that feeds into what Brand Radar tracks.

DataForSEO AI Optimization API

DataForSEO’s AI Optimization API is not a SaaS dashboard — it is a programmatic API that lets you query ChatGPT and Perplexity response data including citation URLs, on a pay-per-call basis.

The practical use case is filling coverage gaps. If your SaaS tool does not cover a specific engine or a specific prompt set, DataForSEO lets you call that surface directly and process the results in your own reporting layer (Google Sheets, Looker, a custom dashboard). For programmes with engineering bandwidth, it is also the foundation for building a fully custom GEO tracking system.

The API returns citation URLs, the full generated response text, and metadata about which model version was queried. From that raw data you can derive citation share, prompt coverage, and — with some text processing — sentiment.

Manual prompt panels

Manual prompting is undersold. For programmes with fewer than 50 tracked queries and a weekly cadence, running each query by hand in ChatGPT, Perplexity, Google (for AI Overviews), and Copilot and logging results in a spreadsheet is entirely viable.

The spreadsheet structure is simple: one row per prompt, one column per engine, a binary citation flag, and a notes field for the cited URL and any sentiment observation. At 50 prompts across 4 engines, this takes about 90 minutes per week with a consistent reviewer.

The limitation is consistency: manual results vary by query phrasing, the reviewer’s account context, and geographic location. Use a logged-out or incognito session, a fixed phrasing for each query, and the same geographic context each week.

How to build the minimal GEO dashboard

A dashboard that covers the essential metrics without requiring a data engineering project.

The prompt list. Start with 30 to 60 queries. Source them from three places: your top 20 GSC impressions queries, your top 20 sales-call questions (verbatim phrasing from transcripts), and 10 to 20 competitor-brand or category queries. The sales-call queries are the most commercially important. If you cannot identify them from call recordings, use your highest-converting landing page keywords as a proxy.

The chatgpt seo playbook covers prompt selection in more detail for the ChatGPT surface specifically.

Engine selection. For most B2B programmes in 2026, prioritise Google AI Overviews and Perplexity first. AI Overviews dominate Google informational search (Seer Interactive Q1 2026: ~36% of informational queries trigger them). Perplexity has the cleanest citation structure and the highest share of technical and professional queries in its user base. ChatGPT and Copilot come second — add them once the first two are stable.

The weekly loop. Run the prompt set through each engine on the same day each week. Log citation flag, cited URL, and any sentiment observation. Calculate citation share per engine. Calculate SOV if competitors are tracked. Flag any prompt where last week’s result had a citation and this week’s does not — engine citation can drop as well as gain, and early detection matters.

The monthly stakeholder output. Three numbers per engine: citation share (this month vs last month), citation SOV vs top competitor, and top 3 cited URLs. Add a qualitative observation about prompt coverage gaps. This is the format that survives a CFO review because it is competitive, specific, and directional.

What the tracking loop surfaces that you cannot see elsewhere

Three patterns that emerge from a structured GEO tracking loop that are invisible in GSC or GA4.

The authority concentration problem. A tracking loop quickly reveals when one or two pages are doing all the citation work. If 80% of your AI citations come from a single blog post, your GEO programme is fragile. One content update by that page’s author, one change in an engine’s citation weighting, and your citation share halves. The fix is expanding entity coverage to more pages — a content gap that GEO tracking makes obvious and classical analytics obscures.

The prompt phrasing gap. Engines respond differently to subtle changes in phrasing. “Best CRM for B2B SaaS” and “CRM software for B2B companies” may trigger completely different citation sets even though they target the same buyer. A tracking loop with carefully varied phrasing across semantically similar queries reveals which phrasings trigger your citations and which do not. This informs the exact language to use in answer-first openings on your content.

The competitive citation pattern. When a competitor is consistently cited and you are not across a cluster of related prompts, the pattern almost always traces to one of three things: their answer-first opening is tighter, they have a structured extraction block you lack, or their entity coverage on the topic is broader. The GEO tracking loop identifies which cluster the gap belongs to. The fix is the same content work described in AI Overview optimization — but now you know exactly where to apply it.

What not to track

A few metrics that look relevant but consume tracking budget without generating actionable signal.

Raw AI-referred sessions in GA4. The attribution is too unreliable. Perplexity passes a referral header more consistently than ChatGPT, but both are inconsistent enough that session counts should not be the primary measurement layer. Use citation share instead, and treat AI-referred sessions as a directional corroboration, not the source of truth.

Keyword rankings on AI Overview queries. Classical rank tracking and GEO citation tracking overlap but diverge. About 40% of AI Overview cited URLs sit outside the classical top 10 for the query (per Semrush 2025 citation correlation study). Tracking rank position on these queries tells you about classical SERP performance; it does not predict or explain citation status. Keep both but do not conflate them.

Total brand mentions across the web. Brand mention tracking tools (not to be confused with citation tracking) report every occurrence of your brand name across indexed web pages, forums, and social. These mentions do not map cleanly to AI citation. Engines cite specific URLs, not brand name occurrences. A brand mention that does not sit on a well-structured page with entity coverage rarely produces a citation. Track citation share, not mention count.

The 90-minute weekly routine

A concrete implementation for a programme starting from zero.

Week 1 setup (one-time, 3 hours). Build the prompt list (60 queries, three sources as above). Decide on tools: one dedicated SaaS tool as the primary tracker, DataForSEO API for gaps, a spreadsheet as the log. Configure the SaaS tool with your prompt list and up to 3 competitor domains. Set the tracking cadence to weekly.

Weekly run (90 minutes). Review the dashboard for each engine: citation share, SOV vs competitors, any prompts that changed status since last week. Flag the 3 highest-priority uncited prompts on high-commercial-intent queries. Add them to the content fix queue. Pull the top-cited URL per engine and log it in the weekly record.

Monthly output (30 minutes, part of the 90). Aggregate the four weekly logs into a one-page summary: citation share trend per engine, SOV vs competitor over the past 4 weeks, top cited URLs, and 1 to 3 content pages in the fix queue. Share with whoever owns the content programme.

The total effort is under 2 hours per week once the setup is complete. The output is the only structured view of AI citation performance most B2B programmes currently have.

What comes after tracking

Tracking is the measurement layer. The intervention layer is content work: fixing the pages that should be cited but are not, by applying the answer-first opening pattern, adding structured extraction blocks, and building entity coverage. The tracking loop generates the fix queue; the content work fills it.

The two loops reinforce each other. Better content increases citation share. Citation tracking identifies where to focus the content work. Programmes that run both in parallel compound faster than those treating them as separate workstreams.

Start with 30 prompts, one tool, and a weekly log. The patterns become visible within 4 to 6 weeks. From there, expand the prompt set and add competitor benchmarking as the programme matures.

The measurement gap is real and it is current. Most teams running GEO programmes in 2026 are doing the content work without any structured view of citation performance. That is like running paid search without a conversion dashboard — the spend is real but the optimisation signal is missing.