Measuring GEO: Tracking Citations in ChatGPT, Perplexity, and Claude

Classic SEO measurement leans on a stable substrate: Google Search Console gives you impressions, clicks, average position, and CTR; rank trackers ping Google's SERP on a fixed query set; GA4 attributes referrals. For Generative Engine Optimization there is no equivalent in 2026.

Four structural problems compound:

• No console. Neither OpenAI, Anthropic, Perplexity, nor xAI ships a webmaster console that reports which queries surfaced your domain.
• No SERP. AI answers are synthesized — there is no consistent "position 1" to scrape.
• Query-dependent rendering. The same question worded slightly differently can produce a completely different citation set.
• Personalization and time decay. Logged-in ChatGPT users with memory get different answers; results shift week to week as models retrain and live web tools update.

The consequence: any number you put on a GEO dashboard is directional, not exact. Build the measurement system with that honesty baked in, and you'll keep credibility with leadership.

There is no Search Console for AI engines, so a market of third-party trackers has sprung up. As of mid-2026 the recognizable names include Profound, Ahrefs Brand Radar, Otterly.ai, ScrubbedAI, Goodie, Surfer's AI Visibility tracker, and Peec AI. Expect this list to churn — half will be acquired or pivot within 18 months.

What each does, broadly:

• Query bank execution: run a defined set of prompts against ChatGPT, Perplexity, Claude (with web search), Google AI Overviews, Grok, and sometimes Gemini.
• Citation extraction: parse the cited URLs and surfaced brand mentions.
• Aggregation: roll up share-of-voice, citation rate, and competitor breakdown.

The ones that scrape via the UI tend to break on engine UI changes; the ones using official APIs (where available) are more stable but miss UI-only features. Most serious teams use one paid tool plus an in-house pipeline for the queries that matter most, rather than trusting a single vendor.

Pick a small KPI set you can defend. Bloated dashboards die.

• Citation rate — of the queries in your bank, what % return at least one citation from your domain? Simple, robust, comparable across engines.
• Share of voice — of all citations across your tracked queries, what % point to your domain vs each competitor's? This is the closest analogue to organic market share.
• Citation context sentiment — when you are cited, is the surrounding generated sentence positive, neutral, or negative? Critical for crypto where "X exchange was investigated by the SEC" is a citation you'd rather not have.
• Query-set coverage — of the buyer-intent queries you care about, on how many engines do you appear at all? Coverage matrices catch the case where you dominate ChatGPT but are invisible in Perplexity.

Resist adding 15 more. Each additional metric increases noise faster than signal at GEO's current measurement precision.

Your query bank is the foundation. Treat it as a versioned asset checked into git, not a Google Sheet that drifts.

Three layers:

1. Long-tail informational ("how does staking work on a centralized exchange", "what is a maker-taker fee"). High volume of variations. Test your educational/SEO content moat.
2. Comparison intent ("Coinbase vs Kraken fees", "best crypto exchange for European users 2026"). Highest commercial value. Most contested.
3. Buyer/brand intent ("is [your brand] regulated in Germany", "how to withdraw EUR from [your brand]"). Reputation-sensitive; also where competitor mentions in your brand queries hurt most.

Target ~300–800 prompts total. Refresh quarterly. Document the intent for each query so when sentiment regresses on one, you can tell whether it's a content gap, a reputation issue, or just a model update. Track the same bank verbatim across all engines so cross-engine comparisons are meaningful.

Three execution paths, each with tradeoffs:

• Official API queries (OpenAI's responses API with web_search tool, Perplexity API, Anthropic's web_search tool). Cleanest, most reproducible, but API behavior can differ from the consumer UI users actually experience.
• UI scraping via headless browsers (Playwright + per-engine session management). Captures the real user experience including UI-only citation widgets. Fragile, expensive at volume, and brushes against ToS — vendors that do this for you are absorbing the legal risk on your behalf.
• Partnered data — agreements with engines or aggregators to receive sampled query-response data. Rare, expensive, but the only way to get population-level (not synthetic-query-level) signal.

Practical setup for a serious team: API where available for the bulk of the bank, UI scraping for ~10% of the highest-value queries to validate that API and UI agree, partnered data for none until you have at least a year of API/UI history to anchor against. Sample each query 3–5 times per cycle to estimate variance — single shots are unreliable.

Closed-loop pipeline from query bank to content update.

When ChatGPT cites coindesk.com/article-about-your-exchange, that's third-party coverage — useful but not your own content. When Perplexity cites your blog post, that's a direct content asset. Your dashboard should distinguish them.

A workable attribution taxonomy:

• Owned: cited URL is on your domain or a confirmed subsidiary.
• Earned: cited URL is a third-party publication mentioning you.
• Paid: cited URL is sponsored content or affiliate.
• Adversarial: cited URL is a competitor explicitly framed against you, or a regulator/lawsuit page.

The practical wrinkle: many AI engines cite a homepage or category page rather than the deep article they actually drew from. Use embedding similarity between the cited text and your full site content to map a citation back to the underlying source page with higher resolution. Without that mapping, you can't tell which content actually earned the citation, which means you can't replicate it.

Citations from engines with web search show up in real time. Inclusion in the base model's training data — which drives the answers when web search is off — is slower and harder to observe.

Proxies that actually work in 2026:

• Common Crawl monitoring: subscribe to the monthly index and confirm your key pages are crawled. Most major training corpora draw from Common Crawl. If you're not in CC, you're probably not in the next training run.
• AI bot allowlist via robots.txt + llms.txt: explicitly allow GPTBot, ClaudeBot, PerplexityBot, Google-Extended, Bytespider. Publish an /llms.txt summarizing your authoritative content (the convention proposed in 2024 and adopted by major engines through 2025–2026).
• Direct memorization probes: occasionally ask each model, without web search, factual questions only your site documents (e.g., your exact fee schedule, a specific governance vote). A correct unprompted answer is evidence of training-data inclusion.

Don't treat any one of these as definitive. Triangulate.

An llms.txt is a plaintext file at your root that tells AI engines what your authoritative content is — analogous to a sitemap, but designed to be read by an LLM, not a crawler.

# Example Exchange

> A regulated crypto exchange operating in 38 jurisdictions, offering spot,
> derivatives, and staking products. Founded 2018. Headquarters Berlin.

## Trust and regulation
- /regulation: regulatory licenses by country
- /security: custody, insurance, audit reports
- /transparency: proof-of-reserves attestations

## Products
- /spot: spot trading
- /derivatives: futures and options
- /staking: supported assets and yields

## Fees
- /fees: maker/taker schedule, withdrawal fees

## Learn
- /learn: educational content (300+ articles)

Keep it under ~1,500 tokens. Link to canonical, evergreen URLs. Update it when product surfaces or licenses change. It is not a replacement for good on-page content; it is a high-signal index that the AI engine can pull cheaply when answering questions about you.

The fastest way to lose executive trust on GEO is to present numbers as if they had SEO-level precision. They don't, and analytics-literate stakeholders will catch it.

A defensible monthly dashboard:

• Citation rate, by engine, with a 4-week rolling average — surface the noise rather than hide it.
• Share of voice on top-30 commercial queries, side by side with your top 3 competitors.
• Sentiment trendline with annotations linking to whatever real-world event (regulator action, exchange outage) caused movement.
• Coverage matrix (queries × engines), heat-mapped.
• Variance band per metric — if you sample each query 3 times and the citation set differs across runs, that variance belongs on the chart.

What to not report: exact ranks ("#2 in ChatGPT"), conversion attribution from AI citations to revenue (unreliable in 2026), and any single-week movement smaller than your variance band. Frame GEO as competitive intelligence and content prioritization input, not as an end-of-funnel revenue channel — yet.

Five mistakes that repeatedly bite GEO programs:

• Vendor lock-in to a tool that breaks. Engine UIs change; vendors lose access. Always export raw query/response logs so you can rebuild metrics if your tool dies.
• Single-shot sampling. One run of one query is noise. Three to five samples per query per cycle, average and report variance.
• Confusing brand mentions with citations. "As one major exchange has done..." is not a citation. Your parser must require an actual linked URL or explicit named mention.
• Reporting only what's good. If sentiment is dropping on "is X exchange safe?", report it loudly with a content plan attached. Leadership trusts dashboards that surface problems before they have to ask.
• Treating GEO as separate from SEO. Most AI engine citations still flow disproportionately to content that ranks well in classic search. Your GEO and SEO programs should share the same content backbone, not run on parallel teams with parallel KPIs.