TL;DR: GEO can't stop at "did an AI cite us." Wire GA4 + Baidu Tongji as dual trackers (two tools: GA4 for overseas AI-engine referrals, Baidu Tongji to cover mainland China), and measure the whole AI citation → landing page → conversion loop — that's how you learn whether visibility actually became business.

Prerequisite: an AI has to read and index you first. If it hasn't yet, clear the first mile → Getting Your Site Read and Indexed by AI Answer Engines.

GEO isn't "did we get cited" — it's "what happened after"

When teams work on GEO (Generative Engine Optimization), the metrics they watch usually stop at the first half of the story: did ChatGPT, Claude, Gemini, Perplexity, DeepSeek and the rest read, index, and cite us? That matters — but it's only visibility, not outcome.

The real questions come next. After an AI cited you, did anyone click through? Once they landed, what did they do? Did any of it turn into a lead, a trial, a paying customer? If all you watch is "were we cited," you're staring at the top of the funnel while the middle and bottom stay a black box.

The way to see that whole chain is analytics instrumentation. This is a hands-on record of how we wired analytics into our own two properties (the 智码航 portal and the 引见 / YinJen marketing site) — and a playbook you can copy for measuring GEO conversion.

The closed loop: AI citation → landing page → conversion

The GEO value chain really has four links, and a break in any one breaks the whole thing:

Readable — crawlers can actually fetch your pages (robots.txt, content signals, indexing).
Cited — an AI mentions or links your brand and content in its answers.
Clicked — a user clicks through from the AI answer onto your site (this leaves a trail in your analytics as "source = perplexity.ai / chatgpt.com …").
Converted — after landing, they complete a key action: book a demo, contact sales, pick a plan, sign up.

The first two links are SEO/GEO engineering. The last two can only be seen with instrumentation. Without analytics, links 3 and 4 are a black box — you can see your rankings and citations, but you'll never know what they're actually worth.

Why we run two trackers: GA4 + Baidu Tongji

We didn't pick one — we run both, because they cover different realities:

Google Analytics 4 (GA4): the most accurate read on referral traffic from overseas AI engines like ChatGPT, Perplexity, Gemini, and Claude, with a flexible event model that lets you flag "key events" (conversions) explicitly. Its weakness: access to google-analytics from mainland China is unreliable, so it drops a slice of domestic data.
Baidu Tongji (百度统计): reliable coverage of mainland traffic, wired into Baidu's indexing and Search Resource Platform; traffic tied to domestic engines like Doubao and Wenxin shows up more completely here.

One overseas, one domestic — complementary. The same instrumentation reports to both at once, and reading them side by side keeps any single channel's blind spot from misleading you.

What we actually track

More events isn't better. Two layers — page views + a handful of key conversion events — are enough. Our list:

Page views (`page_view`): fired on first paint and on every single-page-app (SPA) route change, so every "which page did they see" is recorded.
Key conversion events:
generate_lead — a successful submit of the book-a-demo / contact-sales form (the core lead conversion; flagged as a "key event" in GA4).
pricing_select — clicking a specific plan (carries the plan id), so you can see which pricing tier pulls hardest.
cta_click — clicks on the primary buttons in the hero / nav / footer, measuring their pull.
login — the login entry point being used.
lang_toggle — switching between Chinese and English, an indirect read on audience makeup.
suite_card_click (portal) — tool-card clicks, showing which product the portal funnels traffic toward.

Every event carries an app dimension (portal / web), so the portal and the marketing site can be viewed separately in one backend — or stitched together to see the full "portal → product" path.

Three engineering traps we handled

The worst outcome with analytics is "it's installed but the data is dirty." These are the things we deliberately handled:

Activate only on the real production host. Local development, SSR, and preview environments stay silent — otherwise your own test clicks pollute production and your GEO conversion rate gets watered down by your own traffic. The check is simple: only fire when the host isn't localhost / 127.0.0.1 / a preview domain.
Don't double-count `page_view` in a SPA. Most modern marketing sites are single-page apps; an automatic first-paint hit plus a route-change hit easily counts one visit as two. The fix is to turn off the SDK's automatic first hit and send one yourself on every route change (including the first), so the counts stay honest.
Instrument what you should, skip what you shouldn't. Our ops console (the logged-in internal workspace) is not wired up — it isn't a marketing surface, and tracking it would just mix internal actions into visitor data. Instrumentation should follow "who do you want to measure."

Once it's installed, what to look at

When data starts flowing, the useful views all tie back to the GEO chain:

Source / medium report: filter for referrers like perplexity.ai, chat.openai.com, gemini.google.com — that's real traffic the AI answer engines are sending you.
Conversion funnel: the conversion rate from "AI-referred landing" to generate_lead tells you directly how much your GEO work is worth.
Landing-page performance: the pieces that get cited by AI *and* drive both traffic and conversions are exactly the topics to mine for your next round of GEO.

Copy-paste checklist

Pick platforms: GA4 for overseas audience / conversion modeling, plus Baidu Tongji for reliable domestic coverage — run both.
Track two layers: page views plus a small set of key conversion events; don't over-instrument.
Add a source dimension to every event (which site / which surface) so you can slice it.
Activate only on the production host, send one unified page_view per SPA route, and leave the internal console untracked.
Flag conversion events as "key events," then slice the funnel by AI referrer — that's the step that turns GEO from "looks effective" into "provably valuable."

GEO makes AI see you; analytics makes you see what GEO actually delivers. Do both, and growth finally becomes a closed loop.

Further reading: not indexed by AI yet? Start with GEO Step Zero: Getting Your Site Read and Indexed by AI Answer Engines; to get this measurement loop out of the box, see YinJen.

GEO Needs Analytics Too: Measuring the Full Loop from AI Citation to Real Conversion