Every AEO vendor pitch opens the same way: "here's a dashboard that shows how often your brand is mentioned in ChatGPT." The dashboards are fine. The problem is the number in them is someone else's number. It comes from a proprietary crawler, can't be tied to a specific URL on your site, doesn't persist if you cancel the subscription, and lives outside your analytics. A year in, you can't answer the simple question: did the article I shipped in Q2 actually drive measurable AI traffic?
The workaround is the one AEO metric that lives in your own property: LLM referral visits. When a user reads a ChatGPT answer, clicks a citation, and lands on one of your pages, their session source is chatgpt.com. GA4 sees that. You already own the data. The only reason most sites aren't reporting on it is that there's no default segment for "AI answer engines" — you have to build one, and the list of hostnames you need to include is longer than it looks.
The GA4 LLM Referral Segment does that. Tick the engines you want to track, name the segment, and the tool emits everything you need to install it: a GA4 audience definition, a Looker Studio filter, a GTM trigger + tag for custom events, and a BigQuery SQL query against the GA4 export. No subscription required, no third-party dashboard, no data leaving your account.
Why the referral segment is the only AEO KPI you can trust
Third-party AEO trackers are not wrong, but the number they report is a sample of a sample. They run a fixed bank of prompts against each LLM on a cadence and see whether your brand shows up. That tells you yes/no on prompts you can't verify match what your actual customers are typing. If you change pages, the metric might move, or it might not — the link between the change and the measurement is loose.
The LLM referral count is tight. A visitor clicked a citation; GA4 saw it; it's attributed to the landing page they arrived at. You can:
- Split the count by landing page.
- See which engines cite you most.
- Compare week-over-week after a content change on a specific URL.
- Correlate the visits with downstream events (signups, purchases, time on page).
That's what a KPI looks like. It's observable, attributable, and in your system of record.
What the tool gives you
The tool's default preset covers the five engines that matter: ChatGPT, Claude, Perplexity, Gemini, Microsoft Copilot (Bing Chat). Those five cover the overwhelming majority of AI citation clicks on an English-speaking commercial site.
Toggle the second tier on if you want fuller coverage: DuckDuckGo Assist, You.com, Phind, Brave Leo, Grok, Meta AI, Kagi Assistant, DeepSeek, Le Chat (Mistral), Poe. Each one adds a hostname to the regex the tool builds. You'll get traffic from a couple of those on a normal week, more if you rank for developer-centric or privacy-centric queries.
Google AI Mode is a separate case. It's also opt-in and disabled by default — its referrer is bare google.com, which is indistinguishable from regular organic Google clicks unless you filter downstream by the udm or srsltid URL parameters. Enable it only if you've set up a downstream filter, otherwise it makes the segment noisy and meaningless.
Every selection regenerates the outputs live across six panes:
- GA4 Audience — the regex to paste in Admin → Audiences → Session source → matches regex, plus a step-by-step for the alternative Explorations Custom Segment route (for teams without Audience creation permissions).
- Looker Studio — the filter definition for any connected GA4 property, plus a calculated field for scorecards so you can drop "LLM Sessions" next to "Total Sessions" on a dashboard.
- GTM Trigger — a Referrer-matches-regex trigger plus a GA4 Event tag firing
llm_referralwithllm_source =so you can use the event as a conversion, funnel step, or custom dimension. - BigQuery SQL — daily LLM-referral sessions by landing page, 90-day lookback, parameterised at the top so you can swap the project/dataset in one line.
- Regex only — three flavours of the same match (session-source regex, full-URL anchored, pipe-separated host list) for tools like Plausible, Fathom, Matomo, Cloudflare Analytics, and your CDN's log filter.
- How to install — the 10-minute setup path and the list of things this segment does NOT catch, so you don't claim credit you can't prove.
The things this segment doesn't capture (say this out loud at the kickoff meeting)
Three gaps, honestly:
- In-app AI answers with no click. If a user asks Claude about your product, reads the answer, and never clicks through, you don't see anything. Brand-mention monitoring from seoClarity / Profound / Bing Webmaster is the complement for that signal.
- Desktop or mobile app traffic. Links clicked from inside the ChatGPT macOS app or the Claude iOS app often arrive as Direct, not referral. The workaround is UTM parameters on links you place yourself — in docs, in agent system prompts, in partner-network landing pages.
- Crawler hits from the AI bots themselves. OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended fetching your pages is a different signal — it tells you whether they can cite you, not whether they did. That's a server-log question, not an analytics-referrer one.
If you want the whole picture, you run LLM-referral segment (user clicked) + brand-mention tracker (LLM said your name) + server-log crawler analysis (bots fetched the page). The segment is one of the three, and the cheapest.
How to actually use it week to week
Install it once. Then every Monday:
- Open the Looker Studio dashboard. Check the LLM-sessions scorecard vs last week.
- Drill into landing pages. Which URL caught the most LLM traffic? Is that the page you wanted to be caught?
- For the top three LLM-cited landing pages, run AI Citation Readiness. The pages that cite best are usually the ones with high chunk-density (paragraphs in the 40-150 word band), semantic HTML, and schema coverage. Copy the pattern.
- For pages that should be cited but aren't, run them through the Mega Analyzer and feed the AI fix prompt into Claude Code. Most of the time the fix is one of: empty SPA shell, paragraphs too thin to retrieve, missing schema, or a citation-hostile
robots.txtthat blocks the AI crawlers.
That's the loop: measure referral visits → find pages that work → replicate the pattern on pages that don't → re-measure. It's the same loop classical SEO uses with rankings, adapted to a KPI that actually reflects how AI answer engines send you traffic.
Where this fits with the other tools
- AI Citation Readiness scores a single URL on 14 AI-visibility signals. Pair it with the referral-segment output to see correlation on your own property.
- Mega Analyzer and Site Analyzer now include a chunk-density + SPA-shell check specifically because those two issues destroy LLM referral traffic. If the segment is flat, run those next.
- Funnel Keyword Audit tells you which kinds of queries you should be getting LLM citations for. Cross-reference the top LLM-referring landing pages with the audit's TOFU/MOFU/BOFU split — you'll see whether the AI traffic matches your intent mix or not.
- Entity Citation Radar surveys 16 high-authority sources for brand-entity density — the signal that leads to being cited in the first place.
The short version
AEO dashboards from paid vendors tell you someone else's number. The GA4 LLM referral segment tells you how many of your site's visits came from an AI answer engine, by landing page, over any timeframe, in your own analytics, for free. The only reason most sites don't have it is that building the list of hostnames is a fifteen-minute task that keeps slipping. This tool does it in thirty seconds and ships four paste-ready outputs. Install it once and stop arguing about which AEO vendor has the best methodology.
Related reading
- The $20 Dollar Agency — Chapter 10 (Analytics That Matter) covers the GA4 audience + Looker Studio dashboard pattern this tool extends to the AI-answer-engine case.
- The $100 Network — Chapter 26 (Monitoring at Scale) on rolling up per-site KPIs into a network-level view. The referral segment scales cleanly across a portfolio if you label each property's KPI the same way.