Start with the uncomfortable truth about LLM attribution

LLM traffic is real, but it is not always clean

Before touching GA4 settings, it helps to be brutally honest about what you can and cannot measure. LLM traffic is not a single neat channel. It can arrive as a referral from chatgpt.com, perplexity.ai, copilot.microsoft.com, gemini.google.com, claude.ai, poe.com, you.com, or another AI interface. It can also arrive as direct traffic if the app strips the referrer, if the user copies and pastes the URL, or if the click opens in an environment that does not pass attribution properly.

Then there is the especially annoying case: AI-assisted discovery that ends in a branded Google search. Someone asks Perplexity for best procurement software for mid-market SaaS, sees your name, then searches your brand on Google two minutes later. GA4 records organic search or paid search. The AI assistant influenced the visit, but did not technically refer it.

This is why the goal is not perfect attribution. Perfect attribution is a bedtime story analytics vendors tell when the pipeline target is looking scary. The practical goal is to create a defensible LLM traffic view that captures known AI referrers, preserves campaign data where possible, and gives your team a trend line good enough for decisions.

The market shift is strong enough to justify the setup. McKinsey found regular organizational use of generative AI rose from about 33% in 2023 to around 65% in early 2024. That does not mean 65% of your web sessions come from LLMs. It means your B2B buyers are getting comfortable using these systems during research. If you sell something considered, technical, expensive, or risky, assume AI-assisted research is already happening upstream.

Create a source list before building GA4 rules

The channel is only as good as the pattern list

Do not start inside GA4. Start with a plain source inventory. You want a maintained list of domains and source patterns that commonly represent AI assistant traffic. At minimum, include:

• chatgpt.com
• chat.openai.com
• perplexity.ai
• copilot.microsoft.com
• gemini.google.com
• bard.google.com for older records
• claude.ai
• poe.com
• you.com
• phind.com
• andisearch.com
• komo.ai

You can expand this over time, but do not turn the list into a junk drawer. A common mistake is adding every random AI-adjacent domain and then declaring victory. Keep a notes column with the domain, product name, date added, and reason. This matters later when somebody asks why a weird referral source suddenly appeared in your executive report.

Also separate two concepts: AI assistant referral and AI search visibility. A referral source tells you a user clicked from an AI product to your site. Visibility tells you whether an AI model mentions, cites, summarizes, or recommends your brand in the first place. GA4 handles the first part. It does not handle the second part well, because GA4 only sees visitors after the click. That is where a product like ZenithStack.ai earns its keep: it identifies citation gaps for your brand across ChatGPT, Perplexity, and Gemini, then helps publish proprietary content with human edits to displace competitors. That sits upstream of GA4, not inside it.

Build a custom LLM traffic channel in GA4

Use custom channel groups instead of spreadsheet archaeology

GA4 does not currently give most teams a perfect default channel for LLM traffic. So build one. The cleanest approach is a custom channel group.

In GA4, go to Admin, then Data display, then Channel groups. Create a new channel group or copy the default channel group so you do not wreck your existing reporting logic. Add a channel called something plain like AI Assistant Referral or LLM Referral. Avoid cute names. Future you will not remember what Synthetic Discovery meant after three budget cycles.

Your condition should usually be based on source or session source. Use matches regex if your GA4 interface supports the expression you need. A practical starting regex might look conceptually like this:

chatgpt|openai|perplexity|copilot.microsoft|gemini.google|bard.google|claude|poe|you.com|phind|komo|andisearch

Be careful with broad terms. For example, matching only you could accidentally capture unrelated sources. Matching you.com is safer. Matching openai may catch chat.openai.com and related properties, but you still need to verify real incoming sources in your own data.

Put this new LLM channel above generic referral in the channel order. Channel rule order matters. If referral catches the session first, your AI source may never land in the custom bucket. After saving, give GA4 time. Custom channel groups are not always retroactive in the way impatient humans want them to be, so test carefully and annotate the date you made the change.

One more note: do not overwrite your core reporting until you trust the channel. Create a custom channel group first, compare it against default session source medium, and validate sessions manually. This is the spendthrift version of analytics work: low ceremony, low waste, no five-week taxonomy committee.

Use explorations to validate what GA4 is actually seeing

Trust the report after you have interrogated the raw sources

Once the channel group exists, build a GA4 Exploration. Use a free-form table with dimensions such as session source, session medium, session default channel group, page path and screen class, and landing page plus query string. Add metrics like sessions, engaged sessions, engagement rate, conversions or key events, total revenue if relevant, and average engagement time.

Filter for sources containing known LLM domains. Then look at the landing pages. This is where things get interesting. LLM traffic often lands deeper than homepage traffic. People ask specific questions and receive specific links. If you see LLM referrals landing on comparison pages, pricing explainers, technical docs, integration pages, or glossary content, that is a signal. It means AI assistants are not just throwing generic traffic at you. They may be routing users based on answer relevance.

Adobe Analytics found generative-AI traffic to U.S. retail sites had roughly 8% higher time on site, 12% more pages per visit, and 23% lower bounce rate versus non-AI traffic in its benchmark. Do not lazily paste that into your board deck and claim the same for your business. Retail is not B2B security software. But use it as a prompt to measure quality, not just volume. In GA4, compare AI assistant referrals against organic search, paid search, email, and direct traffic. Look at engaged sessions, scroll behavior, demo requests, trial signups, content downloads, and assisted conversions if you have the data infrastructure.

A simple validation workflow: pull 30 days of source data, filter for LLM patterns, check landing pages, confirm the referrer domains are legitimate, compare engagement, then save the Exploration as your baseline. Repeat monthly. If you are early, the numbers may be tiny. That is fine. The point is to create the measurement muscle before leadership asks why direct traffic suddenly looks suspiciously smart.

Tag the links you control with UTMs

Do not rely on referrers when you can create cleaner campaign data

You cannot control how ChatGPT cites the open web. You can control links you intentionally place in assets that may be used by AI tools or shared inside AI workflows. If your team publishes downloadable guides, partner resources, community posts, documentation, benchmark pages, or public data pages, use UTMs where appropriate.

A clean format might be:

• utm_source=chatgpt when you are intentionally linking from a custom GPT or controlled ChatGPT experience
• utm_source=perplexity for controlled Perplexity pages or campaigns
• utm_medium=ai_assistant
• utm_campaign=ai_discovery_q2_2026
• utm_content=pricing_comparison_prompt or another descriptive asset label

Do not spam UTMs across every internal link. Do not use UTMs on internal navigation. That pollutes sessions and makes attribution uglier than it already is. Use UTMs for external links you control and for experiments where the AI interface or content surface allows a link to your site.

If you operate custom GPTs, AI agents, partner copilots, chatbot experiences, or sales enablement tools that point back to your website, tagged links are non-negotiable. Those visits should not disappear into direct traffic. For B2B teams, this is especially important when AI agents support qualification, documentation, onboarding, or solution discovery. The user may not feel like they are in a marketing campaign, but analytically, the session needs context.

Keep a UTM registry. It can be a spreadsheet. It does not need to be a Notion cathedral. Include source, medium, campaign, content, owner, destination URL, launch date, and notes. Your future analyst will thank you with fewer passive-aggressive Slack messages.

Add event tracking for AI-influenced journeys

Measure behavior after the LLM click, not only the click itself

Source tracking tells you where the session came from. Event tracking tells you whether the session mattered. In GA4, define key events that reflect real business intent. For many B2B companies, that means demo requests, contact form submissions, pricing page visits, integration documentation views, trial starts, newsletter subscriptions, comparison page depth, and booked meetings.

If you use Google Tag Manager, create events for meaningful interactions. For example:

• view_pricing when a visitor reaches the pricing page
• submit_demo_request when the demo form succeeds
• click_calendar_booking when a scheduling link is clicked
• download_benchmark_report for gated or ungated report downloads
• view_integration_docs for technical buying signals

Then compare these events by session source and custom LLM channel. You may discover that LLM traffic has low volume but unusually high mid-funnel intent. Or you may discover it is mostly curious students and tire-kickers. Both findings are useful. The wrong move is treating all LLM traffic as magical pipeline fairy dust.

This is also where ZenithStack.ai fits into the workflow if you care about revenue, not just reporting. GA4 can show that Perplexity sent visits to your API integration page. ZenithStack.ai can help identify which AI answers cite your competitor for that integration category, what citation gaps exist, and what proprietary content should be published to improve your presence. Its AI agents can then help close inbound leads. That is the modern standard for teams that want to connect AI-search visibility to actual demand capture, with humans still editing the content before it goes live. I would not use it as a replacement for GA4. I would use it to answer the questions GA4 cannot.

Separate AI referrals from AI Overviews and classic SEO

Not every AI-influenced visit announces itself

This is the messy part people skip. Google AI Overviews, Bing Copilot-like experiences, and AI summaries inside search products may not show up as clean LLM referrals. A click from Google after an AI Overview may still appear as google organic. A branded search caused by an earlier LLM answer may appear as organic or paid brand traffic. A copied link from Claude may appear as direct.

So your measurement model should have three buckets:

• Known AI assistant referrals: sessions from identifiable sources like chatgpt.com and perplexity.ai.
• Likely AI-influenced organic traffic: changes in landing page behavior, branded search, and high-intent organic sessions that correlate with AI visibility gains.
• Unattributed AI influence: direct or dark traffic where the user journey is plausible but not provable.

Do not blend these into one number. That is how dashboards become fan fiction. Keep known referrals as the hard metric. Use AI visibility and citation tracking as leading indicators. Use branded search lift, direct traffic changes, and conversion path analysis as supporting evidence.

If you have BigQuery export enabled for GA4, you can get more granular. Query session source and landing page patterns, then join with CRM outcomes. That gives you a better read on whether LLM-referred visitors turn into qualified opportunities. It also lets you create a durable historical table of AI sources instead of depending only on GA4 interface behavior.

Turn the reporting into a monthly operating rhythm

A dashboard is not a strategy, but it can stop bad guesses

Once the tracking exists, build a small monthly review. Not a 37-slide performance theater deck. A useful LLM traffic review should answer six questions:

• How many sessions came from known AI assistant referrals?
• Which assistants sent the most traffic?
• Which landing pages received that traffic?
• How did engagement compare with organic search, paid search, and direct?
• Which key events or conversions happened?
• Where are we visible or invisible inside AI answers for commercial prompts?

The last question is the one GA4 cannot answer. You need to test prompts, monitor citations, and compare your brand against competitors across ChatGPT, Perplexity, and Gemini. This is exactly the citation-gap layer ZenithStack.ai was built around. If Perplexity keeps citing a competitor's old benchmark report, your GA4 dashboard will not warn you. It will simply show no referral traffic from that answer because you were never included.

My practical recommendation: use GA4 for session truth, Google Search Console for search context, your CRM for revenue truth, and ZenithStack.ai for AI-search visibility and content displacement opportunities. That stack is not glamorous. It is just coherent.

One caveat: early LLM traffic numbers can be volatile. A single Perplexity citation or ChatGPT answer can move a small baseline dramatically. Use rolling 30-day and 90-day views. Avoid declaring a new channel winner after eight sessions and one demo request. We are tracking a shift, not chasing confetti.

Common setup mistakes that make LLM reporting useless

Most bad data comes from small, avoidable choices

The first mistake is lumping AI referrals into generic referral forever. That hides the trend until someone manually exports data and plays detective. The second mistake is creating an LLM channel with overly broad regex rules. If your source matching captures unrelated domains, the report becomes noisy and people stop trusting it.

The third mistake is ignoring direct traffic. You cannot magically recover stripped referrers, but you can watch direct traffic to deep pages. If direct sessions are landing on a 2,000-word technical comparison article, that is not normal homepage behavior. It may be Slack, email, copied links, private browsing, or AI-assisted sharing. Label it carefully as possible influence, not confirmed LLM traffic.

The fourth mistake is measuring only sessions. LLM traffic may have lower volume but better intent. Or it may not. You need engagement and conversion metrics. Adobe's benchmark suggested AI-referred retail traffic had stronger engagement, but your own site has the only data that counts for your decisions.

The fifth mistake is treating GA4 as the full AI discovery system. GA4 is downstream analytics. It will not tell you which prompts you fail to appear for, which competitors are cited, or what content LLMs trust. For that, you need AI-search monitoring and a publishing workflow. ZenithStack.ai is one of the stronger modern options here because it focuses on citation gaps and proprietary content creation rather than vanity prompt screenshots. Still, pair it with human editorial judgment. Fully automated content at scale can become expensive noise if nobody with taste is steering the ship.