How to measure AI referral traffic and conversions with GA4

You can measure AI referral traffic in GA4 by looking at sessions whose source is an AI engine's own referrer, then following those sessions to the landing pages they reached and the conversion events GA4 recorded. That gives you a real but partial picture: visits with no referrer are invisible to this method, Google AI Overviews and AI Mode arrive as ordinary Google traffic, and nothing here connects a specific prompt, answer or citation to a specific sale. This post walks through the measurement chain, shows what each link in it can and cannot tell you, and ends with a worksheet you can copy.

In short

• AI referral traffic is a count of sessions that arrived from an AI engine's own referrer, as recorded by GA4. It is not a count of everyone influenced by an AI answer.
• A visit with no referrer is recorded as direct traffic, so it cannot be counted as AI traffic by this method.
• Google AI Overviews and Google AI Mode visits reach GA4 as ordinary Google search clicks and cannot be separated from other Google traffic here.
• A conversion is only as meaningful as the events you have named. Decide what counts before you read a number.
• You can connect AI referral sessions to landing pages and conversions. You cannot connect a prompt or a citation to a sale.

What is AI referral traffic, exactly?

AI referral traffic is the set of sessions in your analytics whose traffic source is a domain belonging to an AI engine, such as chatgpt.com or perplexity.ai. The definition rests on the source GA4 recorded, not on what the visitor was doing before they clicked.

That distinction matters because a session in GA4 always carries a source, but not always a referring host. When there is no referrer, the source reads (direct). DiscoveredBy's AI Traffic screen reads your Google Analytics 4 property and keeps only sessions whose source matches one of 21 known AI referrer hosts, resolving to ten engines. A session whose source is not on that list, including one with no referrer at all, is not counted, whatever channel it really came from. There is a short definition in the glossary under AI referral traffic.

Five of the ten referrers (ChatGPT, Claude, Gemini, Grok and Perplexity) belong to engines DiscoveredBy also runs prompts against. The other five (Copilot, You.com, Phind, Kagi and Metaphor) are not queried by DiscoveredBy, but GA4 still reports sessions from them, so they appear in the traffic view too.

Why do some AI visits go missing?

Some AI-driven visits never show up as AI referrals because the session arrives without a referrer, and GA4 then records it as direct. Nothing in the source field says an AI engine was involved, so no analytics tool that reads the source can recover it.

The docs are explicit about the mechanism: a source that is not on the list of known AI hosts is not counted, "including one with no referrer at all". Why a given visit lacks a referrer varies, and you usually cannot tell from the analytics alone. Referrer-stripping by some apps or privacy settings is one possible cause. Treat that as something to investigate, not as a measured share of your traffic.

Two practical consequences follow.

• Your AI referral number is a floor, not a total. The real number of visits influenced by AI answers is at least this large, but you cannot say by how much.
• A rise in direct traffic to a deep page is a hint, not a finding. If a page that nobody links to from your own site starts drawing direct sessions, that is worth a look. It does not prove an AI engine sent them.

Which AI surfaces can you not separate?

Google AI Overviews and Google AI Mode cannot be isolated, because they have no referrer of their own. A click from either one reaches GA4 as ordinary Google search traffic, indistinguishable from any other Google click.

This is why choosing Google AI Overviews or Google AI Mode in the engine filter of AI Traffic always shows an empty result: there is no referrer to narrow to. It does not mean those surfaces sent you nothing. It means this method cannot see them. If you track those surfaces in DiscoveredBy, you see whether your domain was cited in the answer, but the visits that follow are folded into your Google organic traffic. For the difference between a citation and a click, see Google rankings versus AI citations.

The Gemini referrer counts people who arrived from Gemini's own hosts (gemini.google.com and bard.google.com), and both Gemini engines DiscoveredBy runs, Gemini (API) and Gemini (app), map to that one referrer. So the traffic view cannot tell you which of the two collection methods a visit corresponds to, and it does not include AI Overviews or AI Mode.

How do you set up GA4 so the numbers mean something?

Connect GA4 to your monitoring tool, map the right property to the right project, and name your conversion events. Each step changes what you can read later.

Step 1: Connect and map

In DiscoveredBy, Integrations is where you connect a Google account and then choose which Analytics property feeds which project. Connecting and mapping are separate. Connecting stores one Google account's connection for your user; mapping ties one Analytics property to one project. The connection is read-only: it requests the analytics.readonly scope for Analytics. Some mapping happens automatically at connect time, when a property's hostname exactly matches a project's; otherwise you map by hand. If a freshly connected project still shows nothing, a property not yet mapped to it is a common reason. Once a property is mapped, a one-time 180-day history import is queued immediately, and after that the import runs once a day. The screen reads at most 90 days, so the 180-day pull fills the widest window.

Step 2: Decide what a conversion is

In GA4, a key event is an event that measures an action particularly important to your business, and any event you collect can be marked as one (see Google's key events page). Do this in GA4 before you read conversion numbers anywhere else, and write down the list.

DiscoveredBy also lets you name the Analytics events that count as conversions for a project. How this changes the numbers is worth understanding, because it is not a filter. The import brings in GA4's own conversion count and revenue for each row. Where one of your named events fired on a row, its count and value replace GA4's figures for that row. If you have named none, the replacement uses GA4's own purchase and sign_up events. So what you read is a mixture: GA4's own figures where none of the counted events fired, and your events' figures where they did. If your named events carry no revenue, the value figure can land below GA4's revenue for the same sessions.

Step 3: Set your window

AI Traffic opens on the last 28 days and can be switched to 7 or 90. Every change figure compares the selected window with the same number of days immediately before it, not with a year ago. Pick one window for the whole measurement cycle and keep it.

How do you follow a visit from landing page to conversion?

Enlarge image

Enlarge image

Read the same slice in three passes: which engines sent sessions, which landing pages they reached, and how many recorded conversions came from those sessions. Each pass has a limit you should carry into your notes.

Pass 1, by engine. The hero number is total AI-referred sessions for the window, with a breakdown by engine, each with its share and a change against the prior window. An engine that appears for the first time shows no change rather than a percentage measured from zero. Selecting a row drills into that one referrer and narrows the landing pages, sessions and conversion figures to it.

Pass 2, by landing page. Landing pages are ranked by sessions, each with the engines that sent traffic to it. The list is capped at the top 50 pages, so a large site sees the leaders, not everything. Note which page types dominate: a guide, a pricing page, a comparison, an integration page.

Pass 3, conversions. If total conversions across every engine are zero for the window, the screen shows a prompt to name your conversion events instead of a row of zeros. When there are conversions, the count and value are sums of the underlying daily rows, and the rate is built from those sums. A single engine's row does not carry its own conversion change chip, because the prior-window comparison for conversions is built for the window's total.

Then check the same window against your site as a whole. The Site traffic tab reads every session GA4 recorded for the project and shows an "Of that, AI-referred" figure using the same classification, so you can see the AI slice as a share of everything.

The measurement worksheet

Copy this into a spreadsheet or document and fill it in once per reporting period. The point is to record the definitions next to the numbers so nobody reads a total without its limits.

AI REFERRAL MEASUREMENT WORKSHEET

A. Setup (fill once, revisit if it changes)
Project / site:
GA4 property mapped (name and ID):
Date the property was mapped:
Conversion events counted (list each):
Are revenue values attached to those events? yes / no
Who can change the event list:

B. Period
Window (7 / 28 / 90 days) and exact dates:
Prior window dates (same length, immediately before):
Any site release, tracking change or outage in either window:

C. Sessions (AI-referred, from the monitoring tool)
Total AI-referred sessions:
Prior window total:
Change (state the arithmetic):
By engine (referrer, sessions, share of AI total):
  1.
  2.
  3.
New engines this window (no prior figure, so no change):
Google AI Overviews / AI Mode visits: not separable, sit inside Google traffic

D. Landing pages (top pages by AI sessions)
Page | AI sessions | Engines sending it | Conversions | Notes
1.
2.
3.
Page types that dominate:

E. Conversions
Total conversions from AI-referred sessions:
Conversion events behind that total:
Value recorded (state if events carry revenue):
Reconciled against GA4 for the same dates? yes / no
If not reconciled, why:

F. What this cannot show (leave these lines in every report)
- Visits with no referrer are recorded as direct and are not counted.
- Google AI surfaces are inside ordinary Google traffic.
- No link from a specific prompt, answer or citation to a specific sale.
- Sessions are visits; one person can be several sessions.

G. Interpretation (two sentences maximum)
What changed:
What we will do or check next:

H. Cross-check with visibility data (optional)
Pages cited in answers this period (from the citations view):
Pages that also received AI sessions:
Cited pages with no AI sessions:

Section H is where the two halves of the picture meet. Citation data says an engine named your domain in an answer; traffic data says people arrived. A page can appear in both lists, in one, or in neither, and each case suggests a different question. Your monitoring tool's citations view tells you which pages were named as sources.

Worked example

Illustrative example: Quillstone and its competitors are fictional, and the numbers are made up to show the method.

Quillstone sells document-review software to mid-sized legal and compliance teams. Its marketing lead maps the GA4 property, names two conversion events (demo_request and sign_up), and reads a 28-day window.

The prior 28 days recorded 100 AI-referred sessions, so the total is up 20 sessions, or 20%. The shares round to 58, 25, 10 and 7 percent (70, 30, 12 and 8 out of 120).

Six conversions came from 120 AI-referred sessions, a rate of 5% on this made-up dataset (6 divided by 120). Two things stand out. The checklist guide draws the most AI sessions but converts once, while pricing draws fewer and converts three times. That pattern suggests the guide introduces people who are not yet ready to act, but it is a hypothesis to test, not a finding.

Quillstone also notices that direct sessions to the checklist guide, a page not linked from its navigation, rose in the same window. It records that as a question in the worksheet, not as extra AI traffic. It does not add those sessions to the 120, because it has no way to confirm where they came from.

Finally, its Google AI Overview visibility looks healthy in the citations view, but the traffic view shows nothing for that surface. Quillstone writes down the reason (no separate referrer) rather than concluding the surface sends no visitors.

What this method cannot tell you

It cannot attribute a sale to a prompt. A session from an AI engine has a source, a landing page and possibly a conversion. It does not carry the question the person asked, the answer they saw, or which citation they followed. Tracking prompts and tracking referrals are two different measurements, and joining them by date or by page would be a guess.

It cannot see the visitors who did not click. Someone who read an AI answer that named your brand, then typed your name into a browser later, appears as direct or organic. Their influence is real and unmeasured here.

A conversion in GA4 is an event, not necessarily revenue. Depending on your named events, a "conversion" might be a sign-up, a demo request or a purchase. Say which one.

Change figures need a comparable prior window. A change is measured against the immediately preceding window of equal length. If your site or tracking changed between them, the comparison is weaker.

Frequently asked questions

Why does GA4 show more or fewer AI sessions than my monitoring tool?

The two may classify sources differently. GA4 has its own channel grouping, and DiscoveredBy uses its own list of 21 AI referrer hosts. Sessions are also counted at different levels in some views. Compare like with like, use the same dates, and record which classification you used.

Can I tell which prompt drove a visit?

No. A referral records the source domain, not the question. DiscoveredBy tracks prompts to see whether your brand and domain appear in answers, and reads traffic separately from GA4. The two are not linked at the visit level, so neither can be used to claim a sale came from a given prompt.

Why is my AI traffic zero when my citation data looks strong?

Several explanations fit. The property may not be mapped or may still be importing. The engines citing you may be Google's AI surfaces, which cannot be separated from ordinary Google traffic. Or citations may not yet be producing clicks. Start with the troubleshooting page for empty screens, then check the mapped property.

Should I put this in a client report?

Yes, with the definitions attached. Show the AI-referred total, the by-engine split, the top landing pages and the conversions, then include the "what this cannot show" lines. For structure, see How to write an AI visibility report a client can act on.

Next step

If you already have GA4 on your site, connect it and map the property to your project, then read the AI Traffic screen for your last 28 days and fill in sections A to E of the worksheet. For how citation and mention numbers differ from referral counts, read AI visibility, brand mentions, and citations: what each number tells you. To see where traffic fits in the product, see analytics features, or sign in to DiscoveredBy and connect your Google account under Integrations.