The GEO Directory · 20 November 2026 · Published

Your AI traffic number is a floor, not a total.

What your analytics count by default, what they rename, and what never arrives with a source attached. Read the definition before the dashboard.

ONE CLICK FROM AN AI PRODUCT / THREE POSSIBLE BUCKETS

The journey is real.
The label is a choice.

01CountedA click from a browser session on a recognized AI domainlands in the AI Assistant channel
02RenamedA click from Google AI Overviews or AI Modelands in Organic Search, by definition
03UnattributedA tap inside a mobile AI app that passes no referrerlands in Direct
Schematic routing per vendor documentation, not a measurement of any site. Your report shows the bucket. It does not show the journey.

A founder opens the analytics dashboard and asks: how much traffic does AI send us? The answer on screen looks like a measurement. It is closer to a sorted pile of labels. Which clicks land in the AI bucket depends on rules most teams never read, and some AI-driven visits arrive with no label at all.

Before treating that number as the size of AI discovery, read what your tool counts, what it deliberately counts elsewhere, and what it cannot see.

01 / Read the default

GA4 has an AI channel. It has edges.

Google Analytics defines a default AI Assistant channel: users arriving from sources like ChatGPT, Gemini, DeepSeek, Copilot or Grok. Two properties of that definition matter before any trend reading. It excludes Google's own AI Overviews and AI Mode. And default channel groups cannot be edited; custom channel groups are the documented way to draw your own boundaries. [1]

So the AI number in a default GA4 report is a real, citable thing: sessions Google could attribute to a recognized assistant source. It is not "all traffic that involved an AI product." The same report can contain AI-driven visits under other labels, by design.

The counted bucket is also only as clean as its labels: bots and misattributed sources can inflate it. Read the number as an approximate floor (at least this much arrived with a recognizable AI source attached, if the attributions are right), not a guaranteed mathematical bound.

02 / Two leaks

Renamed by definition, dropped by mechanics

The renamed leak is a definition, not a bug. A click from an AI Overview or AI Mode is a Google Search click in Google's channel definitions, so it stays in Organic Search. In Conductor's measurement of 21.9 million US searches over four weeks in September and October 2025, 25.11% triggered an AI Overview. Where AI answers are common, a slice of the "organic" line is already AI-answer traffic; the default AI channel will not show it. [1] [5]

The dropped leak is a mechanism, not a proven share. Conductor describes how LLM mobile apps handle link navigation: unlike a desktop browser, the app may pass no referrer, so the session lands in Direct. ChatGPT was the most downloaded app worldwide in March 2025, which makes the mobile route plausible at scale. Conductor does not demonstrate what fraction of any site's direct traffic is AI-driven, and neither can you from a channel report alone. Treat direct-mobile movement as a hypothesis to investigate, not as hidden AI traffic you have already counted. [4]

Definitions also differ between tools. Semrush tracks Google AI Mode as a separate channel from AI referrals; GA4 folds it into Organic Search. The same visit can therefore be "AI" in one dashboard and "organic" in another, with both tools working as documented. [1] [2]

03 / Scale check

Three panels, three numbers, one direction

Independent vendor panels with different populations, definitions and periods. Do not average them; read them as direction, not as your site's share.

Conductor benchmark

1.08%of sessions were AI referrals

13,770 domains, 10 industries, 3.3 billion sessions, May-September 2025, US. ChatGPT drove 87.4% of the AI referrals; AI referrals grew about 1% month over month.

via Search Engine Land, 13 Nov 2025 [5]

Semrush panel

0.14%of total 2025 visits were AI referrals (rounded to "less than 0.15%" in Semrush's summary)

50,000+ websites, 17 industries; total panel visits were essentially flat over 2025, edging from 546.5 billion down to 544.2 billion (-0.43%). AI referral visits grew 66% over the year, from 462 million to 767 million monthly visits. Google AI Mode, tracked separately, went from 1,600 to 38.2 million monthly visits (0.01%).

Semrush, 27 Apr 2026 [2]

Ahrefs live panel

0.46%AI Assistants share, October 2026

Ahrefs Web Analytics panel of 118,671 sites, read 20 November 2026: AI Assistants 5.4 million visits (0.46%) against Search at 27.55% and Direct at 31.08%; ChatGPT alone held 0.36% of total traffic, the largest share among AI assistants. The publisher notes categories overlap and total 104.1%.

Ahrefs benchmarks, read 20 Nov 2026 [3]

The panels disagree on the level (0.14%, 0.46%, 1.08%) because their populations, channel definitions and periods differ. The one finding they share: AI referrals are a small share of total visits today. Everything else belongs to its source. Growth: 66% over 2025 in Semrush's panel while total visits there were flat (Semrush's headline calls AI the fastest-growing channel; its own breakdown shows paid search growing slightly faster, at 76%), about 1% month over month in Conductor's May-September 2025 benchmark, and +0.03 percentage points in Ahrefs' October 2026 read. Concentration: ChatGPT drove 87.4% of AI referrals in Conductor's panel and held the largest assistant share in Ahrefs' panel; Semrush's study publishes no provider split. A board asked "how big is AI traffic?" should hear each number with its population and period attached: the share, the growth rate, and the definition that produced it.

04 / Instrument

Set the baseline before the story

1. Record the default with its definition. Open the AI Assistant channel and write next to the report: includes ChatGPT, Gemini, DeepSeek, Copilot, Grok; excludes AI Overviews and AI Mode. Every future comparison inherits this definition. [1]

2. Draw your own boundary if you need one. Default groups cannot be edited. If your reporting needs a different source list, create a custom channel group and label it with the exact sources and the creation date, so a number produced this quarter is not compared against a differently-defined number from last quarter. [1]

3. Watch direct mobile like a detective, not a believer. Segment direct mobile sessions by landing page. A spike on deep product or documentation pages, with the homepage excluded, is a pattern worth investigating; it is consistent with AI app referrals and with several other causes. Conductor suggests exactly this segmentation rather than a blanket reattribution. [4]

4. Keep two evidences separate. Crawler visits in server logs and referral sessions in analytics are different events by different actors. A page can be heavily fetched and never referred, or referred and rarely fetched. Report them on separate lines.

5. Compare trend, never just level. Given the floor problem, month-over-month movement inside one unchanged definition is the strongest signal available. A definition change restarts the baseline.

05 / Diagnose

Three meetings, three different mistakes

Hypothetical meeting cards, not observations of any company, tool or dataset.

Level mistaken for trend

"AI is 0.3% of our traffic. Close the file."

The number is a defined subset measured this month. The growth rate inside the same definition, and the leaks outside it, are unanswered.

Next: keep the definition fixed, track the trend, and review the renamed and direct-mobile segments alongside.

Do not report: "AI sends us nothing."

Correlation mistaken for attribution

"Direct mobile jumped 40%. That is ChatGPT."

Plausible, and unproven. App updates, seasonality, untagged campaigns and LLM apps all move direct mobile.

Next: segment by landing page, exclude the homepage, and look for corroborating patterns before naming a cause.

Do not report: "We measured our hidden AI traffic."

Redefinition mistaken for growth

"We built a better AI channel group and AI traffic tripled."

Widening the source list changes the denominator's rules, not the world's behavior. The old series and the new series measure different sets.

Next: restate past baselines under the new definition where the data allows, and mark the break on every chart.

Do not report: "AI traffic tripled this quarter."

The handoff

Ask for the definition with the number

Give me this month's AI Assistant channel sessions with the channel definition written next to them, the direct-mobile trend split by landing page, and the same readings from previous months under one unchanged definition. No AI share claim without its source, population and period.

That request fits in one message and survives every future tool change. It lets a founder take AI discovery seriously without letting a partial counter pose as a total.

Method and sources

What is established, and what is not

This is a documentation- and benchmark-led measurement guide, not an analytics experiment run for this article. No site's data was accessed, and the meeting cards are hypothetical. The three panels cited are vendor samples with the populations, definitions and periods stated next to their numbers; none of them measures the web as a whole, and none measures your property.

The existence and exclusions of the default AI Assistant channel, and the editability rules for channel groups, come from Google's own documentation. The referrer-drop mechanism in mobile LLM apps is described by Conductor; its size on any given site is undemonstrated here and should not be quoted as a percentage. Product descriptions in the shortlist are vendor claims, not tested advantages. [1] [4]

Sources were read on 20 November 2026. Channel definitions, default group rules and panel values change; re-read the definition before reusing any number from this page.

  1. Google Analytics Help: default channel group

    Defines the default AI Assistant channel, its exclusions, and the rule that default groups cannot be edited while custom channel groups can.

  2. Semrush: how AI is reshaping traffic channels

    Published 27 April 2026. Panel of 50,000+ websites across 17 industries in 2025; AI referrals defined separately from Google AI Mode; industry shares are medians across domains.

  3. Ahrefs: AI vs Search traffic analysis

    Live benchmark panel read on 20 November 2026. October 2026 values; the publisher notes the categories overlap and do not sum to 100%.

  4. Conductor: your direct mobile traffic might be from ChatGPT

    Published 22 May 2025. Explains the referrer-drop mechanism in LLM mobile apps; it does not demonstrate what share of any site's direct traffic is AI-driven.

  5. Search Engine Land: coverage of the Conductor AI search benchmark

    Published 13 November 2025. Benchmark of 13,770 domains, 10 industries, 3.3 billion sessions, May-September 2025, US; AI Overviews measured on 21.9 million searches over four weeks.

  6. Google: AI features and your website

    Technical eligibility for AI Overviews and AI Mode; no special AI file or markup is required.

  7. Obi: product overview

    Vendor description, not a hands-on product test.

  8. AirOps: content engineering teams

    Vendor framework for context, workflows and content-team roles.

  9. Profound: enterprise capabilities

    Vendor description of SSO and role-based access.

The GEO Directory. A measured share is a floor with a definition attached, not a total with a promise attached.