AI assistants are now a traffic channel

Google has added a channel grouping to Google Analytics 4 called AI Assistant. From this rollout, visits that originate from ChatGPT, Gemini, Claude, Perplexity, Copilot and the rest of the conversational interfaces get their own row in the dashboard, alongside Organic Search, Social, Email, Direct and Paid. The reference is Google's own page at support.google.com/analytics/answer/15358914.

I have been waiting for a moment like this. Dashboards lag reality, and the dashboard catching up is the signal that the reality underneath has settled.

What the channel actually counts

The AI Assistant channel groups referrers from a list of conversational interfaces that Google maintains and updates: ChatGPT from OpenAI, Gemini and AI Mode from Google, Claude from Anthropic, Perplexity, Copilot from Microsoft, and a handful of others. When a reader follows a link an assistant surfaced and lands on your site, the visit now appears in its own row instead of being scattered across Direct and Referral.

That is the mechanical change. It is small, useful, and overdue. Until this rollout, an analytics view had no honest way to tell you that AI assistants were sending readers; most of the conversational traffic carried no referrer at all, and the rest landed in buckets that hid it.

Why this is more than a reporting tweak

Discovery surfaces produce disciplines. Search produced SEO. Social produced community management. Email produced lifecycle marketing. The reason each one settled into a recognised practice was not the activity itself, it was the moment the dashboard learned to count it. Once a channel has a name, somebody owns it inside the organisation, somebody else gets measured against it, and a vendor category forms around tooling for it.

That tells me three things are now true at once. AI assistants are sending enough traffic for Google to bother. The channel is durable enough to be worth distinguishing. And organisations have a place to put accountability for what happens inside it.

How an AI assistant reads a page

This is the part most analytics conversations skip.

An AI assistant does not browse the way a person browses. It fetches the page, parses what is there, summarises it, and either quotes you, paraphrases you, or sends the reader somewhere else. The decision is made on the basis of what the page declared about itself: in HTML structure, in Schema.org, in metadata, in machine-readable signals. The assistant does not stay long enough to register a scroll. It does not see the hero image. It reads the markup and moves on.

If the page declared nothing, the assistant guessed. If the guess was wrong, your name shows up wrong, or not at all, in someone else's answer.

What the new channel sees, and what it does not

The new channel counts the visits where the assistant decided your page was the right destination and the reader followed the link. Those visits will grow. They are not the whole story.

The much larger population is the one the dashboard cannot see: the reader asked the assistant a question your page answers, the assistant read your page, decided it could not safely cite you, and sent the reader to a competitor instead. The dashboard has no row for that visit, because the visit never happened. You lost the citation in silence, and the only way you would know is to ask the assistant the same question yourself.

So the new channel is a useful floor and a misleading ceiling. The floor is the traffic you are already winning. The ceiling is hidden behind every assistant that read your page and chose not to mention you, and you have to infer the size of that ceiling by hand.

What to do with the new line on the dashboard

Four things, in order.

Watch the share, not the volume. The absolute numbers will be small for a while. The ratio of AI Assistant to Organic Search is the signal worth tracking. The day that ratio crosses one in twenty on a content-heavy site, the traffic mix has changed and the things you optimise for change with it.

Compare what assistants quote against what your page actually says. Pick a question your page is meant to answer. Ask the major assistants. Read what they reply, and check whether the answer matches your page. Note the citations. If you are not in the citation list, the assistant either did not find you or did not trust you. Both have specific fixes, and they are different fixes.

Audit the page the way an assistant reads it. Fetch your own HTML with curl, strip the scripts, and look at what is left. That is what the assistant sees. If the structure is unclear, if the headings do not declare the page, if the schema is missing or thin, the assistant is reading the same gap you are looking at.

Fix the page so the next assistant has nothing to guess. Explicit identity, explicit structure, explicit provenance, explicit machine-readable claims. Not surface markup over a thin body; underlying meaning expressed in a form a machine can verify. This is the work that produces durable lift.

Where MX fits

SEO, GEO and AEO describe how a page presents itself to search engines, generative answer engines, and citation slots. They are surface disciplines, and they keep changing because the surface keeps changing. MX is a different kind of thing. It is the contract underneath the page, the layer that lets a machine verify what it is reading rather than guess from appearance.

Machine Experience, or MX, lives underneath structured data. The MX field dictionary covers identity, state, audience, provenance, governance, and allowed actions. The Gathering is the open community where the dictionary is governed, in a vendor-neutral model that follows W3C precedent: draft notes, public review, ratification stream. When a page carries MX metadata, an assistant reading it does not have to infer who wrote it, when, on what authority, or whether the facts inside are something the publisher actually holds. The page declares those things, and the assistant can check.

Two pillars. MX makes content machine-readable. The signing layer on top of MX makes the same content machine-trustworthy. The combined effect is what the new analytics channel will start measuring whether organisations realise it or not, because AI assistants prefer pages they can verify, and the dashboard will quietly reward the publishers who give them that.

What I would do this week

If I were running content for an organisation today, I would do three things in the next seven days.

Find the AI Assistant row in Google Analytics 4. If it is not there yet, the rollout has not reached the account; check again in a week. If it is there, take a screenshot of the current numbers and the percentage share. That is your baseline.

Pick the ten pages you most want an AI assistant to quote. For each one, ask Gemini and ChatGPT a question that page is meant to answer. Note who they cite. Save the results.

Run an MX audit on the same ten pages. Compare what the page declared against what the assistants quoted. The gap is your work list for the next quarter.

The new channel is a measurement. The work behind it is the same work good publishers have always done: get the facts right, declare them clearly, and take responsibility for what you publish. The difference is that the dashboard can now tell you, for the first time, whether the work is paying.

If you want help with the audit, you know where to find me.