Your best converting visitors are hiding inside Direct. Here is how to find them.

Open GA4. Look at your traffic. Now tell me, with a straight face, how many people came to your site because ChatGPT told them to.

You cannot. Almost nobody can. And that is the single biggest measurement hole in marketing right now.

Here is the ugly bit. The traffic you cannot see is probably your best traffic. Adobe, which tracks over a trillion visits to US retail sites, found that visitors arriving from AI assistants converted about 42% better than everyone else in March 2026. Twelve months before that, the same visitors converted at roughly half the rate of other channels. That is an 80 point swing in a year. Semrush put the cross industry premium at 4.4x. Ahrefs, on its own B2B data, found 0.5% of sessions produced 12.1% of signups.

Different studies, different methods, wildly different multiples. Fine. Ignore the exact number. The direction is not in dispute: the people AI sends you arrive half sold. And most site owners are filing them under "Direct" and calling it a mystery.

Time to fix that. Properly. Four layers, in order, most of it free.

First, get clear on what "AI traffic" actually is

Four completely different things get jammed into one phrase, and that confusion is why most reporting on this is rubbish.

1. AI referral clicks. A human asked ChatGPT or Perplexity something, got an answer with your link in it, and clicked. Real session, real human, measurable in GA4.
2. AI Overview and AI Mode clicks. A human on Google, in an AI generated answer box, clicked your result. Google counts this as Organic Search. Not as AI. More on that landmine below.
3. AI citations with no click. The machine named you, quoted you, recommended you, and the user never left the chat. No session. No GA4 row. Pure brand impact that analytics is structurally blind to.
4. AI crawlers and agents. GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and the rest. Not humans. They never fire your JavaScript, so GA4 never sees them. They only exist in your server logs.

Four things. Four different measurement jobs. If you try to answer all four with one GA4 report you will get a confident number that is wrong. We sort this properly inside the Dojo because getting the definitions right is 80% of getting the reporting right.

Layer 1: the GA4 AI Assistant channel (turn it on, then distrust it)

On 13 May 2026 Google added a native AI Assistant channel to the GA4 Default Channel Group. No setup. No regex. No developer. When GA4 sees a referrer it recognises as an AI assistant, it stamps the session with:

• Default Channel Group = AI Assistant
• Medium = ai-assistant
• Campaign = (ai-assistant)

Find it in Reports, then Acquisition, then Traffic acquisition, with the primary dimension set to Session default channel group. Broad availability across properties landed around 7 June 2026, so if your property looks empty for a few weeks after the launch date, that is rollout lag, not a traffic problem.

Two things to burn into your brain before you show this to anyone.

It is forward only. GA4 does not go back and reclassify history. Everything before 13 May 2026 is still sitting in Referral or Direct. So the channel "appearing" in your chart is a measurement event, not a growth event. Annotate it or some analyst in six months will declare a 400% AI traffic explosion that never happened.

The source list moves. Google's launch note named ChatGPT, Gemini and Claude. The live Default Channel Group documentation has since listed ChatGPT, Gemini, Deepseek, Copilot and Grok. Claude was in the announcement and not in the later published list. Google says these definitions evolve with the market, which is corporate for "we will change this and not tell you". Never hard code a platform list into a client report. Check the live help page the day you publish and stamp it with a date.

Why that channel number is a floor, not a ceiling

Three gaps. They are not the same kind of problem, and knowing which is which tells you whether to fix it, work around it, or just caveat it honestly.

Gap 1: Perplexity is not in it

Perplexity is absent from Google's official definition and still drops into Referral. Which is maddening, because Perplexity is one of the highest intent AI sources going. It shows URLs. It sends people who have already decided to click through. It passes referrer data reliably.

To be precise, because half the internet gets this wrong: Perplexity does not strip referrers. Its problem in GA4 is classification, not darkness. This one is a coverage gap and you can close it completely in about ten minutes.

Gap 2: Google's own AI does not count as AI

AI Overviews and AI Mode clicks are explicitly excluded from the AI Assistant channel. Google routes them to Organic Search. For most normal businesses this is almost certainly the largest chunk of AI influenced traffic they get, and it is completely invisible as AI in standard reports.

No GA4 setting reverses this. It is a routing decision. You measure it in Search Console instead, which is Layer 2.

Gap 3: dark traffic, the big one

Most AI sessions turn up with no referrer at all and land in Direct. Industry estimates put it somewhere around 60% to 70%, and I am not going to pretend that is a precise figure, because it is not. It varies by audience, device and platform.

The causes are structural, not fixable:

• Native ChatGPT iOS and Android apps pass no referrer.
• Mobile in app browsers strip referrer headers.
• People copy the URL out of a chat and paste it into a fresh tab.
• The ChatGPT Atlas browser strips referrer data on handoff.

So your AI Assistant number is the visible slice of a bigger thing. Report it as a floor. Say the word "floor" out loud when you present it. That one habit will save you from the worst conversation in analytics, which is the one where last month's confident number turns out to have been a third of reality.

Layer 2: build the custom channel group (ten minutes, free)

Default channel groups cannot be edited. You cannot shove Perplexity into Google's channel. What you can do is build a custom channel group that runs alongside it and catches the long tail, plus give yourself a bridge back to pre May 2026 data.

Go to Admin, then Data display, then Channel groups. Create a new group. Add a channel called AI Assistants (custom) with a rule on Source matching a regex of the platforms you care about. Start with these:

Now the bit that catches people out. GA4 evaluates channel rules top down. If your AI rule sits below the Referral rule, every AI source hits Referral first and you will swear blind your regex is broken. It is not. Drag the AI channel above Referral. That is the fix, nine times out of ten.

Second move while you are in there: UTM tag every URL you control before you paste it anywhere an AI might read it. Newsletter links, Reddit comments, directory profiles, PDF footers. If you put the tag on it, no amount of referrer stripping can hide it from you. This is the kind of boring half hour that agencies fold into a $2,000 monthly retainer and call "attribution strategy".

Layer 3: Search Console, for the Google AI surface

This is the half of the picture GA4 refuses to give you. AI Overviews and AI Mode clicks land in Search Console as normal Search performance data, which means you can see impressions, clicks and position, you just cannot filter them with a big button marked "AI".

What you can do is read the symptoms, and the symptoms are loud:

• Impressions flat or up, clicks down, position stable. Classic AI answer signature. You are being shown and summarised, not clicked.
• CTR collapsing on informational queries while commercial queries hold. The machine is eating your "what is" and "how do I" traffic and leaving you the buyers. Not the worst trade, if you notice it and stop writing the wrong pages.
• A page that gains impressions across dozens of new long tail questions. That is usually a page getting chunked and pulled into answers.

Set up a monthly export, compare informational versus commercial query groups, and track the ratio over time rather than obsessing over any single week. If you want the full mechanics of getting cited on that surface in the first place, that is the Google AI Mode playbook.

Layer 4: server logs, where the machines actually live

GA4 measures humans who run JavaScript. AI crawlers do not run your JavaScript. So every single bot visit is invisible to your analytics and fully visible in your raw server logs or your CDN logs.

This matters more than it used to. Botify's analysis of roughly 7 billion log entries from November 2024 to March 2026 found OpenAI has tripled its crawl of the web. More interesting than the volume is the mix. Before GPT-5, the ratio of OAI-SearchBot to GPTBot activity sat around 0.95. After GPT-5 it went to about 1.14. In plain English, OpenAI is now spending more crawl budget on live retrieval than on training collection.

That flip is the whole story. Training crawls affect what a model vaguely knows next year. Retrieval crawls affect whether you get cited this afternoon. Here is what to watch for in your logs:

Three questions your logs answer that nothing else will. Are the AI bots reaching your money pages or just your homepage? Are they hitting 404s and redirect chains and burning budget on rubbish? And is crawl frequency going up or down after you publish?

If you are on Cloudflare you already have most of this in the dashboard. If you are on decent hosting, grab a month of access logs and grep the user agents. It is about an hour of work and it is the only place in your stack where you can see the machines making up their minds about you. And if you have quietly blocked half of them in robots.txt without realising, read what actually happens when you block AI crawlers before you touch anything.

The four layer stack, side by side

Read down that table and the funnel appears: cited, then clicked, then converted. Layer 5 sees the mention. Layers 1 and 2 see the click. Your CRM sees the money. Miss a layer and you will end up optimising for the part you can see instead of the part that pays.

The layer everyone skips: are you even being named?

Every layer above measures a click. But the most valuable thing AI can do for you produces no click at all. Someone asks for the best plumber in Newcastle, gets three names, rings one. No session. No referrer. No row in any report. Just a phone call from a stranger who already trusts you.

You do not need a $500 a month platform to check this. Open the four big assistants, ask the ten questions your customers actually ask, and log who gets named. Do it on the first Monday of every month. Same prompts, same order, in a spreadsheet. That is a real trend line for the price of twenty minutes. I broke the whole manual method down in the free AI visibility tracking guide, and members get the tracker plus the live SERP tooling in The Armory.

What the numbers actually look like

Be careful with the headline stats floating around, because they come from wildly different samples. Here is what has actually been published, with the caveats attached, so you can calibrate instead of quoting a number you cannot defend.

Notice what is missing from that table: a number for your business. That is the point. These figures tell you the shape of the thing. Only your own stack tells you the size of it. Anyone selling you an industry average as a forecast is selling you a story.

Timeline: what to expect, honestly

• Day one. Native channel is on already. Custom channel group takes ten minutes. Annotate 13 May 2026 in GA4 so nobody misreads the step change later.
• Week one. Server log pull and a first pass on which bots are hitting what. Search Console export baselined.
• Month one. First clean month of layered data. Resist every urge to draw a conclusion from it.
• Month three. Now you have a trend. Ninety days is the minimum before AI numbers should move your content strategy, because the channel is young, the rollout was staggered and seasonality will lie to you.

Anybody promising you a verdict from three weeks of data on a four month old channel is guessing in a nice font.

Cost: what this should actually set you back

• GA4 native channel: $0. It is already on.
• Custom channel group: $0 and ten minutes.
• Search Console analysis: $0, about an hour a month of your time.
• Log analysis: $0 if you have CDN or hosting logs and can use grep. Dedicated log tooling runs from roughly $50 to several hundred a month depending on volume.
• Citation monitoring: $0 manually, or roughly $50 to $500 a month for a platform.
• Agency doing all of the above: commonly $1,500 to $3,000 a month. For work that is largely a one off setup.

Four of the five layers cost nothing but attention. That is not an accident. Most of what gets sold as advanced AI attribution is a channel group, a log grep and a spreadsheet, wrapped in a deck.

Who this is for, and who it is not

Do this now if: you sell something with a real margin, your buyers research before they buy, you already get organic traffic, or you have watched impressions hold while clicks slide and you cannot explain it.

Do not bother yet if: you get under about 500 sessions a month, because the AI slice will be statistically meaningless and you will chase noise. Fix the fundamentals first. Get the pages indexed, get the answer blocks right, get a few citations, then measure. There is no point building an instrument panel for an engine you have not started.

Five mistakes that wreck the number

1. Reporting the AI Assistant channel as the total. It is the visible slice. Caption it as a floor, every time, or you will be caught out.
2. Putting the AI rule below Referral in a custom channel group. Rule order beats regex. This is the single most common implementation bug.
3. Treating a channel launch as a growth spike. Annotate 13 May 2026. Do it today, it takes fifteen seconds.
4. Counting AI Overviews as AI traffic in GA4. They are not in that channel. They are in Organic. Double counting them is the fastest way to publish a number you have to retract.
5. Blocking the retrieval bots by accident. Plenty of sites tightened robots.txt in the block everything panic and quietly removed themselves from ChatGPT search. GPTBot and OAI-SearchBot are separate controls for separate jobs. Know which one you are switching off.

The alternatives, and where they fall down

All in one AI visibility platforms. Genuinely useful for citation tracking at scale. They do not solve attribution, because they cannot see your sessions. Pay for them for Layer 5, not as a replacement for Layers 1 to 4.

BigQuery export. The right answer if you have the volume and the skills. Join the AI Assistant channel against the unattributed Direct cohort and model your own dark traffic floor instead of borrowing somebody's 60% estimate. Overkill for most businesses.

Doing nothing and waiting for Google to fix it. The list already moved from three named platforms to five inside a month. Waiting for a clean, complete, official AI report is waiting for a bus on a road with no bus route.

The risks nobody mentions

False precision is the main one. The moment you put "AI traffic: 312 sessions" on a slide with no caveat, that number becomes the truth in your organisation and every decision downstream inherits its errors. Report ranges. Report the floor. Say what you cannot see.

The second risk is over rotating. AI referral traffic is still a small share of total web visits for most sites, somewhere around one percent by most published estimates, even after triple digit growth. High intent, small volume. Chasing it while your core organic and local visibility rots is a great way to win an argument and lose a business. If your money is still in the map pack, go and read the local SEO playbook before you spend a month on attribution.

Where this goes next

Three things I would bet on for the rest of 2026 and into 2027.

The referrer gets worse, not better. More agents, more in app browsing, more native assistants. The share of traffic arriving with clean referrer data is heading down. Which means first party signals, UTMs on everything you control, and post conversion survey questions like "how did you hear about us" are about to become serious measurement infrastructure rather than nice to haves.

Logs become a standard part of SEO reporting. When retrieval crawls outpace training crawls, bot behaviour stops being a technical curiosity and becomes a leading indicator. Crawl today, cited tomorrow.

The gap between measured and influenced keeps widening. The most valuable AI outcome, being named without a click, will never appear in an analytics tool. The businesses that win will be the ones that stop demanding a session for every bit of credit and start tracking the mention itself.

Everyone else will still be staring at Direct, wondering where their customers came from.

Do these five things this week

1. Open GA4, find the AI Assistant channel, and annotate 13 May 2026.
2. Build a custom channel group with a Perplexity rule, positioned above Referral.
3. UTM tag every URL you control before it goes anywhere public.
4. Pull a month of server logs and check which AI bots reach your money pages.
5. Ask four assistants your ten customer questions and log who gets named. Repeat monthly.

That is the whole system. Free, roughly two hours, and it puts you ahead of most agencies charging four figures a month to point at a chart they do not understand.

The measuring is the easy half. Getting named in the answer is the hard half, and that is what we drill inside The Dojo: the answer first structure, the entity work, the citations, the listicle placements, the link plays, all of it in order, with the tracker and the tools. If you would rather we just built the ranking asset and handed you the leads, that is Done For You.

Join The Dojo. $67/mo. Cancel whenever. But do not spend another quarter guessing where your best customers came from.