Tracking referral traffic from AI assistants

That is the part vendors skip, and it is the part that decides how you should read the report.

The practical consequence: if your analytics shows a handful of sessions from assistant domains, that is not the size of the effect. It is the size of the visible slice of the effect. Making a budget decision on it is like judging covers by counting people who mentioned the sign.

Why does the number undercount so badly?

Four reasons, stacked, and the first is confirmed by Google.

AI Overview and AI Mode clicks are indistinguishable from ordinary organic. Google states that sites appearing in AI features are included in overall search traffic in Search Console, within the Web search type. No referral parameter, no UTM, no separate analytics channel. The largest AI surface in local search is fundamentally unattributable, and that is a property of the system rather than a gap in your setup.

Referrers get lost. Native app clients, including the ChatGPT and Perplexity mobile apps, frequently strip the referrer. Referrer-policy headers strip it again on some HTTPS to HTTPS hops and inside in-app browsers. Those sessions land in Direct.

The journey has extra steps. People read an answer, then search your business name to check you are real, then click. Analytics credits the branded search, which is the last step and the least informative one.

Citations do not have to be clicks. Someone can read an answer naming your business, decide to call, and never touch your site. Kevin Indig's Growth Memo analysis found that 62% of the time, when an AI system uses your content and links to you, it does not name your brand in the answer text. Practitioner analysis, and Indig sells research in this category, so weigh it accordingly.

Every one of these pushes real influence into channels that look like something else. Direct traffic and branded search absorb it, which is exactly where most people stop looking.

What does Google actually publish?

Impressions, and nothing that resolves this problem.

Google launched Search Generative AI performance reports in Search Console in June 2026. The dimensions are impressions, pages, countries, devices and dates. Clicks, click-through rate, position and the user's prompt are not available, and AI Mode is not separated from AI Overviews inside that view. The data is also still counted inside the overall performance report, so it is a separate view rather than a separate dataset. Do not add the two together.

That is the ceiling. No third-party product has better access than the platform's own reporting, so any tool quoting you clicks or a click-through rate from AI Overviews specifically is generating that number rather than measuring it.

What happens when a metric is structurally unavailable?

The market invents one, and local SEO already ran this experiment.

Nobody can publish local pack click-through rate by position, because Google exposes no local-pack position dimension anywhere. Business Profile performance reports publish views, calls, direction requests, website clicks and messages. No rank. No position. Search Console has no local-pack dimension either. The number is structurally unmeasurable from public data.

So look at what circulates instead.

"Local pack position 1 gets 17.8% of clicks, position 2 gets 13.6%, position 3 gets 10.4%, per BrightLocal." The cited study does not exist. The URL returns a 404 and no such report appears in BrightLocal's research index. The attribution is fabricated.

"Local pack CTR is 24.4% / 13.3% / 8.6%, per the Moz Local CTR Study." Also fabricated, and published by the same aggregator that publishes the first set. Two contradictory figure sets, two studies that do not exist, one page family.

"Local pack CTR is 17.6% / 15.4% / 15.1%." FirstPageSage, which describes its own work as a meta-analysis with no sample size and no method, sourced to a list of blogs including Backlinko, which now cites FirstPageSage in return. Circular.

The one traceable source is a 2018 BrightLocal study using 5,500 Mechanical Turk testers looking at five San Francisco desktop search results pages, which found position one at 16.1%, position two at 10.4% and position three at 2.3% when Local Services Ads were present. Stated preference, five queries, one city, desktop, and a page layout that no longer exists.

Read that pattern and then look at AI referral traffic. Same conditions: a number buyers badly want, no way to observe it, and a vendor incentive to supply it. The AI version of the fabricated CTR table is already being printed. When someone shows you an AI-driven revenue figure, ask what dimension it was derived from, then check whether that dimension exists.

How do you set the tracking up anyway?

Because a floor is still worth having, and it trends.

Create a channel grouping in your analytics for assistant referrers, using the domains of the products your customers use. Keep it as its own channel so it does not get folded into a generic other bucket. Review the domain list quarterly, because the products keep adding and renaming surfaces.

Where you control a link that might get quoted, tag it. A UTM on a menu, a booking link or a page you promote gives you a clean read whenever it is used.

Set up conversion tracking on that channel so you can compare quality, not only volume. Small referral channels sometimes convert well, and volume alone will make you dismiss them.

And keep your ordinary Search Console reporting, since that is where AI Overview clicks actually live. What Google's own documentation says about optimizing for AI explains why: it is one search system, not two, so it reports as one.

What does the honest version of this measurement look like?

Two numbers you can actually produce, and one comparison between them.

The worked example. Over a 90 day window, your analytics shows 14 sessions from assistant domains. Over the same window you asked every new customer how they heard about you and recorded the answer, capturing 180 responses, of which 11 said they asked an AI assistant. That is 6.1%derived of new customers, against 14 sessions in a channel report.

The two numbers do not reconcile, and they are not supposed to. One counts sessions that arrived carrying a referrer. The other counts humans who remember using an assistant, including the ones who never visited your site at all. The intake field found an effect the analytics could not see, and it did it for the cost of one required form field.

What you may not do with them. It is tempting to take the 62% ghost citation finding and gross up your 14 sessions, dividing by the remainder to estimate true exposure. That is not available. The 62% is a share of citations that link without naming a brand, measured over a different population by a different method. It is not a multiplier on your session count, and applying it that way manufactures a number with no source. Report the 14 and the 11 side by side, each labelled with what it counts.

Then add the one validated proxy. Ahrefs published a method for isolating AI Overview effects in Search Console: filter for informational, question-starting, non-branded queries, then compare click-through rate before and after a known rollout date at constant average position. They validated it against confirmed AI Overview keywords. The guessed set showed a drop of 3.76 percentage points, or 42% relative. The confirmed set showed 3.98 percentage points, or 44% relative. The proxy and the ground truth differed by about 2% relative. It is vendor research, it uses data you already have, and it costs nothing.

What proxies are worth watching?

Three, none conclusive alone.

Direct traffic. A rise in direct with no offline campaign running is often referrer loss rather than mystery loyalty. Watch the landing pages: direct visits landing deep on a service page did not come from someone typing your domain.

Branded search volume in Search Console. If more people are searching your business name without any advertising change, something is putting your name in front of them.

Server logs. The only fully deterministic measurement in this subject, verified by IP range rather than user agent string. Logs prove that specific crawlers fetched specific pages, which is upstream of everything else.

Together, moving in the same direction over a quarter, they are a reasonable read. Size the channel honestly while you do it. Sterling Sky reported that for one of its largest multi-location clients, ChatGPT traffic grew from 0.1% of Google traffic in 2025 to 2% in 2026, which is still only 22% of what Bing sends that same client. One agency client is not a benchmark. The order of magnitude is the point, and local businesses should expect much of the effect never to touch a browser at all, since local intent resolves into calls and directions rather than clicks.

What is the most reliable attribution you actually have?

Asking people.

A single question at intake, asked consistently: how did you hear about us. Recorded, not just heard. That one field will outperform your analytics for this specific question, for the reason the worked example shows.

It works if you make it a habit rather than a survey. One line on the booking form, one question on the phone script, one field in the CRM. Staff will skip it unless it is required, so require it.

Older customers are a good source for the same question and for something else you need, since asking for a review from someone who loved you two years ago is a natural conversation to attach it to.

Also, before attributing any traffic change to AI, rule out the boring explanations. A profile suspension will crater your local visibility overnight, and what reinstatement actually involves is a more likely cause of a sudden drop than anything generative.

What to do this week

1. Build an assistant referral channel in your analytics and check it against the last six months of data. You now have a floor. 2. Segment direct traffic by landing page. Deep landings are the tell. 3. Add or enforce a how did you hear about us field, and make it required. 4. Run the Search Console proxy: filter for non-branded question queries and compare click-through rate at stable position across a known date. 5. Pull twelve months of branded search impressions as a second baseline. 6. Get thirty days of raw server logs, or find out that you cannot.

Do not let anyone convince you the fix for weak referral data is more markup. Whether structured data is required for AI features has a documented answer, and it does not change what your analytics can see.

Want the tracking setup without the guesswork?

Tell me which analytics platform you use at eric@seod.com and I will send you the exact channel grouping definition for assistant referrals, the current domain list, and the two segments worth saving. Paste it in and you are done. No charge, no pitch.

I offer this because the setup is fiddly, the documentation is scattered, and once it is built it keeps working without me.

Otherwise, there is more on AI search and AEO here.