Measuring AI Traffic Honestly: What the Numbers Do and Do Not Say

Two statements about AI traffic are both defensible and sound contradictory. It is well under one percent of visits for most websites. And it converts several times better than search traffic. The first makes it easy to ignore; the second makes it easy to over-invest in. Neither should be acted on before you know how it was produced — and the first one, as it turns out, is disputed by a factor of three.

This note is about the measurement rather than the hype: what the published figures say, why they disagree by an order of magnitude, why the direction is nonetheless credible, and how to measure it on your own site instead of borrowing someone else’s average.

What the published figures say

Nearly all of the widely circulated numbers come from analytics vendors and agencies measuring their own client portfolios. Worth reading; not a census.

Two of the larger panels are worth naming, because they show how much the answer depends on who is counting. A study of 101,574 websites across sixteen months to April 2026 put AI referrals at 0.32% of all traffic — up from 0.02% two years earlier, but a third of the “about one percent” figure that circulates in agency reporting. The same study found those visitors spending around 68% longer on the site than visitors from search. On platform mix, StatCounter’s panel of over a million sites recorded ChatGPT at 78.16% of assistant referrals in March 2026, with Gemini at 8.65% having overtaken Perplexity at 7.07% — a reversal from a year earlier, when Perplexity was second at 12.07% and Gemini was at 2.31%.

Note what that implies: second and third place swapped inside twelve months. A tactic built around one assistant is built on a share that has recently moved by a factor of five.

A spread of 1.4× to 20× on the same question is not a detail. It means the honest summary is “assistant traffic appears to convert better, by an amount nobody has established”, and any plan built on a specific multiple is built on a choice of source rather than on evidence.

Why the numbers disagree

Six causes, each moving the result in a predictable direction. Knowing them lets you read any new study in about a minute.

• Whose sites were measured. A vendor’s client base is not the web. Portfolios weighted towards software, B2B services or high-consideration purchases will report high assistant conversion, because those are the categories where people ask an assistant before buying.
• What counted as a conversion. Some studies count purchases. Others count newsletter signups, demo requests, or any goal completion. A 16% conversion rate against “any tracked event” and a 2% rate against “completed order” are not comparable, and the difference is rarely on the chart.
• Tiny denominators. At one percent of traffic, 5,000 monthly visits means fifty assistant sessions. Two purchases is 4%; three is 6%. Much of the reported variance is small-sample noise presented as a finding.
• Attribution loss. Referrer information from assistants is inconsistent — some sessions arrive clean, some as direct traffic. What is lost is lost unevenly across assistants, distorting the share figures far more than the conversion figures.
• What the denominator was. The gap between 0.32% and 1% is mostly about which sites were in the base: a panel of a hundred thousand ordinary websites and a portfolio of AI-attentive clients measure different populations, and neither is wrong about its own.
• Branded intent. Someone who asks an assistant for your address was already your customer. Unseparated, those sessions inflate the conversion rate of a channel that acquired nobody.

Why the direction is still credible

The magnitude is unreliable. The direction is not, and there is a mechanism behind it rather than a correlation.

A person arriving from an assistant has usually already had the comparison conversation — what the options were, what the differences were, what to watch out for — and arrives with a shortlist of one or two. They are near the bottom of the funnel by the time they click, and bottom-of-funnel traffic has always converted better.

Second, the low-intent traffic is stripped out. Assistants answer definitional questions in place — exactly the traffic that historically arrived, read one paragraph and left. Removing it raises the conversion rate of what remains without anything improving: the same accounting effect described in our note on what zero-click numbers actually measure, seen from the other side.

Both mechanisms are real and both mean the same thing: the channel is small, qualified, and partly made of demand that used to arrive through search.

Measuring it on your own site

Your own data beats every published average, and the setup takes an afternoon. Four signals, each catching what the others miss.

One of these signals recently improved: Search Console now reports impressions inside AI Overviews and AI Mode, covering visibility even when no visit follows — with the caveats set out in our note on reading the generative AI report.

Set up the channel group first: a rule mapping known assistant hostnames into one named channel, so reporting is comparable month to month rather than rebuilt by hand each quarter. Then add the free-text question to your enquiry form — in practice it recovers a meaningful share of the sessions analytics recorded as direct.

Then measure three things and only three: sessions from the channel, enquiries or orders from it, and the absolute count of both. Rates are for comparison; counts are for decisions — the same discipline of reading a median rather than a headline that we set out in our note on what a conversion benchmark actually tells you.

One caution on tooling: a channel this small vanishes into noise if half its sessions go unrecorded, so it needs analytics that survive consent choices — the options are in our note on measuring with less data.

Where this traffic actually lands

One column in your analytics is more useful than the conversion rate and almost nobody looks at it: the landing page. Assistant referrals rarely arrive at your homepage. They land on the specific page that answered the question — an old explainer, a specification table, a pricing page, a single FAQ entry buried three levels down.

Which creates a problem worth fixing, because those pages were never designed as entry points. A page written for someone already browsing your site assumes context the new arrival does not have: it does not say who you are or what you sell, and often offers no next step beyond the footer. A visitor who arrives having already decided you are a candidate supplier hits a wall no homepage visitor ever sees. It is the most concrete fix in this note and it costs an afternoon.

Take your top ten pages by assistant referral — or by direct traffic to deep URLs, if referrers are stripped — and check three things on each.

1. Does it orient a stranger in one line? Not a mission statement — one sentence on what the company does and which market it serves.
2. Is there an obvious next step? One clear route onward — the relevant service page, the price list, the enquiry form. Not five equally weighted links, and not only the navigation menu.
3. Can they contact you without hunting? A visible contact route on the page itself, not a phone number four screens below the last paragraph.

These are the same omissions that quietly suppress enquiries from every other channel, catalogued in our note on the seven reasons a site produces no enquiries — assistants simply make them expensive faster, by delivering pre-qualified visitors to your least prepared pages.

There is a second reason to know which pages these are: they are the ones being quoted, so a stale price or a service you no longer offer does the most damage there. Accuracy work is cheap aimed at ten known pages and impossible aimed at a whole site.

What this should and should not change

Three decisions are defensible on current evidence, and two commonly proposed ones are not.

Worth doing: make sure your key pages answer the questions an assistant would be asked, in a form it can quote — a direct answer near the top, specific numbers, plain structure. Keep your factual details consistent everywhere, because assistants reconcile across sources and a wrong price in a directory becomes a wrong price in an answer. And track the channel separately, so that in a year you have your own trend rather than someone else’s estimate.

Not worth doing yet: moving real budget out of channels that produce volume into one that produces thirty sessions a month, on the strength of a multiple that varies twentyfold between studies. And paying a monthly retainer for “AI visibility optimisation” whose deliverable is a dashboard of impressions and whose mechanism is the content work you were already buying. The evidence on what influences citation is thinner than the market implies, and it is ranked honestly in our note on citation factors by strength of evidence.

The proportionate position in 2026: treat assistant traffic as a small, high-quality channel, measure it properly, and feed it with content work you would do anyway. Revisit when your own numbers — not the industry’s — say it is worth more.

Key takeaways

• Well under 1% of visits, several times the conversion rate — the share is disputed by a factor of three (0.32% across 101,574 sites against the ~1% in agency reporting), and the conversion multiple is far less precise than it is usually presented.
• Published multiples range from about 1.4× to over 20×, because studies differ in whose sites they measured, what counted as a conversion, and how much attribution was lost.
• One percent of a small number is a handful of sessions. Decide on absolute counts, not on percentages.
• The direction is credible for two structural reasons: visitors arrive after the comparison stage, and low-intent traffic is answered elsewhere.
• Check where it lands. Assistant visitors arrive on deep interior pages that were never built as entry points; orienting those ten pages and giving them a next step is the cheapest work available.
• Measure your own: a custom channel group for assistant referrers, server logs for crawlers, and a “how did you hear about us” field.
• Do the content and accuracy work; postpone the budget reallocation until your own trend justifies it.