Most of the advice on this question fails in the same way. It either tells you AI is already a third of search, or it tells you AI sends almost no traffic and can be ignored for now.
Both positions are built on real numbers. They are just numbers about different things, and neither one answers the question you asked.
How often does AI answer instead of sending someone to a website?
On roughly four in ten Google searches in the United States.
Similarweb's own page reports a 43%-plus AI Overview appearance rate in the United States. Coverage of their 2026 report put the figure a year earlier near 15%, though that comparison does not appear on Similarweb's own page. Other trackers disagree, and the spread is wide. One recorded 18% in June. Another put it near 48%.
Most of the disagreement is about which searches get counted. Trending searches about sports and celebrities rarely get an AI answer. Question-style and commercial searches get one far more often, and those are the searches your customers run. Four in ten is closer to a floor for the queries that matter to you than an average you should discount.
A field experiment that watched real people browse, rather than crawling a keyword list, observed an AI answer on 36% of participant searches. That is the closest thing to a reality check available, and it lands in the same place.
One distinction is worth getting right, because it is where most articles go wrong. AI Overviews are the summaries that appear above the normal blue links, and those are now everywhere. AI Mode is the separate chat-style tab. Google said it passed a billion monthly users in May 2026, but clickstream data put it at just 0.34% of searches between January and April 2026. When people say AI has taken over search, the thing they are describing is Overviews.
The outcome shows up in how often a search ends without anyone leaving Google. SparkToro, working from Similarweb clickstream data, found that 68% of US Google searches ended without a click between January and April 2026.
SparkToro is careful about what that proves, and the caution is worth repeating. Their earlier measurement used a different panel, so the rise over time is not a clean comparison. The study does not isolate how much of it AI Overviews caused. It also leaves out searches made in Google's mobile app, where the rate is probably higher.
What does an AI answer cost you in clicks?
About 40% of them, according to the only study that tested this rather than inferred it.
Two researchers, at the Indian School of Business and Carnegie Mellon, ran a randomized experiment on 1,065 desktop Chrome users. A browser extension hid AI Overviews from some participants, chosen at random, and left them in place for others. Clicks out to websites fell 39.8% when the summary was shown. Searches ending with no click at all rose 34.5%. Clicks on ads did not move.
The design is what makes it valuable. Every other study on this compares before to after, which cannot separate the AI answer from everything else Google changed in the same window. This one assigned the feature at random.
It also tested Google's defense. Google has said the clicks lost to AI answers were low-quality ones that would not have converted anyway. The researchers found no difference in bounce rate, time on site, or how often people returned to search. The lost clicks looked like ordinary clicks.
Two caveats belong with that number. The paper is still a working draft on SSRN rather than peer-reviewed work, and nobody has replicated it. In its favor, the design was registered in advance, so the researchers committed to what they would measure before they saw any results. It is the best evidence available, not the last word.
The correlational studies point the same direction at different magnitudes. Ahrefs, across 300,000 keywords, found the top result on a query with an AI answer fell from a 7.3% click rate to 2.6%. Seer Interactive tracked 25.1 million impressions and recorded click-through on those queries dropping from 1.76% to 0.61%.
One detail from the Seer study gets left out of almost every write-up of it, and it matters. Queries without an AI answer declined over the same period too, from 2.74% to 1.62%. Clicks are getting harder to win across the board. AI answers made it worse where they appear, but they are not the only thing happening.
The direction is not in dispute. The size sits somewhere between a third and two thirds, and the most trustworthy study puts it near 40%.
Why does none of this show up in my analytics?
Because a click from a Google AI answer is recorded as ordinary organic search.
Nothing tags it as AI. The visitor searched on Google, read a summary, clicked a link inside it, and arrived through the same channel as every other Google visitor. Your reports have no way to tell them apart.
The gaps compound from there. People using the ChatGPT mobile app often arrive with no referrer at all, so they land in Direct. Only a click out of a browser tab on chatgpt.com or perplexity.ai gets filed as an AI referral.
That is why the published figures for AI traffic look so small. Contentsquare, across 99 billion sessions on 6,500 sites, put AI-referred traffic at 0.2% of visits. Other vendors measuring the same thing report 0.32%, 0.43% and 1.08%. A five-fold spread tells you how unsettled the measurement still is.
Every one of those numbers is honest, and every one of them is narrow. They count people who clicked out of a chat window. They do not count the much larger group who read an AI summary on Google. They also miss the group who asked an assistant for a shortlist, then typed a competitor's name into the address bar a day later.
So your dashboard can report that AI sent you almost nothing while an AI answer sits above your listing on half the searches you care about. Both are true. They answer different questions.
Does being named in the AI answer actually help?
The evidence says yes, and it stops short of proving it.
Seer compared queries where a brand was cited in the AI Overview against queries where it was not. Cited brands saw organic click-through of 0.70%, against 0.52% for brands that were not named. Seer states plainly that it cannot prove the citation caused the difference. That caution is fair, since a brand prominent enough to get cited may have been winning clicks anyway.
The click-through is the smaller half of the value regardless. When an assistant answers a question like "what is the best help desk software for a small team" with three names, the person reading it has been handed a shortlist. Whether they click now or search your name next week, you are on that list and your competitors are not.
Is the buying moving inside the chat?
Not so far, and the attempt to move it there went backwards this year.
OpenAI launched Instant Checkout so people could buy without leaving ChatGPT. A Forrester analyst counted roughly 30 Shopify merchants live on it, against the million-plus the launch had implied. Product details were often wrong and merchants proved hard to onboard. By March 2026, OpenAI had moved checkout back out of the chat.
Etsy, one of the largest partners, summarized it neatly. ChatGPT had become a valuable channel for discovery, while purchases made through it stayed low.
That is the useful conclusion. Discovery moved into the assistant. Buying stayed on your own site. The work in front of you is not a payment integration or a new protocol. It is whether the assistant names you when it builds the shortlist.
So should I worry?
Worry about your position in the answer, not about the size of the channel.
The size will take care of itself and you cannot influence it. Your position can be changed, and it is the slow part. Getting named in answers where you are currently absent usually takes two to three months, because the sources an assistant reads have to change first. Review coverage and comparison articles carry most of that weight, and that is where the evidence is strongest.
That is what makes waiting costly. The instinct is to hold off until AI shows up as a real line in the analytics. But that line will never appear, for the reasons above, and the months of work your position needs do not begin until you begin them.
The first step costs nothing and takes an afternoon. Open Google Search Console and compare impressions against clicks for the past twelve months. If impressions held up while clicks fell, something is answering your customers before they reach you. Then ask the assistants what your customers ask. Try "best help desk software for a 20-person company", or "a budgeting app that syncs with my bank", or "protein powder with no artificial sweeteners". Write down who gets named.
You will either be on the list or you will not. Either way you will know where you stand, which is more than your analytics can tell you.