The Short Version
AI assistants send real traffic to small business websites. It is smaller than almost every headline implies, it is growing, and it is tilted in your favour rather than against you.
Everything below carries the name of whoever collected it, its sample and its date. That matters more here than in most fields, because a lot of what circulates about AI search is a correlation sold as an instruction, a 2025 figure with the year quietly changed, or a conversion rate from a single website.
How Much Traffic, Really
Start with the number the industry keeps rounding up.
AI referral traffic is about three tenths of one percent of a typical site's total visits. Ahrefs measured 0.28% across 74,752 websites in March 2026. SE Ranking measured 0.3301% across 101,574 websites in April 2026. Semrush's median across 50,000+ sites was 0.14%. Aleyda Solis, analysing Semrush data across four verticals, found organic search running roughly 108 times larger than AI traffic.
The figure you will see quoted instead is 1.08%. That one comes from an enterprise sample, and its underlying data is from May to September 2025 despite a mid-2026 publication date. It is the ceiling, not the typical case.
The Part That Favours Small Businesses
Almost every scare story about AI search is written about publishers. Your numbers look nothing like theirs.
AI referral traffic as a share of all sessions, by vertical
Previsible's July 2026 study tracked 166 GA4 properties from December 2024 to May 2026. Small business sites went from 0.4% to 1.71% AI penetration, the highest of any vertical they measured. Publishers finished at 0.08%. That is roughly a 21x gap, and it is the opposite of the story you get from publisher-centric coverage.
Ahrefs found the same tilt by site size: sites under 1,000 monthly visitors took 0.37% of page views from AI, against 0.09% for sites over 10,000. Roughly four times the share for the smallest sites. That cut is from 2025 data, so treat it as direction rather than a current level.
Small sites are not winning because they are better optimised. They are winning because an assistant answering "who does X near me" has to name somebody specific, and a small specialist is often the most specific correct answer.
What A Normal Month Looks Like
On the sites we monitor, the median site received 14 visits from AI assistants in August 2026. The 90th percentile site received 96.
AI referral visits per site in a month · share of sites
Roughly 4 in 10 sites sit under 10 AI visits a month. About 1 in 10 clear 100. If you are at a dozen a month, you are normal.
AI referral visits to the typical site · March 2026 = 100
Across a fixed set of sites measured every month, the median roughly doubled between March and August 2026. Using matched windows on the freshest data, the first eight days of September ran about 50% above the same days in August.
We index this rather than quoting totals. Raw totals across a growing set of monitored sites mostly measure how many sites we monitor, not how much AI traffic exists.
Growth Is Slowing, And One Jump Was Not Demand
Two honest caveats on that 2.2x.
First, in share terms the industry curve is flattening, not steepening. SE Ranking's monthly series actually peaked in October 2025 and sat lower in April 2026. Semrush found the number of domains receiving ChatGPT referrals peaked around October 2025 and fell by February 2026. Similarweb's own trend chart shows a plateau from mid-2025 through January 2026.
Second, the biggest single jump of 2026 was a product decision, not a change in what people want. That is the next section.
The Concentration Problem
Between June and August 2026, the top tenth of sites took 65.7% of all AI referral visits we recorded. The 90th percentile site gets about seven times what the median site gets.
This is what you would expect from a system that answers a question once instead of listing ten options. When an assistant picks a business, that business takes the whole click and everyone else gets nothing for that query. Averages are meaningless here, so every figure in this report is a median or a share.
Two structural facts sharpen it. Semrush found that over 30% of all ChatGPT referral traffic goes to just ten domains, and that 21.6% of it goes to Google rather than to a business. So a fifth of what the industry counts as "AI referral traffic" is ChatGPT handing the user back to a search engine.
The Day ChatGPT Rewired Where Its Clicks Go
On 7 May 2026, ChatGPT started turning brand names inside its answers into prominent clickable links instead of burying sources in footnote-style citations.
OpenAI never announced it. There is no blog post, no changelog entry, and nothing in the ChatGPT release-note history about link prominence in 2026. It was detected by people measuring, and independently confirmed by at least four of them.
Similarweb measured ChatGPT referrals roughly tripling week over week, up 157.7%, with homepage referrals up 354.7% and the homepage share of ChatGPT clicks roughly doubling from about 30% to about 60%. A separate GA4 study across 52 sites, using matched six-week windows, found sessions up 97.9%, homepage sessions up 1,316%, and, tellingly, blog traffic down 28%. A third study ran 140,000 ChatGPT answers against a control set of 350,000 answers on Perplexity, Gemini and Copilot, and found only ChatGPT moved.
We can see the same event in our own data, at about half the size.
Homepage share of ChatGPT referral visits · %
The homepage share of ChatGPT referral visits went from 10.7% in March and 17.8% in April to 28.2% in May, and has held between 26% and 28% every month since. Weekly data puts the break in the week of 11 May, the first full week after the change: homepage share stepped from a 19% to 25% range up to a 28% to 32% range, and ChatGPT visit volume rose about 55%.
One confounder to disclose: GPT-5.5 Instant became ChatGPT's default model on 5 May, two days earlier. Nobody has separated the two effects.
The effect is real, it has held for four months, and it is roughly half as strong for small businesses as for the large domains in the panels. That makes sense. The change rewards being a recognisable brand name in an answer.
Two practical consequences:
- Your homepage now has to work as a landing page. A quarter of your AI traffic arrives there cold, having just been told by an assistant that you do something specific.
- One company's UI decision repriced this channel in a day, without telling anyone. Anything you build here should assume another 7 May is coming.
Where AI Traffic Lands
Deep pages take the majority. Blog posts take almost nothing.
Where AI referral visits land, by page type
Classifying every AI referral visit by URL path from March to August 2026: deep and topic pages 61.4%, homepages 23.7%, product and shop pages 9.4%, blog or article pages 3.3%. Checking per site rather than per visit gives the same answer, so it is not one busy site distorting things. For the median site, deep pages take 57.1%, the homepage 25.9%, and the blog 0%.
That median of zero is the finding worth sitting with. For a typical site, the blog contributes no measurable AI referral traffic at all. The 52-site GA4 study above found blog traffic actually falling 28% after 7 May.
This does not mean stop blogging. Blog content still earns the traditional rankings that assistants rely on to find you. It means publishing more short posts is not a route into AI answers.
The assistants also differ, in a way that tells you what each is doing:
| Assistant | Share of visits landing on the homepage |
|---|---|
| ChatGPT | 26.6% |
| Gemini | 18.5% |
| Perplexity | 15.7% |
| Claude | 5.0% |
ChatGPT is the most homepage-heavy, which is exactly what a brand-link change predicts. Claude almost always links the specific source page instead.
Which Assistants Send Traffic
Share of AI referral visits by assistant
ChatGPT accounted for 88.2% of AI referral visits on our sites between March and August 2026, and its monthly share never left the 86% to 91% band. Gemini took 6.6%, Claude 2.6%, Perplexity 2.5%.
Independent measurement puts ChatGPT a little lower, and it is worth showing both rather than only our own:
| Assistant | Our data | Statcounter, Aug 2026 | SE Ranking, 2026 | Sites we saw it reach |
|---|---|---|---|---|
| ChatGPT | 88.2% | 79.4% | 74.8% | 97.1% |
| Gemini | 6.6% | 10.9% | 11.6% | 44.3% |
| Perplexity | 2.5% | 4.3% | 7.2% | 25.6% |
| Claude | 2.6% | 2.6% | 2.6% | 23.1% |
Gemini is the only assistant gaining share, rising from 3.6% of our visits in March to between 6% and 9% since.
You may also read that ChatGPT's share collapsed to around 55%. That is a different metric: share of visits to the assistants' own sites, where ChatGPT fell from about 76% a year ago while Gemini reached roughly 26%. Its share of clicks sent out to websites has not fallen. Gemini answers simply link out far less often per visit. Putting a usage number next to a referral number is the most common error in recycled coverage.
If you ignore everything else here: optimise for ChatGPT, and check Gemini.
The Measurement Gap
The AI traffic you can see is a fraction of the AI traffic you get.
Graphite's March 2026 study, on data covering September 2024 to November 2025, found 83% of AI usage happens in mobile apps worldwide and 75% in the United States, concluding that web-only measurement undercounts AI usage by four to five times.
One caveat, which the study supplies itself: its author writes that they do not have access to mobile app data and cannot independently verify it. The 4x to 5x undercount rests on a figure nobody outside the data vendor has checked. Best available estimate, not a measurement.
Why the referrer disappears, correctly. Most write-ups say the ChatGPT app "strips" the referrer. That is the wrong mechanism, and it changes what you can do about it. In a mobile app the chat is native, not a web page. Tapping a citation starts a fresh navigation in an in-app browser with no referring document, so there is no referrer to send. Nothing is being removed, and no vendor policy change will fix your attribution.
But ChatGPT hands you a way around it. ChatGPT appends utm_source=chatgpt.com to the links it cites. Two published sources flatly disagree about whether it does this, so we checked our own data: between 76% and 89% of ChatGPT-referred visits carried that parameter every month from March to September 2026, and the share is rising. That parameter rides in the URL, not the referrer, so it survives the app handoff. To find ChatGPT traffic that landed in Direct, search your page paths for it.
One warning: never add utm_source=chatgpt to your own canonical URLs. UTM parameters outrank referrer detection in GA4, so you would relabel every visitor from every source as ChatGPT.
The Trap That Will Make Your Report Lie
If you take one thing from this section, take this one, because it is sitting inside measurement guides that are being shared right now.
Several published GA4 setups tell you to match AI traffic with a list of domains, and put search.brave.com on that list next to chatgpt.com and perplexity.ai. Do not. Brave Search sends a strict-origin-when-cross-origin referrer policy, which strips the path, so a click from a Brave AI answer and a click from an ordinary Brave blue link arrive byte-identical in your analytics. There is no hostname, path or parameter separating them. "Brave AI traffic" is not a measurable quantity from referrer data at all.
Put Brave on that list and you book 100% of your Brave organic search traffic as AI traffic. Brave runs over 1.6 billion queries a month, so on a privacy-leaning audience that single mislabelled bucket can outrank Gemini, Perplexity and Claude combined and show up as your second-biggest AI source. It is a reporting artefact with no AI referral behind it, and it is the most common way an AI traffic report ends up badly wrong.
Google's own AI Assistant channel names ChatGPT, Gemini, DeepSeek, Copilot and Grok, and classifies Brave as Organic Search. Follow that.
Stop Building A Custom AI Channel. GA4 Has One.
A lot of agencies have missed this. Since 13 May 2026, GA4 assigns a medium of ai-assistant automatically when the referrer matches a recognised assistant, and reports it under a default channel called AI Assistant. No configuration. If you still maintain a hand-built AI channel group, check you are not double-reporting.
Two blind spots, both from Google's own documentation. It is referrer-based, so referrer-less app visits still land in Direct. And it excludes Google's own AI Overviews and AI Mode, which arrive as google / organic.
Search Console is worth setting expectations on too. Google's generative AI performance report finished rolling out worldwide on 31 August 2026, and it gives impressions only, with no clicks and no CTR, combining AI Overviews and AI Mode. The clicks are already inside your ordinary Web totals with no way to separate them. You cannot calculate an AI Overviews click-through rate in Search Console. The data is not there.
| Method | What it captures | What it misses |
|---|---|---|
| GA4 AI Assistant channel | Clicks from third-party assistants that pass a referrer | App clicks with no referrer, and all Google AI surfaces |
utm_source in URLs | ChatGPT clicks that lost their referrer, including from apps | Assistants that do not tag their links |
| Mention tracking | Whether you appear in the answer text for a question | Whether anyone clicked |
| Crawler log analysis | Which AI bots read which pages, and how often | Humans, since bots are not customers |
Is It Worth Anything?
Probably yes. The honest answer is that the direction is consistent and the size is unknown.
The one study aimed at businesses like yours. Orbit Media looked at 97 B2B lead-generation websites, 28.9 million sessions over the year to June 2026, and measured whether AI visitors actually filled in a form or clicked to call. Visitors from ChatGPT converted into leads at 2.1%, against 0.5% from Google search. That is the closest thing to a service-business number that exists, and it is encouraging.
The retail picture, and the reversal. Adobe reports AI traffic to US retail sites converting 42% better by March 2026 and 54% better by May, having converted 38% worse in March 2025. Eleven straight months of outperformance puts the crossover around September 2025. Adobe also measured 53% more time on site and 53% higher revenue per visit.
Now the caveats, because Adobe is carrying that headline. Adobe never discloses how many AI visits the percentages rest on, never names a single assistant, and never says how it identifies AI traffic. Its comparison group is all non-AI channels pooled, including email and display, which is the most flattering possible baseline. And in the one non-retail vertical where Adobe publishes conversion, travel, AI still converted 28% worse in May 2026, with that gap having widened from 14% in March. For financial services Adobe publishes engagement and time on site but no conversion figure at all. For a service-business audience, that silence is itself informative.
The strongest counter-evidence, and the only peer-reviewed study in the field. Published in Marketing Science in April 2026, it analysed first-party data from 973 ecommerce sites, comparing more than 50,000 ChatGPT transactions against 164 million from traditional channels. Organic search outperformed ChatGPT on conversion by roughly 13%. The abstract: results "contradict widespread expectations of LLM superiority."
And engagement is contested too. Adobe's numbers above come from its retail panel. Ahrefs ran the same comparison across 81,947 sites and found AI traffic worse on two of three measures: 4.0 pages per visit against 5.2 from search, and a 67.8% bounce rate against 63.7%. Only time on site favoured AI. Be careful quoting any single engagement figure as settled.
And the materiality check. Contentsquare, across 99 billion sessions on 6,500 sites, found AI-referred traffic converting at 1.3%, up 55% year over year but still below email at 1.9%, on 0.2% of total visits. A better-than-organic rate on two tenths of a percent of your traffic is not yet a channel.
Three things worth knowing that almost nobody reports:
- The famous number is one website. "ChatGPT converts at 15.9% against organic at 1.76%" comes from a single client site. The same firm's later aggregate study reassigns that 15.9% to Perplexity and puts the organic baseline at 9.77%. It circulates everywhere as an industry benchmark. It is not one.
- Order values are lower, not higher. Two independent studies put average order value from AI referrals 14% and 26% below organic, and one found 85% of AI-referred buyers were first-timers against 42% for the site overall. This behaves like a customer acquisition channel, not a revenue-per-visit channel.
- Nobody has followed it to money. The only study tracking AI traffic through a CRM to closed-won found a median of 0% reaching closed-won across 118 accounts. High intent is real. Booked revenue mostly is not, yet.
The Number That Actually Settles This
Every figure above compares one channel against another, which cannot separate "AI sends better visitors" from "AI only sends the visitors who were already going to buy."
One study does separate them, and it is the most useful thing we found. Researchers used the staggered international rollout of Google AI Overviews as a natural experiment, running a difference-in-differences analysis across 4,658 ecommerce domains in 131 markets, comparing markets that had already received AI Overviews against ones that had not yet. That design measures the effect rather than the correlation.
The result: organic sessions fell 2.8%, while conversion rate rose 2.6% and revenue per session rose 2.1%. Transactions and revenue barely moved, and neither change was statistically distinguishable from zero. In the authors' words, AI interfaces "reduc[e] the number of visits a firm receives while raising the commercial qualification of the visits that remain."
So the qualification effect is real, and measured properly it is about 2.6%, not 4x or 9x. Fewer visits, slightly better ones, roughly the same revenue. That is a far less exciting story than the channel comparisons, and it is the one with the strongest design behind it.
Why the other numbers stay soft. No published study compares AI traffic against branded organic or high-intent paid, which is the only fair comparison. The mechanism works against measurement: people ask an assistant, then search for the brand on Google, then buy, and that conversion gets credited to branded organic. So the visitors counted as "AI referrals" are only those who clicked straight through, a doubly selected group. As Rand Fishkin put it when reviewing one of the bullish studies, the unanswered question is whether those users "would have found these brands anyway."
The Counterweight
Demand for AI-recommended businesses is real but earlier than the hype suggests, and BrightLocal's own two studies show the gap.
Their March 2026 survey found consumers using AI for local recommendations rising from 6% to 45% in a year, putting AI third as a discovery channel, ahead of Yelp. But their July 2026 study, which asked 1,227 people what they actually did on their most recent local search, found something much soberer: 52% started on Google Search and just 8% chose AI first. Of those who started with AI, 43% went back to Google for more. And only 18% felt ready to contact the business the AI recommended.
Read together, that is the real state of things: rising interest, thin behaviour, and a long way from a booking.
The wider ecosystem numbers point the same way. Chartbeat data across 2,576 publisher sites found Google organic search referrals down 33% globally and 38% in the US in the year to November 2025. The Reuters Institute's own summary of what replaced it is blunt: ChatGPT referrals are "little more than a rounding error", and Google delivers 500 times as many referrals from search alone, 1,300 times as many counting Discover.
The strongest evidence here is not a study, it is a set of financial filings. People Inc., the publisher behind Investopedia and Southern Living, told the SEC in August 2026 that its core sessions were down 22% year on year, attributing it directly to "the increasing prominence of Google AI Overviews". Its own investor materials show Google search falling from 55% of its sessions in early 2024 to 25% in early 2026. Chegg reported non-subscriber traffic down 49% and has since stopped publishing the figure at all.
Two caveats keep this honest. The window does a lot of the work: that same Chartbeat dataset shows a smaller decline over two and a half years (21%) than over one (33%), because search referrals rose before they fell. And not every affected company blames AI. Yelp's own annual filing attributes its traffic decline to the economy and lists AI Overviews only as a future risk, while Reddit is facing shareholder litigation alleging it understated the impact. The pressure is real; the attribution is contested.
For your purposes the pattern that matters is this: the search decline hits small sites hardest (small publishers lost 60% of search referral traffic over two years against 22% for large ones), and AI referral is nowhere near replacing it.
On the click side, there is now exactly one randomised experiment. Published August 2026, it gave 1,100 US Chrome users a browser extension for a week and found that forcing every search into AI Mode cut click-through to external sites by 18.8 percentage points, while removing AI features raised it by 8.8. Everything else in this area is correlational.
And be careful with the most-quoted AI Overviews figure. Ahrefs reports that top-ranking pages on AI Overview keywords have a 58% lower click-through rate. But their own control group fell too: over the same two years, position-one CTR dropped 49% on keywords with no AI Overview against 78% on keywords with one. Clicks fell almost as far without AI. Meanwhile Semrush ran the better design, tracking 200,000+ of the same keywords before and after they gained an AI Overview, and found zero-click actually fell, from 33.75% to 31.53%. Their conclusion: this "challenges the assumption that AI Overviews always reduce clicks."
So: a small channel, growing, with better-than-average visitors, that does not replace search. Worth preparing for. Not worth reallocating your budget to.
What To Actually Do
Ordered by strength of evidence, not by how good it sounds. Where a controlled test exists, we used it.
1. Make your facts readable in plain text. Name, what you do, who you serve, where you operate, how to contact you, in text on the page rather than in an image, a slider or a script. This is the single biggest differentiator between our 90th percentile sites and our median ones.
2. Add statistics, direct quotations and cited sources. Do not bother restructuring. Two independent controlled studies agree here. A 252,000-trial study across six models found topical relevance and list position dominate, that adding explicit prices and recent dates helps consistently, and that formatting-only edits had little impact. The academic GEO study found the biggest gains from adding quotations (+42.6%), statistics (+32.8%) and cited sources (+27.7%), while keyword stuffing went backwards. There is no controlled evidence that chunking, Q&A blocks, or answer-first formatting increases citations. Be quotable, not comprehensive: one analysis of AI Overview winners found only 4.4% of slots went to 5,000-word guides.
3. Fix your homepage as a landing page, and give your specific services their own pages. A quarter of AI visits arrive on the homepage promised something specific. Deep pages take the majority.
4. Get your ratings up, and stop replying to reviews for the algorithm's sake. The only causal study in local AI recommendation randomised hotel attributes across 12 models and 61,000 calls. Moving from a 3.9 to a 4.7 guest rating raised the chance of being recommended by 31.7 percentage points. A 47-fold increase in review volume bought only 8.3 points. Responding to reviews measured at 0.11 points, with a confidence interval straddling zero. Reply to reviews because customers read them, not because it wins AI recommendations.
5. Fix the right listing for the right assistant. These diverge sharply. Across nearly 2 million local citations, Google Business Profile was 28.63% of the total, but almost all of those came from Google's own AI surfaces and GBP does not reach ChatGPT's top three at all. For ChatGPT, Yelp appeared as a source in 80% of local answers. Also worth knowing: around 93% of cited domains were individual business websites, so the long tail is your own site.
6. Know which crawler does what before you block anything. This is the most common self-inflicted wound we see. Per OpenAI's own documentation, GPTBot is training only, OAI-SearchBot gates whether you appear in ChatGPT search, and ChatGPT-User fetches on a user's behalf. Blocking GPTBot to keep your content out of training does not remove you from ChatGPT. Blocking OAI-SearchBot does. Most "block the AI bots" advice conflates them. If no AI crawler is reaching your pages at all, content is not your problem. You can check what AI crawlers can read on your site in a couple of minutes.
7. Search your page paths for utm_source=chatgpt.com. The cheapest win here, for the reasons in the measurement section.
8. Add JSON-LD if it helps you, but not for citations. This one has a clean answer now. Ahrefs added schema to 1,885 pages and compared them against 4,000 matched control pages using a difference-in-differences design. AI Overview citations went down 4.6%; AI Mode and ChatGPT moved within noise. The same report found AI-cited pages were nearly three times more likely to carry JSON-LD, which is the correlation everyone quotes. One study, both numbers, and the causal test came back empty. An independent test found that a price existing only in JSON-LD was read by zero of five assistants. Google's documentation states plainly that you do not need special markup or AI text files to appear in its AI features. In the authors' own words, if the only reason you are adding it is more AI citations, "our data doesn't support that bet."
9. Skip llms.txt. Ahrefs checked 137,210 domains: 28% publish a valid llms.txt, and 97% of those files received zero requests in May 2026. Of the requests that did arrive, the largest single category was SEO audit tools. Separately, an analysis of 584,107 published files found 22% contained no links and 10% were just robots.txt files saved at the wrong path. No AI vendor has stated it reads the format. Our own site publishes one, and on this evidence that is habit rather than strategy.
One Thing To Be Sceptical Of
"Brand mentions beat backlinks three to one for AI visibility" is the most laundered statistic in this field. The underlying figure is a rank correlation of 0.664 for branded mentions against 0.218 for backlinks, on a filtered sample, published by the vendor selling brand-mention tracking, with an explicit correlation-is-not-causation note in the original. It is circulating as an instruction. Treat it as a correlation.
Reddit deserves the same caution. It is heavily cited across assistants, but ChatGPT's citation of Reddit went from close to 60% of responses in early August 2025 to around 10% by mid-September. Advice built on citation share has a half-life measured in weeks.
Does Traditional SEO Still Matter?
Yes, but as a floor rather than a mechanism, and the honest version is more interesting than the slogan.
Ranking well is the best-measured correlate of getting cited. seoClarity, across 362,000 keywords and 5.1 million citations, found 94% of AI Overviews included at least one page from the top 20. Seer found a correlation of about 0.65 between page-one Google rankings and being mentioned by an LLM, with backlinks "weak or even neutral."
But rank is neither necessary nor sufficient. Ahrefs, across 863,000 SERPs and 4 million AI Overview URLs, found only 37.9% of citations came from the top 10, and about 31% came from pages not ranking in the top 100 at all. An academic study of 11,500 real queries found source overlap between AI answers and organic results below 0.2 on a Jaccard measure.
So: almost every AI answer contains at least one page that ranks, while a large minority of individual citation slots go to pages that do not. Your existing SEO is what makes you eligible. It does not decide the outcome. Whether SEO is still worth it alongside ChatGPT covers how the two fit together, and if you suspect you are starting from zero, check the signs your site is invisible to AI search first.
One more reason not to optimise for a single assistant. A study of 2.05 million local questions across four engines found all four picked the same top business in just 4% of cases, and four different top businesses in 48%.
How We Measured This
First-party figures come from LovedByAI's own event data, one row per human visit that arrived from an AI assistant, March to August 2026, with September where noted. The assistant is identified from the referrer. Landing pages are classified from the URL path only, never the hostname, so a product code cannot be mistaken for a page type.
Growth uses a fixed set of sites present in every month, then the median per site, indexed to March. Totals across a growing set of monitored sites would measure onboarding, not demand. Distribution figures are shares of sites. The August "typical site" figures exclude sites not already active in July, so part-months do not drag the median down.
One limitation worth stating: our event data only contains AI visits we could identify. Every row has a referrer or a tagged URL. So we cannot measure our own invisible fraction, and we share the blind spot described above. The ChatGPT URL parameter recovers part of it, not all of it.
External figures carry their collector, sample and date every time they appear. Where a source's data is from a different industry or an earlier year, we say so rather than rounding it into ours.
A limit that applies to every per-assistant split in this report, ours included: it covers only the trackable slice. Google's AI Overviews and AI Mode produce no separate referral, so they are absent from every platform share above. Similarweb puts AI Mode's referral rate at 1.6% to 2.5% of queries against 17% to 19% for traditional Google results, so that invisible slice does send clicks, at roughly a tenth the rate. Nobody can currently measure it per site.
