TL;DR: Key Referral Data Insights from the First Three Months of AI Mode Usage (May–August 2025)
- External clicks from AI Mode stay under ~3%. Traditional Google Search sends far more traffic.
- Clicks cluster on Reddit, YouTube, Wikipedia, Amazon, Facebook. Developer hubs like GitHub, Stack Overflow, and Stack Exchange show up.
- Post-click depth from AI Mode is thinner: fewer pageviews and shorter sessions on average than visits from Google Search.
- Two possible hypothetical reasons for short sessions: a) cited links are often broken, stale, or off-target, b) many users are quick-checking a claim from the AI answer and leave after confirming.
- AI Mode query length trends up (roughly 7–8 words in May to 10–11+ in August). ChatGPT stays near ~70.
- Sessions per user dipped, then recovered. No habit loop yet.
Note: Each of the graphs below is interactive. You can hover over them to see precise data or click on legend segments to see a different visualization of the data.
The narrative that AI Search delivers less traffic compared to traditional Google Search is entrenched. In fact, we’ve seen over 50% better conversation rates from AI referrers across all clients that we’ve analyzed this year. Semrush has claimed that AI Search traffic delivered 4.4x value compared to traditional search.
AI Mode specifically generates fewer clicks compared to classic Google Search. The clicks it does send concentrate on big, trusted, and UGC-heavy domains. Query length in AI Mode is trending up, which signals people are learning to abandon ‘keywordese’ and embrace conversational search, not that clicks will surge. Treat AI Mode as a distinct channel with low referral volume and high variance. Build presence where it already exists, and measure citation prominence and assisted impact, not just traffic.
Before diving into the data, let’s review the sequence of events since the last time we checked in.
What changed since last month
August 4 – The Google sales team is lining up AI Mode Ads for aggressive Q4 testing. Shout out to Anu Adegbola for the scoop. Agencies were briefed in early August on conversation-level targeting inside AI Mode. If those units perform, the case for a default or quasi-default experience gets stronger because the revenue model travels with it.
August 21 – Google took AI Mode global. The rollout hit 180+ countries and came with talk of agentic features like bookings and task help. I haven’t seen those live in the wild yet, which suggests staged or Labs-gated releases. Direction is clear. Scale first, deeper capability later.
September 5 – Logan Kilpatrick replied “Soon” to a post pushing for AI Mode to be the default version of search after an announcement that Google had created google.com/ai as a direct straight to the AI Mode interface destination. Momentum looked real.
September 7-8 – Then came the soft pedal. Robby Stein says not to read too much into “soon,” framing the near-term focus as easier access for people who want AI Mode, not a hard switch to default. A day later, Kilpatrick backed that and clarified he was not saying AI Mode replaces main Search right now. Signal remains. Timeline stays murky.
September 18 – Google announces that they are now building AI into Chrome on U.S. desktops.
You get omnibox entry to AI Mode, a side panel that answers questions about the page, multi-tab summaries, and deeper hooks into YouTube, Maps, and Calendar.
You get contextual search.
Now things will get interesting. Friction drops. Will usage rise? And all of this landed after the DOJ remedies phase confirmed Google would not be forced to break off Chrome, which makes the timing notable.
What this points to
Google is removing access friction, seeding usage in Chrome, testing monetization, and expanding reach. Public messaging stays cautious, yet every move increases the odds that AI Mode becomes the primary search experience in the next 6–12 months. That is the working assumption I’m planning against.
Methodology
SimilarWeb provided Elias Dabbas and our team with daily relative data points from over 100K AI Mode users between the timeframe of May 20th (AI Mode’s public launch date) to August 19th on the following metrics:
- Referral Rate: The percentage of searches that led to at least one click to a non-Google site. Each search is counted once: if it triggers one or many external clicks, it qualifies as a “referred” search. (Formula: Referred searches ÷ Total searches.)
- Referral Time on Site (minutes, average): The average time users spent on the external site after a referral click. Calculated across all referred visits that day. (Formula: Total minutes on referred sites ÷ Number of referred visits.)
- Referral Pageviews (average): The average number of pages viewed on the external site after a referral click. Calculated across all referred visits that day. (Formula: Total pageviews on referred sites ÷ Number of referred visits.)
- Referral Share (Top Referred Sites): For a given external domain, the share of all outbound clicks that landed on that domain. Each click is an event in the denominator set. (Formula: Clicks to Domain X ÷ Total outbound clicks.)
- Average Query Length: The average number of words in the initial query that opened the experience. For AI Mode, only the first query is captured; follow-ups create new URLs that aren’t yet query-level trackable. In some cases, the first query may have been entered in classic Google before the user clicked into AI Mode. (Formula: Total words in first queries ÷ Number of first queries.)
This data continues to highlight a contextual moment in time. Google is aggressively promoting AI Mode on social media, through advertising, and with disruptive pop-ups in Search, but it still creates friction due to limited access points.
We now have three months of data from May to August 2025. In our analysis, we are considering the type of person who would push past the friction to utilize the functionality. This data continues to highlight a contextual moment in time. Google is aggressively promoting AI Mode on social media, through advertising, and with disruptive pop-ups in Search. We should also understand that the novelty of it could lead to experimentation, as opposed to genuine, normal search behavior.
That was a major caveat that was addressed in our previous articles on the AI Mode data.
Revisit our initial referral report from the first month of data.
Google Trends: Most People Still Aren’t Searching for AI Mode
Despite all the announcements and updates, people still don’t seem to care about AI Mode. According to Google Trends data (August 2024–August 2025), “AI Mode” has one blip on the radar during the week of June 29–July 5, likely tied to the announcement of deep search and model switching capabilities via their blog.
In contrast, ChatGPT continues to dominate attention, with Gemini in the distance and Perplexity crawling behind.
Sidenote: Has anyone seen Perplexity’s recent advertisements implying people can use it for sports betting analysis and recommendations? YIKES. Bros are going to lose a lot of money. Don’t be like these bros. (H/T to iPullRank’s Lead Relevance Engineer Patrick Schofield)
AI Overviews isn’t on anyone’s radar at this point, and maybe that’s a good thing? If our UX AI Mode study is any indication, people don’t even know that’s what it’s called.
For marketers, this suggests minimal awareness or demand among everyday users. The data that follows reflects a niche user base, not mainstream behavior.
New Perplexity ad on TickTok
What the Charts Show Us (and Why It Matters)
These charts zoom in on how AI Mode actually sends traffic. The feature is tucked away, so small numbers track. The patterns still teach us.
We would expect:
Low referral volume: answer-first design keeps users in the result.
Authoritative-site dominance: clicks skew to Reddit, YouTube, Wikipedia, Amazon, Facebook.
Quicker post-click sessions: if someone does go to the site, they’re either verifying a fact or making a quick purchase.
Volume and behavior volatility: layout tweaks and entry-point friction shake the numbers week to week. People’s familiarity with conversational search may impact behavior over time.
What actually shows up, and the surprises:
CTR stays tiny (<3%) and drifts sideways. This looks like a channel for coverage and credibility, not a traffic faucet.
Developer sites surprisingly punch above their weight. GitHub, Stack Overflow, and Stack Exchange appear often, hinting at early-adopter skew and task types like troubleshooting.
Concentration risk is real. A small set of domains captures a large share of the exits, suggesting that algorithmic biases may be embedded in the model citations.
Longer prompts for AI Mode. Query length rises, indicating a change in search behavior.
Everything here remains volatile: a limited window, a promotion-heavy period, and a user mix tilted toward testers. Treat all of these insights as early signals, not a settled trend.
Referral gap
Clicks out of AI Mode sit below ~3% and drift sideways. No increase over three months of data.
AI Mode delivers significantly fewer clicks to external sites than traditional search results pages.
Implications:
AI Mode is dramatically weaker than traditional search at driving referral traffic. Similarweb’s data backs up what most of us suspected: when Google pushes users into a generative answer box, fewer clicks leave the page. It’s a different mental model for search behavior. Traditional search was already trending toward zero-click behavior. AI Mode accelerates that shift.
People still click out for three main reasons:
- They want to go deeper than the synthesized answer, but that group is smaller than with classic search.
- They want to verify facts and sources, especially users familiar with AI’s tendency to hallucinate.
- They have intent tied to an actual purchase or direct interaction with a business.
Outside those situations, AI Mode captures the curiosity and the clicks.
Where the clicks land
AI Mode referrals show a predictable but revealing pattern. High authority, community driven, and niche expertise sites dominate the top 40% of referrals. This mix of platforms such as Reddit, YouTube, Wikipedia, Amazon, GitHub, Zillow, Healthline, and others reflects how people are using AI Mode both for mainstream topics and for specialized searches like health, real estate, and coding support. Compared to SEMrush and Ahrefs data, the AI Mode list contains fewer narrowly defined sites but more social and news outlets, pointing to shifts in both user intent and the kinds of publishers that win visibility.
Clicks cluster around large trusted ecosystems including social networks, knowledge hubs, and high authority brands, showing how AI Mode rewards broad platforms over niche ones.
Clicks underperform on specialized SEO tracked domains compared to SEMrush and Ahrefs data, signaling a potential gap in how AI Mode surfaces smaller or more targeted websites.
Post-click depth
Semrush
Ahrefs
Google Search consistently drives about six pageviews per referred visit. AI Mode averages closer to three to four pageviews, with much sharper swings. This reflects a thinner, more volatile post-click experience.
AI Mode referral sessions are about 2 pages shorter on average than Google Search sessions.
Pageview counts fluctuate widely, hinting at unstable or experimental user behavior.
Implications:
Fewer pages per visit signals “quick check” traffic or highly qualified traffic. Two opposite ends of the spectrum.
Users coming from AI Mode have already seen an AI summary and often click through to confirm or skim. Assume the content type is informational.
If it’s further down the funnel (including a homepage), these are the pages that are driving the higher rate of AI Search conversions.
Google Search referral sessions hold steady around 5 minutes. AI Mode sessions hover between 2.5 and 4 minutes with periodic spikes, but never settle into a stable pattern.
Shorter dwell times from AI Mode referrals: about half the duration of Google Search visits.
Higher volatility in session time signals mixed intent: some people test, some verify, and a few linger.
Implications:
Shorter sessions suggest issues with link quality (broken or off-target citations), anchor mismatch, or users conducting quick fact-checks before leaving. You may want to build your informational pages for fast trust by placing the cited passage above the fold, adding E-E-A-T signals early on, minimizing distractions, and offering a clear next step. This makes your site friendlier to AI Mode visitors and increases your odds of converting those brief interactions.
Prompt length
Here’s the most compelling example of change in search behavior over time (though not specifically referral-related).
AI Mode query length is growing. In June and July, the average query was around 8 words. By August it jumped to 10 or 11 words. That tells us people are starting to treat AI Mode more like a conversational interface, closer to how they use ChatGPT.
It’s nowhere near the 70-word prompts people use on ChatGPT, but the steady increase matters. Even without AI Mode as the default version of search, users are beginning to change their behavior. That shift forces a rethink of how we approach SEO. Targeting individual keywords based on search volume is losing relevance as queries get longer and more conversational.
Check out how dramatically different it feels when you only look at the AI Mode prompt length.
Engagement Frequency
Google Search stays steady at about 2.6 to 2.7 sessions per user through the period. AI Mode sits much lower at about 1.1 to 1.15 sessions per user but shows a clear rebound starting in July after an early summer dip.
AI Mode usage is still a fraction of Google Search in terms of sessions per user.
Engagement rose in late July and August which suggests new features or better visibility nudged people to come back more often.
Implications:
AI Mode has not yet built habit loops. The rise later in the period suggests that product updates or promotion can move the needle. Until access becomes default or near default, usage is likely to stay in bursts rather than a steady rhythm.
Looking at AI Mode alone, the trendline starts high in May, dips into mid June, then climbs through late July into August, reaching back toward early period highs.
AI Mode’s engagement curve reflects experimentation, friction, and partial recovery. The mid-year trough followed by a rise is consistent with a feature maturing and gaining more exposure, but it’s such a minuscule difference over such a short timeframe that the conclusion feels irresponsible.
I’ll watch how omnibox integration and global rollout influence this number. If AI Mode becomes default or quasi-default, expect a step change upward.
How are people actually using AI Mode?
Around this same period, we conducted the first-ever UX AI Mode Study with Farrah Bostic of the Difference Engine. She interviewed 23 people and had them use AI Mode to perform tasks ranging from finding a local health clinic to researching different credit card loyalty programs.
We hosted a webinar with her and shared our findings.
Since then, Francine Monahan has been diving deeper into each task to see how people use AI Mode across a variety of situations:
AI Mode Impacts on the Finance Industry: Researching Credit Card Loyalty Programs
AI Mode Impacts on Ecommerce: Researching Healthy Breakfast Cereals
AI Mode Impacts on Publishing: Researching Local News and Sports Headlines
For each task, the conclusion was the same: AI Mode does not offer an experience that’s compelling enough to change search behavior.
People were curious and found it to be a novel experience. But many tasks weren’t appropriate given the current state of the experience and what it generated.
What this means for brands right now
AI Mode adoption is minimal. Optimizing for it is even less valuable than ChatGPT. HOWEVER. Over the next few months, as Google tests ads, you can’t afford not to pay attention. When it comes to AI Search in general, you need a strategy because it’s not just SEO.
If you don’t know where to start, we have an entire AI Search Manual for you, along with a webinar walkthrough.
What can you do in the meantime as search behavior slowly evolves to match the conversational search platform?
Calibrate expectations. AI Mode visibility is about presence and validation. Direct traffic will be small for a while (if not permanent).
Earn citations where AI Mode already looks. Reddit threads with credible answers, YouTube explainers with chapters and transcripts, reference-grade pages that define terms and steps, developer hubs if you are B2B tech. Digital PR should continue to be a big part of your organic search strategy. Omnichannel and omnimedia.
Design “confirm-and-go” landings. Assume the visitor already knows the gist. Put the proof element above the fold, surface the exact passage that was cited, and offer one low-friction next step.
Measure the right things. Channel-level citation count (we’re big fans of Profound for LLM channel monitoring), citation prominence within the answer, share of mentions vs peers, and assisted conversions for the visits you do get.
Is AI Mode going to become the default?
AI Mode shouldn’t be ignored. However, it’s not yet impactful. Optimize for being cited, for being credible when clicked, and for harvesting assisted value. Measure input signals to earn the citation, channel visibility to prove presence, and the narrow performance that follows.
