G2 Says AI Agents Were 68% of Its Audience. Session Reports Can't See Them
G2's CMO says humans were 32% of its second-quarter audience. AI discovery drove more than 5% of its referral sessions, more than 20% of referral leads and more than 10% of referral revenue.
TL;DR
- G2's CMO says AI agents made up 68% of G2's Q2 2026 audience, with humans at 32%.
- Session reports don't record buyers' agents, so answer share, server and CDN logs and AI-referral leads show buyer influence better.
- At your G2 renewal, ask how it defines audience, whether human visits grew and whether intent signals count agent visits.
G2 CMO Alex London told Chief Marketer on November 3 that AI agents made up 68% of G2's audience in the second quarter of 2026, with humans at 32%. G2 says it used Profound Agent Analytics and Cloudflare data to separate user-triggered agent activity from crawlers. The report doesn't define whether "audience" means visits, requests or users.
G2's own research points the same way. Its "Evaluation Maze" report found that 82% of B2B buyers had sourced software recommendations from an AI chatbot in the prior two years, Chief Marketer reports. Those buyers can meet a vendor inside an AI answer before they ever visit a page, and a session report can't see that. So answer share, plus the leads and revenue that AI referrals produce, tells you more about buyer influence on a review site than sessions do.
G2's referral funnel shows AI discovery strong on leads and weaker on revenue
London gave a three-step funnel. AI discovery accounted for more than 5% of G2's referral sessions, more than 20% of its referral leads and more than 10% of its referral-attributed revenue. Those are floors, so the exact ratio can't be pinned down. By our math, if each figure sits near its floor, a session that starts in AI discovery yields about four times the leads of the average referral session. At those floors, AI discovery's revenue share is about half its lead share, so each AI-referred lead brings in about half the revenue of the average referral lead. That could mean smaller deals or deals that haven't closed yet, and G2 doesn't say which.
The lead gap has a plausible explanation, though G2 hasn't tested it. An agent pulls the reviews and compares the pricing pages, then narrows a shortlist. The human who clicks through afterward arrives later in the decision and readier to talk. A small slice of sessions can carry a large slice of leads when the research happened before the click. London's standard is "winning the answer," meaning a mention or a citation isn't enough unless the AI recommends you.
G2 has some data on how often it gets cited. A May 2026 post on its sales site analyzed 180 days of Profound data across more than 80,000 products. In aggregate, products drew 4.9 AI citations for every human pageview, and 80% of products were cited more often than they were viewed. "The influenced audience is larger than the visiting audience," the post says. Citations aren't visits, so the 4.9 ratio and the 68% figure measure different things and shouldn't be stacked. The post ran on G2's sales site, and its author, Kevin Indig, used to lead SEO at G2, so treat it as G2's own case.
Session reports were never built to see agent reads
Web-wide data gives a second view, on a different measure. Paid Hosting's compilation of Cloudflare Radar for October 18 to 24, 2026 puts humans at 33.8% of HTML page requests worldwide, with non-AI bots at 54.7% and AI bots at 7.3%. Only 5.2% of AI bot traffic declared itself as user action, against 42.4% declared as training (operators label their own bots, so the split is self-reported). If G2's 68% includes much crawler traffic, most of that "agent audience" would be bots collecting training data. Few would be agents researching for a buyer. G2 says it separated the two, but it hasn't published the split.
Paid Hosting adds, as its own analysis, that agent traffic "generates no ad impressions and no analytics sessions." Analytics tools also filter known bots. So the session report a vendor already reads is probably still mostly human, and bot traffic probably isn't polluting it much.
That doesn't rescue the session as a measure of interest. If agent reads leave no session behind, the report undercounts influence. A buyer's agent can read your G2 profile, your pricing page and three competitors' reviews, then hand the buyer a shortlist, and your session report records nothing. A flat session line can sit next to rising buyer research. The fix for an undercount is a metric that sees the research, which better filtering can't provide.
G2 sells the products its own numbers argue for
Every G2 figure here is company-stated and unaudited, and the method rests on two vendors' tools. G2 sells advertising and buyer-intent products to software vendors, which gives it a stake in a story where review sites matter more to AI answers (Chief Marketer notes this in the interview). The May post goes further and pitches paid profiles as earning more citations. G2 says paid AEO is under 1% of its own paid-search clicks, leads and revenue, even as London urges marketers to invest in it.
London said G2's total traffic is up 35.5% year over year. G2 hasn't said how much of that is agents, so it doesn't show whether human visits to G2 grew. G2 also gave the AI-referral revenue only as a share, with no dollar figure. The habit worth taking from G2 is that it tracks referral sources through to leads and revenue.
What this means for marketers renewing a G2 contract
Before you sign, ask G2 in writing how it defines "audience" in the 68% figure, whether human visits grew and whether its buyer-intent signals include agent visits or only human ones. No public source answers the last of these, and the answer changes what an intent alert is worth to your SDRs.
Add an AI-referral line to your pipeline report. Track sessions, leads and revenue from AI assistants as referrers separately from search and review sites, the way G2 does for itself. If that line converts the way G2's does, its lead share will run well ahead of its session share. That gap is the number to watch. Then check whether revenue follows.
Measure agent reads on your own site. Server logs, your CDN's bot reporting or a tool like the ones G2 used can show which agents read your pricing and comparison pages, even though your session report won't.
Start measuring answer share. Pick a fixed set of prompts your buyers would type for your category, run them each month across the main assistants, and record whether each assistant cites you and whether it recommends you. That gives you a baseline before anyone sells you a citation upgrade.
If your review-site sessions flatten while AI-referral leads rise, don't cut the review budget on the session line. Read the lead and revenue lines first. Before you pay for a premium profile on the citation pitch, ask G2 for your own product's citation counts from the Profound data it used.
Watch whether G2's next audience update gives a human-visit figure on its own. Until it does, the 68% can't tell you whether buyers are visiting more or less.
If agent reads leave no session behind, the report undercounts influence.
Before you renew with G2, ask in writing how it defines audience and whether its intent signals count agent visits, then add an AI-referral line to your pipeline report.
