AI Agents Are Fetching Your Links: How to Tell Bots from Buyers

In June 2026, Cloudflare's CEO announced a milestone most of the web expected years later: automated traffic had overtaken human traffic on the internet for the first time. Some of it is scrapers and security scanners. But the fastest-growing slice is agentic — AI browsers, research agents, and shopping assistants fetching, prefetching, and clicking links on behalf of real users. Your analytics counts much of it as visitors. Your conversion numbers do not. Here is how to tell agent traffic from buyers, why GA4 struggles to separate them, and a practical filtering workflow that does not require an enterprise contract.

The June 2026 Crossover

On June 3, 2026, Cloudflare CEO Matthew Prince announced that bot and agentic AI traffic had officially surpassed human-generated web traffic on Cloudflare's network — roughly 57% automated versus 43% human by Cloudflare Radar's measure of requests to page content, a crossover Prince had previously predicted wouldn't arrive until 2027. Cloudflare is one data source with one methodology (other measures, like Imperva's 2026 Bad Bot Report putting automated share above 53%, land in the same neighborhood), but the direction is not in dispute.

What changed is what the bots do. Human Security's April 2026 "State of Agentic Traffic" analysis found agentic browsers already account for roughly three-quarters of agentic traffic and are growing month over month. These are not scrapers harvesting content — they are agents acting for people: researching purchases, comparing options, summarizing pages, prefetching links a user might click. They follow redirects, load pages, and — critically for your reports — many of them execute JavaScript.

What Agentic Traffic Looks Like

• Agent browsing on a user's behalf. A research or shopping agent visits your page, reads it, maybe screenshots it, and reports back. The "user" never sees your site. From your analytics's point of view, it is a session with a landing page and an exit.
• Prefetching. Some AI surfaces and browsers fetch links in the background before a human commits to the click — link previews, hover-fetches, answer verification. These register as full page loads with no human behind them.
• Verification and monitoring agents. Tools that check whether links in answers, newsletters, or documents still resolve. They hit your redirect, follow it, and leave.
• Hybrid handoffs. The murkiest kind: an agent researches, the human takes over at the last step and buys. Some of your best customers now arrive pre-qualified by a bot that your reports counted as noise.

Why GA4 Can't Separate Them Cleanly

Classic bot filtering assumed bots don't execute JavaScript. Agentic browsers break that assumption: they run a real browser engine, fire your gtag, send plausible user agents (sometimes a real browser's), and generate pageview events indistinguishable from a quiet human reader. GA4 has no "agent" flag, no reliable way to mark a session as autonomous, and its bot filtering was designed for a pre-agent web. Enterprise tooling has started to respond — Adobe, for example, shipped dedicated agentic-traffic dashboards — but if you don't run an enterprise stack, you need cheaper signals. You have them.

Six Signals of Agent Traffic

• Device and OS clusters. A traffic spike dominated by a single browser build, one OS version, or one screen profile — real audiences are messy; agent fleets are uniform.
• Geo spikes in odd places. One link suddenly huge in a country you've never ranked in, while sibling links stay flat. Datacenter egress points leave geographic fingerprints.
• Zero-engagement sessions. One pageview, zero scroll depth, zero events, sub-second duration, 100% bounce — at volumes humans don't produce. A hint, not a verdict: some humans skim exactly like this.
• Odd referrers. AI domains you don't recognize, empty referrers on branded traffic, or aggregator URLs — the referral patterns of automated pipelines rather than people.
• Impossible cadence. Perfectly spaced clicks around the clock, off-hours volume with no weekend dip, or bursts that start and end on exact minute boundaries.
• The conversion disconnect. Clicks up 400%, signups and sales flat. The single most reliable signal in practice — agents fetch, humans convert.

No single signal proves a bot. Two or three together, concentrated on one link or one landing page, is actionable evidence.

The Redirect-Level Checkpoint

Site analytics aggregates agents and humans into the same reports. A tracked redirect gives you a checkpoint before they merge — and, usefully, one link at a time. Every Sinkto.Link short link records each click with country and city, device, browser, referrer, and a time-series, filterable per link. That granularity turns vague suspicion into a specific diagnosis:

• One link spiking in one country or one device family = noise. When a single alias suddenly triples its click rate and the new clicks are 90% one browser build from one region, you are almost certainly looking at an agent — and you know exactly which asset attracted it.
• Deactivate the link being hammered. If an agent has latched onto an old campaign alias, deactivate it (or set an expiration date when you create campaign links) rather than letting it inflate your numbers indefinitely. Reactivate later if the link deserves another run.
• Password-protect links as an agent filter. For links meant only for humans — a partner offer, an internal asset — protect them with a password. Sinkto.Link hashes passwords and rate-limits the protected page, so casual crawlers and prefetchers that can't complete a form stop at the door while people pass through. Be clear-eyed: a determined agent can fill in a password field; this filters the casual majority, not a sophisticated few.
• Compare click-side with site-side. Redirect clicks that never appear as GA4 sessions, or GA4 sessions that never engage — the mismatch itself is the agent signature. The GA4 side of this comparison is covered in our GA4 AI tracking guide.

A Human-First Reporting Workflow

• Report humans to stakeholders. Strip zero-engagement sessions and identified agent clusters from the numbers you present. Your CMO does not care that agents "visited" 400 times.
• Keep a separate agent view. Don't delete the bot data — agents are a leading indicator of AI-surface visibility. A link being fetched by shopping agents often precedes humans buying through them.
• Decide on conversions, not clicks. Judge every channel — AI included — on what humans did after arriving. Pair your redirect analytics with a conversion event on the destination, as covered in how to track link clicks.
• Review anomalies weekly. Ten minutes with your per-link referrer and country breakdowns, looking for the six signals, keeps agent noise from quietly compounding in your reporting.

If AI surfaces are becoming a real channel for you, the same redirect data doubles as attribution — see what changed with ChatGPT's UTM tagging and inline brand links, and the full playbook in our AI link management guide.

Frequently Asked Questions

What percentage of web traffic is bots in 2026?

By Cloudflare Radar's measure, automated traffic passed human traffic in June 2026 for the first time — roughly 57% of requests to page content, announced by CEO Matthew Prince on June 3. Imperva's 2026 Bad Bot Report independently put automated share above 53%. Measures differ by methodology, but on every major count, bots are now half or more of web traffic.

Can Google Analytics tell bots from humans?

Not reliably for agentic traffic. GA4's bot filtering assumes bots don't execute JavaScript, but agentic browsers run real engines, fire analytics tags, and send plausible user agents — so they file as ordinary visitors. There is no agent flag in GA4. You need behavioral signals (zero-engagement sessions, device clusters, cadence) and ideally a redirect-level checkpoint that isolates traffic link by link.

How do I tell if a traffic spike is AI agents?

Look for the cluster signature: one link spiking while siblings stay flat, clicks dominated by a single browser build or OS, a sudden surge from one unfamiliar country, perfectly spaced click cadence, and — most telling — clicks way up while conversions stay flat. Any one signal is weak; two or three together concentrated on one link is strong evidence of agents, not buyers.

Can password-protected links stop AI agents?

Mostly, for casual agents. Sinkto.Link password-protected links use hashed passwords with a rate-limited protected page, so prefetchers and simple crawlers that can't complete a form are filtered out while humans pass through. A determined agent built to fill forms can still get through — treat it as a filter for the casual majority, not a security boundary.

What is agentic traffic?

Web requests made by AI systems acting on behalf of a person or task — research agents reading your pages, shopping agents comparing options, browsers prefetching links, verification agents checking that URLs resolve. Human Security's April 2026 analysis found agentic browsers alone make up roughly three-quarters of agentic traffic and continue growing. It is distinct from classic scraping: agents often execute JavaScript and can look exactly like human visitors.

Next Steps

You cannot stop agent traffic, and you shouldn't want to — agents researching purchases is a preview of how your buyers will shop. What you can do is measure it separately so it stops polluting decisions: one tracked redirect per important link, a weekly anomaly scan, and reporting built on humans.

Create your free Sinkto.Link account →

Related reading: How to track ChatGPT traffic in GA4 · AI link management: the complete guide · Bitly vs Sinkto.Link

Updated: August 2026. Traffic statistics (Cloudflare Radar, Imperva, Human Security) and GA4 behavior verified August 2026 against the named sources; figures shift monthly. Sinkto.Link capabilities reflect the live product. Have feedback? Contact us.