Automated requests passed human ones this summer. The ad industry published its first written response in July. Cloudflare starts blocking agents by default in September. The measurement stack underneath all of it has not been told any of this.
A session arrives on your pricing page. It reads the comparison table, opens two product pages, spends ninety seconds, does not convert, and leaves. Your analytics records an engaged visit from a high-intent account. Your attribution model gives the referring channel credit. Your ABM platform raises the account's score.
Nothing in that chain is capable of telling you whether a person was involved.
What actually happened this summer
On 3 June 2026, Cloudflare's Matthew Prince said automated requests had reached 57.5 per cent of web-page traffic against 42.5 per cent human — machines in the majority for the first time. It is the most-quoted figure in marketing this year and it deserves a caveat almost nobody prints.
Read the number properlyThe 57.5 per cent figure has no dated primary report behind it. It comes from Cloudflare Radar — a live dashboard — surfaced through a chief executive's post. Independent re-derivations of the same underlying data land at 57.4 per cent. More importantly, it measures HTTP requests to web-page content, not sessions, users or impressions, and almost all trade coverage elides that. Prince's own March 2026 framing was a prediction that bots would exceed humans by 2027. The milestone arrived early, which is itself the story — but cite it as a dashboard reading with a date, not as a study.
The better-sourced material sits underneath it. Thales and Imperva's 2026 Bad Bot Report, published around 30 April 2026, put automated traffic at 53 per cent of the internet in 2025 — 40 per cent bad bots, 13 per cent benign — with AI-driven bot activity up 12.5-fold and daily blocked requests rising from two million to twenty-five million.
DataDome's Q2 2026 report, published 16 July 2026 from a base of five trillion signals a day across 400-plus enterprises, counted 17.7 billion AI-agent requests in the quarter, up 45 per cent on Q1's 12.2 billion. One detail in it is worth more than the headline: ChatGPT-User crawls fell 6 per cent quarter on quarter while ChatGPT referrals rose 17 per cent. Crawl volume and referral volume have decoupled. Anyone using one as a proxy for the other is now measuring the wrong thing.
HUMAN Security's monthly agentic-traffic series is the most useful recurring dataset available. Its June 2026 edition put Perplexity's Comet at 47.6 per cent of agent activity, with Claude overtaking Atlas (20.8 versus 16.5 per cent), and — the number that should reframe the whole conversation — 79 per cent of agent activity on product and search routes against 2.3 per cent on checkout.
Agents are researching, not buying
That distribution is the most under-used finding of 2026. The prevailing narrative says autonomous agents are about to transact on buyers' behalf. The telemetry says they are overwhelmingly doing discovery and almost never completing.
For B2B specifically, Octane11's analysis of 400 million B2B website sessions between September 2025 and June 2026, updated 29 July 2026, points the same way with a sharper edge. AI search accounts for under 2 per cent of total B2B search referrals — traditional organic remains above 96 per cent — but that share has grown 35-fold since March 2025. And roughly one in eleven closed-won deals now carries an AI-search touch, peaking near 15 per cent in February 2026, with about one in six of those account touches being exclusive: accounts nothing else in the mix reached.
Two per cent of referrals, nine per cent of closed-won. Whatever agent traffic is, it is not proportional to its volume — which means volume-based filtering will throw away the wrong thing.
One correction while we are here. Adobe's widely-quoted finding that AI-sourced traffic converts 42 per cent better is based on over a trillion visits to US retail sites. It is a retail number. It has been applied to B2B repeatedly this year — including, notably, inside industry guidance that should know better. Do not carry it across.
Why your stack cannot see this
The instinct is to reach for bot filtering. For this class of traffic, bot filtering is structurally obsolete, and it is worth understanding why rather than buying around it.
In-browser agents — Comet, Atlas, the Claude browser extension — run on the user's own machine and inherit a legitimate human session. There is no server to identify, no data-centre IP range, no distinctive user agent to block. They are indistinguishable by every signal conventional filtering uses, by design rather than by evasion.
It gets worse for attribution specifically. Perplexity's Comet arrives in GA4 as a referral from perplexity.ai, which at least leaves a trace. Atlas strips or blocks referrer headers, so those sessions land as Direct / (not set) — the bucket that already absorbs everything unattributable. Discovery that happened inside an AI interface is recorded as if the buyer typed your URL from memory. Behavioural tells do exist — linear cursor paths, uniform timing, movement in fixed sub-pixel increments — but no mainstream analytics stack looks for them.
Visual 1 — Where agent traffic lands, and what it corrupts
System | How agent traffic appears | What breaks |
|---|---|---|
Web analytics (GA4 and peers) | An ordinary session; Atlas as Direct / (not set) | Session counts, engagement rate, channel mix, and the Direct bucket as a diagnostic |
Attribution model | A touchpoint on a real account | Credit assigned to whichever channel the agent happened to arrive from |
ABM / intent scoring | Account engagement, often repeated | Scores inflate on accounts where one person delegated research once |
Paid media | Clicks, occasionally conversions | Spend against non-human clicks; Lunio's June 2026 study put LinkedIn's invalid-traffic rate highest of any major platform |
Content performance | Pageviews with long dwell | Editorial decisions optimised for machine readers |
Conversion rate | Denominator rises, numerator does not | Apparent decline that is arithmetic, not performance — the 79%/2.3% split in one metric |
How to read it: Note the last row. Conversion rate falls automatically as agent traffic grows, with no change in how well the site converts humans. Teams are currently explaining that decline as a creative or offer problem, and reworking pages accordingly.
The industry's first written answer, and its expiry date
On 22 July 2026, dated and versioned, IAB Australia published a bots and crawler decision matrix, launched publicly on 28 July. It sorts every non-human visitor into five classes and four verdicts: discovery and search crawlers allowed; AI training crawlers blocked or licensed; live AI agents allowed with conditions and restricted on ad-funded pages; operational bots allowed if verified; unknown or spoofed challenged or blocked.
That third verdict is the significant one. It is the first time an ad-industry body has said in writing that live agent traffic should not be served on pages funded by advertising — which is, in effect, the industry conceding that agents should not count as impressions. As of writing, there is no equivalent document from IAB US or IAB Europe. IAB Tech Lab's 2026 work has gone into agentic advertising — AAMP 2.3 shipped 30 July — not agent traffic measurement.
Cloudflare moved from guidance to enforcement. Its 1 July 2026 announcement splits bots into search, agent and training classes, and from 15 September 2026, on ad-monetised pages for new domains and free-tier customers, training and agent traffic are blocked by default while search is allowed. Crawlers that cannot separate their functions are blocked outright. A supporting statistic from the same announcement: more than half of AI crawler traffic is re-fetching pages that have not changed.
And here is the part that should give everyone pause. The mechanism the entire agentic-web trust layer runs on — Web Bot Auth, using HTTP message signatures to let a well-behaved agent identify itself — is not an IETF standard. The current architecture draft is an individual submission, dated 2 March 2026, and the datatracker text states plainly that it is not endorsed by the IETF and has no formal standing in the standards process. It expires on 3 September 2026, twelve days before Cloudflare's default-block date.
What to do before September
Decide, per metric, whether a machine visit counts — and write it down. This is a policy question, not a technical one, and it has to be answered consistently across analytics, verification and attribution or the numbers will not reconcile. Sessions probably should include agents; conversion-rate denominators probably should not.
Re-baseline before you accept any 2026 trend. If agent share of your traffic grew across the period, every rate metric with traffic in the denominator has drifted for reasons unrelated to marketing performance. Establish the baseline now; the comparison gets harder every quarter.
Segment the Direct / (not set) bucket and watch it. Growth there without a corresponding campaign explanation is currently the cheapest available proxy for agent-mediated discovery.
Check what your CDN will do on 15 September. If you are on Cloudflare and your pages carry advertising, defaults may change under you. Whether you want agents blocked is a strategic decision — being cited in AI answers requires being readable — and it should not be made by inheritance.
Move the visibility metric from sessions to share of answer. Publishers made this shift this year out of necessity. Given that B2B research is happening inside interfaces that send back one visit for a great deal of reading, the same logic applies — with the Octane11 caveat that the traffic remains small even where the influence is not.
None of this is a bot problem in the traditional sense. Nobody is attacking you. A meaningful share of your audience has simply started delegating the reading, and every measurement system you own was specified on the assumption that reading and visiting are the same act. They are not, and the gap between them is now the majority of the traffic.
Sources and method. A MarketingHubMedia original. The 57.5 per cent figure is a Cloudflare Radar dashboard reading surfaced by Matthew Prince on 3 June 2026, reported via Tech Times, with a competing 57.4 per cent re-derivation noted in Digiday, 21 July 2026; it measures HTTP requests, not sessions. Thales/Imperva 2026 Bad Bot Report via Help Net Security, ~30 April 2026. DataDome AI Traffic Report Q2 2026, 16 July 2026. HUMAN Security, State of Agentic Traffic — June 2026. Octane11, State of B2B AI Search Vol. 2, updated 29 July 2026 (vendor-aggregated B2B panel; methodology summarised, not fully published). GA4 referrer behaviour per MarTech. IAB Australia decision matrix v1.0, 22 July 2026, launched 28 July. Cloudflare, 1 July 2026 (15 September default-block date). Web Bot Auth architecture draft-05, 2 March 2026 — an individual submission with no formal IETF standing, expiring 3 September 2026. Invalid-traffic platform ranking from Lunio's Invalid Traffic Impact Report, June 2026, via PPC Land; Lunio sells invalid-traffic filtering, so the relative ranking is cited rather than the absolute rate. Adobe's 42 per cent conversion figure covers US retail sites only and is cited here solely to warn against its application to B2B. Journalism, not procurement advice. Corrections will be made openly on this article.
