Intelligence report · 2026 edition

The Agentic Traffic Report 2026: AI Agents on Your Landing Pages

Automated traffic now generates more than half of the requests on the networks that measure it, AI agents are its fastest-growing slice, and most paid-media measurement still counts them as people.

ClickTrust ResearchNovember 18, 202614 min read30 sources
1,700%
growth in daily requests from AI agents on Cloudflare's network in the year to September 2026[1]
53%
of all web traffic was bots in 2025, with humans down to 47%[2]
52%
of crawler requests were for AI model training by June 2026, up from 22% a year earlier[3]
1%
of users click through to a source website once an AI summary has answered them[4]

Executive summary

  1. 1Agentic traffic is compounding faster than anything else on the web: daily requests from AI agents on Cloudflare's network grew more than 1,700% in the year to September 2026, according to Cloudflare's report The Internet has a second audience [1].
  2. 2Bots passed 53% of all web traffic in 2025 and human activity fell to 47%, with 40% of bot traffic classified as malicious, according to Imperva's Bad Bot Report 2026 [2].
  3. 3Crawling has shifted toward training: Cloudflare's Content Independence Day report records the share of crawler requests used for AI training rising from 22% in spring 2025 to 52% by June 2026 [3].
  4. 4Agents read far more than they refer. Cloudflare's 2025 Radar Year in Review records Anthropic's crawl-to-refer ratio reaching as much as 500,000:1 [5], and Fastly's Q3 2025 Threat Insights Report finds only 1% of users click through to a source site once an AI summary has answered them [4].
  5. 5Commerce is the destination: Akamai found commerce absorbed 47.9% of all AI bot traffic on its network in the second half of 2025, and commerce organizations left more than 90% of it in monitor mode [6].
  6. 6None of this appears as a separate line in ad-platform reporting. Google says it cut advertiser exposure to invalid leads by an average of 30% so far in 2026 [7], and ClickTrust audits find 14 to 42 percent of paid traffic is invalid depending on channel.
01

What is agentic traffic, and where does it sit in MRC's GIVT and SIVT framing?

Agentic traffic is web traffic generated by software acting on a person's behalf or on its own: AI crawlers collecting pages for model training, fetchers pulling a page to answer a prompt, AI agents completing tasks, and agentic browsers driving a real browser the way a user would. None of these visitors is a customer. All of them land on advertisers' pages, and most look enough like people to be counted as people.

The four differ in intent and in how hard they are to spot. A crawler fetches in bulk on a schedule and usually declares itself. A fetcher pulls one page on demand because a person asked a question, so it arrives when a buyer would. An AI agent completes a task, such as a comparison or a form fill, through an API or a headless browser. An agentic browser is a consumer browser with an AI driver, so it carries real cookies and a real human behind it.

CategoryWhat it doesDeclares itself?Typical MRC bucket
AI crawler (GPTBot, ClaudeBot, Bytespider)Bulk-fetches pages for training or search indexesUsually, via user agent and published IP rangesGIVT: list-based filtration [8]
AI fetcher (ChatGPT-User and similar)Pulls a page on demand to answer a promptSometimes; a growing share reaches pages robots.txt disallowsGIVT when declared, SIVT when disguised
AI agentCompletes a multi-step task: compares, books, submits formsRarely; runs headless or through APIsSIVT: advanced analytics and multipoint corroboration [8]
Agentic browserA person's own browser with an AI driverNo; carries real cookies and device signalsUnresolved: human-adjacent, but not a human reading the ad
Four kinds of agentic traffic and where each sits in the MRC framework

The Media Rating Council's framework predates all of this, but it still fits. MRC's 2015 Invalid Traffic Detection and Filtration Guidelines Addendum defines General Invalid Traffic as traffic identified "through routine and list-based means of filtration", with bots, spiders, other crawlers and non-browser user agent headers as examples, and Sophisticated Invalid Traffic as traffic identified "through advanced analytics, multipoint corroboration, human intervention" [8]. A declared crawler is GIVT. An undeclared agent that passes browser checks is SIVT.

Google's definitions point the same way. Google Ads Help states that "invalid clicks are clicks on your ads that don't represent a genuine interest in your business" and lists automated traffic from "bots or deceptive software" as one of three types alongside accidental and manual clicks [9]. Imperva's Bad Bot Report 2026 describes AI agents as a new category of internet participant alongside good and bad bots [2]. By Google's definition, an agent reading your landing page to answer someone's question is invalid traffic, whatever its intentions.

02

How big is agentic traffic in 2026?

Bots now generate more than half of all web traffic, and the agent-driven share is the fastest-growing part. Imperva's Bad Bot Report 2026 puts bots at more than 53% of all web traffic in 2025, up from 51% the year before, with human activity down to 47% and 40% of bot traffic classified as malicious [2]. Imperva's 2025 report had recorded 2024 as the first year in a decade that automated traffic exceeded human traffic, with bad bots at 37% [10].

53%
of all web traffic was bots in 2025, up from 51% in 2024[2]
47%
of HTML requests on Cloudflare-protected sites came from humans by December 2, 2025[5]
1,700%
growth in daily requests from AI agents on Cloudflare's network in a year[1]

Cloudflare measures a different population and lands in a similar place. Its July 2025 crawler analysis put bots at around 30% of global web traffic [11]. By the 2025 Radar Year in Review, humans generated only 47% of HTML requests on Cloudflare-protected sites as of December 2, non-AI bots 44%, and AI bots averaged 4.2%, slightly less than Googlebot alone at 4.5% [5]. Cloudflare's July 2026 report states that more than 50% of Internet traffic is now non-human, and that in heavily crawled categories human traffic fell as much as 40% in under a year [3].

Bot share of web traffic by measuring network, 2025
Imperva, full year 202553%Fastly, Apr to Jul 202537%Cloudflare, mid-202530%Fastly, Q3 202529%Vercel, one week Aug 202520%F5 Labs, Nov 2023 to Sep 202410.2%

Source: Imperva Bad Bot Report 2026; Fastly Threat Research (August and December 2025); Cloudflare, From Googlebot to GPTBot (July 2025); Vercel (August 2025); F5 Labs 2025 Advanced Persistent Bots Report [2, 12, 11, 4, 13, 14]

The spread in the chart is a measurement story, not a disagreement. Fastly recorded bots at 37% of observed activity in mid-2025 [12] and 29% in Q3 2025, with about 25% of all web traffic classified as unwanted [4]. Vercel saw bots above 20% of traffic in one August 2025 week, about a quarter of it AI crawlers [13]. F5 Labs found total automation at 10.2% of more than 200 billion transactions, yet more than half of all web content page requests came from scrapers [14]. Each network sees a different mix of sites, and none measures paid clicks.

Averages also hide where the traffic lands. Kinsta's study of more than 5,000 WordPress sites found AI bots at 6.13% of bandwidth on average, but the median site saw only 1.57%, the 99th percentile saw 90.3%, and 20 to 27% of sites received no AI-bot requests at all [15]. A given landing page is either barely touched or dominated, and the only way to know which is to measure it.

The trajectory is steeper than the share. Cloudflare's 2026 Founders' Letter says its forecast for automated traffic passing human traffic moved from the second half of 2027 to May 2026, and that if trends continue automated traffic will be 1,000 times human traffic in five years [16]. Its September 2026 report records daily requests from AI agents up more than 1,700% in a year [1], and the Radar Year in Review found "user action" AI crawling, the kind a person's prompt triggers, grew more than 21x from January through early December 2025 [5].

03

Who is sending it: Meta, Google, OpenAI and the crawl-to-refer gap

A handful of companies generate most AI-bot traffic, and the mix depends on whether you count crawlers or fetchers. Fastly's August 2025 threat research found AI crawlers made up almost 80% of AI bot traffic, with Meta at 52% of AI crawler traffic, Google 23% and OpenAI 20%, while OpenAI's ChatGPT bots generated 98% of fetcher requests [12]. Cloudflare's July 2025 analysis shows how fast the mix changes: GPTBot's share of AI crawling jumped from 5% to 30% in a year [11]. In commerce, Akamai names OpenAI, ByteDance and Anthropic as the top three AI bots [6].

Crawling has shifted toward training, which never sends a visitor back. Cloudflare's August 2025 crawl-to-click analysis found 80% of AI crawling over the previous 12 months was for model training, 18% for search and 2% for user actions [17]. Its July 2026 report records the training share of crawler requests rising from 22% in spring 2025 to 52% by June 2026, with mixed-use crawlers at more than 36% of crawler activity [3].

Share of crawler requests used for AI training
22%Spring 202552%June 2026

Source: Cloudflare, Content Independence Day, one year on (July 2026) [3]

The most useful number for an advertiser is the crawl-to-refer ratio: how many pages a company's bots fetch for every visit they send back. For July 2025 Cloudflare recorded Anthropic at about 38,000 pages crawled per referred visit, down 87% from 286,930 earlier in the year, OpenAI at 1,091, Perplexity at 194.8 and Google at 5.4 [17]. By December the gap at the top had widened again: the Radar Year in Review reports Anthropic's ratio reached as much as 500,000:1 during 2025 and OpenAI's up to 3,700:1 [5].

Pages crawled per referred visit, July 2025
Anthropic38000OpenAI1091Perplexity194.8Google5.4

Source: Cloudflare, The crawl-to-click gap (August 2025) [17]

Publishers see the same lopsided exchange, and a landing page is a publisher page as far as a bot is concerned. TollBit's Q1 2025 State of the Bots, presented with FIPP, cited one publisher receiving 200 users for every 200,000 requests from an OpenAI bot [18]. Akamai's April 2026 research found AI chatbots drove about 96% less referral traffic than Google search in Q4 2024 [19]. Fastly found only 1% of users click through to source websites that have an AI summary [4], and Cloudflare still puts Google at roughly 88% of referral traffic [3].

Fetchers are the share to watch in 2026. Cloudflare's Year in Review found retail and computer software together drawing just over 40% of AI crawler activity [5], and Fastly measured fetcher bursts exceeding 39,000 requests per minute [12]. A fetcher lands on the page the buyer would have landed on, at the moment the buyer is researching, and leaves no buyer behind.

04

Why agentic traffic reaches paid campaigns

Agentic traffic reaches paid campaigns because the page an agent reads is the same page a paid click lands on, and every tag on that page fires regardless of who loaded it. The Google Ads conversion tag, the Meta Pixel, the analytics tag and the retargeting tag all execute for an agent answering a user's question. Four leaks follow.

  • Landing pages and analytics. An agent visit inflates sessions, deflates conversion rate and shortens time on page, which pushes bidding and creative decisions the wrong way. Fetchers that load a URL straight from a prompt carry no referrer and land in direct traffic.
  • Lead forms. Agents that complete tasks fill forms. Google's November 2026 announcement said its systems had cut advertiser exposure to invalid leads by an average of 30% so far in 2026 [7], which is also a statement about how much there was to cut.
  • Retargeting pools. Every agent that fires the pixel joins an audience that Meta, Google Ads and TikTok then use as a lookalike seed, so one wave of agentic visits becomes next month's targeting error.
  • Attribution. Agentic browsers carry real cookies, so a later purchase by the person behind them is attributed to a session the person never saw. Agents acting through APIs leave no click at all.

Agents are also becoming buyers. Salesforce's holiday data found AI and agents drove 20% of retail sales and $262 billion of revenue between November 1 and December 31, 2025, and retailers running their own shopper agents grew about 59% faster, at 6.2% against 3.9% year over year [20]. Salesforce measures influence rather than agent-completed checkout, but the direction is clear: an agent researching for a real buyer visits your paid landing page without the buyer. Much of that research is repetitive; Cloudflare's Founders' Letter notes that more than half of what good bots fetch has not changed since their last visit [16]. A suspiciously low CAC is often the first symptom, a point Anvesha Poswalia makes on The Signal.

05

Where it hurts most: commerce, lead gen, affiliate, app installs and travel

Agentic traffic hurts most where a page load, a search or a form fill is the paid event. The evidence is strongest for commerce, and it points at APAC.

Commerce and retail

Akamai found commerce absorbed 47.9% of all AI bot traffic on its network between July and December 2025, with training crawlers at more than 70% of AI bot triggers in commerce; commerce organizations put more than 90% of AI bot activity in monitor mode [6]. Bot activity grew 63% in APAC over the year, against 48% in LATAM, 16% in EMEA and 7% in North America [6]. F5 Labs found more than one in five add-to-cart transactions and almost a quarter of web searches were automated [14]. Every one of those searches on a paid landing page is a session the platform counts.

Growth in bot activity by region, 2025
APAC63%LATAM48%EMEA16%North America7%

Source: Akamai, Commerce Becomes the Epicenter for AI Bot Attacks and Agentic Fraud in 2026 (July 2026) [6]

Lead generation

Lead-gen is paid on the form, and forms are what agents are built to fill. No independent public source publishes a fake-lead rate for paid campaigns, so this report does not quote one. Google's own wording sizes the exposure: a 30% average cut in advertiser exposure to invalid leads so far in 2026 [7], and a Click Quality Form that now covers "invalid clicks, poor-quality form submissions, clicks outside location targeting, or sudden short-term cost increases" [21]. F5 Labs found malicious login attempts averaged 10.6% of web traffic [14]; the automation that attacks a login fills a lead form.

Affiliate and app installs

Affiliate pays on tracked events that automation can produce at will. AppsFlyer's State of Fraud for Marketers 2026 found affiliates account for about 40% of fraudulent installs with a fraud rate near 40% every quarter, against about 1% for self-reporting networks, a 36x gap in Q1 2026 [22]. Organic is now 52% of all fraudulent installs, and physical device farms grew 72% year over year while emulators fell 32% [22]. India accounts for nearly 28% of global Android install fraud, and the affiliate share of India's fraud grew from 57% to 65% [22]. Agents give a dishonest affiliate real-browser sessions at almost no cost, which is why per-affiliate scoring and SDK-level In-App Protection follow the money, not the user agent.

Travel and financial services

Travel is the most scraped vertical because prices change by the minute. Imperva's 2025 report found 48% of all web traffic to travel sites in 2024 was bad bots [10], and F5 Labs put hospitality web traffic at about 44 to 45% automated [14]. Financial services absorbed 24% of all bot attacks and 46% of account-takeover incidents in 2025, according to Imperva [2], and Fastly found 89% of headless-bot traffic targeted financial services and commerce [4]. The agent that fills a sign-up form and the bot that takes over the account are the same technology with a different operator.

47.9%
of all AI bot traffic on Akamai's network went to commerce in H2 2025[6]
36x
gap between affiliate and self-reporting-network install fraud rates in Q1 2026[22]
89%
of headless-bot traffic targeted financial services and commerce in Q3 2025[4]
06

The benign-versus-malicious problem

The hardest part of agentic traffic is that most of it is not fraud. Imperva classifies 40% of bot traffic as malicious [2], so the other 60% of a majority-bot internet is crawlers, fetchers and agents doing something someone asked for. A price-comparison agent working for a real shopper and a scraper harvesting your catalog generate nearly identical sessions. The same bot names appear on both sides: Imperva's 2025 report found ByteSpider alone responsible for 54% of AI-enabled attacks, with AppleBot at 26%, ClaudeBot at 13% and the ChatGPT User bot at 6% [10], and its 2026 report records AI-driven bot attacks up 12.5x year over year [2].

Consent is eroding. TollBit's Q1 2025 data, presented with FIPP, shows the share of AI-bot scrapes bypassing robots.txt rose from 3.3% in Q4 2024 to about 13% in Q1 2025, with more than 26 million disallowed scrapes in March 2025 alone [18]. TollBit's H1 2026 report, as covered by PPC Land, found about 15% of identified AI page-fetchers in Europe reached URLs that sites had disallowed [23].

Most sites do not even ask. Cloudflare found only about 14% of the top 10,000 domains with a robots.txt file addressed AI bots in mid-2025 [11]; by September 2026, 17% of sites blocked training crawlers and fewer than 1% blocked search crawlers [1]. Fastly found only 4% of "wanted" bot requests, AI crawlers and fetchers included, were blocked in Q3 2025 [4]. DataDome's 2025 Global Bot Security Report tested about 17,000 popular domains and found 2.8% fully protected and 61.2% letting every test bot through [24].

APIs are where the line blurs fastest. Imperva found 44% of advanced bot traffic targeted APIs in 2024 [10] and 27% of bot attacks hit API endpoints in 2025 [2]. APIs are rarely instrumented by marketing tags, so an agent that quotes or signs up through one leaves the paid funnel with no record at all. Blocking is not free either: PPC Land's coverage of TollBit's H1 2026 report cites a Rutgers and Wharton study in which publishers that blocked AI crawlers lost about 7% of weekly traffic within six weeks [23]. The answer for an advertiser is not to block agents but to stop paying for them and stop learning from them.

07

What Google, Meta, LinkedIn and Microsoft are doing about it

Platforms filter some agentic traffic before billing, credit some after, and leave the rest to the advertiser. Google Ads Help states that advertisers "aren't charged for invalid traffic detected before billing", that invalid activity found later is credited back, and that investigation requests cover the last 60 days through the Click Quality Form [25]. What Google detects, Google credits. What it does not detect is yours.

Transparency improved in 2026. On November 12 Google announced an Invalid Traffic Overview Card showing "Automated protection", the amount saved, and "Past invalid activity", the credits applied [7]. The Invalid Activity Credit Report breaks those credits down by campaign and network [21]. Google's 2025 Ads Safety Report says it blocked or removed more than 8.3 billion ads and suspended 24.9 million advertiser accounts, with "Abusing the Ad Network" the largest enforcement category at more than 1.29 billion ads [26]. In India, Google blocked or removed 483.7 million ads and suspended 1.7 million advertiser accounts [27].

PlatformWhat it filters or reportsWhat it leaves to you
Google AdsPre-billing filtering; post-billing credits; Invalid Traffic Overview Card; Invalid Activity Credit Report; 60-day Click Quality Form window [25, 21]Undetected agent sessions, conversions and audiences; anything older than 60 days
MetaActioned 692 million fake accounts on Facebook in Q3 2025 and 1.1 billion in Q4 [28]No invalid-click reporting; no agent flag on pixel events or lookalike seeds
LinkedInStopped 88.9 million fake accounts at registration in H2 2025, 99.7% proactively [29]No invalid-click reporting; no agent flag on Insight Tag events
Microsoft Invest (programmatic)Does not bid on user agents in the IAB bots and spiders list; de-duplicates clicks; blocks some cloud providers but not all [30]Undeclared agents and residential-proxy traffic
What each platform filters, reports and leaves to the advertiser
8.3B
ads blocked or removed by Google in 2025, with 24.9 million accounts suspended[26]
1.1B
fake accounts actioned by Meta on Facebook in Q4 2025 alone[28]
88.9M
fake accounts LinkedIn stopped at registration in the second half of 2025[29]

The social platforms report from the account side. In its March 2026 filing under the EU Code of Practice on Disinformation, Meta estimated fake accounts at about 4% of worldwide daily active people on Facebook in Q3 2025 and 5% in Q4 [28]. LinkedIn's Community Report for July to December 2025 records 88.9 million fake accounts stopped at registration and 24.4 million more restricted before any member report [29]. A 5% fake share of the audience your lookalikes are modeled on is not a rounding error.

Microsoft's traffic-quality policy for Microsoft Invest is unusually candid: it subscribes to the IAB bots and spiders list so declared bots are not bid on, blocks "some cloud providers but not all", and cannot guarantee it blocks all unsuitable agents [30]. That is list-based filtration, which is to say GIVT. Nothing in any platform's tooling says whether a conversion came from an agentic browser, and although Cloudflare reports that OpenAI, Google and AWS cryptographically sign their agents' requests [1], no ad platform surfaces that signature next to a click.

08

How to tell an AI agent from a customer

You can tell an agent from a customer by how the session behaves, not by what it says it is. No single signal is conclusive, and because benign and malicious automation share the same bot names, scoring beats blocking. The table lists the signals that separate agentic traffic from human traffic; none carries a measured hit rate.

SignalWhat it indicatesConfidence on its own
Declared AI user agent (GPTBot, ClaudeBot, ChatGPT-User, Bytespider) or a signed agent requestA bot telling you what it is. OpenAI, Google and AWS sign their agents' requests [1]. Filter as GIVT; never bill, convert or retarget it.High
Request from an AI provider or cloud IP range, or a data center ASNAutomation. AWS and Google Cloud alone originated about a quarter of global bot traffic in 2025 [5]. Agents behind residential proxies pass this check.High for crawlers, medium for agents
Headless or automation markers (navigator.webdriver set, missing plugins, viewport mismatch)A scripted browser. Headless agents usually fail; agentic browsers on a real browser may pass.Medium to high
No referrer on a landing page only reachable from an adA fetcher or agent that loaded the URL straight from a prompt.Medium
Page loaded, zero scroll, zero pointer movement, exit in under two secondsContent extraction rather than reading. Fetchers pull the DOM and leave.Medium to high
Form submitted faster than a person could type it, every field valid first timeAutomated completion. People hesitate and make typos; agents do not.High
Many pages fetched under one identity in a tight burst, then silenceCrawling. Cloudflare's crawl-to-refer ratios describe this pattern at scale [17].High
Repeat visits at machine-regular intervals, including the small hours of local timeScheduled automation. People are irregular; scheduled jobs are not.Medium
Passes every browser check but never returns and never converts under any identityPossibly an agentic browser researching for someone. Treat it as a cohort, not a customer.Low alone, useful in aggregate
Detection signals for agentic traffic and what each indicates

Two rules follow. First, score every session on several signals at once and act on the score. Second, split automation from invalidity in your reporting: a declared fetcher serving a real buyer is not fraud, but it is not a conversion either. The steps below turn the table into a first audit.

  1. 1Export landing-page sessions for the last 30 days with user agent, IP, referrer, time on page and form events.
  2. 2Tag declared AI user agents, signed agent requests and known cloud and AI provider IP ranges. That is your GIVT floor.
  3. 3Flag the remainder on behavior: zero-interaction sessions, instant form fills, burst patterns and off-hours regularity.
  4. 4Compare the flagged share with credited clicks and adjusted cost in Google's Invalid Activity Credit Report [21]. The difference is what you are paying for.
  5. 5Push the resulting IP and audience exclusions to Google Ads, Meta and Bing, and remove flagged sessions from every lookalike seed.
09

How ClickTrust helps

ClickTrust classifies every landing-page event against 320 detection traps, each labeled GIVT or SIVT in line with MRC invalid-traffic guidance, and scores it in under 50 milliseconds as a single Traffic Quality Index (TQI). Declared crawlers fall out as GIVT. Undeclared agents, headless browsers and automated form fills are scored as SIVT on behavior.

Three layers share one workspace. Web Protection is a JavaScript tag covering search, social, display, affiliate and direct traffic. In-App Protection ships native iOS, Android and Flutter SDKs for emulators, click injection, repackaged apps and install hijacking. Lite-MMP Attribution adds tracking links, server-to-server postbacks and per-affiliate fraud scores.

IP and audience exclusions are pushed to Google Ads, Meta and Bing automatically, and refund-ready evidence packs support Click Quality Form requests. ClickTrust also runs an MCP server for AI agents, so your own agents can query traffic quality instead of adding to the problem. It is free to start.

What advertisers should do

The playbook

  1. 01

    Separate automation from invalidity in every report

    Add a third column next to human and invalid: automated but declared. Agents and fetchers serving real buyers belong there. They are not conversions, and with Imperva putting 40% of bot traffic as malicious and the rest as not [2], no rule will sort them for you.

  2. 02

    Filter declared AI crawlers at the tag, not only the firewall

    GPTBot, ClaudeBot, ChatGPT-User and Bytespider identify themselves, and OpenAI, Google and AWS sign their agents' requests [1]. Make conversion tags, pixels and analytics ignore them before they fire.

  3. 03

    Quarantine lead forms

    Hold every submission for a behavior check before it reaches the CRM or the conversion tag: fill time, field order, interaction before submit, origin IP. Google says it cut advertiser exposure to invalid leads by 30% on average so far in 2026 [7]; what it did not catch is in your pipeline.

  4. 04

    Rebuild retargeting and lookalike seeds from scored sessions

    Audiences built from raw pixel fires now include agents, fetchers and fake accounts, and Meta estimates fake accounts at about 5% of daily active people on Facebook in Q4 2025 [28]. Rebuild seeds from sessions that passed a behavioral score, and resync exclusions on a schedule.

  5. 05

    Audit commerce and comparison pages as their own line item

    Akamai found commerce took 47.9% of AI bot traffic and that more than 90% of it sat in monitor mode [6]. Product, pricing and comparison pages deserve their own invalid rate, separate from the brand pages agents rarely visit.

  6. 06

    File for credits with evidence, inside the 60-day window

    Google Ads investigation requests cover the last 60 days through the Click Quality Form [25]. Keep session-level evidence for every flagged click, and compare credited clicks in the Invalid Activity Credit Report [21] with the Invalid Traffic Overview Card to see what the platform caught.

Methodology

How this report was built

  • Every figure in this report is drawn from a published source named in the sentence where it appears and listed in Sources. ClickTrust Research did not generate new traffic measurements for this edition, and no ad-verification or click-fraud vendor's data is cited.
  • Primary publisher pages and PDFs were preferred. Where a figure was only available in a partner or trade-press summary (FIPP for TollBit's Q1 2025 data, PPC Land for TollBit's H1 2026 data, Security Boulevard's syndication of DataDome's report), the summary is cited as the source and named as such in the text.
  • Publishers measure different things: HTTP requests, HTML requests, bandwidth, verified-bot traffic, web and API transactions, paid clicks and installs. Rates from different networks are presented side by side, including in the bot-share chart, but never combined or averaged.
  • Dates matter. Cloudflare, Imperva, Fastly, Akamai, Google, Meta and LinkedIn figures cover 2025 to late 2026; F5 Labs covers November 2023 to September 2024; the MRC definitions date from 2015 and remain the governing standard. Where a page carried no publication date, the date is marked as inferred or as fetched.
  • The detection-signal table is practitioner guidance, not a measured dataset. No hit rates are claimed for individual signals. The ClickTrust audit range of 14 to 42 percent comes from ClickTrust's own audits and has not been independently verified.

Sources

References

  1. [1]Cloudflare, The Internet has a second audience (September 2026)
  2. [2]Imperva (a Thales company), Bad Bot Report 2026: Bots in the Agentic Age (April 2026)
  3. [3]Cloudflare, Content Independence Day, one year on: building the business model for the agentic Internet (July 2026)
  4. [4]Fastly, Fastly Threat Insights Report Reveals Organizations are Fighting Back Against the Rise of the Bots (December 2025)
  5. [5]Cloudflare, The 2025 Cloudflare Radar Year in Review (December 2025)
  6. [6]Akamai, Akamai Research: Commerce Becomes the Epicenter for AI Bot Attacks and Agentic Fraud in 2026 (July 2026)
  7. [7]Google, We're introducing a new transparency tool for invalid traffic in Google Ads (November 2026)
  8. [8]Media Rating Council (via PR Newswire), Media Rating Council (MRC) Issues Final Version Of Invalid Traffic Detection And Filtration Guidelines Addendum (October 2015)
  9. [9]Google, Invalid clicks: Definition (Google Ads Help) (Undated, fetched November 2026)
  10. [10]Imperva (a Thales company), 2025 Imperva Bad Bot Report: How AI is Supercharging the Bot Threat (April 2025)
  11. [11]Cloudflare, From Googlebot to GPTBot: who's crawling your site in 2025 (July 2025)
  12. [12]Fastly, New Fastly Threat Research Reveals AI crawlers make up almost 80% of AI bot traffic, Meta Leads AI Crawling As ChatGPT Dominates Real-Time Web Traffic (August 2025)
  13. [13]Vercel, The three types of AI bot traffic and how to handle them (August 2025)
  14. [14]F5 Labs, 2025 Advanced Persistent Bots Report (March 2025)
  15. [15]Kinsta, For AI traffic, the average website doesn't exist (2026 (undated page, September 2026 inferred))
  16. [16]Cloudflare, Cloudflare's 2026 Annual Founders' Letter (September 2026)
  17. [17]Cloudflare, The crawl-to-click gap: Cloudflare data on AI bots, training, and referrals (August 2025)
  18. [18]FIPP (with TollBit), State of the Bots: FIPP and TollBit put the spotlight on AI scraping (July 2025)
  19. [19]Akamai, Publishing Industry Under Attack: Global AI Bot Activity Surges by 300%, Akamai Report Finds (April 2026)
  20. [20]Salesforce, Holiday Season Rakes in Record $1.29T for Retailers, Salesforce Data Shows (January 2026)
  21. [21]Google, About the Invalid Activity Credit Report (Google Ads Help) (2026, undated page)
  22. [22]AppsFlyer, The State of Fraud for Marketers, 2026 Edition (April 2026)
  23. [23]PPC Land, 15% of AI page-fetchers in Europe reached disallowed URLs, TollBit finds (August 2026)
  24. [24]DataDome (via Security Boulevard), Which Industries Are Most & Least Protected from Bad Bots & Unwanted AI Traffic in 2025? (Global Bot Security Report 2025) (October 2025)
  25. [25]Google, About invalid traffic (Google Ads Help) (Undated, fetched November 2026)
  26. [26]Google, 2025 Ads Safety Report (April 2026)
  27. [27]Google India, Gemini is stopping harmful ads before people ever see them (India edition of the 2025 Ads Safety Report) (April 2026)
  28. [28]Meta (EU Transparency Centre, disinfocode.eu), Code of Practice on Disinformation report, March 2026 (covering July to December 2025) (March 2026)
  29. [29]LinkedIn, Community Report, H2 2025 (2026)
  30. [30]Microsoft, Microsoft Invest: Traffic Quality (Microsoft Learn) (March 2025)

Frequently asked questions

What is agentic traffic?
Agentic traffic is web traffic generated by software acting for a person or on its own: AI crawlers collecting training data, fetchers pulling a page to answer a prompt, AI agents completing tasks such as comparisons or form fills, and agentic browsers that drive a real browser with an AI. None of it is a customer, and most of it fires the same tags a paid click would.
How much web traffic is bots in 2026?
Imperva's Bad Bot Report 2026 puts bots at more than 53% of all web traffic in 2025, with humans at 47% [2]. Cloudflare's Radar Year in Review measures HTML requests and also finds humans at 47%, with AI bots averaging 4.2% [5]. Cloudflare's 2026 Founders' Letter says the point at which automated traffic passes human traffic moved forward from late 2027 to May 2026 [16].
Are AI crawlers like GPTBot and ClaudeBot invalid traffic?
Yes, by both definitions that matter. MRC's General Invalid Traffic category explicitly covers bots, spiders and other crawlers identified through list-based filtration [8], and Google Ads Help defines invalid clicks as clicks that do not represent genuine interest in your business, with automated traffic as one of its three types [9]. Declared crawlers should never count as sessions, conversions or audience members.
Does Google Ads charge me for AI bot clicks?
Not for the ones it catches. Google Ads Help says you are not charged for invalid traffic detected before billing and are credited for invalid activity found after the billing period, with investigation requests covering the last 60 days via the Click Quality Form [25]. Clicks and landing-page visits that Google's systems do not classify as invalid are billed and reported as normal.
Should I block AI agents from my landing pages?
Usually not outright. Imperva classifies 40% of bot traffic as malicious, which leaves most automation doing something someone asked for [2], and a block that stops scrapers also stops agents researching for real buyers. The better approach is to score sessions on behavior, keep agents out of conversion counts and audiences, and dispute the clicks you were billed for.
How do I detect AI agents in my analytics without a fraud tool?
Start with declared user agents, signed agent requests and cloud IP ranges, then look at behavior: sessions with no scroll or pointer movement that exit in seconds, forms completed faster than a person can type, bursts of many pages under one identity, and growth in direct traffic to pages only reachable from ads. Score sessions on several signals rather than relying on any one of them.

See what invalid traffic is costing you.

ClickTrust scores every paid click, install and lead in real time and gives you refund-ready evidence. Free to start.