What Is Agentic AI, and What Are the Risks to Paid Media?

Agentic AI has moved from demos to mainstream browsers in a little over a year, and the software now searches, clicks and fills in forms for real people. For paid media teams, that raises an awkward question: what happens when the visitor on the other end of a paid click is a machine acting for a human? This article explains what agentic AI is in plain terms, where it touches your paid search and paid social campaigns, how ad platforms currently classify it, and what to do about it now.

What is agentic AI? A plain-language definition

Agentic AI refers to AI systems that do not just answer questions but take actions. You give an agent a goal, such as "find me a hotel under $200 near the conference venue and book it," and it opens web pages, reads them, clicks buttons, types into forms and moves through checkout, pausing for approval at key steps.

The technology became widely available through a series of launches:

• Anthropic computer use (October 2024). Anthropic introduced a beta capability that lets developers direct Claude to use computers "by looking at a screen, moving a cursor, clicking buttons, and typing text," according to Anthropic's computer use announcement.
• OpenAI Operator (January 2025) and ChatGPT agent (July 2025). OpenAI described Operator as an agent that can "look at a webpage and interact with it by typing, clicking, and scrolling." Six months later, OpenAI's ChatGPT agent launch folded Operator into ChatGPT, with a visual browser, a text browser and a terminal.
• Perplexity Comet (free worldwide since October 2025). Perplexity's AI browser includes an assistant that can act in any tab, and Perplexity announced Comet was free for everyone on October 2, 2025.
• Google Chrome auto browse (January 2026). Google began rolling out an agentic mode in Chrome that can handle tasks like "filling out online forms, scheduling appointments, and managing subscriptions," initially for AI Pro and Ultra subscribers in the U.S., as MacRumors reported on Chrome's Gemini features.

The key difference from older bots is intent. A scraper works for its operator. An AI agent usually works for a specific person who may genuinely want to buy something. That is exactly why agent traffic is hard for advertisers to categorize.

How much agent traffic is there?

Volumes are still small compared with human traffic, but they are growing fast. HUMAN Security's analysis of AI agent traffic found that agentic traffic across its customer base grew by more than 1,300 percent in the first eight months of 2025, with Perplexity Comet and ChatGPT agent accounting for most of it by September. HUMAN also reported that roughly 87 percent of pages agents browsed were product-related. Over the 2025 holiday season, HUMAN measured a 144.7 percent jump in daily agent traffic to shopping sites from the pre-Black Friday baseline to the November 29 peak.

Where agentic AI meets paid media

Agents interact with the same surfaces your ads run on. When an agentic browser runs a search, it sees a results page that can include sponsored listings. When it researches a product, it may land on a URL carrying your UTM parameters and click IDs. When a task involves "get a quote" or "sign up for a demo," it may complete your lead form.

No ad platform has published figures on how often agents click paid placements, so the scale is unknown. The mechanics, however, are straightforward, and each one has a cost for campaigns optimized on clicks and conversions.

The concrete risks of agentic AI for paid search and social

1. Paid clicks with no human on the other side

If an agent clicks a sponsored result or follows a paid link, the click may be billed like any other. The agent might be doing useful research for a real buyer, or comparing ten vendors for a summary, with no person ever seeing your landing page. You pay the cost per click (CPC) either way, and high-CPC categories such as software, legal, finance and travel feel it most.

2. Polluted conversion data and smart bidding signals

Automated bidding learns from what you tell it is a conversion. Google's Smart Bidding documentation describes Smart Bidding as bidding strategies that use AI "to optimize for conversions or conversion value in every auction," using signals such as device, location, time of day, browser and site behavior. If agent sessions trigger micro-conversions such as page views, add-to-cart events or form starts, the bidding model learns to chase the audiences, placements and queries that produce agent activity rather than human buyers.

3. Fake or low-quality form fills and leads

Filling forms is a headline feature of these products. An agent submitting a contact form on a user's instruction is not fraud, but it can still produce leads that are thin, duplicated across many vendors, or submitted without the person ever intending a sales conversation. For lead-gen advertisers who import offline conversions or optimize on form submissions, that inflates cost-per-lead reporting and feeds low-quality signals back into the platforms.

4. Attribution distortion

Agents do not browse like people. They open many tabs, skip pages and may complete a purchase somewhere other than where they researched, which breaks the assumptions behind session-based and multi-touch attribution. A paid click might get credit for a conversion a person made later by themselves, or the conversion might happen inside an agent flow your tracking never sees. If you need a refresher on how models assign credit, see our guide on what marketing attribution is and how it works.

5. Agents that look exactly like Chrome

Many agentic browsers do not announce themselves. HUMAN's technical comparison of ChatGPT Atlas and Perplexity Comet found that both send a standard Chrome user agent string and Chrome-like Client Hints, concluding that "surface-level detection is insufficient." In your analytics and ad reports, these sessions can look like ordinary desktop Chrome visitors.

How invalid traffic rules handle declared AI agents

Ad platforms protect advertisers through invalid traffic (IVT) filtering. The question is whether agents fall inside those definitions, and the honest answer is that the rules were not written with them in mind.

• Google's definition is based on genuine interest. Google's invalid clicks definition covers clicks "that don't represent a genuine interest in your business," and lists "clicks from automated tools, bots, or other deceptive software" as an example. Google states advertisers are not charged for clicks it determines to be invalid. Its overview of invalid traffic does not mention AI agents specifically, leaving open whether an agent acting for a real shopper counts as automated traffic or genuine interest.
• Industry filtering relies on declared bots. The Media Rating Council (MRC) standards split IVT into general invalid traffic (GIVT) and sophisticated invalid traffic (SIVT). GIVT filtering leans heavily on the IAB Tech Lab Spiders and Bots List, which the IAB describes as a core requirement for MRC compliance. That approach works for bots that identify themselves. It does little for an agent presenting a standard Chrome user agent.
• AI crawlers already inflate GIVT. DoubleVerify's Fraud Lab reported an 86 percent year-over-year increase in GIVT in the second half of 2024, with AI scrapers such as GPTBot and ClaudeBot making up a record 16 percent of known-bot GIVT impressions. Those are crawlers, not agents, but the trend is clear.
• Verification vendors are creating a separate category. In November 2025, DoubleVerify introduced AI agent measurement that distinguishes declared AI bots, evasive scrapers, and "automated and agentic AI browsing" such as Atlas and Comet sessions. The fact that agents needed their own bucket tells you standard IVT labels did not fit.
• Signed agents are emerging. Some agents now identify themselves cryptographically. Cloudflare's signed agents program, announced in August 2025, recognizes agents whose requests are signed with Web Bot Auth, including ChatGPT agent. Declared agents like these are easier to segment, but signing is voluntary and far from universal.

The buy side is already reacting. Digiday's reporting on agentic traffic found that a mid-sized advertiser halted spending with Salon after detecting agentic visitors. The takeaway: you cannot assume your ad platform will treat agent clicks as either invalid or valid. You need your own visibility.

What marketers should do now: an agentic AI checklist

You do not need to block every agent. Some carry real purchase intent. You do need to know when they are present and keep them from steering your bidding. Start with these steps:

1. Segment agent traffic in your own data. Classify sessions from known agents and agentic browsers separately from human visitors, and report on them by channel and campaign. Do not rely on user agent strings alone, since several agentic browsers present as Chrome.
2. Check for signed requests. Where your CDN or bot management tool supports Web Bot Auth, use signatures to identify declared agents and decide whether to allow, label or challenge them.
3. Validate conversions before you send them back. Only pass conversions to Google Ads or Meta after checking they came from human sessions. For lead gen, qualify leads before importing offline conversions, and avoid optimizing on micro-conversions that agents trigger easily.
4. Protect your forms. Add server-side validation, duplicate detection and lead scoring. Flag submissions completed at machine speed.
5. Use platform exclusions where evidence supports them. Exclude IP ranges, placements or audiences that repeatedly deliver agent or automated traffic with no downstream revenue, and file invalid click reviews when you have logs to back them up.
6. Monitor landing pages for spikes. Watch for bursts of paid sessions with very short durations, identical navigation paths, or high form-start rates with low completion, especially on high-CPC keywords.
7. Keep your UTM data clean. Consistent tagging makes it far easier to see which campaigns attract agent traffic. Our free UTM builder helps keep parameters consistent across teams.

How Ðeny helps you see agent traffic on paid campaigns

Ðeny classifies every visit to your marketing site in real time. The Ðeny Intelligence Hub sorts traffic into Clean, Good Bot, Residential Proxy, IVT, Out-of-Geo and AI Agent categories, so agent sessions are separated from human visitors instead of being blended into your reports. A Traffic Health Score from A to F and AI-generated summaries show where quality is slipping.

The UTM Breakdown shows traffic quality per channel, so you can see whether a specific campaign is attracting agents or other invalid traffic. Ðeny Bot Shield can block unwanted automated traffic, and Ðeny Enterprise Data provides visitor-level logs and exports you can use to validate conversions or support invalid click claims. Ðeny deploys with a one-line script. See the full Ðeny feature overview for details.

Conclusion: plan for agents, do not panic about them

Agentic AI is a new kind of visitor: automated in how it browses, but often acting for a real person with real intent. Current invalid traffic rules were built for bots that either declare themselves or try to deceive, and agents fit neither cleanly. That leaves paid media teams exposed to wasted clicks, noisy conversion signals and misleading attribution until the platforms catch up.

The advertisers who handle this well will be the ones who measure agent traffic themselves, keep it out of their bidding signals, and make deliberate decisions about what to allow. If you want to see how much of your paid traffic is already coming from AI agents, request a Ðeny demo.