| Scenario | What is that? | What does it do with the ads |
|---|---|---|
| AI crawler | Collects training data on a schedule | Does not load ads at all |
| AI fetcher | One request on a user’s behalf, right now | Usually does not load them, or they are not counted |
| Agent browser | Drives the user’s own browser | Loads the full page, looks like a person, and loads ads |
Your Next Visitor Isn’t Human: What Agentic Browsers Mean for Publisher Revenue
So we’ve more or less adapted to AI Overviews and eventually exhaled. Well, too early: the next scenario is over there.
Here it is: someone asks their AI agent to find the best new TV shows. The agent opens your site, reads the review, summarizes it, and replies. Your content was consumed in full, but your ads remained unseen. Meanwhile, your Google Analytics shows an IVT spike. Yes, that’s Invalid Traffic; you got it right.
And here’s how it looks in numbers:
6.5x
AI traffic grew that much faster than human traffic, January to May 2026
Fastly
45.6%
of all agentic activity lands on media sites, more than any other industry
HUMAN Security, April 2026
+45%
growth in AI agent requests in a single quarter, from 12.2B to 17.7B
DataDome, Q2 2026
Most of that volume is crawlers collecting text for training. The share that loads your page in full, and your ads with it, is the smallest one.
So is it just a temporary mishap, something you can ignore, or a real threat to your profit? We did some research, asked an SEO expert, and are ready to answer.
Who helped us:
Evgeniy Yudin, CEO Rotgar Ai visibility
Before We Go: What Does Agent Traffic Even Mean?
As Evgeniy says, сrawlers collecting data for training, a page loaded on a user’s request, and an agent driving a user’s browser are three different scenarios. A single growth number does not tell you how many people actually handed their own browser over to an agent.
So what are these scenarios in detail? Here is a quick table based on Fastly, HUMAN Security data, and Cloudflare terminology.
What’s the difference, and what does it mean for your profit?
AI Crawler
An AI crawler comes, as Digiday says, ‘often without generating a traditional pageview or ad impression so publishers can’t make money on the visit’. The problem here is not fake impressions or an invalid traffic flag in your report, but the agent used your server, took your content, and paid you nothing. No real harm either, to be fair.
AI Fetcher
A fetcher comes from a datacenter IP, and names itself honestly: ChatGPT-User or something like that. It often doesn’t create any ad impressions, too – but if it did, though, this impression can be marked as GIVT. The same Digiday article underlines that it’s ‘not fraud, so advertisers do not pay for non-human traffic media’.
So what’s the problem?
A buyer gets a report where they see a platform with a growing share of invalid traffic, and they don’t care if it’s just a harmless agent. A red metric is a red metric. That is exactly what Digiday reported at a long-running US news and commentary site: one advertiser cut its spend entirely. According to that report, nothing fraudulent was found; the metric just started to look bad.
Note: it’s too early to panic. Speaking to Digiday, Justin Wohl, VP of Strategy at Adtitude, said: ‘It’s super reactive by the buy side to say we’re not going to spend anymore, because all of my views are agentic visitors’. In the same article, Paul Bannister, CRO at Raptive, summed it up as ‘I think it’s more fear right now than reality’.
Agent Browser
The agent browser is the most unpleasant case. It comes from a home IP address of a user, with normal cookies and a real session. It loads a full page and triggers an ad tag – just like if it were a human user. So, it creates an ad impression, looking very realistic – and a buyer pays for it.
Evgeniy:
Rendering an ad does not mean a paid impression. Google treats automated traffic sources and artificial impressions as invalid traffic. That activity can be filtered, and the revenue can be adjusted. So I would not count agent visits as an equal replacement for human views.
Well, sounds profitable at first sight – you simply get your profit, even with no real users. However, a buyer won’t always look at the reports without checking the results, and eventually they will notice impressions are coming in, but clicks and conversions are not. This might affect your CPM, and buyers might start refusing your site in the end.
To sum up: The crawler costs you revenue today, the fetcher costs you nothing but makes your reports look guilty, the agent browser pays like real traffic until the conversions don’t come, and then your CPM and your reputation go with it.
Does the Industry Suggest Something To Solve It?
Yes. Four options, each covering a part of a problem.
1. Charging for access. Cloudflare’s pay-per-crawl gives you a third option besides allowing and blocking bot traffic. It works like this: you set a price per agent request, and a crawler that will not pay does not get your content.
Here, your inventory is access to your content, priced per request.
2. Charging for use. RSL, or Really Simple Licensing, is a licensing layer on top of robots.txt. Instead of just saying who may crawl your site, you say on what terms: free, credit required, subscription, a fee per crawl, or a fee every time your content ends up in an AI answer. Reddit, Yahoo, People Inc., Ziff Davis, O’Reilly, Medium, Quora, Raptive, and Fastly are on board.
Here, your inventory is the right to use your content in an answer.
3. An industry protocol. IAB Tech Lab is working on the Content Monetization Protocol, or CoMP. It is a standard procedure for an agent to get from ‘blocked’ to ‘allowed under an agreement’.
Here is how it’s supposed to work:
- a bot asks for your content
- there is no agreement, so you refuse and point it to a page with your terms
- the two sides agree on something
- the bot asks again, this time saying what it needs the content for
- you issue it an access token
After that, there is a record of who took what, when, and on what terms. Right now, nothing like that exists on the open web.
Here, your inventory is whatever the agreement says. However, this is not a service you can switch on, and there is no announced final version or publicly confirmed implementation yet. Right now, it describes how this should work.
A share of the browser’s revenue. Perplexity’s Comet Plus collects money from its subscriptions and splits it: publishers get 80%, Perplexity keeps 20%. You get paid in three cases: a reader opens your site from the Comet browser, your content is cited in an answer, or the agent uses your content to complete a task. Fortune, Gannett, The Independent, Time, Der Spiegel, and Blavity are in.
Here, your inventory is the agent’s action itself.
It seems like you have a full pack of solutions, but in reality, most of it is half-built for now, and none of it is something you just switch on. All four have the same catch: they are simply offers. You can set a price, state your terms, but none of that makes an AI company pay.
What’s more, nobody has said what counts as a valid impression when the visitor is an agent. There are no rules for counting those impressions and no separate category for them in invalid traffic reporting.
| What you can actually do today | What you get for it | |
|---|---|---|
| AI crawler | Block it, price it per request, or state license terms in robots.txt | Control over access. Money only if the AI company decides to pay; otherwise, it simply leaves |
| AI fetcher | The same tools, with a weaker grip, since it is harder to tell apart from a person and there is no agreed price or unit to count | Control over access again, and a license only if you are big enough for someone to negotiate with you |
| Agent browser | Spot it by behavior, and get paid for it through one browser’s partner program you have to be invited into | Mostly just knowledge that it happened. No standard way to count the impression, no category for it in reporting, and no one else paying for it |
As you see, you can more or less handle the crawlers, but all the rest have pretty vague guidelines. Anthony Katsur, CEO of IAB Tech Lab, is blunt about the timeline: ‘Nothing truly scales until 2027 or 2028’.
So, to answer the question: the industry has built plenty, but almost none of it puts money in your pocket this year. As Evgeniy says,
My strategic position is to give AI access to useful content and learn to work with this channel. I think presence in assistant answers will only get more important, so I prefer to adapt now. Access can still be regulated: allow certain bots, limit how often they come, keep part of the material behind a login, and set the terms of use separately.
So What Can a Publisher Actually Do?
You cannot rely on the industry to solve this right now, so what are your real and urgent moves? In short: separate agent visits from the rest, then work on your content and your monetization instead of fighting the technology.
1. Start by understanding what your traffic is made of
Your traffic number might lie to you: bots go up, people go down, but the total stays flat. Meanwhile, your revenue might drop for no visible reason. Splitting the traffic tells you which of the two is happening: you are losing readers, or you are just collecting machines.
Where to look? Two places:
- Server logs, or your CDN dashboard. This is the only place crawlers and fetchers exist. They do not run JavaScript, so Google Analytics never sees them. Look for the names: GPTBot, ChatGPT-User, ClaudeBot, PerplexityBot, Meta-ExternalAgent. Count how many requests they made and which pages they visited.
- Your analytics. It shows people and agent browsers together, and there is no name to filter by, because an agent browser is a real browser on a real person’s laptop. It gives itself away by how it behaves, not by what it calls itself. The cursor jumps straight to the button with nothing in between – people do not move a mouse that way.
Evgeniy:
You cannot tell who is in front of you from a single click, a scroll or the time on page. A person can read the answer fast and leave without clicking once, and an agent can run a long sequence of actions. Even a captcha passed by a human does not prove the person does the rest of the work. I would look at the whole set of signals, keep a ‘could not determine’ category, and count bot requests separately from human visits.
Once you catch even the cleverest agents, you clearly see what share of your sessions were human, and your revenue per human session. If most of your traffic is crawlers taking content, that is a licensing question, and you already know the answer to that one from the section above. If the human half is what is shrinking, no amount of bot detection will help you, and the next two steps are the only ones that will.
Evgeniy tracks this on his own projects, and here is his conclusion:
We track requests from known AI bots to our pages and to the files we prepare for them. To me that shows the material is accessible and in demand with these systems. It is not yet proof of commercial value. Separate Markdown or JSON files can be convenient for particular systems and tasks. But there is no general rule that all neural networks prefer such files. Google states directly that special AI files and Markdown are not required to appear in its generative search. So technical experiments need checking: a file being downloaded does not prove it was used in an answer, or that clicks went up.
2. Adapt the content
If your page only exists to give a short answer, and AI can give that answer instead, the reader never comes. What holds up is content you cannot squeeze into a summary: your own data, a named expert, tools people use instead of reading – anything that gives a reason to come back.
Evgeniy:
‘In my view the user’s task and the way the site makes money matter more than the topic itself. If the person came for a short factual answer, the AI can satisfy that need right in the chat. For an information site that earns on ads or affiliate clicks from the article, that is a problem: the person got the information but never reached the point of monetization.
You have to give the person a real reason to continue: use a tool, watch a video, see the raw data, read the full opinion of the author, join a discussion, or take an action such as booking an appointment. Unique data on its own does not guarantee a click, because the assistant can retell it too. What matters is what else the user can get on the site after the short answer.’
Note that even if AI shows your content far more often than people click through to it, that presence can still pay off. It depends on your niche. Evgeniy sees this on medical projects: as presence in AI answers improves, branded search goes up. The likely path is that a person picks an organization with the help of an AI tool, then looks it up in Google or Maps and gets in touch.
Evgeniy:
‘I sometimes call what is happening the arrival of an internet for neural networks. The point of that metaphor is that the site now has one more reader. A person understands a lot from context, while an assistant needs that context spelled out.
A local resident knows how insurance works, so on a clinic page they read prices with cover, without cover, and under different plans. Comparing that with a competitor who lists one minimum price, the AI can wrongly conclude that the first clinic is more expensive. In our practice we try to reduce those readings: we spell out the conditions, what the price includes and the limits.’
Just do not expect to prove it – as Evgeniy puts it:
It is hard to separate that effect from other channels, so you cannot attribute all the growth in branded search to AI.
So, adapt the content either way. What changes is what you should expect back: in some niches the mention alone brings people in; in others nothing happens until they actually open your page.
3. Create a pool of loyal users
AI broke that simple deal: where Google sent you a visitor, they watched an ad, and you got paid; now AI answers the question itself, so the visit might never happen.
So it doesn’t make sense to chase more visits, as you probably did before. Instead, a good fix might be loyal readers who come back on their own – and no AI agent can take those readers away from you.
You can obviously add more monetization channels, like subscriptions, paid content, or affiliate links, but all of it only works on someone who is already there. It’s not really different with the ads: formats that reach a reader who opted in, like push subscriptions, do much better work on a returning audience.
Evgeniy says the same thing when he explains what to measure:
Open access and extra files do not guarantee survival on their own. You have to check whether presence in AI answers turns into an audience, into a relationship with that audience, and into income.
Instead of The Summary: What Should You Do This Quarter?
Publisher checklist
What to do about agent traffic this week
Fourteen things you can do without waiting for the industry to finish building. Tick items as you complete them.
01
Find out what your traffic is made of
02
Decide your terms for crawlers
03
Watch the number your buyers see
04
Give readers a reason to come back
Agents are growing, but most of them are crawlers: Fastly counted 85% in May 2026, and a crawler never loads an ad. The human half of the web is still there, still clicking. So, ads are not going anywhere – there are still real readers on your site, and we are ready to help you monetize them.
