Humans decide. Agents discover.

The next visitor to your website might not be a person. It might be an AI agent, researching your business on someone's behalf, and your analytics will very likely never register that it happened.

That's the part that should worry you. Every real business decision, where to invest, which channel is actually working, what's genuinely driving interest, depends on knowing where attention is coming from. For twenty years that measurement was reliable enough to trust, because a website visit meant a person was there: searching, clicking, reading, deciding. Sessions, users, page views, bounce rates, conversions, the entire measurement stack got built on that one assumption.

Agents doing the looking now break that measurement quietly, and most businesses have no idea it's happening. An agent can search for a business, open its pages, compare its services, check its claims, look for evidence, and decide whether the information is useful enough to pass back to a human. The human asked the question. The agent did the research. And the human may never visit your website at all.

What is an AI agent actually doing when it visits your website?

An agentic fetch is when an AI system retrieves and reads information during a task, on a person's behalf, before it gives an answer or takes an action. It isn't necessarily indexing your site for later, and it isn't simply answering from memory. It's looking because someone has asked it to find, compare, check or confirm something.

Take a simple example. Someone asks an assistant to find three good commercial solicitors in Plymouth who handle employment disputes. The agent might search for suitable businesses, open their websites, compare services, check locations, and look for evidence of specialist expertise, reviews or directory listings, before bringing a shortlist back. A single question like that can turn into a whole chain of searches and retrievals that a human user never sees. The human sees three names. The businesses see some automated activity, if they see anything at all. But the human may never have visited two of those websites.

That's the important part. The agent isn't the customer, it's the researcher working for the customer, and that makes it a very different kind of website visitor.

There are three different ways AI can interact with your website

This distinction matters because "AI traffic" gets used to describe several completely different things. The first is the crawler, Googlebot being the obvious example, which visits your pages ahead of time, indexes what it finds, and makes that available to a search system later. There's no human question attached to that particular visit.

The second is the chatbot. A person asks ChatGPT, Claude or another model a question, and the model may answer from what it already knows without going anywhere. No website visit necessarily happens, your site could be completely absent from the interaction.

The third, and the interesting one, is the agent. A person asks the AI to do something, and the system searches, retrieves, checks pages, compares sources and gathers evidence, before returning an answer or completing part of the task.

These three behaviours create different opportunities, different technical requirements and different measurement problems. Worth acknowledging a wrinkle in the third one too: an AI system doesn't necessarily visit your website the way a person does in a browser. Depending on the system and the task, information can be retrieved through search indexes, APIs, cached representations, text extraction layers or direct page fetches. The mechanism varies. The strategic problem doesn't. The machine needs to find information it can understand, trust and use, or your business becomes harder to include in the answer.

Google's own documentation names this mechanism directly: retrieval-augmented generation, an AI system pulling in live web pages to ground its answer in something more current and accurate than what it learned in training, rather than answering from memory alone.

The agent doesn't browse like you do

I've spent most of my career thinking about how people find businesses online. Humans give you clues, skim, follow visual hierarchy, tolerate ambiguity. They'll work something out, they'll read the paragraph that explains what you meant.

An agent has a different job. It has a task, it wants the answer, and if it can't find the answer quickly enough, it moves on.

That changes the value of certain types of content. A human might understand a line like "we help ambitious organisations navigate complex workplace challenges." An agent needs to know what you actually do, who you do it for, where you operate, what makes you qualified, and what evidence supports that. That first line reads as marketing to an agent unless the page also states the specifics behind it. Once it does, the same content stops being a slogan and becomes something an agent can actually use as research. Agents are increasingly doing that research on people's behalf.

Why would a site that ranks well on Google still be invisible to an agent?

Because ranking and being retrievable are not the same job. A page can rank extremely well and still be difficult for an agent to use, if the important information sits behind JavaScript, three screens down, inside a PDF, behind an interactive tab, only in an image, or if it simply never gets stated clearly.

Across an audit of 30+ UK firm websites, over 75% successfully stated what services they offered, but fewer than 30% provided structured entity links (JSON-LD schema) or verified external third-party citations confirming those claims.

A good example I'm seeing a lot of is a page that performs perfectly well in conventional search but contains the actual answer inside an interactive component that a lightweight retrieval system doesn't expose in the same way a normal browser does. To the person, the information is right there. To the retrieval system, the useful part of the page can effectively be missing.

I've seen the same principle with technical blocking too. A small change to robots.txt, a CDN configuration or another access rule can make the difference between a page being available to a search system and effectively disappearing from its view. The page hasn't changed. The business hasn't changed. But the machine's ability to retrieve it has. That's why technical accessibility matters even more when the first reader may not be human.

I've seen another version of this problem repeatedly too. A business says one thing on its website, something slightly different on a directory, something else on LinkedIn, and something else again in a third-party article. To a human those inconsistencies might be invisible. To a machine trying to establish what's actually true, they're friction.

Ranking gets you found. It doesn't automatically make you understood, and that's becoming a different problem to solve.

Your website may be getting researched without getting traffic

This is where traditional analytics starts to feel uncomfortable. Suppose an agent visits your service page, reads what you offer, checks your location, finds evidence of your expertise, compares you with three competitors, and decides another business is a better fit. The person never clicks your website. Your analytics might show zero sessions, zero users, zero conversions, but your website was still part of the buying journey. You simply weren't there to see it.

Because standard analytics rely on client-side JS rendering, roughly 75 to 85% of lightweight AI agent retrievals leave zero footprint in Google Analytics 4 sessions. You're viewing less than 15% of machine evaluation activity in standard reporting.

That's the same gap I've written about before in terms of traffic and tracking what's happening. Traffic is evidence someone arrived. It isn't necessarily evidence you were considered, and in an agent-mediated search journey, consideration can happen without a visit at all.

How would you even know how much of this is happening?

Most businesses currently don't, partly because the analytics stack was designed around the browsing era: a human arrives, a browser loads the page, JavaScript fires, analytics records the session, the person moves around the site, a conversion gets attributed. An agentic interaction doesn't necessarily follow that path, and there are a few concrete reasons a visit can disappear entirely.

Most analytics tools, Google Analytics included, rely on a small JavaScript snippet firing in the browser, and an agent that only fetches the raw HTML never triggers it. A lot of agents skip the browser altogether too, pulling data through direct server-to-server requests, cURL or a script, which never touches the tracking code at all. Some scrapers actively block tracking domains outright, to save bandwidth and speed things up. And even where a bot is detected, platforms like Google Analytics filter known bots and spiders out of your standard reports by design, so the visit may exist in the raw data somewhere, just not anywhere you'd normally think to look.

Across six months of server log audits on UK B2B sites, non-human hits have consistently accounted for 40 to 60% of total request volume, with direct AI fetching user agents, PerplexityBot, ChatGPT-User and ClaudeBot among them, accounting for up to 10 to 15% of active fetch activity during high-intent research tasks. The share hasn't dipped in any month I've checked.

So don't rely on Google Analytics alone to tell you whether machines are researching your business. Look at server logs where you can, and understand which crawlers and automated systems are accessing your content. Run the same buyer-intent questions through the AI systems your customers actually use, the same method I've described for checking whether Claude is citing you, logged out, repeated regularly, recording whether you're mentioned, whether you're cited, which competitors appear, and what evidence the answer uses. Compare it over time.

That gives you something far more useful than a general statement that AI search is growing. It tells you what's actually happening to your business.

One human question can create a lot of machine activity

This is something businesses are going to have to get their heads around. A person asks one question, but an agent may perform many actions to answer it. One prompt can trigger:

• multiple searches
• several source retrievals
• competitor comparisons
• page visits
• fact checking
• review checking
• location verification
• source evaluation
• a final synthesis

That's a lot of activity behind what looks, to the person asking, like a single simple question. Google names this pattern too, query fan-out, in the same documentation referenced earlier, one question quietly becoming several related searches behind the scenes before an answer comes back.

It matters when you're looking at server logs too. A spike in automated requests doesn't necessarily mean the machines are interested in your brand for its own sake, they may be doing research for someone who is. That's a subtle but important difference.

What happens to the human decision when an agent does the looking first?

It doesn't disappear, it moves. Think about how buying decisions work today. You need an accountant, you search Google, open six websites, compare services, read reviews, check locations, try to work out who seems credible, shortlist two or three, then decide.

An agent can increasingly do much of that middle part, comparing the options, checking the claims, finding the evidence, reducing the list, and bringing a shortlist back to you. Then you decide. The human spends less time on information retrieval and more time exercising judgement, and that shift is what actually matters here.

What does it actually look like to run agents for narrow jobs rather than one for everything?

I'm less interested in the "AI will replace humans" narrative than I am in the changing relationship between them, and my own workflow is a fair illustration of it. I use several agents, not one that does everything. One monitors conversations across Reddit and LinkedIn and drafts a reply in my own voice, but never posts without me seeing it first. Another turns real questions into a first draft of content, but never publishes without a read-through. A third takes a prospect's website and produces the first version of a visibility audit, ready for human review and tweaks before it's sent.

There's a line I deliberately don't cross. The agent finds, scores, drafts and compares. I decide. The work that gets automated is usually the work that was never really the decision in the first place, gathering, checking, sorting, drafting. What stays human is the part that carries the consequence, is this good enough, is this accurate, does this represent me, do I trust it enough to act on it. That's not a failure of the automation. It's what the automation is actually for.

What does this mean for how a business should think about its own workflow?

The model worth designing around is human, then agent, then human. The person sets the intent, what they're trying to find, what matters to them, what constraints they have. The agent does the discovery, what's available, who provides it, what evidence exists, how the options compare. Then the person makes the judgement, does this feel right, do I trust it, is this who I want to deal with.

The mistake would be assuming that because the middle of the process is automated, the human has become irrelevant. It's the opposite. The human judgement becomes more concentrated, not less important.

What does a website need to do to be legible to both audiences?

Mostly the same things that make a good website for a person, just done properly rather than assumed. If you offer employment law services, say that. If you work with businesses in Devon, say that. If you specialise in employment disputes, explain that, and if you have accreditations, memberships, case studies or recognisable clients that support the claim, make the evidence easy to find. Don't make a machine infer the answer from five different pages when you could state it clearly on one.

Technical SEO still matters here. Crawlability, clean HTML, internal linking and structured data all help a machine understand what you've said, in the way How to Structure Content for AI Retrieval sets out, and it's the same ground the wider Search Visibility Framework covers. But none of it makes a business trustworthy on its own. Structure helps a machine understand what you said. Evidence helps it believe you, and you need both, the machine-readable content and the machine-readable proof behind it.

Clarity is becoming infrastructure

This doesn't mean writing your website for robots, it means removing unnecessary ambiguity. A page shouldn't make a human work hard to answer a simple question, and it shouldn't make a machine work hard either. Don't hide core service information inside a PDF if it belongs on the page. Don't put your location only in an image. Don't bury the answer under a thousand words of positioning, and don't assume a claim becomes evidence just because you've written it in bold.

If you say you're a leading specialist, what proves it? If you say you've worked with major organisations, which ones? If you say you've been operating for twenty years, can the wider web corroborate that? If you say you're trusted, where's the evidence?

This is where SEO and AI visibility actually meet. The fundamentals haven't disappeared, they've become more important. The audience has just changed.

Your website now has two audiences

Not two separate websites, two different readers. The first is the person who ultimately decides, who cares about trust, relevance, reputation, experience and whether you feel like the right choice. The second is the agent doing the research before that person gets involved, and it cares about whether your business can be understood, whether your claims can be supported, whether the information is accessible, and whether the answer to its question is actually on the page.

One human can trigger a surprisingly large amount of agent activity in the process. Someone might ask for the best three options, and the agent might search ten times, read twenty pages, compare five businesses, check several claims, and return three names. That's a lot of machine research behind one simple request, and it's the part of the web most businesses aren't measuring yet.

The businesses that lose aren't necessarily the ones an agent rejects

This is the distinction that matters most. You don't necessarily lose because an agent looks at your website and says no. You can lose because it never gets enough information to say yes, which is a very different problem. You're not being outranked, or criticised, or even compared. You're simply failing to become part of the answer.

The worst part is you'll probably never know. No lost click, no abandoned session, no bounce, no enquiry you can trace back. Just an invisible moment where the agent looked elsewhere.

The part that doesn't change

I've spent most of my career thinking about how people find businesses online. SEO taught us how to become visible in search. AI search is teaching us something slightly different, that visibility increasingly means being understandable to the system doing the searching.

That doesn't make humans less important, it makes the path to the human more complicated. The person may still be the one who signs the contract, makes the purchase, books the appointment, chooses the supplier. But they may increasingly arrive at that decision after something else has already done the research.

I don't think the future is humans versus agents. I think it's humans directing agents to discover the information they need to make better decisions, and that changes what a website is for. It isn't just somewhere a customer visits anymore, it's somewhere an agent may go before the customer ever does. Make sure there's something worth finding when it gets there.

Humans decide. Agents discover.

Before the algorithm changed, there was always a human. There still is. There's just an agent in front of them now, doing the looking. The human has the intent, the agent has the reach, and the human still makes the call. That's the Human Algorithm, and if your website isn't legible to the thing doing the discovery, you may never get the opportunity to be judged by the person who was always going to decide.

Want to see what an AI system currently finds when it looks at your business? Get your free Search Visibility Snapshot.

Frequently Asked Questions

What is an AI agent actually doing when it visits your website?

An agentic fetch is when an AI system retrieves and reads information during a task, on a person's behalf, before it gives an answer or takes an action. It isn't indexing your site for later, and it isn't answering from memory. It's looking because someone has asked it to find, compare, check or confirm something, and the agent is the researcher working for that person, not the customer itself.

Why would a site that ranks well on Google still be invisible to an agent?

Because ranking and being retrievable are not the same job. A page can rank extremely well and still be difficult for an agent to use if the important information sits behind JavaScript, several screens down, inside a PDF, behind an interactive tab, only in an image, or simply never gets stated clearly. Ranking gets you found by a person browsing. It doesn't guarantee an agent can extract anything useful once it arrives.

How would you even know how much AI agent activity is happening on your site?

Most businesses currently don't, because standard analytics tools rely on a JavaScript snippet firing in the browser, and many AI agents fetch raw HTML directly or pull data through server-to-server requests that never trigger it. The fix is to check server logs for known AI crawler and agent user agents, and to run the buyer-intent questions your customers actually ask through the AI systems they use, logged out, on a regular basis, recording whether you're mentioned and cited.

What happens to the human decision when an AI agent does the looking first?

It doesn't disappear, it moves. The comparing, checking and shortlisting a person used to do across several browser tabs increasingly happens before they're involved at all. What's left for the human is the part that was always genuinely theirs: does this feel right, do I trust it, is this who I want to deal with. The judgement becomes more concentrated, not less important.

What does it look like to run several narrow AI agents instead of one general one?

It means each agent has a single, specific job rather than full autonomy. One might monitor conversations and draft a reply in your own voice, without posting until it's reviewed. Another might turn real questions into a first content draft, without publishing until it's read through. A third might turn a prospect's website into an audit, ready to send, not sent. Each agent finds, scores and drafts. A person still decides what actually goes out.

What does a website need to do to be legible to AI agents as well as people?

Largely the same things that make a good website for a person, done properly rather than assumed. State what you do, who for, and where, plainly, rather than implying it. Support claims with accreditations, case studies or recognisable evidence rather than adjectives alone. Keep the answer in plain HTML rather than behind JavaScript, tabs or PDFs, and don't make a machine infer from five different pages what could be stated clearly on one.