AI Browsers Business Traffic: After Comet and Atlas

Something quiet is happening to your AI browsers business traffic, and most business owners will not notice until the numbers go sideways. Perplexity shipped Comet. OpenAI shipped Atlas. Google put Gemini inside Chrome and Microsoft put Copilot Mode inside Edge. Each one puts an assistant between the customer and your site, and in some cases removes the customer from the loop altogether.

This post is about what AI browsers do to business traffic: what actually changes, what your analytics will and will not show you, which parts of your website quietly stop working when software is the visitor, and what to fix first. No predictions about 2030. Just what is already live, what I am seeing in client accounts, and the practical list.

Key Takeaways

• AI browsers like Comet and Atlas are already changing business traffic: fewer visits, better qualified, and much harder to attribute correctly.
• Google users who see an AI summary click a traditional result link about 8% of the time, against 15% when no summary appears.
• Agent visits often land in analytics as direct traffic or nothing at all, so a falling session count can hide rising demand.
• If your prices, hours or service areas only appear after JavaScript runs, an AI browser agent reads none of it and skips you.
• Structured data, plain HTML service pages, opening hours and clear contact details are what an agent reads before it recommends anyone.
• Agent-assisted buyers arrive having already compared three suppliers, so response speed decides the job more often than ranking does.
• 78% of deals go to the business that responds first, which matters more when the same agent has queued enquiries to your competitors.
• Track calls, form fills, branded search and “how did you hear about us” answers rather than raw sessions from here on.

What an AI Browser Actually Is

A regular browser fetches a page and shows it to a person. An AI browser fetches the page, reads it, and can act on it. That is the whole difference, and it is a big one.

Perplexity released Comet in mid 2025, initially to paying subscribers and later more widely. OpenAI released ChatGPT Atlas in October 2025, macOS first. Both are built on Chromium, so they look and feel like Chrome. The Browser Company shipped Dia. Google added Gemini into Chrome directly. Microsoft added Copilot Mode to Edge. Five serious players inside about eighteen months, all pointed at the same idea.

The sidebar assistant

The mild version is a panel sitting beside the page. The customer is on your pricing page and asks the assistant, “Is this cheaper than the other two I looked at, and do they cover Albany?” The assistant reads your page, reads its memory of the other tabs, and answers. Your page still got the visit. It just did not get the attention. Your carefully written third paragraph got summarised into eleven words by something that has never met you.

Agent mode

The serious version is agent mode. The customer says “find me three plumbers who do emergency call-outs in East Auckland tonight and send them all the same enquiry.” The browser then does it. It opens tabs, reads pages, fills in forms, and reports back. Nobody read your homepage. Software read your homepage, decided in about four seconds whether you qualified, and either submitted your contact form or moved on.

Agent mode is unreliable today. It gets stuck on cookie banners, mis-clicks dropdowns, and gives up on anything that needs a login. Anyone telling you it is finished is selling something. But “unreliable today” is exactly what mobile search looked like in 2009, and the direction of travel is not ambiguous.

Why this is different from a chatbot

A chatbot answers questions about the world. An AI browser acts inside your session, with your logins, your tabs and your history. It is the difference between asking someone for directions and handing them your car keys. That shift is the reason this matters commercially rather than just technically. The assistant is no longer an information source. It is a buyer’s agent, and it is standing between your marketing and your customer.

AI Browsers, Business Traffic, and the Zero-Click Squeeze

You have already lived through a smaller version of this. Google’s AI Overviews rolled out, and a lot of informational traffic evaporated. The Pew Research Centre studied this behaviour in 2025 and found that when an AI summary appeared, users clicked a traditional search result about 8% of the time, compared with about 15% when no summary appeared. Roughly half the clicks, gone. Only around 1% clicked a link inside the summary itself.

AI browsers push that further, because now the summarising happens on your page rather than before it.

The clicks you lose are the ones you were never going to convert

Here is the part that gets missed in the panic posts. The traffic that disappears first is informational. “How much does a new hot water cylinder cost in NZ.” “Do I need a building consent for a deck?” “What is the difference between Invisalign and braces?”

That traffic was mostly people gathering information. It fed your content marketing reports and some of it converted eventually, but it was never your best traffic. The assistant handles it now.

What survives the filter is transactional. Someone who wants a quote, a booking, a call-out, a consultation. Those visits are worth more each, and there are fewer of them.

The new shape of the numbers

So the pattern I keep seeing in client accounts through 2026 looks like this. Sessions down. Pages per session down. Time on page down hard. Enquiries flat or slightly up. Conversion rate up sharply, because the denominator shrank.

If you judge yourself on sessions, that graph looks like a business in decline. If you judge yourself on booked jobs, nothing bad happened at all. Plenty of owners are about to fire an agency over the first graph without ever looking at the second.

Where the pain is real

It is not all comfortable. Three real losses.

You lose the browse. Somebody used to land on a blog post, poke around, see a case study, and remember your name. The assistant strips that to a single answer with no wandering attached.

You lose the brand impression. Your design, your photos, your tone of voice, the reason you sound different from the other four. An agent reading your HTML gets none of that.

You lose top-of-funnel reach, which means the businesses that only ever competed on informational content are genuinely exposed. If your entire strategy was ranking for questions, that strategy is thinning out. I wrote about the related shift on the buying side in agentic commerce and AI checkout, and the same logic applies to services.

AI Browsers: Business Traffic Your Analytics Cannot See

This is the section I would read twice if I were you, because the measurement problem arrives before the traffic problem and it makes people panic about the wrong thing.

Why agent visits go missing

Google Analytics 4 runs on JavaScript. A lot of agent activity either does not execute that JavaScript, executes it and gets filtered as non-human, or arrives with no referrer at all and lands in the direct bucket alongside people who typed your URL. Some agents identify themselves properly in the user agent string. Some identify themselves as Chrome, because they are Chrome.

So you end up with three flavours of invisible. Visits that never register. Visits that register as direct. Visits that register as a browser on a Mac in the United States, when the customer is in Hamilton.

What to look at instead

Stop treating sessions as the health metric. Start with outcomes, then work back.

Phone calls, tracked properly, with the source recorded. Form submissions by form, not just in total. Booking requests. Branded search volume in Search Console, which tells you whether people are being told your name somewhere you cannot see. And the oldest tool in the box: ask every single enquiry how they heard about you, and record the answer in your CRM rather than in somebody’s head.

Then go one level deeper and read your server logs, or your Cloudflare bot analytics if you sit behind it. That is where agent traffic is actually visible. You will see request patterns that no human makes: eleven pages in four seconds, no images loaded, straight to the pricing page and out.

The attribution answer nobody likes

Multi-touch attribution was already wobbly. With agents in the mix, it is finished for small businesses. The honest replacement is a self-reported source field on every enquiry, matched against total spend and total revenue at the month level. Less precise, considerably more true.

Most owners I meet cannot answer “where did last month’s twelve best leads come from” without opening four systems and guessing. That is not an AI browser problem. That is a data problem you had already, now exposed by a new variable. Getting every number into one place is the fix, and it is worth doing whether or not a single agent ever visits your site.

Five Things That Break When an Agent Visits Your Website

Here is the practical list. I have run these checks on client sites for the past few months and the hit rate is depressingly consistent. Most sites fail at least two.

1. Content that only exists after JavaScript runs

If your prices, service areas, opening hours or phone number are injected by a script, a framework, or a third-party widget after page load, assume a meaningful share of agents never see them. Everything that matters commercially should exist in the HTML that arrives with the first response. This is the single most common failure and usually the cheapest to fix.

2. Booking and contact behind a widget

Calendars embedded in iframes, chat bubbles that are the only path to a human, forms that require a drag-to-confirm slider or a multi-step wizard. Humans cope. Agents do not. Keep one plain contact form on a plain URL, with normal labelled fields, that works without a mouse. Think of it as an accessibility requirement, because it is literally the same requirement.

3. No structured data

Schema markup stopped being an SEO nicety and became the machine-readable version of your business. LocalBusiness with your address, phone and opening hours. Service for each thing you sell. FAQPage for your questions. Review and AggregateRating if you have them honestly. This is what an assistant reads when it decides which three businesses to shortlist. Missing it does not penalise you exactly. It just makes you harder to pick.

4. Key information locked in PDFs and images

Price lists as PDFs. Service areas as a picture of a map. Phone numbers as graphics to stop scrapers. Every one of those hides the information from the thing now doing the choosing. If a fact would help a customer decide, it belongs in text on a page.

5. Nothing that answers the comparison question

Agents are asked comparison questions constantly. “Who is cheapest.” “Who can come tomorrow?” “Who does commercial as well as residential.” If your site never states your response time, your service area, your minimum job size, or who you are not for, the assistant has to guess, and guessing usually means leaving you out.

The fix list in order

Get the facts into plain HTML. Add proper structured data. Put a simple form on a simple URL. Write one page that states plainly what you do, where, for whom, how fast, and at what rough level of investment. Then check your robots rules so you are not accidentally blocking the assistants you want reading you while leaving the training crawlers you do not want wide open. Those two decisions are separate, and most sites have them backwards.

None of that is exotic. It is the same discipline that makes a site good for people who use screen readers and people on bad rural connections. The agent just made it urgent.

The Advantage Is Not the Website, It Is What Happens Next

Now the part that actually decides whether this hurts you or helps you.

When an agent shortlists three suppliers and sends the same enquiry to all three at 9:40 pm on a Tuesday, the website stops being the competition. The response is the competition.

Enquiries arrive pre-researched and in parallel

The customer who reaches you through an AI browser has already had your pricing page summarised, your reviews read, and your competitors compared. They are further along than a normal enquiry and considerably less patient. They also sent the same message to two other businesses about ninety seconds ago.

The old rule holds and gets sharper: roughly 78% of deals go to whoever responds first. When three enquiries land simultaneously, first response is not an advantage; it is the entire contest. Most businesses take over four hours. A lot take until the next morning, which is the same as never.

Which is a capacity problem, not a marketing problem

This is where I have to be blunt with people. You cannot fix a 9:40 pm enquiry with better SEO. You fix it by having something that answers at 9:40 pm.

That means an AI receptionist that picks up the phone when nobody is there, an automated first response that goes out in under ninety seconds with a real question in it rather than “thanks, we’ll be in touch”, and a booking link that works without a human in the middle. One dental client, Dr Claire, had two receptionists and was still missing close to half her calls at peak times. After putting an AI receptionist on the overflow and after-hours, missed calls went to zero and booked appointments rose 44%. Nothing about her marketing changed. The capacity to answer changed.

I have written more on the mechanics of that in after-hours call answering and AI appointment booking, because the two together cover most of what an agent-driven enquiry needs on the other end.

The compounding version

The businesses that come out of this ahead are not the ones with the best-optimised pages. They are the ones where the business itself can think and act without the owner sitting at a desk.

That means the business knows its own facts well enough to publish them clearly, which is what the agent reads. It means someone or something answers every enquiry immediately, which is what wins the job. And it means the owner is not personally the bottleneck between an enquiry arriving and a person getting a reply, which is what makes the first two sustainable.

An AI brain that holds your context and an AI workforce that does the responding is not a separate project from all this. It is the same project. The website changes are the visible half. The ability to respond is the half that gets paid. If you are earlier in that thinking, AI agents for small business and AI adoption for SMEs are the sensible starting points.

What I Would Actually Do This Month

Do not rebuild your website. Do not panic about AI browsers business traffic or sessions. Do this instead.

Check whether your prices, hours, service areas, and phone number exist in plain text on the page. Add proper structured data if you have none. Put a simple contact form on a simple URL. Write one comparison page that answers the questions an agent gets asked. Then turn your attention to the response side, because that is where the money is: measure how long it currently takes you to reply to a new enquiry, at 10am and at 10pm, and be honest about the second number.

AI browsers do not remove the need for a website. They change who reads it and what they do with it. The reading is now done by software that wants facts, and the deciding is now done in minutes rather than days. A business built out of one person’s head cannot move at that speed. A business with a brain of its own can.

Book a 30-Minute Discovery Call

If you want a straight answer on where your business is exposed to this, book a 30-minute discovery call. No deck, no pitch. I will ask what happens when an enquiry lands at 9 pm, look at how you currently know where leads come from, and tell you which of the fixes above is worth your time first. If the answer is “nothing yet, you are fine”, I will say that too.

Frequently Asked Questions

About Octavius