Say a customer who lives in Xinzhuang, New Taipei, wants someone to come and clean the air conditioners at home this Saturday. If they ask ChatGPT, what they get back is a list of businesses; if they hand the job to an AI agent, an AI assistant that goes online to look things up, compare prices and place orders for people, what they get back is a booking that’s already made.

The difference is in the middle. With ChatGPT, the calls, the questions about price and the decision on who to go with are still up to the customer. With an agent, the customer only has to say, “Find someone who can come to Xinzhuang on Saturday afternoon and clean two split-type air conditioners, book them, then let me know,” and the agent does the looking, comparing and booking itself, then reports back with who’s coming and when.

For you, as the business owner, who does the choosing in that middle step decides which hurdle you have to clear. When customers choose for themselves, a mention from AI can be enough for them to call and ask; when an agent chooses for them, being mentioned isn’t enough any more. As we see it, an agent first has to be sure you meet the customer’s conditions before it will pick you.

Ask AI and the customer gets a list; hand the job to an agent and they get a result

When a customer asks AI, its job ends at the answer. It reads what’s online about these businesses and writes a summary; after reading it, which ones to compare, who to call and who to book are all still up to the customer. There’s still a fair amount of room for you along the way: even if the AI describes your business in one line, say that you “offer home cleaning”, an interested customer will still call to ask whether you have a slot on Saturday, and on the phone you can win the job.

When a customer hands the job to an agent, the choosing and the booking go with it. Following the customer’s request, the agent opens the businesses’ websites one by one and reads them, picks one, and books that business for the customer if it can, at most asking them to give the nod at the last step. What the customer ends up seeing is the business the agent chose; the customer usually won’t know which others it looked at or why it passed them over.

This kind of agent isn’t something from the future. OpenAI, the company behind ChatGPT, has already put one into ChatGPT, called ChatGPT agent, available to users on paid plans. According to OpenAI, it opens a browser on its own, clicks buttons, fills in forms and moves from one website to the next to get things done. When a site needs a login, or before a high-impact action, it stops and asks the user to take over or give the go-ahead.

Booking, asking prices, paying: agents are already doing it

Google is doing it too. Google Search has an AI Mode where you can keep asking questions as if in a chat, and it can now find restaurant tables for people. A customer tells it the party size, date, time, area and the kind of food they want, and it checks several booking platforms and restaurant websites on its own and finds the options that can still be booked; only the very last step, actually making the booking, is left to the customer, who follows a link to finish it.

This booking feature started in the US and expanded to more places in April 2026. According to tech press reports, Hong Kong and Singapore are among them; Taiwan wasn’t on the list.

In the US, Google does something even more direct: it phones businesses for customers. Since 2025, a customer in the US who searches for pet groomers nearby can answer a few questions, such as what kind of pet they have, which services they need and when, and Google will call around to ask about prices and availability, then pull the results together. Every call opens by saying it’s an automated system from Google, calling on behalf of a user.

In Taiwan, the examples we can find are ones companies have built on their own platforms. The restaurant booking platform inline launched an AI assistant on LINE in 2025: a customer types something like “Any suggestions for dinner for two on Friday night?” and it recommends partner restaurants that still have tables, and can book one directly. Taiwan Mobile, for its part, announced a partnership with Visa in June 2026 to let holders of its co-branded credit card buy things and pay through an AI agent at its online store.

The air-conditioner example at the top is still only hypothetical. But finding open slots, asking prices, booking and paying are, one by one, already being done by agents; going by that trend, our guess is that home services just haven’t had their turn yet.

From answering to doing, customers do less and less themselves

Looking back makes the change clearer. Google results used to be a column of links: customers clicked and read for themselves, and your web page did the talking. Then Google put an AI-written summary at the top of the results, called an AI Overview, and what customers read became the few lines the AI wrote about you. With agents, customers don’t even need to read those lines; they just wait for the result.

With each step, customers do a little less themselves, and you have a little less chance to win them over with your web pages or on the phone. A list can have five names on it; only one gets booked.

An agent picks the business it can be sure of, and you can’t see where it gets stuck

So how does an agent decide which business to book? ChatGPT and Google haven’t published any rules, so this section is our own view, worked out from how agents operate: an agent will only pick a business it can be sure meets the customer’s conditions.

An agent is working for someone else, and at the end it has to be able to tell them: “This one meets your conditions.” The conditions are the customer’s. The customer at the top packed several into one sentence: Saturday afternoon, Xinzhuang, two split units. The agent has to find something online showing that you match each of them before it can say that line.

When customers are shopping around themselves, a website that’s a bit vague can still get by. Say your site reads “Serving parts of Taipei and New Taipei”: a customer who isn’t sure whether Xinzhuang is included will call to ask. If it says “We quote once we’ve seen the job”, they know to ask about that too.

An agent reading those same two lines can’t be sure whether Xinzhuang is covered or roughly what it’ll cost. It could stop and go back to the person it’s working for; but if another business nearby spells all of this out, the agent’s easiest choice is that one.

Vagueness isn’t all that leaves an agent unsure; so do details that differ from one place to another. Your website isn’t the only source of information about you: there’s also your Google Business Profile, the listing that shows your business details on Google Maps and in Search. Suppose your website says you’re open on Saturdays but the profile says you’re closed: faced with two versions like that, an agent will usually just move on to someone else.

So being mentioned by AI and being a business an agent can be sure of are two different things. When AI answers a question, one line on your site saying you “offer home air-conditioner cleaning” is enough to put you on the list; when an agent is trying to get a job done for a customer, that line isn’t enough for it to decide.

Asking AI and handing jobs to agents will probably run side by side for a long while. The good news is that the things an agent needs to be sure of are just as useful to AI when it answers questions. Say a customer asks ChatGPT today, “Who can come to Xinzhuang on Saturday to clean air conditioners?” If the AI can read that you’re open on Saturday and cover Xinzhuang, it has a chance of giving those details when it describes you.

Writing things clearly doesn’t mean fixing everything in stone, though. You don’t have to publish every quote; how much detail to give on pricing, and what to leave for the phone, is a business decision for you. The catch is that it’s hard to know whether that decision is costing you work, because you can’t see where an agent gets stuck.

Whatever conditions a customer gives an agent never reach you. The customer at the top wanted Saturday afternoon, Xinzhuang and two split units; the next might live in a top-floor flat, and the one after that might want someone first thing tomorrow. Every request is different, and a website that answers one customer’s may not answer the next.

Nor can you see who an agent is comparing you with. As we said above, if just one business nearby spells things out more clearly than you do, the agent will pick it; but which businesses get lined up against each other, and what each of them says, is something only the agent knows.

When an agent skips you, nobody tells you what was missing

A business that an agent passes over won’t find out afterwards either. When a customer calls, even if nothing comes of it, you’ve at least taken the call and know someone was after a Saturday slot. When an agent skips you, there’s no call, no message and no word on why it moved on. This, too, is worked out rather than something any platform has said; our reasoning is that an agent working for a customer has no need to go back and tell a business it didn’t choose.

Losing those calls also means losing a way to learn. When customers used to call to ask whether you cover Xinzhuang, enough of those calls told you what customers care about and what your website was missing; once customers hand the job to agents, those calls stop coming, and the missing line stays missing.

The calls may stop, but AI keeps reading the information about you, such as your website, your business profile and what others write about you online; the agents to come will read the same things. That information is what GEO (generative engine optimisation) deals with. Our managed GEO service starts right there: our team fixes your site’s content, brings what’s said about you elsewhere into line, and regularly checks how each AI engine describes you; you just review the results.

For the regular checks, what we put to major AI engines such as ChatGPT, Gemini and Perplexity are the things customers used to ask on the phone. For air-conditioner cleaning, that means questions like whether you have a Saturday afternoon slot, whether you can do a top-floor flat, and whether taking the unit down and putting it back costs extra. What we look at is who they recommend and whether you’re among them; and, when they mention you, whether they can state your availability, service area and pricing, and whether what they say matches how you work now.

When an agent picks a business for a customer some day, you won’t be in the room. But whether AI can already speak to the things an agent needs to be sure of is something you can hear for yourself in a demo: a session with one of our consultants where they show you live. The consultant picks a few of those questions and puts them to ChatGPT and Gemini, then looks at which things they can say about you when they mention you, and which they can’t. Going by our reasoning above, the things they can’t say are also the likeliest to make an agent turn to someone else. Beyond how AI talks about you, you’ll also see how your website scores on the GEO audit. The GEO audit is geoweb.tw’s free website audit, which checks how well AI can understand your site and scores it across 12 areas.

Further reading