Search in the Agentic Age, Part 2: What Happens Before AI Names a Business
Someone saw a ghost in the back room and asked AI who to call. It wasn't Ghostbusters.
In seconds, the AI weighed that question against what it found about you and your competitors. If your website didn't have the answer, it delivered that customer to your competitor, and you'll never know. Here's what happens behind the scenes.
Why this matters more than it used to
For twenty years, search gave people a list of ten links. If you were on the first page, you had a chance. AI tools work differently. They usually give one answer, sometimes with a short list of businesses. If you aren't named, the customer doesn't scroll down to find you. There's nothing to scroll.
What the AI already knows
The large language models behind today's AI are trained on data: books, the internet, social media and more. That training gives the system context and basic knowledge. It learns what a furnace is and what an HVAC company does.
That knowledge has a cutoff. It tells the model what things are, not what is true right now. It doesn't know which companies are open this week, which ones have new reviews, or which one just added furnace repair to its services. So when you ask who can fix a furnace that won't start, the model pivots. It uses what it learned in training, extracts the meaning from your question, and searches the internet based on that meaning.
How it searches
Here's the part most people never see. The model can break your question into parts and search each one separately. One part might be the service, another the location, another the quality signal. A single question like "Who can fix a furnace that won't start near me?" might become several searches: furnace repair companies in your area, emergency heating service, and reviews of local HVAC companies.
Behind those searches, many systems likely turn the text of your question and of web pages into numbers called vectors. Each vector captures meaning, so the search can find similarly matched meanings, not just matching keywords.
Here's what that looks like. "Furnace repair" and "my heater stopped working" share no words at all. A keyword search treats them as unrelated. A search built on vectors sees that they mean almost the same thing. The same goes for "no heat in the house" and "heating system service," or "AC blowing warm air" and "air conditioning repair."
This cuts both ways. It means a customer doesn't have to guess your exact wording to find you. But it also means the AI is matching the meaning of their problem against the meaning of your website. If your site only says "HVAC services" and never describes the problems you solve, there isn't much meaning there to match.
Specific questions raise the stakes. Say a parent asks AI who performs tonsil surgery for children. One practice's website says it performs tonsil surgery. Another says it provides extensive pediatric care, including tonsil surgery for children. Both do the procedure, but only one answers the question the parent actually asked. The second practice is the closer match, and the first may be left out.
Your customers rarely ask general questions. They ask about their child, their furnace, their situation. Every specific service you offer but never describe is a question your website can't answer.
What it checks
Then it curates what comes back. It uses its training to understand that it needs to look at reviews, know what type of work you're interested in, and find companies matching those services and that location. It checks many sources, like review sites and the Better Business Bureau, before it answers.
In practice, that means it's looking for:
- Reviews and ratings, across more than one site. A strong rating on one site and nothing anywhere else gives the AI less to go on than steady reviews in several places.
- Clear descriptions of the services you offer. "We repair and replace furnaces, boilers and heat pumps" tells the AI far more than "full-service HVAC."
- Where you work and which areas you serve. If the customer's town isn't mentioned anywhere, the AI has to guess whether you cover it.
- Whether your website and your listings tell the same story. If your website lists one phone number and your Google profile lists another, or one says you install heat pumps and the other doesn't, the AI has a reason to doubt both.
None of this is new to good businesses. What's new is that a machine is reading it all at once, comparing you with everyone else, and making the call in seconds.
From answers to action
The line between chat and agent is blurring fast. A chatbot retrieves and answers. An agent retrieves, acts, and verifies. It can browse your site, check a booking system, compare your listing against your website, and give a recommendation. Chat finds answers. Agents finish jobs.
Picture the furnace question handed to an agent. It searches for local companies, reads their websites and reviews, and checks which ones offer emergency service. Then it might look for open appointments and come back with a recommendation, or even a booked visit. Every step depends on what it can find and read about you.
I saw this firsthand. In the Grok chat window, it decided on its own that my question needed deeper research. It handed the work to a separate Grok Bot I had set up, and the bot researched in the background. Then Grok brought back what it found.
Since August, xAI, Meta and OpenAI have launched these agents: Grok Bot, Muse and dots. They can research the question, who can answer it, and the competitors, and they can use the tools you connect to them.
As these systems evolve, it will increasingly be bots talking to bots. An agent working for a customer may read your website, or even connect to your booking system. The more complete your information is, the more effectively these systems can connect customers with your services and products.
Most business websites aren't ready for this. Keywords and backlinks alone are no longer enough.
Try this today
You can see what the AI sees in about ten minutes:
- Open ChatGPT, Gemini, Claude or Grok. Try at least two, since they don't always agree.
- Ask the question your customers would ask, in their words, not yours. For example: "Who can fix a furnace that won't start in [your town]?" Swap in your own service.
- Ask a specific version, the kind a real customer asks: "Who repairs heat pumps for older homes?" or "Who does emergency furnace repair on weekends?" Specific questions are where gaps show up first.
- Ask it the way a worried customer might: "My heat went out tonight. Who should I call?"
- Note what you find. Did it name you? What did it say about you, and was it accurate? If it named someone else, what does their website say that yours doesn't?
That last question is the one that matters. It's the gap between what your customers are asking and what your website is answering.
So go back to that customer's question. Can your website answer it?
What changes will you make to your website as search enters this new phase? Tell me in the comments.
This is part 2 of a 5-part series. Read part 1.
Sources
- Google, AI Mode and query fan-out: blog.google
- xAI, Grok Bot launch, August 11, 2026: x.ai
- Meta, Muse launch, September 8, 2026, via Reuters/Fox Business: foxbusiness.com
- OpenAI, dots launch at DevDay, September 29, 2026, via TechCrunch: techcrunch.com