A facilities manager in Cleveland has a chiller that keeps tripping offline. It is nine at night. She does not open Google and start clicking through ten blue links. She opens ChatGPT and types something like "best commercial HVAC contractors in Cleveland for a building over 200,000 square feet."
Something answers her. Three or four companies get named. Maybe a couple of links sit underneath.
That was the first round of a deal, and you were not in the room for it. You did not get an impression, a click, or a form fill. Nothing showed up in your analytics. But a shortlist was created, and either your name was on it or it was not.
This is the part that owners underestimate. We are not talking about a new traffic source. We are talking about who gets considered at all.
The Shortlist Gets Built Before You Know There Is a Deal
None of this is really new. B2B buyers have been doing their homework privately for years, which is something we have written about before. The research happens in the dark, the committee forms without you, and by the time somebody fills in your contact form they have already made up most of their mind.
What has changed is who narrates that research.
Two years ago your buyer typed a question into Google and got a page of results. She did the summarizing herself. She read your homepage, in your words, with your framing, and formed her own opinion. Your message reached her more or less intact.
Now something reads the internet on her behalf and hands her a summary. That summary was assembled from whatever the model could find about you, most of which you did not write. It might be current. It might be three years out of date. It might quietly leave you out because it never found enough about you to feel confident naming you.
One of the marketers we listened to while researching this piece described it as playing telephone in the school yard. The message that arrives at the end is not the message you sent. Except in this version, the game happens thousands of times a week and the outcome is which vendors get on the list.
For a business selling five-figure and six-figure work, this matters more than the raw traffic numbers suggest. Dreamdata's 2026 benchmarks report, built on more than 3.5 million completed customer journeys, puts the average B2B buying journey at 272 days and the typical deal at ten stakeholders. Nine months, ten people. And buyers spend the first 220 of those days educating themselves before they ever enter a sales pipeline. A shortlist formed early in that stretch gets carried through all of it.
Ranking Gets You Found. Getting Named Gets You Shortlisted.
Here is the shift in plain terms, and it is worth being precise about because most advice on this topic is not.
Traditional search is a competition for a click. You rank, somebody clicks, they land on your page, you get your shot. Everything you know about SEO is built to win that click.
AI search is a competition for a mention. The answer gets delivered inside the chat. Frequently nobody clicks anything at all. So the thing you are competing for is not a visit, it is whether your business name appears in the sentence.
There are two distinct ways to show up, and they are not the same prize:
- Cited. Your web address turns up in the little source list under the answer. Useful, and it can send you a visitor.
- Mentioned. Your business name appears in the body of the answer itself. This is the better outcome by a distance, because being named is what puts you on the list. A citation without a mention is a footnote nobody reads.
Getting both is the goal. If you have to pick, take the mention.
Now the correction, because there is a lot of panic being sold on this subject. AI search is not replacing Google. Ahrefs tracked referral traffic across 76,000 websites and found that for every visitor ChatGPT sends, Google sends about 190. The volume is not close.
But look at the same study from the other side, because this is the whole point of the article in one statistic. ChatGPT accounts for something like 12% of Google's total query volume while producing barely 0.2% of website traffic. Enormous share of the questions. Almost none of the clicks. People are getting their answer and never leaving.
What is actually happening is more interesting, and it changes what you should do. Buyers start in an LLM to orient themselves, because it is a fast way to ask the questions you feel silly asking out loud. Who does this kind of work? What should I be comparing? What is a fair price? Then, once they have names, they go to Google and to your website to check whether any of it is true.
AI search is not eating traditional search. It is feeding it. Which means the answer is never to abandon one for the other. It is to make sure the first conversation includes you, and that the verification that follows holds up.
Before You Change a Thing, Go Ask It About Yourself
Almost every article on this subject jumps straight to tactics. Skip that. The first move costs nothing and takes an afternoon, and without it you are optimizing blind.
Go find out what the machines currently say about you.
Write ten questions your buyer would actually ask. Not what you wish they searched for. What a person with your customer's problem types at nine at night. Split them into two groups.
Discovery questions, with no brand names in them, to test whether you surface at all:
- "Best commercial cleaning companies in Tampa for medical offices"
- "Who supplies food-grade stainless fittings in the Midwest"
- "What should I look for in an industrial electrical contractor"
Brand questions, with your name in them, to test what it says and whether it is accurate:
- "Is [your company] any good"
- "[Your company] versus [your main competitor]"
- "What do customers say about [your company]"
Weight these toward the middle and bottom of the funnel. Broad educational questions might get your blog cited, but it is the "who should I hire for this" questions where a name gets said, and the name is the thing you want.
Ask each one across the four platforms that matter: ChatGPT, Google's AI Overviews, Perplexity, and Claude. Between them you cover the overwhelming majority of this behavior.
Turn off the personalization. Use temporary or incognito mode in each one, and log out. Otherwise the model is partly answering based on your own history, and you will get a flattering result that no buyer will ever see. This step gets skipped constantly and it invalidates everything after it.
Put it in a spreadsheet. Queries down the side, platforms across the top, paste the full answer into each cell. Where you can see the model's sources, hover over them and note what it pulled from.
Then read your own spreadsheet for four things:
- Do you appear at all? Count your mentions across the unbranded questions only. The branded ones will obviously mention you.
- Is it accurate? This is the one that shocks people. Wrong service area, a service you dropped two years ago, an old phone number, a claim you never made.
- What is the tone? Positive, neutral, or quietly damning by omission.
- Which sources keep coming back? This is the most valuable column on the sheet. The sites the models quote over and over in your category are your actual target list. Everything in the next section flows from it.
Re-run the whole thing in a month or two. It is the only way to know if anything you did worked.
Most of What AI Knows About You Isn't on Your Website
Here is the finding that reorganizes the whole problem, and it comes from a company with no reason to flatter itself.
Ahrefs audited where their own brand mentions in Google's AI Overviews were coming from. The top ten pages driving mentions of Ahrefs did not include ahrefs.com. Their own site appeared as a secondary source, down at position twelve. On ChatGPT they were not mentioned at all in that test.
Ahrefs is a large, well-known company with a genuinely excellent website. If their own site is not the main thing feeding AI answers about them, yours is not either.
Their larger study across 75,000 brands put a number on it. Plain brand mentions across the web correlated with AI visibility more than three times as strongly as backlinks did. That is a real departure from classic SEO. The model does not need a link. It needs to have seen your name discussed, repeatedly, in the same breath as the thing you do.
So the work moves off your website. For a small or mid-sized B2B business, the practical list looks like this:
- Industry and trade association directories. The ones for your sector that nobody thinks about. Models sample structured databases heavily, and these are often thin enough that a complete listing stands out.
- Your Google Business Profile, filled in properly. Categories, service area, services, hours, description. It is a structured, trusted, machine-readable statement of what you do and where.
- Review platforms your category actually uses. Not just Google. The vertical ones your buyers already read.
- The "top 10 [category] companies in [place]" lists. These get cited constantly. Many were written by your competitors, who put themselves at number one and gave you a lukewarm paragraph. Getting into the ones written by neutral publications is high-value and unglamorous work: find them in your source column, then pitch your inclusion.
- Trade press and local business press. A single article in a publication the models trust does more than another blog post on your own site.
- Reddit and YouTube. Ahrefs' research puts both of them in the top three domains that AI answers quote most often. Be genuinely useful in the subreddit where your buyers hang out. Do not go in and advertise, which does not work and will get you removed. A demo video with a clear title and a full description is a surprisingly durable asset here.
None of this is exotic. It is the unglamorous credibility work that a business with a decent reputation and no marketing department has usually never got around to.
Your Site Still Has to Survive the Click
Your own website is not the main source, but it is the verification step, and this is where deals get lost quietly. The buyer has your name from an AI answer. She clicks through. She is now checking whether the summary was true.
Make your site easy for both her and the machine to read:
Answer the question in the first two lines. Put the conclusion at the top of the section, then explain. If a page needs three paragraphs of throat-clearing before it says what you do, a model has nothing clean to lift and a human has already scrolled past.
Write the boring specifics down. Where you work. Who you are for. What you do not do. Rough price band. The size of job you take. Businesses hide all of this and then wonder why they get unqualified enquiries and never appear in "best X for Y" answers. That specificity is exactly what makes a model confident enough to name you for a narrow question.
Use real FAQs in your buyer's words. "Do you work with medical facilities?" answered plainly with "Yes" and a sentence is a quotable unit. This is not schema trickery, it is just writing down the questions your phone already asks you.
Put comparison information in a table. Structured content gets parsed cleanly and quoted.
Keep the facts in text, not locked inside images. If your service list only exists inside a graphic or loads through JavaScript, the model may never read it. Alt text and video transcripts matter for the same reason.
Check your robots.txt is not blocking the AI crawlers. Some sites block them by default and nobody notices. OpenAI's search crawler is called OAI-SearchBot, and if you are blocking it you cannot appear in ChatGPT's results no matter how good everything else is. Worth five minutes with whoever runs your site. While you are there, make sure you are indexed in Bing, which matters again because ChatGPT's search still leans on Bing's index.
Keep it fresh. Ahrefs analyzed 17 million citations across seven AI platforms and found cited pages run about 26% newer than what ranks in classic organic search. Roughly half of the pages being cited had been published or updated within the previous 13 weeks. A page you refreshed this quarter genuinely beats one last touched in 2023.
There is not much room at the top, which is why this is worth doing properly. Surfer studied more than 405,000 searches and found an AI Overview cites about five sources on average. Five slots. One of them can be yours more than once if you cover a subject thoroughly enough.
AI Repeats What Your Customers Said About You
Every model pulls heavily from reviews and third-party accounts of your work, which produces a slightly uncomfortable conclusion: the most influential marketing copy about your business was written by your customers, and you did not brief it.
Remember who is reading the answer. That committee of roughly ten people mostly never speaks to you. Several of them will never visit your website. Their entire impression of you is the summary plus whatever validation they can find in ninety seconds. When one of them asks the group "has anyone actually used these people," the answer needs to exist somewhere public.
Which makes a boring review program one of the best-value things on this list. Ask every satisfied customer, at the moment the job goes well rather than six months later. Make it one link. Spread them across the platforms your category reads instead of piling them all on Google.
Case studies do the same job in longer form, and they are more useful here than they used to be, because a detailed write-up naming an industry, a problem and an outcome gives a model something specific to match against a narrow question. "We do great work" matches nothing. "We replaced the chiller in a 240,000 square foot medical office building over a weekend without closing the building" matches a very particular search.
Don't Bet the Budget on This. Do the Cheap Parts Now.
Now the honest part, because there is a growing industry selling panic about this and charging retainers for it.
The traffic is still small. Google sends vastly more of it, and will for some time. If someone tells you to redirect your marketing budget into AI visibility this quarter, they are selling something.
But three things make the cheap version of this work worth doing today.
It converts unusually well. Ahrefs published their own numbers on this: AI search was half a percent of their traffic and produced 12.1% of their signups, a conversion rate roughly 23 times higher than their traditional organic traffic. That is one company's data and your mileage will differ, but the direction makes sense and agencies watching their own analytics report the same shape. Someone arriving from an AI answer was pre-qualified before they clicked. Something they trusted already told them you were worth a look, which is a warmer introduction than any ad you can buy.
It is growing, and the compounding is slow. Third-party mentions, reviews and directory presence take months to accumulate and to be re-crawled. Starting when it becomes urgent is starting two quarters late.
And almost none of it is wasted if you are wrong. Look back at the actual list. Clearer pages. Specific service information. Real FAQs. A working review program. Getting listed in your industry's directories. Being present in the trade press. Every one of those helps your conventional SEO, your sales calls, and your close rate, whether or not a single buyer ever asks a chatbot about you. That is what makes this a safe bet rather than a gamble.
So do it in that order. Run the audit this month, because it is free and it will probably surprise you. Fix the factual errors you find, because those are actively costing you. Then work down the source column from your spreadsheet, one entry at a time.
What you are really doing is old-fashioned. You are making it easy for anyone, human or otherwise, to find out what your business does, who it does it for, and whether the people who hired you were glad they did. That has always been the job. The audience just got a new member.
We spend our days helping small and mid-sized B2B companies get found and get chosen, and this is one of those areas where a couple of afternoons of unglamorous work goes a long way. If you would rather have someone handle it, that is what we do. If you would rather run the audit yourself this week, do that instead. You will learn something either way.
