Hello Lunch Break. Chase cut the ribbon on its first branch in our market last Wednesday, out on Highway 17 in Hampstead. Inside you’ll find a business banker, a private client banker, a home lending advisor, a wealth management advisor. Next year they'll open in the old Wild Wing Cafe at Landfall, then Calabash, then Ogden.

This is the biggest bank in the country, the one whose app is already on half the phones in town, spending money placing humans inside local buildings. Their market director told WilmingtonBiz the branches are there for advice, education and help. Let’s be real. Nobody drives to Hampstead to deposit a check. You drive there when something's wrong, or when you're about to sign something big and you want a person across the desk.

That's what I’m talking about this week. People moved the searching part of buying to AI. They didn't move the trusting part.

Alright. Clock's ticking. Back to work.

THE GAME CHANGED

When somebody types one broad question into Google's AI search, the system splits it into a set of smaller searches and builds the answer from what those turn up. That's query fan out. 

On September 15, Google Research published Retrieve for Train, a small model trained to copy a much bigger model's fan outs. Google's own example: someone searching "camping gear" doesn't want ten versions of the same tent. They want the tent, the sleeping bag, the stove and the headlamp, so the model is trained to make every sub question cover a different angle instead of repeating synonyms.

For you, this means the customer who types "roofer Wilmington NC" gets an answer built from the searches they never typed: what it costs, whether you work insurance claims, how fast you come out, what happens if it leaks next spring. A page that only answers the headline question gets pulled for one of those searches out of several.

What to do about it: open your main service page and make sure it answers four things: the price range, who it's for, how long it takes, and what happens if something goes wrong. One page, four answers. Don't spin up a separate page for every sub question.

Sources: Bypassing inference bottlenecks: Accelerating complex AI search with Retrieve for Train, Google Research, September 15, 2026. Google Announces New Query Fan Out Framework: R4T Diffusion, Roger Montti, Search Engine Journal, September 24, 2026.

THE WINDOW IS NOW

PYMNTS Intelligence surveyed 2,191 US consumers in August and found 23 percent now start their retail shopping research with AI, about 61 million people. Then the money went where it always goes.

PYMNTS told shoppers an AI assistant would find the identical item at the identical price, with identical shipping and identical returns, from a marketplace, a department store or the brand's own site. 40 percent still picked the marketplace. 22 percent picked the department store, 16 percent the brand, and the rest let the AI choose. The top reason was easier or more reliable returns, at 37 percent, from people who had just been told the returns were the same. PYMNTS CEO Karen Webster's read is that trust goes to whoever owns the bad day.

Why the window is open right now: AI moved where the search starts. It hasn't moved the trusting, and in your category trust is still up for grabs. You can't beat Amazon on shipping. You can be the business that answers the phone, comes back out to fix it, and puts that promise in writing where an AI can read it and repeat it: your guarantee, your warranty, how fast you call back, who handles problems. Most local sites bury that in a footer or leave it out entirely.

Sources: Agentic Commerce Needs Its Febreze Moment, Karen Webster, PYMNTS, September 23, 2026. The 50 Million Consumer Migration, PYMNTS Intelligence, August 2026.

YOUR COMPETITORS ARE ASLEEP

Later, the creator marketing platform, published a guide this month arguing AI answers lean on other people talking about a business far more than on the business talking about itself. From its own campaigns, it reports one in three AI answers cite creator content, and 57 percent of the YouTube videos AI cites come from channels under 100,000 followers.

What gets cited: Follower count isn't what gets a video cited. Covering the exact topic, making it machine readable, and several voices saying the same thing are. One executive in the guide describes a dermatologist’s brand had never worked with who kept showing up in AI answers and shaping how the brand got described.

What to do about it: find people nearby who already make videos about your category. The Wilmington real estate channel, the inspector who films walkthroughs, or Fisherman's Post, whose weekly Carolinas fishing reports reach about 5,300 YouTube subscribers. Give them something worth filming. Put it on YouTube with a full transcript and a title that says the exact question it answers.

Source: Marketer's Guide: How to Win AI Search with Creators, Later, 2026.

WHAT’S THE PROMPT

Kevin Indig and Ramp's George Bonaci made the case in Growth Memo this week that click attribution was built to measure paid media and has never been good at anything else. Their fix is triangulation: watch three signals with different blind spots, how often AI names you, what customers say brought them in, and what they buy. When all three move together, trust it. When they split, that's what you study.

This prompt handles the first signal and writes the second for you. Paste the "how did you hear about us" question it gives you into your contact form and intake paperwork this week. The third signal you already have: your sales.

Run the prompt

After you run the prompt, read the answer to question four. If ChatGPT couldn't find your guarantee, your customer can't either.

Source: The usefulness of attribution for AI Search, Kevin Indig with George Bonaci, Growth Memo, September 28, 2026.

CITED

I asked ChatGPT for the best roofer in Wilmington and got six names, led by Best Choice Roofing at 4.9 stars from 391 reviews.

Then I asked the way somebody asks when they care about trust.

Best Choice disappeared. Eight roofers came back, sorted by what their own sites promise in writing, and ChatGPT picked one to call first: Howard Roofing & Construction, which hadn't shown up in the first answer at all.

Here's why. Howard's roof repair page says its workmanship warranty covers the labor on every repair, and it answers "Do you offer a warranty on repair work?" as its own question. Another repair page promises that if the repaired spot leaks again because of their work, they'll come back and make it right at no charge. ChatGPT cited Howard's own site, not a directory.

That's a roofer who confirmed what you really want to know in the words of a skeptical homeowner.

Want to know what your website should be saying? Request the AI Visibility Audit and let us help you uncover your website gaps.

See you next Thursday.

- Anna

THE POLL