Hello Lunch Break. Met a friend out this weekend and the conversation shifted to AI. Not hard to do if you’re around me but the questions were the same ones I had in a Wilmington Chamber meeting just a month back. Governance, privacy, security. “How should I be using AI?,” they ask.

Understandable especially after last week’s main headline was AI has the potential to wipe out civilization and also when you read how a three person team used an AI model to break into accounts at OpenAI through its public help forum in under three days.

Alright. Clock's ticking. Back to work.

THE GAME CHANGED

Josh Blyskal at Profound pulled 3.5 million ChatGPT citations and found that starting around August 8, ChatGPT began running about 70 percent more background searches per answer.

Interestingly, the extra searching ChatGPT was doing wasn’t finding more sources. ChatGPT was checking things. In his data, 64 percent of ChatGPT's second searches carry two or more ideas forward from the first, which means it's reading, reasoning, and searching again. His read is that ChatGPT has stopped retrieving and started evaluating: comparing brands itself instead of trusting whoever wrote the roundup. 

If this sounds familiar, it should. In Issue 21 of Cited, I reported PromptWatch measuring ChatGPT's searches per answer jumping from about 1.08 to 1.83. A 69 percent increase from a completely separate dataset. 

Two companies saying the same thing.

What this means for you: ChatGPT is getting smarter. To be a source, you’ve got to be smarter as well. You can’t just say you’re the best and hope to make it into the answer. You’re going to have to demonstrate “why”.

Source: Josh Blyskal, Profound, LinkedIn, September 21, 2026 https://lnkd.in/p/eNZQnBgj. Corroborating date: ChatGPT site: operator fanouts, PromptWatch, August 2026.

THE WINDOW IS NOW

Blyskal broke this change down from ChatGPT like this: what kind of page ChatGPT cites, and who owns it. Listicles are down 38 percent since August 8, and social posts are down 91 percent. Commercial pages are up 35 percent, and he counts pricing pages, product pages, and landing pages there. Documentation is up 84 percent. By owner, first party sources are up 30 percent and institutions up 58 percent, while third party sources fell 36 percent and user generated content fell 71 percent.

Here’s what you need to know: At the end of August, in Issue 21, I told you to spread out across directories and third party sites, and that still gets you onto the list. Blyskal’s new research challenges this recommendation. For a local business, a ChatGPT check is likely to cover what a consumer might want to know. Things like what a job costs, what the warranty covers, how the process runs, how long it takes, and what license you hold. 

Most local sites don't have that page and that is to your advantage. Publish it, let ChatGPT come and check and if you’ve proven you’re an expert, you’ll have the potential to get cited.

Source: Josh Blyskal, Profound, LinkedIn, September 21, 2026 https://lnkd.in/p/eNZQnBgj.

YOUR COMPETITORS ARE ASLEEP

A security firm called Hacktron published how three researchers got into OpenAI employees' ChatGPT and Codex accounts in under 72 hours, starting with an image upload. They entered through OpenAI's public help forum, community.openai.com, which runs on outside forum software. A flaw in the library that reads iPhone photo files let them take over the forum, and a problem with "Sign in with OpenAI" turned that into access to the accounts of anyone who logged in there. One of those accounts belonged to an OpenAI employee whose Codex was connected to the company's code on GitHub, so they had it open a harmless pull request in OpenAI's internal repository to prove it, then stopped and reported it. OpenAI fixed its side within about 14 hours and paid a $6,500 bounty.

Here’s what you need to know: Attacks like these took incredible skill and months of time before AI. This whole two month campaign, which also hit Slack, Meta, and GitHub Enterprise through the same image library, cost under $3,000 in tokens. Of every company they tested, they say only Shopify noticed, even after thousands of images crashed the image processors.

What you should consider: The potential security threats in this new era of AI have largely been unconsidered for most small business owners. But, here’s one way to begin considering your risk. Open ChatGPT's settings and look at every app you've connected, like Gmail, Drive, and your calendar. The researchers could reach GitHub because an account was connected to it. Minimize your risk and disconnect anything you don't use weekly.

Source: Hacking OpenAI, Harsh Jaiswal, Mohan Pedhapati and Rahul Maini, Hacktron, September 13, 2026. HEIF Heist, Hacktron.

WHAT’S THE PROMPT

If ChatGPT is doing the checking and comparing now, a useful test is to watch it compare you.

Run the prompt

Run this, and read the "not found" cells. Every one of them is a fact a competitor's site states and yours doesn't, which is exactly what the evaluator is out there looking for.

Here’s how to evaluate: If you made the three with few blanks, your site is doing its job, so be sure to keep the facts current. If you made the three with lots of blanks, you're winning on reviews and directories but this won't hold as ChatGPT leans harder on first party facts. And, if you missed the three: start with the third answer. This is what a potential customer needs to choose you.

CITED

I asked ChatGPT for the best vet in Wilmington and it answered from two review roundups, LinkToReview and ThreeBestRated, with LaCroix Veterinary Hospital first on a 4.9 from about 268 Google reviews. 

Then I asked in a more conversational way specifically about a puppy exam. B All four were the vets' own pages (a pricing page, a new clients page, an emergency care page, a page for pet owners), and LaCroix didn't make the comparison at all.

The pick was Wilmington Animal Healthcare on Gordon Road, and every reason came with a citation to its own site: new patients welcome, a puppy care page, urgent care during business hours, AAHA accreditation, and a clear handoff to the 24 hour emergency hospital after close. It didn't win on stars. It won because it had the answers to the questions I asked.

Here’s the best part. ChatGPT said most Wilmington vets don't publish a puppy package price, then found one that does: Seacoast Animal Health Center, $428 for the whole puppy series, which it called the standout on price transparency. But, Seacoast is in San Marcos, California!

ChatGPT flagged that it couldn't confirm the location, but still built a whole section of a Wilmington answer around a vet across the country, because that clinic was the only one that published pricing for a puppy package!!!

A Visibility Tip for Vets: The first Wilmington vet to publish a puppy package price takes that spot from California.

See you next Thursday.

- Anna

Want to know where you actually stand in AI search? The AI Visibility Audit tells you exactly what the engines say about your business and what to fix. topofsearchmarketing.com/audit/

THE POLL