Hello Lunch Break. Driving down South 17th Street and notice the hospital parking lot is full. Last week U.S. News put Novant Health New Hanover Regional Medical Center at number 12 in North Carolina in its new Best Hospitals rankings. That surprised me. Wilmington is known for a lot of things. Healthcare has never really been one of them. And if you’ve spent time as a patient at Novant Main, you know it has been understaffed for years which has a real impact on a patient’s experience. But real patients in Wilmington didn’t decide this. U.S. News did. 

Which brings me to influence. Three separate studies this week show how little from your own website ChatGPT is willing to read, what’s influenced it already, and how little paid media can do about it. The model deciding whether you're any good is not deep diving into your website. 

Alright. Clock's ticking. Back to work.

THE GAME CHANGE

A Paris agency built a browser extension that records the raw data stream ChatGPT sends to your browser, then ran an experiment that found ChatGPT turns out to have its own index, called labrador. Everyone had previously thought ChatGPT used Google or Bing to build an answer so that was news. And, the agency shared when a Chatgpt user is in the free instant mode, nine answers out of ten never open a single webpage. 

Instead, this labrador model answers from a record that knows three things about you: your URL, your page title, and a snippet cut at 202 characters. That snippet is anchored on your H1 and grabs whatever visible text happens to sit next to it. So every question about your business gets the same 202 characters ChatGPT reads until the next crawl.

Source: Olivier de Segonzac, RESONEO, in Search Engine Land, August 17, 2026. Inside ChatGPT's retrieval stack

THE WINDOW IS NOW

Suganthan Mohanadasan watched the search queries ChatGPT writes for itself, and found brand names were in the model's very first query in 21 of 27 conversations. Then he checked what happened to those brands. 

A brand the ChatGPT model named in its own query showed up in the final answer 68.9 percent of the time. A brand that only got fetched afterward showed up 2.1 percent of the time. That's a gap of roughly 33 times, and it means the brand names showing up are from what the ChatGPT model already believes about your category, not from what it finds when it goes looking in a search.

The fix to be in the category vocabulary ChatGPT has gets built off your domain, slowly, in reviews, comparisons, roundups and local coverage, which is exactly why the window matters. The businesses that will get named are building toward it now.

Worth saying: This is from 27 conversations, one ChatGPT Plus account, a sample weighted toward software and AI tools. The author himself calls every percentage in it directional rather than measured, and asked for replication across multiple accounts. Read it as directional.

Source: Suganthan Mohanadasan, ChatGPT decides before it searches, July 2026

YOUR COMPETITORS ARE ASLEEP

SE Ranking ran 50,006 commercial prompts across 20 niches on a single day in July and found sponsored placements on 25.94 percent of them, close behind Google's AI Mode at 29.45 percent. 

Then they checked whether buying the ad got the advertiser into the answer sitting directly above it. Spoiler: it did not. 

Advertiser domains appeared as a cited source 3.63 percent of the time, and the exact advertised URL appeared 0.09 percent of the time. 

You can buy the slot. You cannot buy the answer.

Source: SE Ranking, 50,006 prompts, July 23, 2026, via Search Engine Land

WHAT’S THE PROMPT

Pick the page you'd most want ChatGPT to describe correctly, usually your main service page, and find out what the first 200 characters after your H1 really say. What you see is the entire impression the ChatGPT model has of that page for the nine out of ten answers where it never opens anything.

Run the prompt

Then do the one thing the answer tells you to do. Add a sentence in the first 200 characters that says what you do, where you do it, and who you do it for.

Want us to run this across every page on your site instead of one? That's the first thing our audit checks. Start with the audit.

CITED

I asked ChatGPT and Gemini to recommend the “best wellness studio in Wilmington NC” and I got the same response. Renoja, out at Mayfaire, which the engines own answers put at 4.9 star rating.

Then I asked the way the engines like someone with a Saturday free asks, infrared sauna and cold plunge, walk ins, close to downtown, and the initial businesses disappeared. 

ChatGPT put 20 Orange Spa first, a 4.0, the lowest rated business anywhere in either answer, and Gemini named a different studio entirely, so the two engines ended up in disagreement. 

When you step back and analyze, you can see what 20 Orange has that the 4.9 does not. They have a page sitting at 20orangespa.com/thermotherapy-infrared-sauna whose copy says the contrast room combines an infrared sauna with a cold plunge, in those words. Renoja offers cryotherapy, which is the same idea under a different name, and on this query the name is what counted.

Worth saying: neither engine confirmed the walk-in part. Gemini said outright that most Wilmington studios run appointment first, 20 Orange's own page lists no hours and no walk-in policy, and ChatGPT pulled the hours it quoted from Yelp. 

Two constraints in the prompt out of three were answered and the third quietly disappeared, which provides its own lesson about how much weight to put on any single AI answer.

See you next Thursday.

- Anna

Want to know what your own site is telling them? The AI Visibility Audit checks exactly this.

THE POLL