Ask an AI agent about your company today, and it’s answering from a photograph of the internet taken sometime last spring.
One research team measured the lag. Every major model shipped since July was trained on data that was 2 to 5 months old on its launch date. Whatever you published this month appears in the search results. It is not in the memory.
And the memory is writing the sentence.
Hi, I’m Jeff Payne. You’re listening to The Jeff Payne Show, Episode #104: It Looked You Up and Still Didn’t Say Your Name.
Most of what gets sold as AI visibility right now is about one thing: getting cited. Get your page into the handful of sources the answer links to, and you’re in.
That’s the citation. It’s real, and it matters. But there’s a second thing happening in every AI answer that almost nobody measures, because it doesn’t show up as a link.
Before the model searches anything, it already has an opinion of you. It read the internet once, months ago, and came away with a fuzzy recollection of who you are, what you do, and whether you’re worth mentioning. That recollection is baked in. It doesn’t update when you publish. It updates when the next model ships.
Then, sometimes, it searches. And here’s the part that changes how you should think about this. The search doesn’t replace the opinion. It confirms it.
The research team ran the same questions through an AI two ways:
once with search turned off, so the model could only answer from memory, and once with search turned on.
Then they looked at the brands that were actually sitting on the pages the AI cited.
They narrowed it to brands that appeared on at least 15% of the cited pages. These are companies that did everything the citation playbook asks.
Their name was in the room
Now sort those companies by how often the model mentioned them from memory, without any search.
The ones the model rarely named on its own, when it didn’t search, were named about 11% of the time when it did search.
The ones the model already knew well got named 99% of the time.
Same pages. Same citations. Same question.
The difference was what the model walked in believing.
Think about what that means for a CEO reading a visibility report.
The report says you’re cited.
The dashboard says you’re in the sources.
And the customer who asked the question never heard your name because the model read the page that mentioned you and, from memory, decided you weren’t worth repeating.
A while back on this show, I talked about the opposite problem: the brand that gets named in the answer without a single page cited. I called that being named from memory.
This is the same mechanism, just run in the other direction. Memory can put you in the answer with no citation. And memory can keep you out of the answer with one.
The citation is a ticket. It gets your page in the room. The seat at the table was decided before the door opened.
My client, Daniel Goodwin at Provident 1031, is the clearest example I have of building the memory, not just the citation.
Daniel wrote a book on 1031 exchanges that hit the #1 Best Seller list on Amazon. He has a standing byline at Kiplinger, every two weeks, for years now. His Masterclass series has returned $5 for every $1 invested.
None of that was built to win a single search result. It was built so Daniel’s name and his subject appear together, consistently, in many different places, over a long stretch of time.
That is exactly the shape of what the research describes.
A model doesn’t learn who you are from one page. One page is enough to get cited. It takes many pages, on many sites, saying the same thing about you, to get remembered.
So when a buyer asks an AI about 1031 exchanges or Delaware Statutory Trusts, Daniel Goodwin isn’t relying on search to find him. The search finds a reputation that was already there.
Here’s the shift.
Stop counting only where you’re cited.
Start counting where you’re described.
Every place your company is described in words, by someone other than you, is a deposit into the next model’s memory of you.
A trade publication.
A client telling their story.
A podcast transcript.
A review that says what you actually do.
Those don’t appear as citations on a dashboard. They appear months later as the model’s opinion.
And the lag works in your favor if you start now. What gets written about you this quarter is what the model shipped next spring will believe. What you’re counting on next spring has to be written this fall.
Two questions for the next time you look at an AI answer in your category.
If every page that cites you vanished tomorrow, would the model still know your name?
Secondly, where is your company described, right now, in places you don’t control? Is it described the same way in each of them?
I’m Jeff Payne. Thank you for listening. I will see you tomorrow.
The search was the same. The memory wasn’t.
Ask an AI about your company today, and it’s answering from a photograph of the internet taken sometime last spring. Whatever you published this month is in the search results. It is not in the model’s memory. And the memory is writing the sentence.
That is the finding behind Episode 104 of The Jeff Payne Show, and it changes what “AI visibility” should mean to a CEO.
Two things happen in every AI answer. Most people measure one.
Most of what gets sold as AI visibility right now is about getting cited: get your page into the handful of sources the answer links to, and you’re in. Citations are real, and they matter.
But something else is happening in every answer that almost nobody measures because it doesn’t show up as a link. Before the model searches anything, it already has an opinion of you. It read the internet once, months ago, and came away with a recollection of who you are, what you do, and whether you’re worth mentioning. That recollection is fixed. It doesn’t update when you publish. It updates when the next model ships.
Then, sometimes, the model searches. The search doesn’t replace the opinion. It confirms it.
The search doesn’t replace the opinion. It confirms it.
Cited on the page. Missing from the answer.
A research team at Graphite ran the same questions through ChatGPT two ways: with search disabled, so the model could answer only from memory, and with search enabled. They then checked which brands actually appeared on the pages the AI cited.
They narrowed the set to brands with substantial mentions on at least 15% of cited pages — companies that had done everything the citation playbook asks. Their name was in the room. They then sorted those companies by how often the model mentioned them from memory, with no search at all.
The model named brands it rarely knew on its own in roughly 11% of answers, even when search was on, and their pages were cited. The model named brands it already knew well 99% of the time. Same pages, same citations, same questions. The difference was what the model walked in believing.
Picture a CEO reading a visibility report. It says the company is cited. The dashboard shows the pages in the sources. And the buyer who asked the question never heard the company’s name, because the model read the page that mentioned it and decided, from memory, that it wasn’t worth repeating.
The citation is a ticket. It gets your page in the room. The seat at the table was decided before the door opened.
The mirror image of being named from memory
Back in Episode 85 I talked about the opposite problem: a brand that gets named in an AI answer without a single one of its own pages being cited. I called that being named from memory. This is the same mechanism, run the other direction.
Memory can put you in the answer with no citation. And memory can keep you out of the answer with one. Both are the same lesson: the citation is evidence; recognition is the verdict.
This is one of the best written, most informative email newsletters I’ve ever read!
Thank you for sharing your expertise, and hope you’re doing well.
Beth Zimmerman
Founder, Executive Director
PROOF: A REPUTATION THE SEARCH ONLY HAD TO FIND
My client, Daniel Goodwin at Provident 1031, is the clearest example I have of building memory, not just citations.
Daniel wrote a book on 1031 exchanges that hit #1 on Amazon. He has had a standing byline at Kiplinger every two weeks for years. His Masterclass series has returned $5 for every $1 invested. None of that was built to win a single search result. It was built so that Daniel’s name and his subject appear together, consistently, in many different places, over a long stretch of time.
That is exactly the shape the research describes. A model doesn’t learn who you are from one page. One page is enough to get cited. It takes many pages, across many sites, saying the same thing about you to be remembered. So when a buyer asks an AI who understands 1031 exchanges, Daniel doesn’t depend on search to find him. The search finds a reputation that was already there.
One page is enough to get cited. It takes many pages, on many sites, saying the same thing about you, to get remembered.
The Move: count where you’re described, not only where you’re cited
Every time someone other than you describes your company, it’s a deposit into the next model’s memory of you.
A trade publication.
A client telling their story.
A podcast transcript.
A review that explains what you actually do.
Appearing as an Amazon Best Seller.
Videos published on different marketing channels.
Published articles by third-party sites.
Branded Newsletters.
Informative Guides.
Those don’t show up as citations on a dashboard. They appear months later as a model’s opinion. And the lag works in your favor if you start now.
The same research measured the gap between a model’s training cutoff and its release date: two to five months for every major model shipped since July 2026.
What gets written about you this quarter is what the model shipped next spring will believe. You’re counting on next spring has to be written this fall.
TWO QUESTIONS TO ASK
If every page citing you vanished tomorrow, would the model still know your name?
Where is your company described right now in places you don’t control — and is it described the same way in each of them?
Sources: Gregory Druck, PhD, Ethan Smith, and Bevin Benson, Graphite — “Parametric Knowledge Optimization” (October 16, 2026). Full dataset: 3,115 entity–question pairs across 57 questions, logged-out ChatGPT, collected August 2026. The 11% / 99% figures are medians for entities mentioned on 15%+ of cited pages; the authors report two independent replications with comparable effect sizes. Knowledge-cutoff gaps computed from vendor model documentation. https://graphite.io/five-percent/research/parametric-knowledge-optimization
Similarweb, The 2026 Generative AI Landscape — about 7% of U.S. ChatGPT responses across all prompt types carry a citation, as cited in the Graphite paper. https://www.similarweb.com/corp/the-2026-generative-ai-landscape/
COMPLETE THE FORM TO
BOOK A STRATEGY CALL
"*" indicates required fields
COMPLETE THE FORM TO
BOOK A STRATEGY CALL
"*" indicates required fields
Subscribe and Share – WE APPRECIATE YOUR SUPPORT
