How I Build Brand Memory So AI Recommends Me

The machine had five of me

When I went to create a Wikidata entry for myself this month, there were already five Michel Fortins in there.

A French actor. A mathematician. A politician. A Quebec archeologist. A dentist.

None of them wrote The Death of the Salesletter. None of them coined Power Positioning. But if you asked a model who Michel Fortin is, that’s the lineup it had to choose from. Thirty five years of work, and the machine couldn’t tell me apart from a dentist.

That’s not an SEO problem. My pages rank. It’s a memory problem, and it’s the one most of us are ignoring while we argue about llms.txt files.

What brand memory is

An AI answer has two sources. What it retrieves right now, and what it already knew before you asked. The second one is parametric memory, and it’s where brand lives.

When someone asks ChatGPT or Google’s AI Mode for a fractional CMO who understands expert-led firms, the model doesn’t start from zero. It starts from whatever it absorbed about the names in that space during training. Then it goes looking for pages to back that up. If your name isn’t in the first part, the second part rarely saves you.

Cyrus Shepard put a number on this in September. His 131 experts rated brand or entity in the model’s memory at +2.08, third of 13 factors. Cross-web consensus scored +1.81 and publisher reputation +1.78. Structured data, the thing most agencies are still selling, came in at +0.80.

Dan Petrovic’s comment in the survey is the one I keep coming back to. Getting into a model’s training data, he said, is “much harder than link building.”

He’s right. And it’s the reason I stopped treating my website as the whole job.

The page is only the floor

I’ve written before about how I structure content so AI cites it. SOME, the four moves, still holds. Summarize, Organize, Modularize, Emphasize. Three of Shepard’s top four factors map straight onto it.

But SOME fixes the page. It does nothing for what the model believes about the author of the page.

Around the page there’s a second layer of work, and I’ve started calling it the four Fs. Fan-out means covering the whole question set around your topic, so the model finds you on the subqueries it runs and not just the headline query. Fidelity means the same brand facts, identical, on every surface a machine cross-checks. Frameworks means naming your thinking so it can be cited by name. Footprint means earning the third-party surfaces you don’t own. SOME makes a page quotable. The four Fs make the author worth quoting.

Fan-out is mostly a content architecture job, and I cover it in the AI citations playbook and the Triple-P coverage layer. The other three are what this article is about, because they’re the ones that build memory.

In my Triple-P Model, that’s the difference between Presence and Prominence. Presence is what you claim and own. Prominence is where you show up when you’re not the one talking. A firm can have flawless Presence, every page citable, and zero Prominence, because nothing off its own domain ever mentions it.

The model reads Prominence as trust. And trust is the part you can’t write yourself.

Step one, give the machine one of you

Here’s what I did on September 11.

I created a Wikidata item for Michel Fortin the marketer. Label in English and French. A description that no other Michel Fortin could claim: Canadian marketing strategist, author, and fractional executive. Occupation statements for marketing consultant, author, and copywriter, each with a reference. Country, languages, official website, and my LinkedIn, X, and Facebook identifiers so the item points at the same profiles my site does.

Then I went back to my About page and added the Wikidata URL to the sameAs array in my Person schema. Now the site points at the entity and the entity points at the site. A machine crossing between them finds the same name, the same facts, the same links on both ends.

That’s Fidelity, the second F. Same brand facts, everywhere a machine cross-checks. It sounds trivial until you realize how many experts have three bios that disagree on their own title.

Does a Wikidata item guarantee anything? No. Nobody can promise you a knowledge panel, and I won’t. What it does is give Google’s entity graph a canonical node to attach your schema to, which is the prerequisite for being disambiguated from the dentist. Before it existed, the machine had to guess. Now it doesn’t.

Step two, name your thinking

A model can’t remember a paragraph. It can remember a label. That’s the third F, Frameworks.

This is the reason I’ve spent decades naming frameworks, long before anyone called it entity SEO. Power Positioning. The Bullseye Method. OATH, QUEST, FORCEPS, IDEAL, SOME, Triple-P. I wrote about the habit in Why I Brandify Categories Instead of Branding Products, and I’ll admit the habit started as a memory aid for my own ADHD, not as a strategy.

It turned out to be one. A named framework is a citable unit. When I ask an AI about “power positioning,” my name tends to come back attached to it, because the phrase and the person have appeared together in enough places for enough years. Ask about the underlying idea without the name and the answer anchors to nobody.

Shepard’s experts made the same point from the other side. Geoff Kenyon noted that a specific number that’s already everywhere earns nothing. Specific and unique together is what gets cited. A framework with your name on it is specific and unique by definition.

If you have a way of thinking you explain the same way on every client call, it needs a name and one canonical page. That page becomes the address every future mention points to.

Step three, earn the surfaces you don’t own

This is the slow part, and it’s the part that moves the needle.

Everything above happens on surfaces I control. Brand memory gets built on the ones I don’t. That’s Footprint, the fourth F, and the one the data says counts most. Shepard’s experts scored those off-page signals higher than almost every on-page tactic, and my own experience says the same thing. What has worked for me falls into three groups.

Things I make that live elsewhere

Original research. Publishing your own data makes you the primary source other people cite, which is the one third-party mention you can set in motion on purpose. Andy Crestodina’s 13-year survey of content marketers (Orbit Media, September 2026) found that publishing original research raises the odds of strong results by about half, and Shepard’s experts scored unique first-party information at +1.85. It’s the tactic I push hardest in advisory work, because a study with your name on it gets quoted for years.

Books. Power Positioning and The Death of the Salesletter sit on Amazon, on Goodreads, on publisher and reviewer pages. Every one of those is a domain I don’t control describing what I wrote and who I am. A book is the oldest form of entity corroboration there is, and most experts who have one never think of it that way.

Things other people make about me

Podcasts. I’ve been a guest on more than 50 shows. Every one produces a page I didn’t write, on a domain I don’t own, where a host says in their own words what I do. That’s a third party describing the entity, which is exactly what a model needs to corroborate what my site claims. I pitch shows where the host’s audience overlaps with my buyer, not shows with the biggest download counts.

Guest articles and bylines. My author page at MarketingProfs dates back to the years I was teaching marketing management part-time at Algonquin College. When I created my Wikidata item, that page was one of the references I used to prove the occupation statement, because it’s independent of me. A byline on someone else’s site does double duty: a reader finds you, and a machine gets a second source for who you are.

Digital PR. Being quoted is different from being published. When a journalist or a newsletter writer cites you by name on a domain the model already trusts, the trust transfers. I don’t chase volume here. One citation on a source the model weights heavily beats ten on sites it ignores.

Speaking. I’ve given more than 200 presentations, to rooms as small as a board of directors and as large as the 10,000 people at the Wembley Conference Centre in London. A talk produces an event page, a speaker bio, a program listing, sometimes a recording and a recap. Four or five third-party surfaces from one afternoon.

Reviews and communities. Review platforms, association member pages, community profiles. Places where someone else describes what I do, in a format machines already know how to read. Lower effort than everything above, and it’s the Fidelity check in practice: every one of those profiles has to say the same thing, or the machine learns the disagreement.

Things I do with other people

Partnering with other creators. When Eli Schwartz or Cyrus Shepard publishes something I build on, I credit them by name and tag them. Some of them engage, and the exchange lives on both our profiles. The reverse happens too. Every one of those exchanges is a co-occurrence of two names in the same context, which is how models learn that the names belong to the same field.

Showing up in professional circles. Substantive comments on peers’ work, being tagged back, attending the events where the podcast hosts and editors are. There’s a limit, and Steve Toth measured it. He tracked 60 LinkedIn posts ranking in Google’s top three for SEO terms and found 37% gone within six days and only two still there after six months (AI Notebook, September 2026). The posts themselves don’t build durable memory. The relationships do, because they produce the podcast invitations, guest bylines, and co-authored pieces that last. Social is the door, not the room.

None of these are new. What’s new is the reason. We used to do them for reach. Now we do them because the model is reading all of it and forming an opinion about who we are.

The test I run

Ask an AI assistant your money question. The one a buyer would ask when they need what you sell. “Who should an expert-led firm hire as a fractional CMO?” “Who wrote about positioning for expert-led firms?”

Does your name come back without prompting?

If it does, the memory is there and the job is to keep feeding it. If it doesn’t, no amount of on-page work will fix it. The fix is off-site mention and repetition. More surfaces, same facts, same name, same frameworks.

I run this on my own name every month, alongside the brand monitoring I keep in Ahrefs and Semrush, where “power positioning” is one of the tracked terms. The tools tell me where the phrase is showing up. The AI question tells me whether the phrase is still attached to me. They’re different checks, and the second one is the one most firms skip.

It’s humbling, and it tells me where to spend the next month.

The order matters

If you take one thing from this, take the sequence.

Make the pages eligible and citable first. That’s SOME, plus Fan-out so the model finds you on the questions around the main one. Then fix Fidelity, one entity, one set of facts. Then name your Frameworks, one canonical page each. Then earn Footprint, the surfaces you don’t own.

Most firms run it backward. They hire someone for “AI visibility,” which means the last step, before the ones in front of it exist. The model has nothing to remember, and no way to confirm it if it did.

In Power Positioning terms, this is the Multiply pillar doing its real job. Multiply means the same position showing up in enough independent places that the market, and now the machine, stops needing to be reminded. Posting more on your own site doesn’t get you there.

The dentist is still on Wikidata. So am I. The difference is that now the machine can tell us apart.

If your firm’s name disappears the moment you ask an AI a question you should be the answer to, that’s the diagnostic I run. The work that follows it, the entity, the naming, the surfaces you don’t own, is what I’ve come to call the Authority Accelerator. Let’s talk.

---

Frequently Asked Questions