AGENCY GROWTH CLUB
Mary-Anne Da’Marzo on the BRAIN Framework Behind 9M LinkedIn Views
Guest Mary-Anne Da’Marzo
Guest Mary-Anne Da’Marzo
Mary-Anne Da’Marzo built a custom GPT trained on six years of her own content, then used it to script a single LinkedIn video that hit 9 million views.
On this episode of the Agency Growth Club, she breaks down the BRAIN framework behind that system, the client who got badly burned by “vibe coded” development work with no human QA, and why good SEO is good GEO when it comes to showing up in AI Overviews and LLM answers.
Pulled from the transcript’s real per-line timestamps, matched to where each topic starts.
Mary-Anne has created marketing content consistently for six years, starting from a deliberate 30-day challenge to get comfortable on camera when her business had no customers yet. When AI tools became mainstream, she tested the standard advice directly, prompting ChatGPT to “act like a viral video scriptwriter”, and found the results inconsistent and often built on outdated source material.
Her fix: rather than asking AI to improvise from a generic persona, she exported and analyzed six years of her own video scripts, identifying the consistent frameworks and patterns behind what actually worked, then loaded those frameworks into a custom GPT. Early results were decent but missing something specific, the GPT didn’t sound like her and didn’t understand her actual target audience. Layering in her company’s vision, mission, values, value proposition, avatars, and tone of voice fixed that gap directly.
For the viral post specifically, she used BuzzSumo to spot a trending topic (AI influencers), fed the linked article into her trained GPT, and had a finished script within minutes. “I then just filmed the script to camera and we posted it, and it got 9 million views. I was like, wow, this works.”
Mary-Anne’s core thesis: “If you give it a brain, then it can do work to your standard.” The five components:
Mary-Anne shares a direct cautionary example: a client brought a development project to her agency, then took a competing developer’s pitch at roughly a third of her agency’s price. Two months later, the client came back, the front end looked fine, but the back end was broken, the developer had essentially let AI generate the code without properly understanding or verifying the underlying architecture.
“Yes, you can get AI to do the job, but if you don’t really know what’s good versus bad, then who’s the judge?” Her point extends beyond development specifically: the same risk applies to AI-generated copywriting, if you can’t personally judge good copy from bad, you can’t verify AI’s output either, regardless of how confident it sounds.
Actionable advice:
Mary-Anne’s honest answer traces back to her own agency’s origin: starting with roughly £2,000 in the bank, she could only afford to hire hungry graduates rather than senior talent, so she spent early mornings (6-9am, in a Starbucks on Wardour Street) building extremely detailed, click-by-click process documentation specifically so junior staff could operate like she would once client work started at 9am.
It wasn’t until partway through this year that she realized those existing frameworks could be directly reformatted and loaded into AI systems. Her practical advice for agency owners: start with processes you’ve already documented, that’s the fastest win, since it’s a straightforward reformatting exercise rather than building from scratch. For areas without existing process, she recommends slowing down first to genuinely assess the time savings and other benefits before investing in building it out.
She also flags a specific tool for this exact gap: Guide (guide.com), which records every click and scroll during a normal run-through of a repetitive task and automatically generates a step-by-step SOP document from that recording, which can then be fed directly into an AI system.
Beyond Guide, Mary-Anne’s other favorite tool is Whisper Flow, a voice-to-text tool that goes beyond standard dictation: pressing a function key transcribes speech while actively cleaning up filler words, correcting phrasing, formatting text, and even generating bullet points on request. “I can’t remember the last time I typed an email.”
Mary-Anne highlights ChatGPT’s updated voice mode specifically, noting the earlier version felt glitchy and start-stop, while the updated version listens continuously and responds naturally. Her example: drafting an entire marketing strategy out loud during a 90-minute drive across LA, arriving back at her desk with a fully reformatted document ready for team review. She’s used the same feature for lighter tasks (asking for nearby pizza recommendations while driving) and for a genuinely significant one: brain-dumping her full 2026 company vision during a hike, then asking the AI to organize the resulting stream of ideas into a coherent structure and flag any gaps or missing questions.
She’s currently writing a book, voice-dictated in the same way, on authentic intelligence versus artificial intelligence, focused on how brands can maintain a genuine presence through the current AI growth era, directly connected to the philosophy behind the BRAIN framework itself.
Josh shared his own related process directly: he fed every Agency Growth Club podcast transcript (32 guests at the time) into NotebookLM, asked it to extract every quote related to hiring, leadership, and delegation, and used the resulting quote list as nightly writing prompts, complete with suggested angles to take each one, turning book-writing into a genuinely enjoyable two-hour nightly habit rather than a dreaded year-long slog.
Mary-Anne describes NotebookLM’s core distinction from a general AI chatbot directly: it only draws from documents you explicitly upload, rather than pulling from the open web, which makes it uniquely suited to fact-checking. Her clearest client example: a construction industry client with heavily regulated content, previously requiring a larger internal team to manually fact-check standards and regulations. Running content through NotebookLM against the client’s own compliance guidelines now surfaces contradictions automatically, adding a review layer without adding headcount.
Mary-Anne’s favorite recent NotebookLM use case: onboarding a new team member (Faith) by loading case studies, slide decks, and relevant podcast appearances into a dedicated notebook, then generating a roughly 23-minute personalized audio “podcast” (NotebookLM addresses the new hire by name throughout) covering client profiles and processes specifically tailored to that one person’s onboarding needs.
Mary-Anne’s simple discovery method: whenever she hits a genuine day-to-day problem, she assumes there’s likely already an AI tool built to solve it and searches directly, rather than waiting to hear about tools secondhand. A secondary benefit of years of consistent content creation: tool founders now reach out to her directly before public launch specifically for early feedback, meaning some of what she shares publicly has already been tested privately for months beforehand.
Fathom, an AI meeting notetaker, is Mary-Anne’s other core daily tool, used for generating content ideas, following up with the team, and tracking which recurring discussion points haven’t actually moved forward. A Zapier integration automatically pulls action items from every recurring meeting (one-on-ones, client calls) directly into the agency’s Asana board, closing project-management gaps without manual note-taking.
Her clearest example of a fast, high-value use case: after a genuinely inspiring conversation with another founder, she fed the Fathom transcript into a custom LinkedIn copywriting GPT, generated a full post drawn from the other founder’s own story, and sent it over immediately, a task that took seconds but wouldn’t have happened at all without the tool removing the usual time barrier.
Mary-Anne credits Gary Vaynerchuk’s “document, don’t create” principle directly as the foundation of her content approach, sharing what she’s actually doing day to day (her Whisper Flow and NotebookLM content, for example) rather than manufacturing a separate, artificial content calendar disconnected from her real work.
Her diagnosis of most AI content failing on LinkedIn: it’s not really an AI problem, it’s a dissonance problem, people trying to sound like an expert they aren’t yet, rather than genuinely sharing lived experience and real knowledge. “The more you try and fake being something before you are, the harder it is to create content.”
Mary-Anne’s core explanation of how LLMs actually work: every word is converted into a numerical representation, with statistical relationships (word affinities) determining what’s likely to appear together, which is how these models predictively generate answers. The practical implication: consistent naming and phrasing across every channel (PR, SEO, social, your own site) reinforces those word associations, making a brand more likely to surface when someone asks a related question.
She cites Rand Fishkin (a past guest on this podcast) as someone who deliberately practices this: whenever invited to speak or appear on a podcast, he specifies exactly how he wants his name, bio, and key terms phrased, not out of organizational preference alone, but specifically to reinforce consistent brand association across the web. “The more people that point back to Google or point back to the LLMs [with the same name and phrasing], they’re going to consistently build that profile.”
Mary-Anne is direct that the GEO tooling space is currently “the wild wild west,” with many tools overselling guaranteed AI Overview or LLM rankings that don’t hold up to how these models actually work. Her audit process for any new tool: research the CEO directly (Google them, check LinkedIn history) to assess genuine authority in the space versus opportunistic timing, and critically evaluate whether the tool’s specific claims are technically plausible given how LLMs function.
Her clearest red flag: traditional keyword tracking works because Google’s index is well-structured and well understood, but the sheer number of ways a single conversational query can be phrased makes a blanket guarantee of ranking for “every variation” implausible on its face.
Actionable advice:
Josh shared a direct observation from a recent London SEO event with senior practitioners from major brands and agencies: despite plenty of confident-sounding content online, the honest consensus in the room was that everyone is still actively experimenting, with real budget behind genuine tests rather than proven playbooks.
Mary-Anne’s response reframes this positively rather than as a problem: SEOs are natural problem-solvers who thrive on ambiguity, and the current uncertainty in AI search isn’t fundamentally different from SEO’s historical relationship with Google, some principles are well-established, others are actively being tested in real time. Her practical grounding technique: directly Google your own brand and products to see where AI answers are actually pulling information from, then reinforce or improve those specific existing reference points rather than chasing every new, unproven tactic.
Yes, unambiguously, according to Mary-Anne, who’s led numerous client rebrands. The “B” (brand assets) component of her team’s “brain sprint” process for clients essentially is a rebrand exercise, vision, mission, values, value proposition, and tone of voice, precisely because a business’s AI-facing presence needs to actively reflect how it wants to be found and represented going forward, not remain anchored to outdated information and phrasing.
Mary-Anne’s agency uses AI, including “vibe coding,” across most development work, but always pairs it with real human QA from an active developer to verify the output makes sense. Her reasoning ties directly back to the earlier client story: someone without real development experience can get AI to produce something that looks functional on the surface while missing structural issues only genuine expertise would catch, the same risk applies directly to copywriting and any other AI-assisted output.
She references a comment from an OpenAI developer she heard on a separate podcast: the real danger point isn’t AI making mistakes, it’s when an AI gets stuck in an unproductive loop and neither the AI nor the human overseeing it notices or can resolve it. Her position isn’t anti-experimentation, she encourages pushing boundaries and learning through trying, but with a clear-eyed expectation that substituting AI for genuine expertise inevitably surfaces gaps neither party anticipated.
Mary-Anne is candid that video comes naturally to her personally, so her advice may not generalize perfectly, but the underlying principle does: build a personal brand around your actual strengths, not the format currently trending. If you’re a confident, natural communicator on camera, the only real barrier holding you back is fear, worth pushing through directly. If writing and storytelling is your real strength instead, lean into that format (adding photos or AI-generated visuals to support it) rather than forcing yourself into video just because it’s what’s currently working for others.
Her own honest admission: despite carousels currently performing extremely well on Instagram, she avoids the format entirely simply because she doesn’t enjoy creating them, a deliberate trade-off in favor of consistency and genuine enjoyment over chasing every high-performing format.
Brand assets (vision, mission, detailed tone of voice), Request frameworks (documented SOPs), AI instructions (how the AI should reference other documents), Interaction directives (prompt sequencing, including self-scoring), and Next-level examples (feeding proven wins back into the system as new benchmarks).
Mary-Anne recommends researching the tool’s founder directly (their background and authority in search specifically) and being skeptical of any tool promising guaranteed rankings across all conversational query variations, since the sheer number of ways a question can be phrased makes blanket guarantees technically implausible.
She trained a custom GPT on six years of her own video scripts, identifying consistent frameworks and patterns, then fed it a trending news topic (found via BuzzSumo) to generate a script, which she filmed and posted directly, with no manual scriptwriting involved.
A client hired a developer at roughly a third of her agency’s price who used AI to generate a project’s code without properly verifying the underlying architecture. The front end looked functional, but the back end broke within two months, since no one had done real human QA on the AI-generated work.
Whisper Flow (voice-to-text with automatic cleanup and formatting), Guide (auto-generates SOPs from screen recordings), NotebookLM (fact-checks content against uploaded source documents only), and Fathom (an AI meeting notetaker integrated with Zapier and Asana for automatic task tracking).
According to Mary-Anne Da’Marzo, since LLMs work on statistical word association, consistently using the same brand name, product names, and phrasing across every channel (PR, social, your own site) reinforces the associations these AI models draw on when generating answers, the same foundational discipline as traditional SEO.
With 95% of roles filled within 25 days, our approach is built for pace and backed by a 1 year guarantee. Get access to the systems and frameworks that power elite SFH teams.
