SEO consultant Marie Haynes today published a 44-minute YouTube video on how she has spent the months since Google released the Open Knowledge Format building a paid business on it. She says her OKF "brain" of roughly 600 markdown files, connected to Gemini 4 Argon and a server that logs Google algorithm updates, now drafts the traffic drop, helpfulness and post-update reports she sells. She also says it "made its first $$$ this week."

In Short

Marie Haynes, who has spent years diagnosing why websites lose Google traffic, put everything she knows into hundreds of small, labelled text files that an AI can read. She now uses those files with Google's newest model to write paid website reports faster and cheaper than she could by hand, and she says the AI's work is often as good as hers. If that holds up, expert knowledge could be sold as something an AI looks things up in, not only as hours of a consultant's time, though Haynes is the only source for how well it works.

The video and the post

The video, titled "How I Am Building & Making Money with OKF (Open Knowledge Format) and Gemini 4 Argon," went up on the Marie Haynes YouTube channel today. The channel has 9,810 subscribers, and the video had drawn 104 views within hours of release. Haynes promoted it in a LinkedIn post the same day, writing that she had "put together a video to answer the most common questions I've had about OKF and show the cool things I've been building with it."

The video is split into 18 chapters. It mixes a tutorial with product demos: a general knowledge base she calls her OKF brain, a health and fitness tracker, a voice-cloned "Self-OKF" she talks to, a chores app, an automated SEO report pipeline, and a chatbot for members of her paid community. It ends with Haynes reflecting on artificial general intelligence and work.

Nearly every figure in it comes from Haynes herself, about her own systems, costs and clients. None of it has been checked independently. She also says plainly that the reports are for sale. "This is not supposed to be a sales uh pitch for for my reports," she said in the video. "I mean, I love if you want to order them."

What OKF is, and where Haynes's account differs

The format is Google's. Open Knowledge Format version 0.1 was published on June 12, 2026, by two Google Cloud tech leads, as PPC Land reported when it described Google's OKF bid to become the lingua franca for AI agent knowledge. It is a specification for markdown files with YAML frontmatter, and the only field it requires is type. An optional index.md lists what a directory holds, so an agent can see what is available before opening a single document. Google's own examples were analytics datasets: GA4 e-commerce data, the Stack Overflow public dataset and the Bitcoin public dataset.

Haynes opens the video by saying "it's been three months since we first heard about OKF." By the date of the specification, the gap is a little over four months. She describes the format much as Google's specification does. "Every OKF has an index.md file, just markdown," she said. "And the index acts kind of like a card catalog if you were going into the library." On frontmatter, she said: "I believe that the documentation says the only thing that's necessary in the YAML front matter is to have some type of a type." That matches what the specification requires.

She also says she is using the format for something other than what Google built it for. "Now, what we've been building, I think, is maybe not what OKF was designed for," she said. "The examples that Google gave were gathering analytics data and then using OKF as a standardized way to uh to tell agents how to display it and what data to pull."

Her description of a verification marker as "something that got added in uh OKF version 0.2" fits the record too. Version 0.2 followed on July 24, 2026, and added provenance, trust, lifecycle and attestation fields, according to PPC Land's OKF explainer. When Google showed how a 9-concept OKF bundle becomes 17 access-controlled catalog entries in Cloud Knowledge Catalog on August 26, Haynes was among the first to comment. That evening she posted on X that "Github is still the way to go for most of our uses."

"More specified rag"

The question Haynes says she gets most is how OKF differs from retrieval augmented generation, in which a model pulls passages from an indexed store before it answers. "It actually is rag, but it's more specified rag," she said.

She told a story to show why the difference matters to her. Early on, she set up a Vertex AI system that indexed her website as a vector space and asked it what she thought about buying links. Haynes started in SEO cleaning up sites with manual actions for unnatural links, and she opposes link buying. According to Haynes, the answer was: "Marie Haynes is well known as a seller of links." She put the error down to retrieval across a very large vector space, which she said "will hallucinate or will confabulate will make up information." OKF, in her account, avoids this by breaking knowledge into small, labelled files that an agent can choose between.

She also tied OKF to Andrej Karpathy's "LLM wiki" pattern, saying the two are "essentially the same thing, just that OKF standardizes the uh front matter." Google's own announcement credited Karpathy's gist for that pattern, as PPC Land reported in June.

Inside the brain

Haynes's main knowledge base has, by her estimate, "about maybe 600 or so files." It holds every blog post she has written, broken into concept files, playbooks, entities and a separate store of news items drawn from a newsletter she says she started around 2012. Each file carries tags and a visibility value. Public files are shared with "most of my clients." Private files hold her own procedures, such as how she writes invoices and proposals. Each client also has a folder with its own memory.

She says she did not design this structure. An agent did. "I didn't make any of this," she said of the index frontmatter. "This just anti-gravity did it." Google's agent-first development platform is spelled Antigravity, and its 2.0 release headlined Google I/O 2026 alongside Gemini 3.5 in May. She also runs a periodic clean-up she calls "lint," which she says she borrowed from Karpathy.

Scale has limits, and she says so. "If you're doing this for enterprise, then what I'm doing here is not going to scale," she said. "Your agents are going to be taking forever trying to go through all of this documentation." She said she had looked at Google's Knowledge Catalog and decided "it's just for enterprise." She named Neo4j, an open-source graph database whose links between concepts carry their own labels, as something she expects to study.

Costs, as Haynes states them

The chat front end for the brain runs on Gemini 3.8 Flash, which she called "quite fast" and "very inexpensive." Using her systems "sometimes all day long," she said she pays $5 or $6 a day at most, and that queries generally cost "just pennies per query." She keeps the files in Google Cloud storage buckets, which she said "costs $0" at her scale, while many members of her community keep theirs on GitHub.

The interface was built by an agent too. According to Antigravity's own summary, which Haynes read out on screen, it is a React and Next.js app that preloads her profile and master index and gives the model tools to search files, read them and save to memory. She said she knows "a little bit of JavaScript uh but not enough that I could build this on my own," and that she runs Antigravity with a "turbo mode" setting "which is not recommended, but uh it basically gives anti-gravity permission to do whatever it wants." She added that anyone working with enterprise data probably should not do the same.

Where Argon comes in

Haynes credits one model with turning the project into income. "Gemini 4 argon was the thing that turned my OKF from a cool project into something that actually is making me money today," she said.

Google released Gemini 4 Argon on October 13, 2026, at an introductory $2 per million input tokens for paid users only, with output at $10 per million. Free users have no access. That matters for one of her cost claims: she says projects not attached to billing get free API use, but the model she credits with the money-making work sits behind paid tiers. Haynes says she uses a paid key for her health app and adds: "when you use a paid API key through Google, they do not train on your data." That is her statement of Google's policy, not something the video demonstrates.

She also said that "when Gemini 4 Argon came out, Google said Demis said that it's really good at knowledge work." PPC Land's coverage of the Argon launch contains no such quote from Demis Hassabis, and Haynes gives no source for it.

The audit pipeline

The business case in the video is a product Haynes has sold by hand for years. She said she has analysed sites hit by Google updates since the Penguin and Panda era of 2012. Demand once pushed her prices up until some clients "were willing to pay um you know high five figures for some of my reports." That left out small site owners, the people she says she prefers to help. "I spend three weeks making tens of thousands or I may I spend uh three weeks making a few hundred dollars," she said. "It doesn't make any sense to work on uh the smaller."

The new pipeline joins three pieces: Gemini 4 Argon, the OKF brain, and a list of Google algorithm updates she maintains, which she has turned into an MCP server that agents can query. In one real case, she said, the agent found an enquiry in her email through the Workspace CLI, worked out that the site had a pure spam manual action, read its sitemap, and found "programmatic, templated, thin content." It then listed the pages likely responsible and drafted clean-up advice. "Where did that advice come from? It came from my OKF brain," she said.

On data, she said the agent pulled Search Console where she had access and used the Ahrefs API where she did not, "not always as accurate." It also read update dates from her MCP to mark charts. When she disagrees with something in a draft, such as a label or a font, she corrects it, then runs a /learn command so the agent records what she changed for the next report. After each review she removes identifying details and feeds the lessons back into the brain.

On accuracy, the only check is her own: "I've been manually checking all of the information and it's right. It's not hallucinating." The demo reports in the video were for a site she described as "completely madeup," so viewers cannot compare a real report with the data behind it.

Orders now run through Stripe and an intake form, and she generates each report in Antigravity. According to Haynes, Gemini suggested that "once we've done 10 or 20 of these," correction would no longer be needed, after which reports could be emailed automatically. She said one helpfulness assessment was ordered "a few hours ago."

Selling access, not files

Haynes says her view of how to make money from OKF has changed since her first videos. Zipping up markdown files and selling them hands over the intellectual property along with the knowledge. "I don't think it's a good idea to sell your knowledge bundles," she said. "Instead, you sell access to your bundles."

There are three versions of that in the video. Clients get a "Marie's brain" chat. Members of her paid community get "Marie's Agents," including an "OKF on OKF" chatbot trained on her posts, video transcripts and answers about the format. She said each conversation costs "a few cents," and "maybe if somebody has a long one, it might cost me 20 cents like at at most." The third is the automated reports.

She also suggested that consultants could use an OKF chatbot to answer set questions and then direct visitors to book a call, and that people without paying clients would find it hard to make money this way. "In order to make money with OKF, you need to have knowledge and you need to have people who are willing to buy the knowledge as well," she said near the start. "So, it's not an easy ticket to becoming a millionaire."

Near the end, she referred to Google features for business customers announced, she said, "just today, as I'm recording this": a panel for connecting external MCP servers. She suggested this could let people sell OKF knowledge as a skill or tool inside Gemini Business without handing over the whole bundle. The video gives no date for that announcement and names no product page. PPC Land reported on October 8 that Google's new Gemini agent works with any MCP server and that SMBs are in early access.

The personal apps

Two demos are about her personal life, and the health one comes with numbers. Haynes said she built a health and fitness OKF to "flush out" theories about health that "don't really align with traditional medicine," and logs meals and Fitbit statistics into it every day. She showed it working out her protein intake for one day: about 14 grams at breakfast, 38 at lunch and 32 at dinner. She said her resting heart rate had fallen and her heart rate variability had risen since she started logging, and that she was "down um 22 pounds." She did not say the app caused those changes, and nothing in the video shows that it did. She said it was not something she wanted to sell.

The memory design is the reusable part. Haynes said it slowed down after about a month because one memory file had grown too long. She replaced it with three layers: a core memory of permanent facts and rules, a recent memory covering 14 days, and a vault of older entries the system can search when needed. Her voice-cloned "Self-OKF," which she said was built on a Google voice cloning feature she covered in an earlier video, shares that memory. She demonstrated it in a live call.

Why this matters for marketers

The video is one practitioner's account, but it gives a working example of something PPC Land has been tracking since June: Google's OKF specification moving out of BigQuery examples and into consultancy workflows. Three points stand out.

The first is the economics. Haynes describes moving from a few high-priced reports to many low-priced ones, with model and storage costs she puts at pennies per query and nothing for storage. If that pattern spreads, the price of diagnostic SEO work could fall for small businesses. A consultant's edge would then sit in a curated, proprietary knowledge base rather than in hours billed. Earlier this year Haynes argued that Google now buries competent pages that add nothing new. Her own business case leans on the same idea: what an agent cannot copy from the web is the knowledge she wrote down.

The second is her view of generative engine optimisation. "I view generative engine optimization especially tracking brand share of voice a major distraction," she said, citing an article and study of her own that the video does not show in detail. That puts her at odds with the vendors selling AI visibility tracking, a category PPC Land has covered repeatedly.

The third is what she leaves unresolved. Haynes ran an agent with blanket permissions, through her email, over client data, and said she does not read the code. She checks the reports herself and says the system is "not hallucinating," and the only evidence is her own review. The money involved is small so far: a first dollar, in the video's own words, and reports ordered in recent days.

AGI and jobs

The video closes on how these systems might change knowledge work. "My OKF plus my MCP is a better consultant than I am," Haynes said, adding: "I don't think it's long before it learns to do this without me." Earlier she said that "we are reaching the point soon where AI can do things better than any consultant," and that some automated reports "were better than what I would do."

She said Demis Hassabis had predicted AGI by 2030, and cited "an article in the Wall Street Journal just uh yesterday" saying he had moved that up to two to three years. She gave no headline or link, and the claim has not been verified. Google's own leadership has framed timing differently. When Sundar Pichai was asked about AGI in May, he put a rough window of three to five years on it and said "whether you and I call it AGI or not doesn't matter."

Haynes expects "a very messy time where a lot of people don't have jobs." Her LinkedIn post ended on the same question: "I can see the day where most knowledge work is done by AI. I wonder what that means for humans?!?"

Timeline

Summary

Who: Marie Haynes, an SEO consultant who sells traffic drop, helpfulness and post-update assessments and runs a paid AI and search community.

What: A 44-minute YouTube video and a LinkedIn post describing how she uses Google's Open Knowledge Format to run a knowledge base of about 600 markdown files. Combined with Gemini 4 Argon and her algorithm update MCP server, it now drafts paid SEO reports, and she says it made its first money this week. She argues that experts should sell access to their knowledge bundles rather than the files.

When: The video and post were published on October 23, 2026. OKF v0.1 dates from June 12, 2026, and Gemini 4 Argon from October 13, 2026.

Where: On Haynes's YouTube channel and LinkedIn. Her systems run in Google Antigravity, with files in Google Cloud storage and access sold through her website and community.

Why: Haynes says the setup lets her sell lower-priced reports to small site owners she could not afford to serve by hand. For marketers, it is an early example of a consultant pricing proprietary knowledge as something an agent can query. Its accuracy, so far, rests only on her own checks.