Advanced: Let an AI agent build a self-publishing site that AI search wants to recommend

What you'll get

Build your own "GEO content factory" with the open-source GEOFlow. GEO means getting AI search tools like ChatGPT, Doubao and DeepSeek to cite you when they answer users. It covers the whole chain. You feed in your company's existing material. The AI writes articles and topic pages in your voice, runs automatic quality checks, and sends approved pieces to your website and channels after a human review. Then you see how often AI search mentions you. Your AI agent handles installation, setup and troubleshooting. This article includes 3 prompts you can copy and paste directly.

Your AI agent sets up the system and drafts content. You vet the source material and approve what gets published. This system amplifies the real material you already have, so you need material to amplify.

Why a full chain, not just a writing tool

When AI search picks sources to cite, it looks at more than whether an article is well written. It also checks whether the content is credible, whether it's kept up to date, and whether the site structure is easy to read. If you only use AI to mass-produce articles, you end up with a pile of filler that nobody trusts. GEOFlow links several checkpoints together:

• Start with credible material: articles draw from your own knowledge base, and every claim traces back to a source document.
• Then run quality checks and human review: the AI first scores content on evidence and data, then checks for advertising language. Content that fails doesn't go through. Content that passes still needs a person's approval.
• Review after publishing: see which questions made AI mention you, and which competitors got cited. Then fill those gaps in the next round.

According to the project's repository, it's the 3.2.1 stable release. It's open source under the AGPL-3.0 license, free for individuals and companies to use commercially, and has about 4,000 stars on GitHub.

Step 1: Check whether it fits you

This system suits teams that have real material, someone to review it, and plans to run it long term. Examples include a company website, a brand's knowledge hub, or an industry resource site. Before you start, get four things ready:

• A computer or server that can run Docker: try it on your own computer first. Move it to a server for real use.
• An LLM API key: writing, quality checks and image generation all call a model. These costs are billed by usage and are separate from the software.
• A set of real material: product manuals, case studies, FAQs and industry reports. You can import PDF, PPT, Word and Markdown files. Scanned files and image-only pages need to be converted to text first.
• A review lead: a person has to read every AI-written article before it goes live.

Step 2: Have your AI agent install the trial version on your computer

Start by running it on your own computer so you can see the interface. Don't touch a server yet. Send this to your AI agent:

I want to try the open-source project GEOFlow (https://github.com/yaojingang/GEOFlow) on this computer. This is a local trial only, not a server deployment. Please help me:
1. Read the repository README and docs/deployment/DEPLOYMENT.md first. Check whether my computer has Docker, and whether the memory and disk space are enough. Tell me the result in plain language.
2. List everything you plan to install, and wait for my approval before installing anything.
3. Start it using the Docker development method in the README. Tell me the front-end and back-end addresses, and where the docs say the default admin account is.
4. The repository includes a GEOFlow Agent Skill (.agents/skills/geoflow). If you support Skills, read it before you act.
5. If startup fails, check the logs and explain the cause and the fix in plain language. Retry the same error at most twice.
6. Run it only on this machine. Don't expose any ports to the internet, and don't change any files outside this project.

After it's installed, log in to the admin panel. The home page is an AI workbench that asks "What would you like to do today?" You can describe a task in one sentence and start right away.

Step 3: Feed in material and configure the model

What GEOFlow can write depends on what you give it. In the admin panel, import your material into the Knowledge base, and enter your model's API key under AI configuration.

Import material in batches. Start with the most authoritative documents and the questions people ask most often, such as product descriptions and pricing, plus FAQs and customer cases. Remove outdated or conflicting material first. Otherwise the AI will state wrong claims with confidence.

Step 4: Have your AI agent help you pick the first topics and rules

Once the system is set up, what really drives results is what you write and the standards you write it to. In this step, your AI agent does the groundwork first. Then you fill everything into the admin panel together:

Why list questions before writing? Because AI search is triggered by questions. Users ask, and then AI looks for an answer. Item 5 runs two test articles first. That shows you problems in the prompt before mass production, so you don't have to wait until 30 articles are off-target and then redo them.

Step 5: Review, then publish

Articles that pass the quality checks go into pending review. The reviewer focuses on three things: whether the claims match the material, whether the data has a source, and whether the text sounds like how your company talks.

Approved articles can go to these places:

• Your website or hosted site: after publishing, the sitemap and llms.txt (a site guide written for AI) update automatically. That helps AI crawlers read your site.
• WordPress or other systems: pushed through an API.
• Toutiao, Sohu and NetEase publishing accounts (Chinese content platforms): the system creates a manual task. An operator uses a Chrome operations assistant to do the posting, and a person must confirm the final publish step.

Step 6: Check AI visibility, then decide what to write next

After you set up the AI search query interface under AI visibility, the system regularly asks AI your keywords. It records whether you're mentioned and where you rank. It also records which sources AI cites and how often competitors appear.

When you review, don't just check whether your mention rate went up. Two things are more useful. First, look at which questions cited competitors but not you. That's where your content is missing. Second, look at which sources AI cites. That shows which types of sources it trusts. Then aim your next round at those types.

Before going live, hand the job to your AI agent once more

Once you're happy with the trial and want to run it long term on a server, have your AI agent follow the official deployment guide:

I want to deploy GEOFlow to a server [server details I will provide in the terminal. Do not ask me to paste my password into the chat]. Please follow the production deployment method in docs/deployment/DEPLOYMENT.md:
1. Before deploying, list what I need to prepare: a domain, an HTTPS certificate, a database backup plan, and a persistent storage directory. Confirm each item with me before you start.
2. Use a fixed, officially released version tag. Don't deploy the latest code from the main branch.
3. I will set the admin password myself. Don't use the default value.
4. After deployment, run a health check. Tell me how to back up, how to upgrade, and how to roll back if something goes wrong.
5. Don't turn off any security settings. If something requires opening a port or changing the firewall, ask me first.

The official docs recommend a fixed version tag. The main branch changes constantly, so deploying it directly may pull in changes that haven't been fully tested. For upgrades, follow the official process too: stop the service, back up, then migrate. Don't just pull the latest code and restart.

Notes

• It amplifies your material. It doesn't create content from nothing. With little material and no review, mass generation only produces a pile of low-quality pages, and that can hurt your brand.
• Model costs are separate. Writing and quality checks call the model, and visibility monitoring calls the search API. Start with a small batch, check a week's bill, and then scale up.
• Read the AGPL license. You can use it yourself and offer services to clients for free commercially. But if you modify the code and let others use it over a network, you must provide the source under AGPL. Ask first about complex closed-source integrations.
• Third-party platform rules still apply. Assisted publishing only saves you clicks. Whether content complies and how often you post still has to follow each platform's rules.
• To work out the basic GEO groundwork your website needs first, read the tutorial Get AI search to cite your site: a checklist for your AI agent.

References

• GitHub: yaojingang/GEOFlow (AGPL-3.0)
• Website: geoflow.me
• Deployment guide: DEPLOYMENT.md