James Coyle (@OpenDeMind) header
James Coyle (@OpenDeMind) profile picture

I run AI agents on real local SEO. 17 home service sites. Building SEO Genius. Founder @RizeDigitalUSA

San Francisco Bay Area
Joined February 2014
Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)Photo by James Coyle (@OpenDeMind)
You installed Claude Code, Codex, Gemini CLI or Cursor and still use it like a chatbot. 35 GitHub repos that make LLMs materially better at marketing, SEO and PPC: No archived repos. No mystery links. No pretending every public repo is fully open source. No unofficial ad-account MCP asking for production write access. Important distinction: some of these install directly as skills or MCP servers. Others are libraries and platforms your agent can operate through code. A MARKETING STRATEGIST marketingskills Dozens of practical skills covering CRO, copywriting, SEO, paid ads, analytics, launches, pricing, growth and lifecycle marketing. → github.com/coreyhaines31/mar… google/skills Google’s official collection of Agent Skills, including Google Ads API troubleshooting, setup and analytics-related developer workflows. → github.com/google/skills social-media-skills Skills for finding angles, writing hooks, creating posts, auditing profiles, developing short-form video ideas and analyzing social performance. Some workflows require external API accounts. → github.com/charlie947/social… humanizer Finds repetitive AI writing patterns and forces a more natural editing pass. Use it as an editor, not as some magical “AI detector bypass.” → github.com/blader/humanizer agency-agents A large collection of specialized agency roles covering strategy, paid media, content, design, research, analytics and operations. Copy the roles you need. Do not dump the entire collection into every context window. → github.com/msitarzewski/agen… AN SEO TEAM claude-seo A dedicated SEO skill and agent suite covering technical SEO, schema, content, local SEO, international SEO, ecommerce, GEO, AEO and reporting. → github.com/AgriciDaniel/clau… advertools A serious Python toolkit for crawling, sitemaps, robots.txt, SERPs, keywords, log-file analysis, advertising data and search-marketing analysis. → github.com/eliasdabbas/adver… siteone-crawler A cross-platform website crawler that checks SEO, performance, accessibility, security and technical quality. Useful for giving an agent structured evidence instead of asking it to “look at the website.” → github.com/janreges/siteone-… SEOnaut A self-hosted SEO auditing crawler for finding technical issues and running repeatable site scans without relying entirely on SaaS tools. → github.com/StJudeWasHere/seo… chrome-devtools-mcp Lets an agent inspect a live Chrome browser, performance traces, network requests, console errors, rendered pages and screenshots. Run it in a clean browser profile. The agent can see what that browser can see. → github.com/ChromeDevTools/ch… playwright-mcp Microsoft’s browser-automation MCP. Useful for checking rendered pages, forms, conversion funnels, mobile states, JavaScript behavior and landing-page QA. → github.com/microsoft/playwri… firecrawl-mcp-server The official Firecrawl MCP for web search, scraping, crawling and structured extraction. Good for competitor research and turning websites into usable LLM context. Hosted usage can cost money. → github.com/firecrawl/firecra… crawl4ai A local-first, LLM-friendly crawler and scraper that converts websites into cleaner, structured information for agents and RAG systems. → github.com/unclecode/crawl4a… firecrawl The core Firecrawl platform for searching, crawling and extracting web content at scale. Useful when you want to build the research layer yourself instead of relying on a black-box scraper. → github.com/firecrawl/firecra… crawlee A mature browser-automation and crawling framework built around Playwright, Puppeteer, proxies and repeatable extraction jobs. Better for custom data pipelines than one-off prompting. → github.com/apify/crawlee A RESEARCH AND CREATIVE TEAM gpt-researcher An autonomous research agent for producing multi-source reports using different LLM and search providers. Useful for market research, competitor analysis, category research and content briefs. → github.com/assafelovic/gpt-r… markitdown Microsoft’s utility for converting PDFs, Word files, PowerPoints, spreadsheets and other documents into Markdown an LLM can actually work with. This is not the marketing-skills repo, despite what several viral lists claim. → github.com/microsoft/markitd… docling A more structured document-processing system for PDFs, DOCX, PPTX, XLSX, tables and complex layouts. Useful when simple PDF-to-text conversion destroys the information you need. → github.com/docling-project/d… OpenCut An MIT-licensed open-source video editor. Useful for turning LLM-generated scripts, footage and creative assets into ads, reels and short-form content. → github.com/OpenCut-app/OpenC… ComfyUI A node-based image-generation backend for building repeatable creative workflows instead of generating every ad image manually. Only install custom nodes from maintainers you trust. → github.com/Comfy-Org/ComfyUI A PPC AND MEASUREMENT TEAM google-ads-mcp OFFICIAL / REPORTING-FIRST Google’s official MCP server for querying Google Ads accounts, discovering available fields and metrics, and listing accessible customers. Its current tool set is read-oriented, but your account data is still being exposed to the connected model. → github.com/googleads/google-… google-analytics-mcp OFFICIAL / EXPERIMENTAL / READ-ONLY Runs GA4 core reports, funnel reports, realtime reports and property queries through the Analytics read-only scope. One of the safest direct marketing-data connections on this list when configured correctly. → github.com/googleanalytics/g… google-ads-python OFFICIAL / WRITE-CAPABLE Google’s official Python client for the Google Ads API. Use it to build reporting, budget monitoring, anomaly detection and controlled account-management systems. Start with test accounts and put human approval in front of every mutation. → github.com/googleads/google-… facebook-python-business-sdk OFFICIAL / WRITE-CAPABLE Meta’s official Python SDK for Marketing APIs. It can support reporting, audience workflows, creative operations and campaign management, but it should never receive unrestricted production credentials by default. → github.com/facebook/facebook… tiktok-business-api-sdk OFFICIAL / WRITE-CAPABLE TikTok’s official SDK for integrating with its Business API. Useful for reporting and campaign tooling, but the same test-account and approval rules apply. → github.com/tiktok/tiktok-bus… Meridian Google’s media-mix-modeling framework for measuring channel contribution and modeling budget allocation. This is actual marketing science, not another ad-copy generator. → github.com/google/meridian Robyn Meta’s experimental marketing-mix-modeling package. Useful for modeling diminishing returns, response curves and budget allocation across channels. → github.com/facebookexperimen… GeoLift Meta’s geo-experiment framework for estimating the incremental lift caused by advertising. Useful when platform-reported conversions are not enough to prove causality. → github.com/facebookincubator… PyMC-Marketing A Bayesian marketing toolbox covering media-mix modeling, customer lifetime value, retention and buy-till-you-die models. → github.com/pymc-labs/pymc-ma… GrowthBook An experimentation and A/B-testing platform for testing landing pages, offers, onboarding flows and product changes. Useful for proving whether an LLM-generated variation actually performs better. → github.com/growthbook/growth… AN OPS AND QA TEAM PostHog Product and web analytics, session replay, funnels, feature flags and experiments, with agent and MCP support. Useful for showing an agent what visitors actually did after clicking the ad. → github.com/PostHog/posthog promptfoo Tests prompts, agents and RAG systems across models. Use it to regression-test ad copy, prohibited claims, brand rules, factual accuracy and required output formats before anything reaches a client. → github.com/promptfoo/promptf… Langfuse LLM observability, evaluations, prompt management, traces, datasets and performance monitoring. Useful for finding out why a marketing agent made a bad recommendation instead of merely seeing that it failed. → github.com/langfuse/langfuse n8n A fair-code workflow-automation platform with AI, MCP and hundreds of integrations. Useful for connecting research, approvals, content, analytics, CRM and reporting into one operating system. → github.com/n8n-io/n8n Activepieces An AI workflow and MCP automation platform with a large connector library and self-hosted options. Useful for teams that want a more guided alternative to building every integration from scratch. → github.com/activepieces/acti… THE SEVEN I WOULD START WITH marketingskills google/skills claude-seo chrome-devtools-mcp google-ads-mcp google-analytics-mcp promptfoo Then add tools only when a real bottleneck requires them. Before installing anything: Pin a release or commit. Do not blindly run @latest in production. Read the README, package scripts, install scripts and every SKILL.md. Run browser tools in a separate profile or container. Use read-only OAuth scopes wherever possible. Put human approval in front of bids, budgets, ads and campaign mutations. Never paste client credentials directly into prompts. Do not install all 35 just because someone made a list. All 35 repositories are public. That does not mean every hosted service, API request, model, integration or commercial use is free. Public is not the same thing as open source. Source available is not the same thing as unrestricted. And “official” still does not mean “give it production write access.”
16
2
15
2,930
i would bet 80–90% of SMBs running google or meta ads have junk in their conversion data. Spam calls, bots hitting chat widgets, good ol’ form spam. If you’re counting those as leads, your attribution can be spot on and you’re still bidding on garbage. How many agency owners and PPC managers send clients a monthly report and have to guess how many of those “leads” were real?
getting lead tracking DIALED in 🤩
1
13
1) Google and Meta bidding improves a lot when you feed them the actual closed deals and revenue instead of form fills and calls. Most Level 1 and 2 owners never send that data. The upload paths already exist. The hard part is getting your client to record the sale at all. Curious what others are doing about that.
1
7
3) Google takes GCLID or hashed email/phone through Data Manager. Meta takes it through the Conversions API. Low volume can even go through Sheets.
1
2
4/ Has anyone found a low friction way to get this data back into Google and Meta from people who won’t open a CRM? Social engineering might be necessary to get these busy clients to take this seriously.
1
This thread is WILD. Incredible detective work @TheMidasProj
🚨NEW: We just uncovered a covert influence operation using right-wing creators to promote Nvidia chip sales to China. Earlier this year, we reported on similar campaigns, and Republican lawmakers went after Nvidia. But the story didn’t end there. 🧵
51
Gotta stay on your toes. Google can’t seem to make up its mind on AI overview Citations.
Heads-up: Google has again made changes to how AI overview citations show in search results. This is the first time in a while that they have made an update that actually makes the citations more visible compared to the default display. Normally, the first citation would display as clear, then a fade would be applied to lower citations. And then, last month, Google removed the citations from desktop completely. It looks like Google has now reversed this change and is enforcing a default state of having three citations, with there being no fade at the bottom (making the third citation more visible). Google has gone back and forth with changes related to AI overview citations in recent months, and it looks like they are finally using a display that is actually beneficial to websites.
68
Sure, it could be Automated Ai hacking agents breaking the foundation of the internet. But bear with me. It’s far more likely that the people responsible for keeping servers up for our global infrastructure actually just requested off to play World of Warcraft forever. Nobody at these companies can fix the issue, except Dave, who requested the week off to grind his Horde warlock up to level 60.
136
Delete the phrase “Here’s why that matters” and all similar variations from your writing vocabulary. It’s a dead giveaway your text was ai generated. I see tons of humanizer and anti-slop skills and writing guidelines miss it.
116
Don't ship ranking changes during a Google core update. You can't measure them. Especially right now when all position tracking tools seem to be so inaccurate. One started November 10. It can run two weeks. It's the third core update this year, on top of four spam updates. Any title or copy change you ship inside that window is contaminated. Rankings move, and you can't split your edit from Google's. What I changed the day it started: 1. Froze rewrites of titles, H1s and body copy on every site we run. 2. Kept correctness fixes shipping: indexing errors, broken links, schema errors. 3. Set the clean date at completion plus 7 days. No verdict on any site before then. 4. Wrote the hold into every agent loop. Each run reads it before it picks a change. Number 4 was the real work. An agent wants to ship. So a held change gets logged as held, not skipped and not failed. A run that ships nothing is a correct run. Why I'm strict about it: in August a spam update finished, and a contractor site I track lost 97% of its search impressions that same day. It still ranked first for its own name. It had 220 money/ location pages, 87.9% identical. If we'd been rewriting pages that week, I couldn't have told you which one did it. The hard part of automating SEO isn't getting agents to do more. It's giving them a reason to wait.
1
1
153
Git tells you what changed. It doesn't tell you why. When your agents ship dozens of changes a day, the why is the first thing you lose. In SEO that gets expensive. Rankings drop 3 weeks later and you can't tell which edit did it. Keep a change log. What, when, why.
53
Google started a core update on November 10. Two days later it made Gemini 4 Hydrogen the default in AI Mode. That model is now handling the hardest questions in AI Overviews. If you run a service company, both of those hit in the same week. Rankings can move while the system answering the hard questions changes underneath them. The temptation is to tear the site up because the graph looks ugly. I would wait. Google has said this update can take two weeks. The comparison that means something is the week after it finishes, set against the week before November 10. We are almost a week into a rollout that can run two. This is a bad time to rebuild the site. I use this stuff constantly, and I am not against the new model. Hydrogen is supposed to connect people to a broader set of links than before. That only helps if the page it can link to describes the job. The town you work in. What was wrong when you showed up. What the customer was worried about. A homepage that introduces the company and stops there does not give a newer model much to work with. So this week I am not buying a recovery package, and I am not rewriting pages because a line moved. I am reading the service pages the way a person reads them when they have to hire someone. If the page would not help me pick the company, it will not help the model either. Search Console can wait until the update is actually done. Then I will compare that calm week with the week before November 10, and change the pages that were thin before any of this started.
2
4
158
X just dropped the ranking weight for copying a post’s link from 20 to zero. If you’re growing a small account, you can take “copy the link to boost this” out of your posts. A share through DMs still has a weight of 5. Give somebody a reason to send the post to their business partner. Here’s what X’s public For You code shows after the November update. -Copy link went from 20 to 0 -Replies, quote posts, and DM shares are each weighted at 5 -A follow is weighted at 4 -A like is weighted at 0.708, up from 0.66 earlier this month -On an original post, the reply weight can reach 20 for a viewer who follows you and whom you follow back These are weights on what X predicts each person will do when they see your post. If X thinks someone is likely to reply, that prediction carries a weight of 5. The score is personalized to the viewer, so dividing 5 by 0.708 won’t tell you how many likes to trade for a reply. The part I’d spend time on is getting people to finish the post. X says it trained the system that finds posts from accounts you don’t follow to surface things people read all the way through. You can give someone the answer in the first line and still give them a reason to keep reading. Show them how you got there. Say you run Google Ads and the dashboard says you’re getting leads, but some of them never show up in the CRM. Show how you’d trace one test submission from the form to the thank-you page to the CRM. Show where it can break and what you’d check first. That’s the kind of post someone sends to the person who runs their ads. It also gives them a reason to follow you. They know what they’ll get from you the next time their numbers look off. Same with video. Opening a video has a weight of zero on its own. If you’re showing a broken lead form, put the form on screen in the first few seconds. Walk me through what happens after you hit submit. I can figure out who you are after I see whether you know how to fix the problem. And I use AI constantly, but I don’t see the point of using it to pump out “Great insights!” under 50 posts. Use it to clean up a screen recording or turn your notes into a draft. When you reply to somebody, add the thing you’d actually say if you were talking shop with them. If you’ve got 30 minutes to spend on X today, pull up something you worked on this week. Show what went wrong, what you checked, and what you’d do first if it happened again. Then stick around long enough to answer the people who have questions.
149
The model doing your SEO work gets swapped out every few months. Crawl data goes stale in a week. The record of what changed, why, and what happened after is the part that should get more useful with age. Those who are tracking what was done by AI, why it was done, and what affects it had, will have a goldmine of data to ensure the next future model can perform at max performance.
1
143
🚨 Many major services appear to be down or experiencing disruptions, with user reports spiking on Downdetector: • Microsoft Outlook • Microsoft 365 • Microsoft Store • Microsoft Copilot • Microsoft Teams • Amazon Web Services (AWS) • Microsoft Azure • Gmail • Microsoft OneDrive • Google • Microsoft Account • Cloudflare • Zoom • GoDaddy • Xfinity • Xbox Network • Steam • Amazon • Calendly • Salesforce • Claude AI • Yahoo Mail • X (Twitter) • Facebook • Archive of Our Own • Slack • Microsoft Dynamics • Microsoft Entra The scope and cause remain unclear.
214
1,541
390
4,497
984,461
The actual home upgrade market, stripped of marketing theory: 60-70% of the click traffic and impression traffic comes from people over 65. By a large margin. They own the houses (often paid off), they have the equity, and they’re the ones aging in place or doing the real renovations. The algorithms and creative teams keep chasing younger cohorts who mostly rent or do small projects. The money doesn’t.
109
Jensen Huang the matrix has you
89
“Why don’t you just show up and act like nothing happened?”
🇺🇸🇨🇳 Huang reportedly returned to work like NOTHING happened after admitting to an affair with a woman suspected of spying for China. He apologized to executives on Monday, then immediately moved on to the usual business agenda. Meanwhile: •⁠ ⁠Nvidia has reportedly lost over $400 BILLION** in market value this week •⁠ ⁠The company's board has launched an investigation •⁠ ⁠The FBI is reportedly examining the relationship And a U.S. senator is already calling for Huang to resign over national security concerns... Source: NY Post / Writer: Samuel
70
LinkedIn is dead
154
Verified my ChatGPT business profile. Easier than Google. No live video. But adding products is about all it does. So far...
72