Finding Turf Twitter

I scraped my Twitter network to find high-signal turfgrass people. Here are the top 100 accounts posting original content, participating in conversations, and maintaining positive tone. I’ll walk you through how I did this.

Turf Twitter is dead

Many have said Twitter feels dead or less useful for turfgrass conversations. The majority believe Turf Twitter lost its “zeal.” Turfnet published multiple articles lamenting the decline. Micah Woods used to run annual analyses of #TurfTwitter accounts tracking the most active contributors, but X’s new, restrictive API plans make this impossible. I stopped tweeting for 2.5 years. I used Twitter to stay informed of industry changes, learn about ongoing projects, and build connections. This behavior was common across industries. During this decline, Linkedin and Bluesky came up repeatedly as alternatives. And of course, the scientists have written a paper about this.

Finding high-signal people

I wanted to see if I could find people still sharing interesting things on Twitter/X. This is a signal vs. noise problem. High-signal posts show me something useful. For me, a high-signal feed looks like people:

1. Talking about turfgrass
2. Contributing a mix of original content and retweets/replies/reposts
3. Being positive

Noise is everything else, like politics, sports, religion, or hobbies. Some accounts are noisier than others.

Inspired by Swyx’s scraping your Twitter network and Micah’s Turf Twitter work, I set out to refresh my feed. While Micah tracked influence across Turf Twitter and identified top accounts, I’m focused on content quality. I scraped my Twitter/X network and used LLMs to score accounts on topic relevance and sentiment. Results are in dropdown tables near the bottom.

Graphs and network analysis

I read up on network analysis. Graphs are everywhere and everything is a graph. Social networks are graph structures: nodes (people) connected by edges (follows). My first hop connections are accounts I directly follow, and second hop connections are accounts they follow (“friends of friends”). I can map my existing network, then expand outward to find high-quality accounts already vetted by people I trust.

Linkedin uses graph databases to recommend people you may know based on mutual connections, shared employers, etc. The same approach applies to Twitter. If ten turfgrass people I trust all follow the same golf course superintendent, that’s a strong signal I should follow them too.

Methods

I’ll walk through how I identified quality turfgrass accounts. To summarize: I scraped my network and expanded to second-degree connections, classified accounts by turfgrass relevance, filtered for activity, then scored based on content quality and sentiment.

1. Scrape first hop accounts

I started by mapping my immediate network, then expanding one degree out. I used the open-source twscrape API to scrape data on the 1,128 accounts I already follow, including their handle, name, bio, url, location, following/follower count, and if verified or not. This took 31 minutes and cost nothing except time.

2. Filter to turfgrass accounts

I took each account’s information for handle, name, and bio and used this to classify accounts as turfgrass-related or not. I used the LLM gpt-4.1-mini to do this with the prompt below. The LLM labeled each account as turfgrass, not_turfgrass, or unknown, plus a category label for industry, academic, golf, sports, lawn_care, business, media, or other. This took 21 minutes and cost $0.43.

You are an expert in the turfgrass industry and classifying social media accounts. Your task is to classify Twitter accounts based on their relevance to turfgrass.

TURFGRASS CONTEXT:
Turfgrass refers to grasses used for lawns, golf courses, sports fields, and other managed grass areas. Related fields include:
- Turfgrass breeding, genetics, pathology, entomology
- Urban grasslands
- Golf course/sports field management
- Lawn care and landscape maintenance
- Turfgrass equipment and chemical suppliers
- Academic research in turfgrass science
- Extension education for turfgrass professionals

CLASSIFICATION GUIDELINES:
- yes: Clearly turfgrass-focused (mentions turfgrass, golf course management, sports turf, lawn care professionals)
- no: Unrelated to turfgrass (personal accounts, other industries, general agriculture without turf focus)
- unknown: Uncertain or insufficient information to determine turfgrass relevance

ROLE CATEGORIES (select the single best fit):
- industry: Turfgrass industry professionals, consultants, chemical or seed company employees, employees of professional organizations like USGA or GCSAA, must be turfgrass industry related; must be an individual human in the name field, not a commercial/business account
- academic: University researchers or scientists, professors, graduate students, university research centers, extension groups or individuals, must be turfgrass academic related
- golf: Golf course superintendents, assistant superintendents, equipment managers, spray techs, must work on a golf course
- sports: Sports field managers, stadium groundskeepers, sports field employers, also includes parks and community sports fields, must work on sports/park fields
- lawn_care: Lawn care operators, landscape maintenance, university or campus grounds manager, must work on lawns only
- business: General business/commercial accounts related to turfgrass (like Toro, DLF Seed, Syngenta, USGA); must be a commercial or company account in the name field, not individual person's account
- media: Turfgrass publications, journalists covering the golf and turfgrass industry, podcasts and podcast hosts
- other: Turfgrass-related but doesn't fit other categories
- non_turfgrass: Clearly not turfgrass related, could be software and tech people, fashion, pest control or gardening only, media that is not turfgrass or golf oriented

Analyze the account information and provide a structured classification with the single most appropriate role.

Please classify this Twitter account for turfgrass relevance, role, and reasoning:

Handle: @[handle]
Name: [name]
Bio: [bio]
Location: [location]

When that was done, I took 10 minutes to skim the labeling results and manually corrected about 50 misclassified accounts. Fifty out of 1,128 total means the LLM was over 95% accurate. Good enough. My first hop (who I’m following) included 448 turfgrass-related accounts, or 40% of all accounts I’m following. I otherwise follow a lot of non-turf scientists and tech people.

3. Scrape second hop accounts and filter

Next, I took that subset of 448 turfgrass accounts and scraped everyone they follow (second hop). The second hop included 115,298 unique accounts after deduplication. This is the power of network analysis, where I can access 250 times more potential connections. Getting the profile data for 115k accounts took over 57 hours (2 days and 9 hours) of active scraping. I ran sessions overnight for a week.

I labeled these accounts as turfgrass, not_turfgrass, or unknown, plus category, using the same LLM approach from step 2. This took 6 hours and cost $23.27 using the OpenAI batch endpoint. This resulted in 16,383 turfgrass accounts collected in the second hop, or 14% of the total. You can see a wider net captures less relevant accounts.

Here’s the category breakdown in a table, for first and second hop accounts. Golf dominates the Turf Twitter population, and I overindex on academics.

4. Scrape recent tweets from turfgrass accounts

Profile data alone isn’t enough to assess quality, I need actual tweet content. I used the twscrape API to get recent tweets for all turfgrass accounts. The API returns 10-80 tweets per account depending on activity and pagination. At this point I dropped private or empty accounts because I could not access their tweets. I fetched a total of 495,077 tweets from 15,724 turfgrass accounts. Tweet scraping took 28 hours.

I cleaned the tweet data to include text, date, and tweet_type (original, reply, retweet, or quote). The scraped data included parent tweets, which are the quoted or replied-to tweets not authored by the target account. I dropped these parent tweets and filtered to only the last 15 tweets for each account. After filtering I had 222,113 tweets for 15,724 turfgrass accounts

5. Apply hard filters

I applied some hard filters to drop accounts that are not active or engaged:

1. Drop any account that has not posted in the past 6 months (since April 2025).
2. Drop any accounts that did not have at least 1 original tweet. I want accounts that post unique content and not just retweets.
3. Drop any accounts that did not have at least 1 reply, retweet, or quote. I want accounts that are participating in conversations with others.

Only 60% of the turfgrass accounts I follow have posted in past 6 months. That means I’m following almost 200 inactive accounts. I went ahead and unfollowed them. After the hard filtering step, I had 4,776 turfgrass accounts that passed the hard filters.

6. Apply relevance scoring

The last step was to score accounts based on content quality to find the highest-signal users. I wanted a single metric combining turfgrass relevance (% turf tweets) and sentiment (% positive/neutral).

I used gpt-4.1 to classify each account’s 15 recent tweets as turfgrass, non_turfgrass, or unknown. Ambiguous tweets, like one-word replies, get labeled unknown. I used batch endpoints again, so it took 4 hours and cost $9.56. All accounts had at least 1 turf tweet in their recent 15, and an average of 8.5 turf tweets per account.

I analyzed user tone to figure out who is more positive or negative. Negative people make feeds unpleasant regardless of topic. I used Twitter-roBERTa-base to analyze each account’s 15 recent tweets. This NLP model is trained on Twitter data and outputs positive, neutral, or negative sentiment for each tweet. This is a public model and took 1 hour and 12 minutes to run. The count below shows seniment across all tweets. People are rarely negative.

I computed a final score as the proportion_turfgrass_tweets × proportion_positive_or_neutral_tweets. Each account gets a score between 0 and 1, and higher is better. I could do some sophisticated weighting or include other metrics, but this is a good starting point.

Results

Here are the top-ranking accounts to follow today. These accounts are on topic, sharing unique information, participating in the digital town square, and positive. When I viewed all the data together, the top-ranking accounts were dominated by businesses. It makes sense, a company social media account would stay on topic and be professional (positive), though the content is more generic. I’m really interested in the individuals and non-business accounts, so I split the results to separate them.

The rankings had many ties at the top scores. I used mutual following (how many people I follow also follow that account) as a tie breaker to rank accounts with identical scores. This goes back to the idea of network effects and ranking accounts by people I trust. The tables below show the top 50 accounts in each group.

Giving Turf Twitter another chance

Is Turf Twitter dead? Not completely. This analysis surfaced 100 accounts still sharing quality content, so there’s signal if you know where to look. I’m going to follow these folks and give it a year. If you’re looking to rebuild or curate your own Turf Twitter feed, start with the lists above. Will this bring back the value Twitter used to provide? I’m cautiously optimistic about the platform’s direction. Twitter/X is still eroded, but now I have a refreshed starting point instead of relying on whatever the feed decides to show me. Let me know who I’m missing from the list.