Building LLaMaRush 🚀 AI SEO & content automation Sharing what I learn building SaaS from 0 → 1
Bharat
Joined September 2025
- Tweets149
- Following49
- Followers14
- Likes598
Pinned Tweet
I'm building LLaMaRush - an AI SEO engine for founders who don't have an SEO team.
I'm documenting:
→ what works
→ what fails
→ SEO experiments
→ AI search
→ getting our first customers
Currently: $80 MRR.
Building from here ↓
1
41
Jenish retweeted
We scanned 500+ prompts across ChatGPT, Perplexity, Gemini, Copilot
Here’s who is winning in AI SEO 👇
aiseotracker.com/rankings
🧵 I think AI is creating a weird problem for SaaS SEO:
Content is becoming cheap. Deciding what NOT to publish is becoming valuable.
Anyone can generate 2,000 words now.
The harder question is:
“Will this page actually help my business?”
1
1
14
I think this is where SEO changes.
When content was expensive, the advantage was producing more.
When content becomes nearly free, the advantage shifts to better selection, better information, and better distribution.
1
1
1
Whoever did this, either he is genius or dumb 😂
He is making money by just giving first spot to product.
I checked in the morning it is showing me to pay $14k to claim #1 😂
i just spent $420 to take #1 rank on this site
lets see if it pays off... (300 current active visitors on the site and 150+ clicks in 15 minutes)
1
267
🧵 We helped PhysioCarePMS increase organic impressions by 40% in 28 days from 11.5K → 16.1K.
During the same period, demo bookings also increased by ~60–75%.
The interesting part?
We didn't start by publishing more content.
Here's what we changed ↓
1
1
27
Instead of immediately creating dozens of new articles, we focused on the existing site:
→ Fixed orphan pages
→ Improved internal links
→ Restructured important pages
→ Improved on-page elements
→ Aligned content with search intent
Then we measured the changes.
1
1
3
Good advice by Neil Patel about how to crack marketing and what factors affects the marketing and trust of company.
Must watch!!
Marketing feels complicated. Funnels, algorithms, SEO, ads, AI tools. It seems like you need to learn 20 things before you're even allowed to start. But after a decade managing over $1 billion in marketing spend, I can tell you it really comes down to three things.
In this video I teach marketing from scratch: how to get found where your customers are actually searching, why "click the link in my bio" is quietly costing you sales, and the kind of proof that makes someone actually pull out their wallet and pay you.
This is the same approach my team at NP Digital uses to help some of the biggest brands in the world grow.
1
16
This is actually the workflow I built LLaMaRush around.
Instead of manually digging through GSC for every page, it uses your real Search Console data to find the keywords you’re already close to ranking for, then turns those opportunities into content.
#1 SEO hack to get more traffic:
1. Open GSC → Performance.
2. Select 3 months and open any URL.
3. Filter high impressions keywords.
4. Note those keywords you haven't used.
5. Use them on your existing content.
You will see the traffic growth...
1
14
Your next SEO article might already be hiding in Google Search Console.
Look for keywords where:
→ You're getting impressions
→ You're ranking around positions 8–30
→ The query is relevant to your business
Don't start with a blank keyword list.
1
1
15
Then ask 3 questions:
Does this query match what we sell?
Do we already have a page targeting it?
Can we improve that page instead of creating a new one?
This alone can give you a much better content backlog.
1
1
5
This is where I see the missing layer: data → insight → execution → feedback.
Finding an SEO opportunity is one thing. Turning that signal into content, publishing it, and feeding the results back into the next decision is where the real automation starts.
kimi K3 openrouter, your search console data, google analytics data, and data for SEO API key the most powerful agentic marketing stack
how it works
pipeline pulls the data on a schedule.
warehouse holds it and does the joins.
agent queries the warehouse. that's the only thing it touches.
the agent never calls an api for bulk data. ever. that's the whole trick.
the pipeline
three extractors on a daily cron.
gsc → searchanalytics.query, dimensions [date, page, query, device], rowLimit 25000, paginate on startRow. backfill 16 months once. data lags 3 days so always re-pull the last 7.
ga4 → runReport, dimension landingPagePlusQueryString, metrics sessions + keyEvents + revenue.
dataforseo → not on the cron. on demand only, cached 14 days. serp calls are the recurring bill.
every row passes through one url normalizer on the way in. lowercase, drop protocol, drop www, drop params, drop trailing slash. this is the load-bearing function in the entire system. skip it and gsc never joins ga4 and the agent confidently reports numbers that don't exist.
the warehouse
postgres is fine. duckdb is fine. clickhouse if you're doing this for 50 clients.
three tables:
gsc_daily (date, url_clean, query, clicks, impressions, position, device)
ga4_daily (date, url_clean, sessions, conversions, revenue)
serp_snapshots (fetched_at, query, position, competitor_url, has_ai_overview)
one view on top that joins all three on url_clean. the agent only ever sees the view.
upsert on natural keys so re-pulls are idempotent. you'll re-pull constantly.
the agent
two tools. not four apis.
run_sql — read only role, schema in the system prompt, 500 row cap
fetch_serp — one keyword, live endpoint, writes back to serp_snapshots
openrouter: moonshotai/kimi-k3, routing exacto. without exacto you get routed across providers on different quants and tool calls fail at random.
first job. ship this one only.
monday 6am:
from the joined view, pull pages ranking 5-20 with 500+ impressions and below-average ctr for their position. fetch serps for the top 10. tell me who's above us and whether an ai overview is present. flag anything with zero conversions in 90 days. post to slack.
then add one per week
decay — clicks down 25% mom by url → serp fetch → competitor or layout change
cannibalization — one query, 2+ urls, split impressions
internal links — urls with impressions and no inbound links
rules
aggregate in sql, not in context. 1m context is a cost trap, not a feature.
output goes to slack or a ticket. never a chat window.
log every query the agent writes. that's your eval set.
the model is the least differentiated part. the pipeline and the joins are the whole product.
5
Here is key takeaways and Important points.
Today we released the August 2026 spam update. We'll update our ranking release history page when the rollout is complete: status.search.google.com/inc…
21