How I marketed LLM Gateway with AI agents: the routines, the reports, the skills, the brand and the pages that kept compounding. June 2025 → October 2026.
IG
Ismail Ghallou · Smakosh. Self-taught designer and developer, open-source person. Co-founded LLM Gateway with Luca Steeb in May 2025: Luca on the gateway core and infra, me on product surfaces, billing, the playground, SOC 2 and everything growth.
Claude Code sessions4,641 commitsPostHog via MCPSearch ConsoleDiscord #trafficadmin dashboard
Sources: admin dashboard (Oct 3 2026) · theopenco/llmgateway git history · Claude sessions · smakosh.com/blog
The journey · paying customers, from admin
Eleven slow months. Then +3,000 in four.
JUN 8 2025
First paying customer. 43 sign-ups.
JUL 2025
Launch: 9,441 people. 10 paying by Jul 29.
NOV 18 2025
100 paying customers.
DEC 2025
New brand, logo system, model timeline.
JAN 2026
DevPass launches quietly. OpenCode guide.
MAY 20 2026
500 paying. Search starts compounding.
JUL 7 2026 1,000 · DevPass jump
AUG 3 2,000
SEP 3 3,000
OCT 3 20263,649
Paying customers + sign-ups: admin, cumulative, read Oct 3 2026 · traffic: "17 months of marketing" post
The thesis
Almost all of the marketing was built by agents. I reviewed it.
The SEO audit moved to a Claude routine. The traffic report stayed a plain GitHub Action: HogQL plus formatting, nothing to hallucinate. When I want a story, I paste it into Claude and get an artifact (Oct 1, September report).
Traffic Bot APP
📈 Monthly Traffic Report · September 2026
vs previous month
Product
Visitors
Δ
Gateway
●●●●●●●
▲
DevPass
●●●●●
▲
Lounge
●●
·
AirSide
●●
▲
Docs
●●●
▲
Illustrative layout · real values stay in Discord
Routine 3 · Search Console → posts
Google compounded. Every quarter.
People from Google per quarter. 33.4K clicks from 2.33M impressions over 16 months.
1 · Export (manual, for now)
Queries, pages, 404s and snippet warnings from Search Console.
2 · Claude reads the gaps
Queries we show for but don't answer → topic clusters.
House style, written down once, reused by every agent.
.claude/skills/blog · #2742
blog
Read the docs page and the shipping PR first
Pick one primary SEO keyword: title, summary, early paragraph, an H2
Frontmatter checked by content-collections
gpt-image-2 OG image
Never invent prices, limits or quotes
.claude/skills/changelog · Jun 2026
changelog
Dated entry in apps/ui/src/content/changelog
Confirms plan gating: free, Pro, Enterprise
Exact field names and error codes from the code
Hands back the OG image prompt
Tags + pagination later (#4207)
Installed skills
The marketing kit
marketing-skills · product-marketing
seo-audit · ai-seo
frontend-design · brand-guidelines
A portable marketing skill distilled from the retrospective, reviewed by a second model
you › publish new articles regarding AirSide, one of them is how to make money with inference through AirSide. use marketing-skills, seo-audit and ai-seo skills. create a PR once done · Sep 13 → #4057
A month of prompts
One line from me. A PR from the agent.
Sep 5 redesign the model census page… keep the airport/passport theme… run an seo-audit and ai-seo audit, a11y check, then create the PR → #3950
Oct 1 make an artifact with details about this: 📈 Monthly Traffic Report · September 2026
Sep 18 redesign og images for providers pages… use the model family in model page og image
Sep 20 update all our comparison pages and migration guides, search the web and keep things up to date → #4262
Sep 22 extend the SCX promo 15 days and make an announcement blog post → #4193
Sep 22 the H3 rapper-swap video trend… Steve Jobs and Bill Gates → #4192 blog guide
Sep 27 finish the landing redesign… make the copy FTC compliant, use marketing-skills and frontend design → #4261
Oct 3 fix OG images… kimi-k3 shows Bedrock regional pricing instead of global
The new channel
AI assistants are the fastest-rising source.
0 → 0
people / month from ChatGPT, Gemini, Claude, Perplexity, Copilot, Grok
0
people tagged utm_source=chatgpt.com that referrers miss
0
monitored prompts for AI search visibility
May 2026 · #2427
llms.txt everywhere
Enriched llms.txt + sitemap page, then per product (#3415).
Aug–Sep 2026
Agent readiness
#3769 and #3951: discovery files, API contracts, markdown pages agents can read.
Distribution
Inside the tools devs use
Built-in provider in OpenCode, guides for Claude Code, Conductor, VS Code, TanStack AI.
Then AI audited it all
17 months of marketing, checked by an agent, checked by me.
0
editorial files read
0
workstreams found
0
slide public report
0
financial data used
Good at
HogQL against PostHog via MCP, month by month from May 30, 2025
Every number traceable to a source
Rebuilding report and deck from one dataset
Needed me
MCP hit the wrong PostHog project: only 11 days of data
First draft compared one week to the last. Accurate, useless.
Sign-up event fires on ~16% of paths, so no conversion rates
Where it landed · Oct 3 2026
From 1 to 3,649.
0
paying customers · first on Jun 8 2025
0
sign-ups · 20,659 verified
0M
API requests served, all time
0
people / month in Sep 2026 · 1,522 in Jun 2025
Jun 2025 → Dec 2025
1 → 160
Jan → May 2026
198 → 571
Jun → Aug 2026
890 → 2,900
Sep → Oct 3 2026
3,612 → 3,649
Two jumps. The launch faded. The one built on search, DevPass and a comparison article held and kept growing.
admin dashboard, all time, cumulative at month end · traffic from PostHog via the report
What I keep
"AI marketing output is cheap to produce and just as cheap to get wrong. The review is the part that matters."
next
Keep feeding search
Catalogue-backed pages refreshed as models ship. Rework stalled compare pages around real queries. Automate the Search Console export.
next
Fix the signals
Sign-up event on every auth path, real options in the source survey, AI assistants as one channel next to Google.
next
One real experiment
Matched cohorts, a fixed window. No more tea leaves from week-over-week charts.