--- title: 17 months of marketing LLM Gateway, audited by AI date: 2026-09-27T15:00:00.000Z edited: 2026-09-27T15:00:00.000Z author: Ismail Ghallou (Smakosh) canonical: https://smakosh.com/blog/llm-gateway-marketing-takeaways reading_time: 5 min (1175 words) Tags: ai, growth --- # 17 months of marketing LLM Gateway, audited by AI ![People per month across every LLM Gateway site, May 2025 to September 2026](/assets/blog/llmgateway-marketing-report.jpg) Most of the marketing behind [LLM Gateway](https://llmgateway.io) was built with AI agents: the blog posts, the comparison pages, the model catalogue pages, the docs, the launch videos, the SEO audits. So when I wanted to know what actually worked, it felt right to hand the audit to AI as well, and then check its homework. The result is a public report: **[LLM Gateway marketing retrospective](https://smakosh.github.io/llmgateway-marketing-report/)**, plus an [18-slide deck](https://smakosh.github.io/llmgateway-marketing-report/presentation.html). This post is the short version. ## How the report was made I gave a coding agent two sources and one rule. **The codebase.** Everything we shipped for marketing lives in the monorepo: 98 blog posts, 109 changelog entries, 23 setup guides, 5 migration guides, 6 DevPass comparison articles and 143 product documentation pages. The agent read all of it and grouped the work into 16 workstreams, from positioning and naming to referrals, lifecycle email and provider promotions. **PostHog, through its MCP server.** Instead of screenshots of dashboards, the agent queried PostHog directly with HogQL, month by month, from the first tracked day (May 30, 2025) to today. Traffic by product, referrers by quarter, content by page type, AI-assistant referrals, UTM tags, self-reported sources and product milestones. **The rule: no financial data.** Revenue, spend, purchases and paid-plan starts stayed out. Everything in the report is people, pageviews and events. The same session also turned the lessons into a reusable marketing skill for agents, which a second model (Claude Fable 5.1) reviewed before it shipped. The report itself I reviewed by hand. It says plainly what it can't prove: these are observations, not experiments, and none of the numbers are causal. ## The headline In June 2025, **1,522 people** visited our sites. In the first 27 days of September 2026, **37,011** did. There were two jumps. The July 2025 launch brought 9,441 people, and two months later we were back to 3,148. The second jump came in July 2026, with DevPass and a coding-plans comparison article, and that one held and kept growing. The difference between those two jumps is basically the whole report. ## 1. Google compounds. Social doesn't. People arriving from Google, per quarter: | Quarter | People from Google | | ------- | ------------------ | | Q3 2025 | 2,985 | | Q4 2025 | 5,618 | | Q1 2026 | 9,132 | | Q2 2026 | 18,163 | | Q3 2026 | 40,071 | It grew every single quarter. Social and community referrals did the opposite: 5,812 people in the launch quarter, then roughly 2,400 a quarter for a year. Launch posts and threads feel like marketing. Search was what actually kept growing. ## 2. Pages built from live data beat one-off posts Grouping llmgateway.io pages by type, people per month, February vs September 2026: | Page type | Feb 2026 | Sep 2026 | | ------------------- | -------- | -------- | | Model pages | 1,909 | 9,276 | | Blog | 480 | 5,426 | | Tools | 22 | 5,396 | | Provider pages | 441 | 2,441 | | Guides | 141 | 1,254 | | Compare & migration | 709 | 611 | "Tools" means the model timeline, rankings and cost calculators. They're generated from the live catalogue, so every model we add makes them better without anyone writing a word. Same for model and provider pages. The comparison and migration pages are the opposite: hand-written, static, and flat since February. We have a rule in the repo that docs and marketing copy never hardcode lists of models or providers, because they go stale the day the catalogue changes. The data says the same rule applies to whole pages. This is also where AI earns its keep. An agent can write a hundred comparison pages in an afternoon, and it would be wasted work. What was worth the effort was building pages that read from the live catalogue, so they stay current without anyone touching them. ## 3. Launches and newsletters are spikes Tagged newsletter traffic over the whole period, in people: - TLDR Founders: 2,092 - ConvertKit: 1,366 - TLDR WebDev: 516 - The Neuron: 366 Each one was a single burst. Nobody kept coming from those links after the issue went out. So now I plan launches and mentions around a page that keeps earning search traffic afterwards, and I send every spike visitor to one clear first step. ## 4. AI assistants are the fastest-rising source People referred by ChatGPT, Gemini, Claude, Perplexity, Copilot and Grok went from **10 a month** in June 2025 to **450** in September 2026. Gemini alone went from 10 to 180 in six months. The referrer count also undercounts. ChatGPT tags its outbound links with `utm_source=chatgpt.com`, and that tag alone brought **2,237 people** the referrer data misses. If you track AI discovery, merge both signals into one channel. It's still small next to Google. But Google took a year to compound, and this looks like the start of the same curve. ## 5. Fix the signals before you trust a funnel This part was humbling. The audit found that: - PostHog's sign-up event only fires for about **16%** of the sign-ups in our admin panel. Most auth paths never send it. - **52%** of answers to "where did you hear about us" are "Other", because the options don't include Google, AI assistants or newsletters, which the referrer data shows are our biggest sources. - Click tracking only started in March 2026, so earlier funnels can't be rebuilt. That's why the report has no conversion rates. Any rate built on those events would have looked precise and been wrong. ## What AI was good at, and where it needed me The agent was great at the tedious parts: reading 245 editorial files, writing the HogQL queries, keeping every number traceable to a source and rebuilding the report and deck from the same data. It also made mistakes a human might not catch. The PostHog MCP connected to the wrong project at first, and the only reason we noticed was that "May 2025 to now" returned eleven days of data. The first version of the report compared just one week to the week before, which was accurate but told me nothing useful. Both got fixed because I read the output instead of trusting it. Treat AI marketing output the same way: it's cheap to produce and just as cheap to get wrong. The review is the part that matters. ## What we're doing next 1. **Keep feeding search** with catalogue-backed pages, refreshed as models ship, and rework the stalled comparison pages around what people actually search for. 2. **Fix the signals**: fire the sign-up event on every auth path, add real options to the source survey, and track AI assistants as one channel next to Google. 3. **Then run one real experiment** with matched cohorts and a fixed window, instead of reading tea leaves from week-over-week charts. ## The full report Everything above, plus the 16 workstreams, the timeline, what the audit can't prove and the raw data, is in the report: [Open the report in a new tab](https://smakosh.github.io/llmgateway-marketing-report/) · [View the slides](https://smakosh.github.io/llmgateway-marketing-report/presentation.html) · [Download the data](https://smakosh.github.io/llmgateway-marketing-report/analytics.json) --- _Originally published at https://smakosh.com/blog/llm-gateway-marketing-takeaways_