there's a KDD paper on GEO that almost nobody is talking about. it found that letting AI rewrite your blogs cut citations by 22%. this paper uses 171k real web pages, 2,700 queries, 9 domains, and a full retrieval → reranking → answer pipeline. tl;dr: GEO optimizations to content are bullshit. To titles, meta descriptions, schema, maybe not 👀 What to actually do, with their numbers: 1. put the page's hard numbers in the meta description. retrieval +29%, citations +7%, the best single tactic in the study. the same stats added to the body cut citations 4%. 2. name the entity in the title. "Panel Clarifies Advice" → "IOM Panel Clarifies Vitamin D Guidance". retrieval +27%. 3. use plain language. simpler wording in metadata: retrieval +27%, citations +5%. adding jargon to the body: retrieval -14%. 4. never let AI rewrite your pages. retrieval -36%, reranking -42%, citations -22%. 5. answer in the first paragraph, use the product name instead of "it", and make every claim stand on its own. the paper's best method was built on rules like these. it got cited 58% of the time vs 50%, and a full position earlier in the answer. 6. only the top 10 results ever reach the model. 5.8% of pages "optimized" themselves from #10 to #11 and disappeared. If a page already gets cited, be careful touching it. Personally, I'm always skeptical of stat-stuffed titles and descriptions, but I'll be trying some of these out.

here's the cold email approach that took us from 0 to 30K mrr, and growing, in 2 months during YC. you're cooked if you do mass spam outreach. i only do super-personalized outreach (10 emails/week), that makes my prospects say "timing is insane, it's a top1 problem for me". here's how: 1/ you need to be clear on who your ICP is. and you need to be really specific with it. for instance, we help devtool companies with devrel and AEO. that's about as niche as it gets. could we help all B2B SaaS with AEO? Yeah. but we don't because specificity really sets you apart, especially now that anyone can build anything. 2/ don't try to convince people not looking to buy i think that's the misconception most people hold about sales. i'd rather convince a hot lead to go with me, than to convince someone random why they have that problem and why we are the solution for it. it's double the work. I try to find signals that show me my ICP is facing the problem and they're ready to buy. 3/ Sales like dating is about timing and chemistry timing = is this a burning pants problem for them? chemistry = is your email/offer actually good? on timing, the signals i look for are specifically funding in <3 months AND they have bad docs AND they're hiring a devrel/educator/etc. these signals are proof my ICP is aware of the problem we solve (devrel) and they have the money to solve it. 4/ chemistry the best emails are insanely helpful and don't sound like slop. I do personalized audits for each company i'm targeting and send it to them. we even built a tool to automate this process: docs[dot]review then, I scour the buyer's twitter/linkedin etc. for hours to find some unique point that i can use in my email. sometimes it's things like they're tennis fans, or sometimes its more business-y like their opinion on a specific topic. my email is also very value-driven. i typically just send: " hey xyz, i was oging through your docs and found 3 errors: - error #1 - error #2 - error #3 we do docs for ABC, can send an audit if you're interested " it's short, doesn't sound like slop, and establishes credibility. -- the ICP point again is really important cause you can only get the timing and offer right when you can think like the person you're selling to. i send max 10-15 emails a week and generate replies >50% of them. my first-meeting to pilot rate also hovers around 80% because the prospects that come in are already qualified. you just need to be gritty and keep doing it, even when it's boring.

if you're cracked, come work with us

Happy #MarTechMonday. Each week I spotlight a tool reshaping the modern GTM stack. This week's question: is your product invisible to AI? We talk about GEO for content, but rarely for technical docs. @manicule (YC P26) is pioneering "AgentRel," structuring API docs so AI agents can discover, read, and integrate them without a human in the loop. If agents are making buying and integration calls in 2026, your docs need to be machine-legible. Check it out: manicule.dev Who else is building for AgentRel?

You can read a diff and still not know what it means for the person using the product. The answer is in a map you built before the change landed, or it's nowhere.

good thing we don't publish claudeslop for customers.

A hypothesis we're testing right now is that agents tend to recommend tools/approaches that are highly extendable, since the main value prop of agents is the personalization they bring to their user's situation. maybe it's true - we have a feeling it will be. brb with the numbers. what this also means is if you market your tool as the most extendable/easy to customize, agents might recommend you more.

reddit + smart keyword targeting + brand mentions is really all you need to be the #1 AI recommended tool. not to mention the ability to do this for 3-6 months consistently.

or just hire manicule to do everything and prove roi to u

"I'll do the selling. Just get me a meeting." That's what @NamanBansal0611 asked @shreyansj, then a founding engineer @supermemory, for. No product, no raise, just samples and one warm intro. He wrote their entire blog, ran SEO and GEO, created their docs. It went viral, and the inbound started: @greptile, @reductoai Naman Bansal on how one customer became @manicule

we've come up with a structure that seems to work great for writing agent-optimized docs. it took one of our customers from 38% coding agent success on docs to 92%. here's how it goes: (before any of this you need to be clear on what outcome you want the agent to be able to achieve from that particular page) --- # Title description + key outcome. ## Do this - one snippet combining all the code needed to get to the decided outcome. (ideally cURL). ## A table with SDK snippets for the same. ## Facts a bulleted list of facts about how the api works and what it accepts + common gotchas. ## Not supported mine your existing data to see where agents are erroring most commonly, and include those common errors here. ## Verify how can the agent verify if it was able to accomplish the outcome on this page? ## Related links to other pages you want the agent to visit. --- real-time code verification is a big part of this. our system goes through each customer's docs, runs every code snippet, and then works on fixing each error. if you serve wrong code to agents, you're cooked anyways.

Agent-optimized technical docs are all the rage right now. We recently re-did @surgeapi's docs and made them more agent-optimized. Their agent traffic has since exploded 4x.

We're hiring an awesome X/Reddit content marketer. Somebody who's written great articles and threads on X and Reddit, ideally in devtools. Looking to close in <24 hrs; contract to start; negotiable pay. Send your best X and Reddit work to naman[at]manicule[.]dev

After going through YC, I can attest that a large part of building a startup is being lucky, and anyone who says otherwise is suffering from survivorship bias. For instance, five years ago, I started with creative writing (wrote my first book after reading Harry Potter!). Then, I somehow got into software engineering. Those felt like two unrelated paths until I combined them, stumbled into DevRel, and eventually started @manicule. Good DevRel requires people who understand engineering deeply enough to earn a developer's trust, but also have the marketing intuition and writing ability to make technical ideas interesting. There's a reason companies are paying $250K+ for this role. Most people spend years getting great at one side. Very few accidentally become good at both. I got lucky that my weird path forced me to develop that combination. Then, I got lucky again with the team. @shreyansj was already a friend and is genuinely the best engineer I know. He pushes ~15 PRs a day, to the point where we got hit with a $1,300 @greptile bill yesterday 😭 I found @ritiksahni22 almost randomly. I read his writing on Twitter, reached out, and he turned out to be great at the exact combination we needed. @knlmsh and @ninadwrites came together through similarly unplanned collisions. This is a large part of why I think @manicule is the best team to solve this specific problem of high-quality technical content. Everyone here has lived both halves of it. We can understand the product deeply, turn that understanding into writing developers actually trust, and figure out how to get it in front of them. So yes, startups are built on luck. But the part you control is how often you put yourself in situations where luck can find you, and how you act on it. Write before there's an obvious payoff. Learn skills that don't neatly fit together. Reach out to people whose work you admire. Take enough intelligent bets without knowing where each one leads. Most of those bets will go nowhere, and a few might end up becoming a cool company.

extremely bullish on digital ads on everything. I love this @manicule

I made GPT believe it's Claude inside Claude Code by setting up CLIProxyAPI routing: - Fable -> Fable - Opus -> 5.6 Sol (high) - Sonnet -> 5.6 Sol (medium) - Haiku -> 5.6 Terra (medium) and it auto-uses API credits (for my YC bros) after you exhaust limits! Setup Prompt 🧵

@manicule is neat. Docs are really hard. So much content to create, prose and code snippets to keep in sync, diverse audiences to consider… Docs are also pretty fun to work on if you give yourself the time. There’s a particular joy in explaining things well.

hiring hiring