A growing share of your potential users will never see your Google ranking, because they didn't Google anything — they asked Claude, ChatGPT, or an agent embedded in their editor "what should I use for X," and took the answer. Agent SEO is the practice of making your product that answer. It's an emerging field, which means 2 things are true at once: nobody has definitive data on what works (anyone claiming a proven "rank #1 in ChatGPT" playbook is selling a course), and the builders who start now are optimizing for a channel their competitors haven't noticed. This post is our current best read — opinionated about direction, hedged on specifics — on how to be the product the machines recommend.
What's actually changing — and how much
First, calibrate. Traditional search isn't dead; Google still handles billions of queries a day. But the trendline is hard to miss: AI assistants now answer a meaningful and growing slice of "what tool should I use" questions, AI Overviews sit on top of a large fraction of Google results, and agents increasingly take actions — researching, comparing, sometimes even signing up — on a user's behalf. Exact market-share numbers vary by study and change monthly, so we won't pretend to know them; the direction is what matters.
The structural shift is this: a search engine returns 10 links and lets you judge. An assistant returns 1–3 recommendations and the user rarely asks for the long list. Being the answer matters more than ranking among answers. Recommendation engines have winner-take-most dynamics, which is exactly why an unknown product should care — there's no page 2 to survive on.
How do assistants pick? Mechanisms differ — some answer from training data, most now also do live retrieval (which leans on conventional search indexes), and agents may crawl your site directly. That mix produces the to-do list below.
Make your site legible to machines
Agents are impatient, literal readers. Optimize for that reader.
Fast, crawlable, server-rendered pages. If your content only exists after a JavaScript framework hydrates, some retrieval systems will see an empty div. Server-render or statically generate anything you want machines to read. Agent crawlers also tend to have tight time budgets — a page that takes 8 seconds to be readable may simply get skipped, though crawler behavior varies and is poorly documented.
Structured, extractable answers. Models lift answers from pages that state them plainly. Clear H2s phrased like real questions, a direct answer in the first sentence under each, tables for comparisons, lists for steps. The writing advice for agent SEO is suspiciously close to the advice for good technical writing — that's not a coincidence, and it means none of this work is wasted on human readers.
Schema markup. JSON-LD for your product, pricing, FAQs, and articles gives machines unambiguous facts instead of inferred ones. Whether assistants consume schema directly today is debatable; search indexes they retrieve from certainly do, so the bet costs an hour and is hedged by design.
llms.txt. An emerging convention — a markdown file at your domain root summarizing what your product is, what it does, and where the key docs live, written for LLM consumption. Adoption by the model providers is uneven and the standard may evolve or fizzle; it's also a 30-minute task that doubles as the best concise description of your product you'll ever write. Do it for the forcing function alone.
Public, indexable docs. Documentation is disproportionately represented in what models read and cite. Docs locked behind a login are invisible to this entire channel.
Be present where agents actually look
When an assistant does live research on "best X for Y," it reads roughly what a diligent human would: review roundups, comparison pages, Reddit threads, Hacker News, directories, docs. Your job is to exist — accurately — in those sources.
Reddit and HN are heavyweight sources. Reddit's licensing deals with AI companies and its dominance in "best tool for..." threads make it plausibly the single most influential surface here, though nobody outside the labs can quantify it. The play is not astroturfing — it's being genuinely present: answer questions in your niche, mention your product where it honestly fits, and make the product good enough that strangers name-drop it unprompted. Strangers' mentions are worth far more than yours, to humans and machines alike.
Directories and listicles. Product Hunt, AlternativeTo, G2, GitHub (an agent reading a README is agent SEO), curated "awesome" lists, niche directories. Individually minor, collectively they're the corpus. An hour of submissions is cheap coverage.
Honest comparison pages on your own site. This is the highest-leverage single tactic we know of right now. When users ask "X vs Y" or "alternatives to X," assistants hunt for comparison content — and most of what exists is either the incumbent's marketing or thin affiliate sludge. Write "YourProduct vs BigIncumbent" pages that are actually honest: a real feature table, and a section on who should pick the competitor instead. Honesty isn't just ethics here; it's strategy. Models are specifically tuned to discount one-sided promotional copy, and a page that concedes real trade-offs reads as the trustworthy source. Lying to a machine that has also read everyone else's page is a losing game.
Consistency across all of it. If your site says one thing and your Product Hunt page another, you've fed the machines contradictions, and contradictions get resolved against you or omitted. Same name, same one-line description, same category, everywhere.
Strategy for a product nobody's heard of
Models can't recommend what they've never encountered. For a new product, agent SEO is mostly classic distribution with a new reader over your shoulder — the groundwork covered in distribution before you build feeds this channel directly.
Three directional bets we'd make:
Own a narrow phrase. You won't be the answer to "best project management tool" this decade. You can plausibly become the answer to "changelog tool for solo developers" within months, because the corpus there is thin. Win the narrow query, expand from the beachhead — the same topical-cluster logic as classic SEO, compressed.
Write the category content yourself. In a young niche, 5 solid articles can be a meaningful share of everything retrievable on the topic. Early, honest, well-structured content punches absurdly above its weight while the corpus is small.
Don't abandon classic SEO. Retrieval-augmented assistants lean on search indexes, so traditional ranking partially feeds agent visibility. The timelines and tactics in how long does it take to rank in SEO? still apply — agent SEO is a layer on top, not a replacement.
The prerequisite nobody mentions
Everything above assumes a live product with a real domain, real docs, and a real pricing page for machines to read. An unlaunched repo has no agent SEO, no classic SEO, no Reddit mentions — the corpus can't cite what doesn't exist, and every month unlaunched is a month the niche's content gets written by someone else. That's the gap LaunchBuddy closes. LaunchBuddy is a launch studio: submit your unlaunched project, and if it's picked, we build it onto the harness, ship it live, and operate the growth — the launch plumbing and the distribution legwork this whole channel depends on. You keep ownership; you pay a flat fee or share revenue.
Get the free honest assessment. If it's a no, you get the why. 60 seconds to submit at launchbuddy.app.