Write for AI Overviews, not around them
Google now answers a large share of queries above the blue links, and the pages it pulls from are the ones that answer the question directly instead of warming up for six paragraphs. None of this removes SEO. It raises the bar on being clearly useful. Every page you want quoted should:
- Answer the question in the first two sentences, before the context and the story
- Use clear sections and definitions a model can lift whole
- Carry examples, data or product insight nobody else has
- Keep claims specific and easy to cite: numbers, dates, named tools, versions
The pages losing to AI Overviews are the ones that were padding to hit a word count. The pages winning are the ones a reader would have skimmed for the answer anyway.
Rank across AI engines, not just Google
ChatGPT, Perplexity, Claude and Gemini all read your pages, and they all reward the same thing: content that can be summarized without guessing. Five things make a page easy for a model to use:
- Headings that say what the section answers
- A direct answer under each heading, not a build-up
- Original examples that don't appear on ten other sites
- Consistent entity names: same product name, same category words, same spelling everywhere
- A plain statement of who the product is for, and who it isn't for
The exercise worth running this week: paste one important page into a model and ask it to summarize what the product does and who should buy it. Whatever it gets wrong or hedges on is exactly the part to rewrite.
Publish an llms.txt, and keep it honest
llms.txt is a plain markdown file at your root that tells a model what your product is, what your key pages are, and how to describe you. It is not a ranking factor and no engine promises to read it, but it costs an afternoon and it removes the guesswork when one does. Treat it as the summary you would hand a journalist.
- One paragraph on what the product does and who it is for, in the words your buyers use
- A linked index of the pages worth quoting: docs, pricing, comparisons, guides
- The category words you want associated with you, used consistently everywhere else too
- An llms-full.txt with the actual content inlined, if your docs are worth reading whole
- Regenerate it from the same source as your docs, so it cannot drift into being wrong
Never write anything in llms.txt you would not put on the pricing page. A file that oversells gets contradicted by your own site, and a model that finds a contradiction resolves it by hedging or naming a competitor instead.
Get listed where the models already look
Models cite aggregators far more than they cite vendors, because a roundup reads as neutral and a homepage reads as a claim. So the highest-leverage AEO work is often not on your site at all. Being in the lists is what puts you in the answer.
- GitHub "awesome" repositories for your category, which are scraped, mirrored and quoted constantly
- Comparison and alternatives pages on sites that are not yours, including the ones listing your competitors
- Reddit threads answering the question your buyer asks, where you contributed something real
- Directories and leaderboards with editorial standards, not link farms
- Your own docs, if they are public and crawlable, because docs answer specific questions cleanly
Open a pull request to the awesome list with a one-line entry in the same format as the others, and only where you genuinely belong. Maintainers reject self-promotion instantly, and a rejected PR is public. The Authority and SEO chapters cover the outreach side of getting into other people's lists.
Be usable inside the assistant, not just cited by it
Getting named is the first half. The second is being reachable without leaving the chat. The ChatGPT plugin platform that shipped in 2023 is retired, so ignore any guide still telling you to build one; the current surfaces are custom GPTs, apps inside ChatGPT, and MCP servers that connect your product to ChatGPT, Claude and the coding agents at once.
- An MCP server is the portable option: one implementation that several assistants can call
- A custom GPT is the cheapest experiment, and it is a distribution surface with its own store
- Expose the actions someone would actually want mid-conversation, not your whole API
- Publish an agent-facing docs page, so a model wiring you up gets it right first time
This only pays off if the assistant can do something useful with your product in one call. A connector that requires four steps of setup before it returns anything gets abandoned in the chat where it was installed.
Measure it, or you are guessing
AI referrals arrive as ordinary traffic with a telling referrer, and the rest is invisible unless you go looking. Two habits are enough to stop flying blind:
- Segment referral traffic from chatgpt.com, perplexity.ai, claude.ai and gemini in your analytics
- Run the same five buyer prompts across the engines monthly, and log whether you are named and described correctly
- Track the wrong answers specifically: a model repeating a stale price or a dead feature is a page to fix
- Ask new signups where they heard about you, because assistant referrals often arrive with no referrer at all
Key moves
0/8Reading is warm-up. Check these off as you actually ship them; progress saves in your browser.
Models quote brands that keep publishing
The engines read what is current. Viraloop keeps your content and your channels active, so the answer a model gives about you is this quarter's, not last year's.