Getting cited by Google AI Overview: the check I run first, and it uses curl
When someone asks me why a page never gets cited in generative answers, the first thing I do isn't opening an SEO tool. I open a terminal and download the page without executing JavaScript. Half the time, the conversation ends right there.
TL;DR
- Many crawlers — classic and AI — don't execute JavaScript, or do it later and not always.
- You can check it with one command, in ten seconds, and it needs no paid tool.
- The rest (structure, direct answer, structured data) matters, but afterwards: if the text isn't there, there's nothing to structure.
- Beware of certainty: nobody has the numbers on what actually makes a generative system pick a source. Anyone handing you those numbers made them up.
The check, in full
curl -sL https://yoursite.com/page | sed 's/<[^>]*>/ /g' | tr -s ' ' | head -40
It downloads the page the way a crawler that doesn't open a browser receives it, strips the tags and shows you the first lines of text. Two outcomes:
- You see your content. Good, the problem is elsewhere.
- You see three words and a menu. Then your page, to anyone not executing JavaScript, is blank — and all the keyword work was done on something that reader never saw.
It's worth trying even on a page you believe is fine. The most common surprises are sections that load on scroll, FAQs inside a client-mounted accordion, and tables built by a script.
The second check: is the answer near the top?
A generative system that summarises an answer needs to find it, not deduce it. If your page puts the answer in the third paragraph, after the introduction about "a changing world", you're asking for extra work from something that has twenty sources to choose from.
The test is trivial: read the first three sentences cold. If after those you still don't know the answer to the question in the title, the order is wrong.
The third: one question per page
A page that answers five questions wins none of them. It's the oldest advice in SEO and it survived the move to generative systems intact, for a simple reason: a citation points at an address, and that address either answers or it doesn't.
What can be measured, and what can't
| Can be measured | Can only be guessed |
|---|---|
| The text is there without JavaScript | How much speed weighs in source selection |
| Structured data is valid | Whether the declared author actually counts |
| The answer is in the first lines | Why they cite you today and not tomorrow |
| The page is indexed | How long this rule will hold |
You work on the left column. On the right you make assumptions, and it's worth saying so rather than dressing them up as method: as of today Google has published no criteria for what gets a source selected for the generative box. Anyone selling you a service with a success percentage is selling an invented percentage.
That doesn't mean nothing can be done. It means you do the things that are right anyway — reachable content, the answer up top, correct structured data — and you measure what's visible: impressions, citations, traffic.
The mistake I find most often, and it isn't technical
Perfect pages nobody ever linked to. A document that can't be reached from any other page on the site doesn't get discovered, and undiscovered content doesn't have a quality: it doesn't exist.
I saw it starkly on a site whose data I follow: twenty-five published pages, three indexed, nineteen that Google said it didn't know at all. It wasn't a content problem — no already-indexed page pointed at the others. It was fixed with an internal index, not a rewrite.
FAQ
Can I force an appearance in AI Overview?
No, and you can't make it measurably likely either. You can make the page readable, direct and reachable: those are the three things you need regardless.
Do Core Web Vitals matter for generative answers?
Nobody knows publicly, and anyone claiming yes is extending by analogy from classic search. A slow page is a problem for visitors: that alone is reason enough to fix it, without attaching a promise to it.
Ideal length?
Whatever it takes to answer. Eight hundred words that answer beat two thousand that circle — not least because a system that summarises needs to find the sentence, not leaf through.
Do I need llms.txt?
It's a proposal, not an adopted standard. I wrote about it without enthusiasm here: does llms.txt actually do anything?.
If you'd like us to look at your situation with the numbers in hand, there's the SEO, GEO & MarTech service. For the strategic reasoning rather than the checklist, continue with why SEO is becoming GEO.
Related articles
- When the posting and the interview describe different rolesA hiring process produces four documents — the posting, the technical exercise, the final interview, the rejection email — and each says something different about the real role. How to read them before investing weeks.
- When the interview measures a conflict that isn’t yoursIt wasn't the first time I'd seen this pattern at a large, well-run company: whoever writes the posting wants one thing, whoever evaluates you needs another. How to spot it before you sign.
- The worst bug does not break the feature. It breaks the view26 tools connected and working in production, the dashboard showed zero. The bug had been there for months for another product too, and nobody had seen it because whoever could notice had admin permissions.