While we were running a GEO audit for an exterior home improvement company this summer, we analyzed the citation export prompt by prompt, and one row stopped me.
The prompt was "how long should residential windows last before replacement?" Google AI Overviews cited a page from The Window Experts, a regional window installer in the Greater Houston area. Twice, on that one prompt.
The organic SERP for that question has Pella on it. The Spruce. Window World. Ring’s End. First American Home Warranty. Every one of them has more domain authority, more budget, and more content operation than a Houston installer with a blog.
Then I opened the page that got cited. And by any content-quality standard you’d apply on a client call, it’s kind of a mess.
This learning has broad application across anyone looking to improve their citation quantity and visibility in Generative Engine Optimization (GEO) right now.
Why did Google AI Overviews cite a local installer over national brands?
Because the first 200 words answer the question with light qualifiers which provide depth. This simple formula allows extraction to happen quickly with a nibble of intelligence.
Here’s the shape of the opening on their window lifespan page:
Sentence one is a single line of human framing about expecting your windows to outlast your mortgage. One sentence. Not four paragraphs.
Sentence two gives you the number: "The average window lifespan for residential units is typically 15 to 30 years," with the variables that change it named right there in the same sentence.
Then a material breakdown, immediately. Fiberglass 30 to 50 years. Vinyl 20 to 40. Wood 15 to 30, with maintenance called out. Aluminum 15 to 30.
Then a signs-you-should-replace list. Drafts. Unexplained energy bills. Condensation between the panes. Windows that won’t open or lock. Visible cracking or rot.
That’s it. That’s the whole citable asset. A question, a direct numeric answer, a segmented breakdown, and a diagnostic checklist, all inside the first screen.
A model retrieving against "how long do windows last" gets everything it needs before it has to work for anything.
So what’s wrong with the page?
Almost everything else.
First, the broken image right at the top. It would stop a human cold, did it stop the robots? Heck no!
Once you scroll past the first 200 words, the page goes on a journey. It pivots into hail resistance. Then hurricane and impact windows. Then NFRC performance ratings, Design Pressure, Performance Grade, High-Velocity Hurricane Zone certification. There’s a section on storm shutters and window film. There’s sales copy about their installation crews dropped directly into the middle of the article.
And there’s a raw LaTeX artifact sitting in the body copy, an unrendered $\approx 140\text{ mph}$ where a plain "140 mph" should be. Which is a pretty good tell about how at least part of the page got written.
If a client sent me this page for review I’d have notes. A shit ton of notes.
It’s getting cited anyway.
What does that actually prove?
That page quality and citation eligibility are two different scoreboards, and most of us have been reporting on the wrong one.
The engines are not grading your page. They’re looking for a passage that resolves the query, and they take it wherever it sits. A strong 200-word block on a sprawling page beats a beautifully structured 2,000-word page that makes the model hunt for the answer.
Which cuts the other direction too, and this is the part worth sitting with. Your best page, the one with the custom illustrations and the tight narrative arc and the three rounds of stakeholder review, is losing this query to a Houston installer… if your answer is in paragraph six.
I’d bet a lot of money that most of the national brands on that SERP have better content than this page. That isn’t what’s being measured.
This is consistent with what the research shows. In 252,000 controlled trials across six LLMs, Vishwakarma, Kumar & Jamidar (2026) found topical relevance and list position were the strongest citation drivers, while formatting-only changes had little effect. Precision on the question, positioned early. That’s the whole mechanism, and The Window Experts page satisfies it by accident on a page that fails almost every other test.
One note on reading this, because it matters for how you audit your own stuff. Two AIO appearances on one prompt is a real finding and a small one. AI Overviews are inconsistent between two runs of the same query and sensitive to minor rephrasing, which Grossman et al. (SIGIR 2026) measured directly. So this page may or may not be holding that citation the day you read this, and if you run the query yourself and see something different, that’s the system behaving normally rather than the finding being wrong. Snapshots are what we have. Take them per engine and take them more than once.
The second finding: they’re competing with themselves
While I was in there I found a second page on the same domain targeting the same question. An older post that opens with a full paragraph about how a house without windows would be a bunker, then works through reasons to replace windows, and finally lands the material lifespans roughly 600 words down.
Same domain. Same intent. Same question. Two pages.
The older one is the traditional shape: warm up the reader, build to the payoff. It’s honestly better writing. It’s also the one not getting pulled, and it’s splitting whatever topical equity the domain has on this question.
If they consolidated those two pages into one, kept the answer-first opening, and moved the hail and hurricane material to its own post where it would rank for its own queries… they’d probably own this question outright instead of holding a slice of it.
One more thing about that SERP
Above all the brand pages, two Reddit threads own the #1 and #2 spots. They feature homeowners asking each other. On a question where six companies have published dedicated content, the top of the page belongs to strangers on forums.
This shift toward low-click and community horizons is reflective of broader trends where less than one-third of Google searches result in an organic click, making AIO citation the primary vehicle for brand impressions.
That’s the earned-media pattern showing up in a category nobody thinks of as community-driven. If you sell to homeowners and your GEO plan is entirely on-site, the forums are eating the top of your funnel and you can watch it happen in a browser.
What to do with this
Four things, and none of them take a quarter.
- Pick your highest-value question and read your own page like a model would. Where’s the answer? If it’s below the fold, you’ve found the problem.
- Move it up. One sentence of human framing, then the direct answer, then the segmented breakdown. You can keep the good writing. Put it underneath.
- Search your own domain for the same question twice. Two pages on one intent means neither one wins. Consolidate.
- Look at who’s actually above you. If it’s Reddit and Quora, no amount of on-page work fixes it, and you have a different investment to make.
The Window Experts did not do all of this on purpose. I doubt anyone there has the phrase "answer-first extraction" in a deck. They wrote a page that gets to the point in the first 200 words, and that turned out to be worth more, in the one place a homeowner is now getting their answer, than everything the national brands brought to the same question.
As my father, the carpenter and window installer always taught me: Measure twice, cut once. Then put the answer up top.
Frequently Asked Questions
What is an answer-first content structure? An answer-first structure involves placing a direct, concise response to a primary query at the very beginning of a page. This typically consists of one sentence of context followed by specific data or a checklist, making it easier for AI models to extract information for users.
Do LLMs prioritize brand authority or content structure? While brand authority matters for traditional search rankings, evidence suggests that Generative Engines often prioritize passage relevance and structural clarity over domain strength. A small site can outrank a national brand in AI Overviews if its content is better formatted for direct extraction.
How can I audit my site for GEO citation eligibility? Start by identifying your highest-value questions and checking if the direct answer appears above the fold. Review your domain for duplicate pages targeting identical intents and consolidate them into single, authoritative assets that use answer-first patterns.
