What 90 days of AI Overviews data taught us about the new SERP
Ninety days of citations, scraped, sorted, and second guessed.
Aditya Rao
Head of Generative Search
The setup
We picked twelve accounts across SaaS, fintech, healthcare and ecommerce. For each one we tagged a hundred priority queries and pulled the AI Overview block daily. No scraping shortcuts, no synthetic prompts. Just the same SERP a buyer would see, captured at the same hour every morning.
Then we logged who got cited, who got paraphrased without a link, and who quietly disappeared after a model refresh.
Five patterns that held up
- Pages cited in AI Overviews almost always rank in the top ten classic results. The Overview is not a parallel universe. It is a remix of what already wins.
- Specificity beats authority on long tail queries. A 900 word teardown by a niche operator routinely beat a 4,000 word guide from a household name.
- Tables, definitions and step lists got pulled verbatim. If your page has a clean answer block above the fold, you have a real shot.
- Brand mentions in third party sources matter more than we expected. Pages from sites that were also mentioned across forums and review platforms got cited at twice the rate of equally optimised pages without that off site footprint.
- Citations are unstable. Roughly a third of cited URLs swapped within thirty days. Reporting on this weekly is a waste of everyone's time.
Three things that turned out to be folklore
Schema is not magic. Adding FAQ markup to thin pages did not earn citations. Removing it from genuinely useful pages did not lose them either.
Word count is not a lever. The longest page rarely won. The page that answered the exact sub question in the first paragraph won.
Author bios alone do nothing. A photo and a short paragraph at the bottom of the article had no measurable effect. What did move the needle was the same author appearing across multiple credible sites talking about the same topic.
What we are doing differently now
We brief writers around the answer first, then the context. We treat the first 120 words as the only words a model might quote. We invest in earned mentions across the same topic cluster, not just links to the page itself.
And we have stopped promising clients a fixed AI Overview share. The right metric is qualified traffic and assisted pipeline. Citations are a leading indicator, not a KPI.
Work with us
Want this kind of thinking on your next quarter?
We take on a small number of new engagements each quarter. If the brief is interesting, we will say so. If it is not, we will tell you that too.
Start a conversation