How AI search builds an answer: 4,222 answers traced to their sources
Written by
Odin Lund
Category
AI search
Date
Share this article
What we measured, and how
The data set is one legal category in New York: motor vehicle accident representation. We wrote 60 prompts the way customers phrase them, from first question (“who handles motorcycle accident cases in Buffalo”) to decision (“is this firm worth calling”), and ran them daily on ChatGPT, Gemini and Google AI Overviews against 14 tracked firms. For every answer we recorded which businesses were named, in what order, and every URL the model retrieved or cited. Over 30 days to 17 September 2026 that produced 4,222 answers and a citation record for each.
Two limits are worth stating. The set is one category and one geography, so the proportions will differ elsewhere. And “retrieved” means the model fetched the page while building the answer; it does not mean the page was quoted. We count both, separately. All figures below come from our own tracking at aisearchmonitor.com unless a source is named.
Where the claims come from
For the firm we manage, its own website was cited in 1,164 answers, 27.6% of the set, drawing on 63 distinct pages. Third-party sources mattered more by volume: 46 domains carried the firm and produced 18,742 retrievals in the same period, roughly four times the retrievals from the firm’s own site. Directories and review platforms made up most of that count. Press was a small share of retrievals but appeared in the answers with the highest recommendation rate, which fits what other researchers have found about authority signals.
Concentration is the pattern to notice. The firm’s site drew 4,319 retrievals in 30 days; the nearest competitor’s drew 1,296. Models do not sample the web evenly. They lean on a short list of domains per category and return to them, so being on that list, with the same facts as everyone else on it, matters more than publishing volume.
What decides who gets named
Three factors separated the businesses that were named from the ones that were not. Consistency: businesses whose facts matched across their site, their listings and their reviews were repeated; businesses with conflicting facts were hedged (“some sources say…”) or dropped. Density: a single page that answered one question fully was retrieved more often than a large library of partial pages. The firm had published more on motorcycle cases than any competitor and was almost never cited for them until those pages were restructured, one question per page. Structure: pages with schema markup and plain statements of fact were retrieved at a higher rate than pages of equal length without it.
This matches the one controlled study of the question we know of. Aggarwal et al. tested nine content changes across 10,000 queries on generative engines and found that adding citations, quotations and statistics raised a source’s visibility in answers by up to 40%, while keyword stuffing did nothing or made it worse (“GEO: Generative Engine Optimization”, KDD 2024). Our field data points the same way: the models reward pages that state checkable facts and agree with the other sources they read.
What this means in practice, and sources
Audit the facts: pick the ten claims you need the models to repeat and check that every source states them the same way. Audit the pages: for each question a customer asks, there should be one page that answers it completely. Audit the sources: find the directories, review platforms and publications the models cite in your category and be present on each with the same facts. Then measure again. The answer is rebuilt continuously, and a position that is not maintained slips.
The stakes are set by the audience. Google reported that AI Overviews reach more than 2 billion monthly users (Alphabet Q2 2025 earnings call, July 2025). OpenAI put ChatGPT at 800 million weekly users (OpenAI DevDay, October 2025). Pew Research Center found that when a Google result page carried an AI summary, users clicked a traditional result in 8% of visits versus 15% without one (Pew, July 2025). Customers are reading the paragraph and acting on it.
Sources
SourceNode tracking, aisearchmonitor.com, 60 prompts, 14 firms, 30 days to 17 September 2026.
Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande. “GEO: Generative Engine Optimization.” KDD 2024 (arXiv:2311.09735).
Alphabet Q2 2025 earnings call, 23 July 2025: AI Overviews monthly users.
OpenAI DevDay, 6 October 2025: ChatGPT weekly active users.
Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results,” 22 July 2025.
Share Blog
