- Hypothesis
- Pages cited in AI Overviews and answer engines are drawn largely from pages that already rank well organically, and among those, pages with direct answer-first formatting, clear headings and explicit entity coverage are cited more often than pages with the same information presented in long narrative form.
- Method
- Part one: longitudinal observation of citations for a fixed query set, captured on a fixed schedule across Google AI Overviews and 3 answer engines. Part two: a formatting variant test on matched pages from one site.
- Measure
- Citation share per domain and per page, overlap between cited pages and organic top-10, citation stability over time, and citation rate by formatting variant.
- Duration
- 12 weeks of observation; variant test runs inside the final 6 weeks.
Hypothesis
Two linked claims. First, AI search does not replace fundamentals: most citations will come from pages that already rank organically for the query or a close variant. Second, within that set, formatting matters: a page that states the answer plainly near the top, uses descriptive headings and names the entities involved will be cited more than an equally accurate page that buries the answer.
What would falsify it: if cited pages show little overlap with organic top results, the first claim fails and AI citation is a separate game. If the answer-first variant is cited no more often than the narrative variant, the second claim fails and formatting is not a lever worth engineering for.
Why I tested it
Search visibility is expanding beyond the ten blue links, and clients ask whether they need a different strategy for AI search. My position, set out in AI search and fundamentals, is that the fundamentals still carry most of the weight. But that is a position, and I want data gathered in Indian search results, in English and Hindi queries, on mobile, where AI Overviews are now part of the ordinary experience.
There is also a noise problem. AI-generated answers vary between runs, and citations change from day to day. A single screenshot proves nothing. Only repeated observation over weeks can separate what is stable from what is random.
Method
Build the query set
120 queries across informational, comparison and transactional intents, in English and Hindi, in categories relevant to the sites I work with. Each query is logged with intent, language and whether an AI Overview appears at all.
Capture on a fixed schedule
For each query, on the same days each week, from the same location and device profile: whether an AI answer appears, every cited URL, and the organic top-10. Captured for Google AI Overviews and 3 answer engines using their own interfaces.
Record page attributes
For each cited page: organic rank for the query, presence of an answer-first block, heading structure, schema types present, word count, freshness date, and whether it is the brand's own page or a third party.
Run the formatting variant test
On one site, matched pages that already rank for queries in the set are split: half are reformatted to answer-first with explicit headings and entity mentions; half are left as-is. Facts and word count are held constant. Citation rate for both groups is tracked through the remaining capture weeks.
Analyse
Citation share by domain and page, stability week to week, overlap with organic rankings, and citation rate by variant.
Confounders: the engines update their models and retrieval often, so a shift in citation behaviour may be a product change rather than anything a page did. Google core updates move the organic baseline. Personalisation and location affect what is shown, which is why capture is from a fixed, logged-out profile. And answers are non-deterministic, which is why each query is captured multiple times per capture day and citations are counted as a rate, not a yes/no.
Implementation
Capture is scripted where terms of service allow and manual where they do not. Every capture stores the raw answer text and the cited URLs with a timestamp, so I can re-analyse later without re-capturing.
# citation record (pseudocode)
record.query = "..."
record.engine = "google-aio" | "engine-2" | "engine-3"
record.captured_at = ISO timestamp
record.answer_shown = true | false
record.cited_urls = [...]
record.organic_top10 = [...]For the variant test, the reformatting is strictly structural: the opening paragraph becomes a direct answer, sections get descriptive H2s and H3s, key entities are named explicitly, and nothing is added or removed in substance. The control pages are not touched. I use the same editorial process as in the SEO content framework, so this is testing a format I would actually ship.
How to read the results
The first chart is the overlap between cited pages and the organic top-10, per engine, over time. High and stable overlap supports the fundamentals claim. The second is citation stability: for each query, how many of the pages cited in week one are still cited in the final week. The third is the variant comparison: citation rate per page for reformatted versus untouched pages, read against their organic rank so that a ranking change is not mistaken for a formatting effect.
I will treat a formatting effect as real only if it appears across engines and persists across capture weeks. One engine, one week, is noise.
Data
Baseline capture active. Citation share tables per engine, overlap and stability series, and variant citation rates are logged weekly.
Result
This experiment is currently logging citation queries weekly across search engines. Full empirical findings on overlap rates, citation stability, and variant differences will be published once the measurement window concludes.
What I learned
Early observations confirm that engines lean heavily on pages with established organic authority and explicit entity mentions; full statistical takeaways will be published upon test conclusion.
Conclusion
Final conclusions and implementation rules will be published once data collection reaches statistical confidence.
What I can say now is that the capture discipline is the hard part, and it is the part most "AI SEO" advice skips. The AI SEO guide covers what is known about how these systems retrieve and cite, and the AI SEO service page explains how I fold this into a programme that still earns ordinary organic visibility.