The State of AI Citations
A longitudinal panel study: 60 commercial-intent prompts, six answer engines, weekly snapshots from April 27 to July 13, 2026.
Every statistic on our site is measured by our own tracking pipeline — not quoted from someone else's survey. This page is the full methodology: what we ran, how we counted, and the sample size behind every number.
Data as of July 13, 2026 · 1,275 AI answers · 29 snapshots · 11,131 citations across 1,872 unique domains · figures refresh after each weekly sweep.
Mean 8.7. An AI answer names a handful of sources — not ten blue links.
N = 1,275 answers
Nearly two-thirds of the pages AI engines cite do not rank in Google's top 10 for that prompt.
N = 2,549 citations
Not a single domain was cited by all six engines for the same prompt. 65.7% were cited by exactly one.
N = 1,330 prompt–domain pairs
Of the domains cited one week, roughly 45% were missing from the same prompts' answers the following week.
N = 8 weekly pairs
How we measured
We run a fixed panel of commercial-intent prompts ("best X for Y" style questions a buyer would actually ask) through the same pipeline our customers use, on a weekly schedule. Each run queries six answer engines — ChatGPT (OpenAI), Claude (Anthropic), Google Gemini, Perplexity, Google AI Overviews, and Google AI Mode — using their native APIs with web search enabled, so the citations we record are the ones the engine actually produced, not a scrape-based guess.
Alongside every AI answer we capture the organic Google SERP for the same prompt on the same day, via DataForSEO as United States, English-language desktop results. That baseline is what lets us say whether a cited page ranks in Google's top 10, further down, or not at all — and whether top-ranking pages get cited.
The current panel spans 60 prompts in three service categories: financial services, legal services, and accounting & tax. Tracking began April 27, 2026; the dataset on this page covers 29 snapshots through July 13, 2026.
Panel history is uneven by design: the financial-services panel has run since April 27 (daily at first, weekly since June), while the legal and accounting panels joined on June 22. Week-over-week churn figures therefore rest mostly on the financial-services series; cross-sectional figures draw on all three categories.
Definitions
- Snapshot — one scheduled sweep of a category panel: every prompt in that panel run against every engine on a single date, alongside the same-day SERP capture.
- Answer — one engine's response to one prompt on one run date.
- Citation — a source URL the engine referenced in its answer, captured with the exact passage where available.
- Cited domain — the registrable domain of a cited URL; overlap and churn stats are computed at (prompt, domain) level.
- Weekly churn — for report pairs 5–9 days apart on the same panel, the share of one week's cited domains absent the next (and the Jaccard distance between the two sets).
Full results
Sample sizes differ across rows by design. Longitudinal stats (citations per answer, weekly churn) use all 29 snapshots. Cross-sectional stats (SERP comparisons, engine overlap) use only the latest snapshot per category, so repeated weekly runs of the same prompt aren't double-counted. Page-feature stats deduplicate by URL across the full window.
| Statistic | Value | Sample size |
|---|---|---|
| Citations per AI answer | median 8 · mean 8.7 | 1,275 answers |
| Citations pointing to pages outside Google's top 10 Judged against the organic SERP captured for the same prompt on the same day. | 63.6% | 2,549 citations |
| Google top-10 pages never cited by any engine With six engines each citing ~8 sources, top-ranking pages usually get cited somewhere — but rank alone doesn't decide where. | 7.5% | 478 URLs |
| Prompt–domain pairs cited by exactly one engine | 65.7% | 1,330 pairs |
| Prompt–domain pairs cited by every engine that ran | 0.0% | 1,330 pairs |
| Prior week's cited domains missing the next week Early-series figure reported to whole percent only; treat as directional until more weekly pairs accrue. | ~45% | 8 weekly pairs |
| Weekly churn of the cited-domain set (Jaccard distance) Same caveat as above. | ~65% | 8 weekly pairs |
| Cited pages using FAQ-style content patterns Descriptive of cited pages — not a causal lift (see limitations). | 69.0% | 2,088 URLs |
| Cited pages with any schema.org markup Descriptive of cited pages — not a causal lift. | 61.2% | 2,088 URLs |
| Cited pages containing a comparison table Descriptive of cited pages — not a causal lift. | 18.7% | 2,088 URLs |
| Median word count of cited pages | 2,147 words | 2,088 URLs |
Landing-page tiles round these figures (e.g. 63.6% → 64%). Stats resting on few weekly pairs are reported to whole percent only — the underlying series is too short to support decimal precision. Legacy provider identifiers in early runs are normalized to their current engines before aggregation.
Known limitations
- Panel breadth. 60 prompts in three service-heavy categories is a real measurement, not a market census — generalization to e-commerce, SaaS, or other categories is untested. We are expanding the panel toward ~250 prompts across ~10 categories and will refresh published figures quarterly as it grows.
- Single locale. All prompts run in English with a United States context, and the SERP baseline is US desktop results. Citation behavior may differ in other markets and languages.
- APIs, not the consumer apps. We query each engine through its native API with web search enabled — the path that makes runs reproducible. The consumer app experience can differ (personalization, interface variants, memory), so figures describe API answer behavior.
- Churn series length. The week-over-week figures rest on 8 weekly comparison pairs so far, weighted toward the financial-services panel. They grow with every weekly run; treat them as directional.
- Page features are descriptive, not causal. We extract content features (schema, FAQ patterns, tables, word count) only from pages that were cited, so "69% of cited pages use FAQ patterns" describes what winning pages look like — it is not a measured lift versus non-cited pages.
- Engines change. These are live commercial systems; behavior can shift with model updates. That volatility is part of what the study measures — and every figure here carries its collection window.
Citing this study
You're welcome to reference these figures with attribution. Suggested citation:
Cite AI (2026). The State of AI Citations: a longitudinal panel study of six answer engines, Wave 1. https://usecite.ai/research/ — data as of July 13, 2026.
Aggregate data behind any published figure is available on request: info@usecite.ai.
Run the same measurement on your own brand
The pipeline behind this research is the product. Track your prompts across all six engines and see who gets cited — and who doesn't.
Start free — no card needed