The State of AI Citations, Q3 2026

By Abiot Y. Derbie - Founder, RankEcho - Published July 10, 2026 - Updated September 13, 2026 - CC BY 4.0

Method and scope

The full prompt battery is listed below, and the exported domain observations are available as a downloadable CSV.

• Battery: 48 prompts across 8 commercial categories (SaaS, Ecommerce, Health, Finance, Legal, Home Services, Education, Travel), spanning four buyer intents (category, comparison, problem, brand). No prompt mentions RankEcho or any site we operate.
• Engines: Claude, Perplexity, Gemini, Google AI Overviews. OpenAI/ChatGPT was not included, so the findings do not describe that engine.
• Sampling: every prompt x every engine, run twice, 24h apart (July 9, 2026 and July 10, 2026). 384 observations. Run 0 measures the cross-sectional questions; run 1 exists to measure stability.
• Statistic: cited-domain sets are compared with Jaccard similarity (|A∩B| / |A∪B|). Cross-engine agreement is the mean pairwise Jaccard per prompt; stability is the mean Jaccard between the two runs of the same prompt and engine.
• Analysis rules: engine-infrastructure redirects (e.g. unresolved grounding URLs) are excluded from every domain statistic and reported separately. Skipped observations are flagged and excluded from domain counts; the CSV does not disclose their reasons. Non-skipped observations with no exported domains are retained, but the CSV does not establish whether an answer was present.

Finding 1 — The top 10 domains accounted for 18.0% of run-0 citations.

The ten most-cited hostnames contributed 230 of 1,275 publisher-domain occurrences (18.0%) in run 0. Each hostname is counted once per prompt-and-engine observation. The denominator excludes 34 infrastructure occurrences, such as unresolved grounding redirects.

Reddit was the most-cited hostname, with 64 occurrences (5.0%). The residual other category contained 1,035 occurrences (81.2%). These are hostname-based classification labels: other includes hostnames that did not match another category, and does not verify ownership, independence, provenance or source quality.

The source counts were recalculated from the CSV retrieved on September 13, 2026. The study dates remain July 9–10, 2026; this section uses run 0. Several hostnames tie at the top-ten cutoff. Alphabetical tie-breaking selects ten names without changing their combined count.

Interpretation for this dataset: Run-0 citations span 670 distinct hostnames. The domain categories do not establish how many independent publishers or original sources produced the cited material. This distribution does not estimate citation concentration for other prompts, categories, locales, or dates.

Finding 2 — Conditional mean cited-domain overlap was 0.195.

Across identical prompts in this study, mean cited-domain overlap was 0.195 - about 20% - when both engines in a pair returned sources. When a comparison with an absent source answer was assigned zero overlap, the pooled mean was 0.138. Roughly four of five cited domains in the observed responses were unique to one engine. The run-specific conditional means were 0.195 on the first date and 0.201 on the second; their similarity is descriptive, not an independent replication.

Interpretation for this dataset: A blended result would conceal the engine-level differences observed in this battery. Per-engine results are therefore more informative for these observations, but the study does not establish how strongly the engines differ for other prompts or dates.

Finding 3 — Among observed pairs, Gemini and Google AI Overview had the lowest overlap.

Among the available paired comparisons in this study, Gemini and Google's AI Overview had the lowest conditional mean cited-domain overlap at 0.129. Google AI Overview and Perplexity had the highest observed pairwise value (0.295). These rankings describe the available comparisons only; the study did not test why the systems differed.

Caveat, stated up front: the AI Overview appeared for a minority of queries, so this pair rests on 15 comparisons. Treat it as a narrow, high-uncertainty result rather than a settled cross-system pattern.

Finding 4 — Mean within-engine stability after 24 hours was 0.671.

Across matched prompt-engine observations on the two study dates, mean cited-domain-set stability was 0.671. The engine-specific values varied: Perplexity and Claude were near 0.77, while Gemini was 0.361. These values compare two runs 24 hours apart and do not measure longer-term volatility.

Interpretation for this dataset: A single run would omit the day-to-day variation observed in this battery. The two-run design cannot estimate longer-horizon variation or prescribe a monitoring cadence for other prompts. This full cited-domain-set Jaccard statistic is not manual exact-URL citation persistence across consecutive post-change windows.

Finding 5 — AI Overview appeared for 40% of this study's prompts.

Within this 48-prompt battery, Google's AI Overview appeared for 40% of prompts. No AI Overview was recorded for the remaining prompts. When one appeared, it cited a mean of 7.7 sources. The report therefore keeps appearance rate separate from citation count conditional on appearance. Neither estimate describes AI Overview coverage for commercial queries beyond this battery.

Limitations

Read these before citing anything above.

• 48 prompts across 8 categories is a probe, not a census. US-English, consumer-commercial intent only.
• Two runs, 24 hours apart, one week. Stability over longer horizons is unmeasured.
• Four engines. OpenAI/ChatGPT is outside this dataset.
• Finding 3 (Gemini vs AI Overview) rests on 15 paired observations. The observed ranking is descriptive, not definitive.
• We measured which domains are cited, not whether a brand was named in prose without a citation. Answer-text brand mentions are outside this dataset.
• AI engines change without notice. This dataset is a dated snapshot of two runs across one week; it does not measure longer-horizon patterns.

The data

The CSV contains prompt IDs, categories, intents, engines, run numbers, skipped flags, cited hostnames, source labels and exported positions. Positions are indices in the exported domain list, rather than ranks of original source URLs. Reuse the data under CC BY 4.0 with attribution; the prompt battery below supplies the study questions.

The September 13 data check retained 2,764 rows representing all 384 planned observations, including 16 skipped observations and 54 non-skipped observations without domains. The CSV omits answer text, skip reasons, model versions and original URLs. A no-domain row therefore cannot distinguish an absent answer from an answer without citations.

Download the dataset (CSV)

Method note: this report puts its bounded answer before the detailed method and uses Dataset and FAQ structured data. Those presentation choices do not affect the calculations.

Frequently asked questions