Shanghai, September 21, 2026 — The 23SEOGEO team has released State of GEO Q3 2026, the first edition of its quarterly public benchmark tracking how often continuously monitored brands appear across AI engines under one deduplication method.
The observation window runs from 2026-08-11 to 2026-09-09 (UTC) and covers five continuously monitored brands, nine AI engines and 5,536 deduplicated checks. The sample recorded 1,430 citation hits, for an overall citation rate of 25.8%.
These figures come from 23SEOGEO's own monitoring sample. They are not official engine statistics and should not be read as an all-industry average.
What this edition publishes
The benchmark publishes more than headline figures. It includes the calculation method, known limitations, an anonymized aggregate dataset and a reproduction script. Readers can download the files and independently recompute every engine rate and the overall result.
- Read the complete quarterly benchmark and methodology
- Download the anonymized aggregate dataset (CSV)
- Download the reproduction script
- Browse the State of GEO benchmark series
Three results worth monitoring
1. The four leading engines are tightly clustered
Claude, Google AI Overview, Perplexity and ChatGPT land between 28.5% and 30.4%, a spread of under two percentage points. In this sample, concentrating on one engine shows no clear advantage; sustained cross-engine coverage is the more useful variable to monitor.
2. The Chinese-engine sample is lower, but remains too small for strong claims
DeepSeek records 22.4%, while Doubao and Kimi each record 20.8%. Each has only 192 checks, below the report's low-sample threshold of 300. These results are directional only and cannot establish a market-wide pattern or a fixed preference of any engine.
3. Low samples and short windows must travel with the numbers
Grok records 6/294 (2.0%). Gemini records 0/209 and only joined monitoring partway through the quarter. We do not translate “no hit in this sample” into “will not cite,” and we do not use shortened coverage windows for strong cross-engine comparisons.
Deduplication method
One check is counted for each tracker, engine, question and UTC calendar day. Repeated hits on the same day count once. Failed calls, timeouts and empty answers are excluded from the denominator, preventing scan frequency or service failures from distorting the rate.
A hit means the answer contains the registered brand name or alias, or a citation link points to the brand domain. This measures visibility, not endorsement, sales or ranking improvement.
Known limitations
- The monitored brands skew toward B2B and cross-border services, not an all-industry sample.
- Check volume varies by engine; engines below 300 checks are directional only.
- Gemini has a shorter observation window, so this edition draws no conclusion about its citation capability.
Why publish a quarterly benchmark
GEO is growing quickly, but many public figures still omit sample size, data cutoff, deduplication rules and reproducible evidence. State of GEO is designed as a more restrained public record: every number ships with its sample, time window, method and limitations, and can be independently recalculated.
The next edition, State of GEO Q4 2026, is expected in early January 2027. Historical figures will not be silently rewritten; any methodology change will carry a version and a change note.
— the 23SEOGEO team