AI Citation Share Benchmarks: What Good Looks Like in 2026
Reference numbers for AI citation share and citation rate, from measured audits: per-engine anchors, why domain authority does not buy share, and the zero-sum math of 3 to 6 sources per answer.
Every AI answer is a small auction with 3 to 6 winners. That is the entire reason citation share matters: unlike a search results page, there is no page two, and every source an engine cites displaces one it did not. Teams asking "what is a good citation share?" today find almost no published reference numbers, and most of the few that exist were measured with the tool bug that the citability.dev calibration methodology documented: unforced web_search calls that report 0% ChatGPT citation for sites ChatGPT cites constantly. This page publishes the reference numbers we actually measured, with dates and methodology, so there is something concrete to benchmark against.
The Reference Anchors#
All numbers below were measured with forced-search invocation and a calibration receipt (the methodology is open-sourced in the AI Visibility Readiness Framework). Dates are stated because engines move.
| Anchor | Engine | Measured result | When |
|---|---|---|---|
| Wikipedia (ceiling reference) | Perplexity | 80 to 100% citation rate on brand-recognition queries | early 2026 |
| Wikipedia (ceiling reference) | ChatGPT (forced search) | mid-seventies citation rate | early 2026 |
| Wikipedia (ceiling reference) | Claude (forced search) | 80 to 100% citation rate | early 2026 |
| Small low-authority site, clean infrastructure | ChatGPT (forced search) | mid-teens citation rate | early 2026 |
| Ahrefs (DA 92) | 3-platform panel | 100% brand visibility, 5% citation rate | April 2026 |
| chudi.dev (DA 28, pre-fix baseline) | 3-platform panel | 29% visibility, 0% citation rate | April 2026 |
| Same site, post-infrastructure-fix | Microsoft Copilot (Bing WMT first-party count) | 671 verified citations over 90 days | mid 2026 |
| Invented .invalid domain (fabrication check) | all three engines | 0 citations in 18 tested responses | early 2026 |
Three readings from the table. The ceiling is real but reserved: Wikipedia-class sources saturate every engine, and nothing below internet-scale training presence gets there. The floor is fixable: the same site that measured 0% in April earned 671 first-party-verified Copilot citations over the following 90 days after structural fixes, before its authority moved. And exact zeros are suspicious: a 0% ChatGPT reading next to healthy Perplexity numbers is the signature of the measurement bug, not of invisibility.
Domain Authority Does Not Buy Citation Share#
The six-site benchmark (April 2026, 10 infrastructure checks plus 3-platform citation testing) is the clearest data we have against the default assumption that authority wins AI answers:
| Site | DA | Infrastructure score | AI outcome |
|---|---|---|---|
| ahrefs.com | 92 | 10/10 | 100% visible, 5% cited |
| semrush.com | 91 | 10/10 | partial visibility and citation |
| chudi.dev | 28 | 10/10 | 29% visible, 0% cited (pre-fix baseline) |
| reddit.com | 97 | 7/10 | failed infra checks for extraction |
| medium.com | 95 | 7/10 | failed infra checks for extraction |
| x.com | 96 | 5/10 | failed infra checks for extraction |
The three highest-authority sites scored lowest on exactly the checks AI extraction depends on: answer-first content, JSON-LD, and dateModified freshness. They still get cited, but through training-data presence, a path that only exists at internet scale. For everyone else, the citation path runs through infrastructure, and that is why a DA 28 site can out-earn a DA 92 site on citations per answer in its own topic: the full analysis is in Benchmarking AI Visibility Across 6 Sites.
The Zero-Sum Math#
A response set of 100 answers at 4 to 5 citations each carries roughly 400 to 500 citation slots. That is the whole market for that query panel. If your domain holds 20 of them, your citation share is 4 to 5%, and the only way to 10% is taking slots someone else holds today. Two practical consequences:
- Competitor tracking is not optional at the share level. Citation rate can be measured alone; citation share needs the full denominator, which means recording every cited domain in every response, not just yours.
- Small panels lie about share. At 20 queries and ~90 citation slots, one displaced competitor citation moves your share by more than a percentage point. Report share with the panel size attached, and treat single-period share moves the way the velocity post treats rate moves: annotate first, act on trends.
The Integrity Caveat That Invalidates Most Benchmarks#
Any ChatGPT citation benchmark published without stating its search-invocation method should be treated as unverified. The default API behavior answers from training data and returns empty citation arrays; a benchmark built on that default reports structural zeros that have nothing to do with the sites being measured. Before comparing your numbers to anyone's published figures (including ours), run the 60-second known-positive test from the pillar post: measure Wikipedia on ChatGPT, and if the tool reads 0%, discard the tool's numbers, not your site's prospects.
Benchmark Your Own Site#
The reference anchors above all sit downstream of one precondition: infrastructure that does not cap your numbers at zero. Run the free scan to check the 10 structural signals first, then follow the HowTo on this page to produce share numbers comparable to these anchors. For the per-engine citation behavior behind the anchor gaps (why Perplexity and ChatGPT cite differently even when both are measured correctly), see Perplexity vs ChatGPT Citation Rules on chudi.dev.
Chudi Nnorukam
AI-Visible Web ArchitectBuilds chudi.dev and citability.dev. Authored the AI Visibility Readiness Framework. Contributor at freeCodeCamp /news.