What Is a Good AI Citation Rate? 31,581 Measured Placements Say the Median Is 0.88%
The Machine Relations Index publishes 31,581 segment placements. Here is the full distribution of AI citation rates, the percentile bands, and what each rate is worth in rank terms.
Across 31,581 published segment placements in the Machine Relations Index, the median AI citation rate is 0.88%. The top decile begins at 4.55%. The top 1% begins at 16.67%. The highest rate any domain holds in any measured segment is 58.67%.
That is the benchmark almost every AI visibility conversation has been missing. The standard answer to "what is a good citation rate?" has been that the question is unanswerable because the prompt set defines the difficulty. That is true, and it is also a dodge. The difficulty can be measured. This page publishes the measurement.
Every number below is read from the public Machine Relations Index release mri_score_v2.0+2026-09-19+0cad03121f60, observation window 2026-05-10 to 2026-09-19, 126 days, six answer engines, 912 eligible buyer queries, 15,883 answer runs, 125,115 source events, 22,213 cited source domains. The machine-readable file is at machinerelations.ai/data/machine-relations-index.json.
The benchmark bands
A segment placement is one domain's citation rate inside one published segment, where a segment is one subject category paired with one buyer question shape. The Index publishes 88 segments and 31,581 placements in them. Sorted from lowest to highest, they fall out like this.
| Percentile | Citation rate |
|---|---|
| 10th | 0.16% |
| 25th | 0.32% |
| 50th (median) | 0.88% |
| 75th | 1.90% |
| 90th | 4.55% |
| 95th | 7.00% |
| 99th | 16.67% |
| Maximum | 58.67% |
Mean rate: 1.86%. The mean sits above the 75th percentile, which is the signature of a long right tail — a small number of very frequently cited domains pulling the average away from the typical one.
Read the table the other way and it becomes a scorecard. Find your own rate in the left column; the right column is the share of all measured placements you are beating.
| Your citation rate | Placements below it | Share of all 31,581 placements |
|---|---|---|
| 0.5% | 9,670 | 30.62% |
| 1.0% | 19,241 | 60.93% |
| 2.0% | 24,101 | 76.32% |
| 3.0% | 26,582 | 84.17% |
| 5.0% | 28,904 | 91.52% |
| 7.5% | 30,106 | 95.33% |
| 10.0% | 30,680 | 97.15% |
| 15.0% | 31,181 | 98.73% |
| 20.0% | 31,381 | 99.37% |
| 30.0% | 31,537 | 99.86% |
A 3% citation rate, which most dashboards render in a colour that suggests failure, is better than 84.17% of every measured placement in the index.
Why the median is that low
Two structural facts, both readable in the release.
The rooms are crowded. The smallest published segment carries 101 distinct cited domains. The largest carries 4,025. The typical segment carries a few hundred. When a few hundred domains divide the citations from roughly a hundred observed runs, most of them land on one or two runs each, and a sub-1% rate is the arithmetic outcome, not a verdict on the brand.
News questions and buying questions are different markets. Seven of the 88 published segments cover the legacy news-topic bucket, and they alone hold 12,028 of the 31,581 placements — the widest, thinnest rooms in the index, with a median rate of 0.17%. Strip them out and the 81 segments covering the six buyer question shapes hold 19,553 placements with a materially higher floor.
| Population | Segments | Placements | Median | 75th | 90th | 95th | 99th |
|---|---|---|---|---|---|---|---|
| All published segments | 88 | 31,581 | 0.88% | 1.90% | 4.55% | 7.00% | 16.67% |
| Buyer question shapes only | 81 | 19,553 | 1.46% | 2.94% | 6.06% | 9.42% | 20.00% |
| News-topic bucket only | 7 | 12,028 | 0.17% | — | — | — | — |
If your prompt library is built from category and comparison questions rather than news questions, benchmark against the middle row. Your realistic median is 1.46%, and 6.06% puts you in the top tenth.
What a rate is actually worth: rank, not percentage
A percentage with no room attached is not a benchmark. The same rate buys wildly different standing depending on which segment it is earned in. Taking each of the 81 buyer-shape segments and asking where a fixed rate would place:
| Fixed citation rate | Best rank achieved | Median rank | Worst rank | Segments where it reaches the top ten |
|---|---|---|---|---|
| 1% | 72 | 119 | 413 | 0 of 81 |
| 3% | 27 | 56 | 96 | 0 of 81 |
| 5% | 14 | 32 | 55 | 0 of 81 |
| 10% | 3 | 12 | 24 | 31 of 81 |
| 15% | 1 | 5 | 15 | 73 of 81 |
| 20% | 1 | 3 | 10 | 81 of 81 |
A 10% citation rate is rank 3 in the toughest room and rank 24 in the easiest. That eightfold spread in standing, from one identical number, is the strongest argument in this dataset against reporting citation rate as a bare percentage. It is also the reason a vendor comparison of two brands' citation rates across different categories tells you almost nothing.
The thresholds themselves vary the same way. Across the 88 published segments:
| Position | Lowest rate that achieves it | Median rate | Highest rate required |
|---|---|---|---|
| Rank 1 in the segment | 9.87% | 26.34% | 58.67% |
| Top three | 8.28% | 18.14% | 33.85% |
| Top ten | 4.12% | 10.69% | 19.57% |
The leader of the easiest segment holds 9.87%. The leader of the hardest holds 58.67% — huntress.com on cybersecurity best-tools questions, where a buyer asking for the best endpoint security platform gets an answer built substantially from one vendor's own material.
Where the bar sits by question shape
The six buyer question shapes are not equally contested. Rates below are medians across every placement in that shape, with the median winning rate across that shape's segments.
| Question shape | Segments | Placements | Median rate | 90th percentile | Median rank-1 rate |
|---|---|---|---|---|---|
| Top lists | 14 | 3,877 | 1.53% | 6.11% | 28.38% |
| Best tools | 14 | 3,413 | 1.52% | 7.02% | 34.09% |
| Comparisons | 13 | 2,593 | 1.52% | 6.86% | 28.00% |
| Is it worth it | 13 | 3,024 | 1.46% | 6.00% | 27.74% |
| How buyers choose | 14 | 3,638 | 1.01% | 5.38% | 25.24% |
| Problem-first research | 13 | 3,008 | 0.99% | 4.63% | 20.19% |
| News topic (legacy) | 7 | 12,028 | 0.17% | 1.00% | 11.35% |
Two readings matter for planning. Shortlist-shaped questions — best tools, top lists, comparisons — have the highest winning rates, meaning engines concentrate their answers on a handful of sources when a buyer wants a recommendation. Problem-first and how-buyers-choose questions are flatter: the winner holds less, the field is wider, and a mid-tier rate is worth more rank there than the same rate on a best-tools question. If a brand is invisible on shortlist questions, the flatter shapes are the cheaper entry point.
Where the bar sits by category
Across the six buyer question shapes, by subject category.
| Category | Segments | Placements | Median rate | Rate for top ten | Median rank-1 rate | Segment leader |
|---|---|---|---|---|---|---|
| Emergent prosumer | 6 | 1,038 | 2.00% | 9.50% | 21.88% | youtube.com 37.89% |
| Enterprise software | 6 | 1,241 | 1.87% | 8.42% | 22.24% | erpresearch.com 33.66% |
| Healthcare services | 3 | 636 | 1.87% | 9.80% | 24.51% | omnimd.com 32.71% |
| Family software | 6 | 876 | 1.85% | 12.02% | 23.93% | apple.com 37.04% |
| Fintech | 6 | 1,107 | 1.85% | 10.29% | 23.83% | airwallex.com 38.89% |
| Deep tech and hardware | 6 | 1,311 | 1.65% | 10.43% | 25.58% | vlsishuttle.com 35.24% |
| Education and training | 6 | 1,363 | 1.54% | 9.00% | 21.00% | reddit.com 33.96% |
| Cybersecurity | 6 | 1,481 | 1.53% | 10.53% | 27.21% | huntress.com 58.67% |
| Consumer health | 6 | 1,632 | 1.52% | 13.36% | 32.82% | healthline.com 43.18% |
| Consumer finance | 6 | 1,318 | 1.45% | 15.53% | 35.49% | forbes.com 39.39% |
| AI infrastructure | 6 | 1,440 | 1.30% | 13.21% | 32.70% | medium.com 46.75% |
| AI security and privacy | 6 | 1,758 | 0.94% | 12.24% | 32.20% | arxiv.org 47.37% |
| AI visibility and GEO | 6 | 2,515 | 0.88% | 10.78% | 27.14% | youtube.com 33.33% |
| Consumer products | 6 | 1,837 | 0.76% | 11.24% | 26.38% | forbes.com 29.77% |
The gap between the "top ten" column and the median column is the real cost of entry in each category. In consumer finance a domain needs 15.53% to reach a top-ten place against a median of 1.45% — roughly eleven times the typical placement. In enterprise software the same jump is about four and a half times. Consumer categories are winner-concentrated; B2B software categories are flatter and cheaper to climb.
What a good rate is for your kind of source
Citation rate is conditioned on what kind of source you are. The Index classifies every domain into one of nine source roles, and the typical rate differs by an order of magnitude between the top and bottom of that list.
| Source class | Placements | Median rate | 90th percentile | Highest rate held | Top-ten places held |
|---|---|---|---|---|---|
| Search or media platform | 104 | 7.66% | 27.83% | 37.89% | 46 |
| Community and social platform | 392 | 3.51% | 19.07% | 33.96% | 103 |
| Academic and government source | 883 | 0.94% | 4.63% | 47.37% | 29 |
| Editorial publication | 2,631 | 0.88% | 5.30% | 46.75% | 118 |
| Other observed source | 23,781 | 0.88% | 3.96% | 43.18% | 453 |
| Vendor-owned source | 1,775 | 0.84% | 5.66% | 58.67% | 95 |
| Wire and press-release distribution | 49 | 0.82% | 2.80% | 5.88% | 0 |
| Market and company database | 1,269 | 0.76% | 3.31% | 33.66% | 21 |
| Analyst and consulting research | 697 | 0.33% | 3.03% | 18.63% | 15 |
Three findings a buyer can act on.
Platform and community sources are a different weight class. YouTube, Reddit and their peers carry median rates four to nine times the rest of the index. A brand comparing its own site against a Reddit thread's citation rate is comparing across classes, and the comparison flatters nothing.
Vendor-owned content can win outright, but rarely. The single highest rate in the entire index belongs to a vendor-owned domain, and vendor-owned sources hold 95 top-ten places. Yet the class median, 0.84%, sits below the editorial publication median. Owned content is a lottery with one enormous prize and a thin expected value.
Wire distribution and analyst research are the weakest classes measured. Across all 88 published segments, the best placement any wire or press-release domain holds is rank 24, businesswire.com on fintech top-list questions, cited in 5.88% of that segment's runs. Zero top-ten places, across every wire domain and every segment. Analyst and consulting research carries the lowest median of the nine classes at 0.33%. Both are historically expensive line items. Neither converts into AI citation at a rate that matches its budget share, which is consistent with the earned versus owned citation analysis published on the same instrument.
Most domains have no rate to benchmark at all
The most important context for any citation-rate benchmark is how few domains qualify for one. The Index publishes an overall citation rate for a domain only once it clears an evidence floor of at least 10 observations across at least 7 distinct run dates.
Of 22,213 cited source domains in this release, 503 clear that floor and carry a graded overall rate: 14 graded A, 58 graded B, 431 graded C. The remaining 21,710 are marked collecting — observed, but not yet on enough distinct days to publish a stable rate.
Among the 503 graded domains, overall citation rate across the whole index runs:
| Percentile | Overall citation rate |
|---|---|
| 25th | 0.23% |
| 50th (median) | 0.29% |
| 75th | 0.43% |
| 90th | 0.77% |
| 95th | 1.23% |
| Maximum | 12.67% |
The ceiling for a whole-index rate is reddit.com at 12.67%, followed by youtube.com at 9.02%, linkedin.com at 6.79%, medium.com at 5.08% and forbes.com at 4.14%. Any vendor reporting a brand's "AI citation rate" above 13% across a broad query set is either measuring a narrower room than the index does, or counting something other than what the index counts. Ask which.
Collecting is also not zero. A domain with no published rate has not been shown to be uncited; it has been shown to lack enough distinct observation days for a stable number. Treating the two as the same is the most common misreading of any AI visibility dashboard.
How to use these numbers
- Name your room before you name your rate. Category plus question shape. A rate without both is not comparable to anything, including your own rate last quarter.
- Benchmark against 1.46%, not 0.88%, if your prompts are buying questions. The all-segment median is dragged down by the news-topic bucket.
- Convert the rate to a rank. Your segment's top-ten threshold is published on its own page, for example consumer finance best-tools questions. The rank is the number that moves a buying decision; the percentage is the number that moves a dashboard.
- Compare within your source class. An editorial placement at 3% and a community thread at 3% are not equivalent achievements. The community thread is below its class median; the editorial placement is above the 75th percentile of its class.
- Check the evidence grade before you act on a change. A rate that moves from 1.2% to 2.4% on a domain still marked collecting is more likely to be sampling than progress.
How this compares with the independent literature
The shape of this distribution is consistent with what independent measurement of generative engines has found, which is worth stating because a single instrument's distribution is only as useful as its agreement with others.
A large-scale empirical study of six LLM-based search engines and two traditional search engines across 55,936 queries found that generative engines cite a more diverse set of domains than traditional search, with 37% of domains unique to the LLM-based systems (Source Coverage and Citation Bias in LLM-based vs. Traditional Search Engines). That breadth is exactly what produces rooms of several hundred cited domains and a sub-1% median.
Breadth and concentration coexist. Work on citation selection across eleven models from three vendors found that the top decile of candidate sources receives 23.3% to 30.2% of citations against 15.6% under indifferent selection, with cross-vendor agreement nearly matching within-vendor agreement (When AI Writes, Who Gets Cited?). A long tail with a heavy head is the same structure the percentile table above describes.
A two-stage measurement framework separating citation selection from citation absorption, built on 602 controlled prompts and 21,143 search-layer citations, found that Perplexity and Google cite more sources on average while ChatGPT cites fewer with higher average influence per fetched page (From Citation Selection to Citation Absorption). Citation rate measures selection. It is a necessary metric and a partial one.
An audit of Google AI Overviews on high-stakes queries drawn from MS MARCO Web Search found AI-generated documents cited more frequently than human-authored ones even after controlling for retrieval rank (Auditing Citation Behavior in AI-Generated Search Summaries). Any benchmark of a brand's rate over time has to assume the eligible source population is changing underneath it.
Two further points of reference matter for anyone building an internal benchmark. The AI Visibility Lifecycle Framework, an IETF informational draft, models the path from crawling to visible placement as a sequence of stages, which is the structural reason a crawled page and a cited page are different populations and carry different denominators. And the scholarly world reached the same conclusion about measurement discipline first: DataCite and Make Data Count publish a Data Citation Corpus precisely so that reuse can be counted from actual citation records rather than estimated. A citation rate is only as trustworthy as the record it is counted from.
FAQ
What is a good AI citation rate?
Against the 31,581 published segment placements in Machine Relations Index release mri_score_v2.0+2026-09-19+0cad03121f60, the median is 0.88% and the 90th percentile is 4.55%. On buyer question shapes alone the median is 1.46% and the 90th percentile is 6.06%. A rate above 4.55% is top-decile across the whole measured market. A rate above 10.69% is typically enough for a top-ten place in a given segment.
Why is the median citation rate under 1%?
Because each measured segment contains between 101 and 4,025 distinct cited domains dividing the citations from roughly a hundred observed answer runs. Most domains appear in one or two runs. The long tail is the market structure, not a measurement flaw.
Is a 3% citation rate good or bad?
A 3% rate sits above 84.17% of all published segment placements. In rank terms it lands between rank 27 and rank 96 depending on the segment, and does not reach a top-ten place in any of the 81 buyer-shape segments. It is a strong percentile and a weak rank, which is why both numbers belong in the same report.
How does citation rate differ by source type?
Median rates run from 7.66% for search and media platforms and 3.51% for community platforms down to 0.76% for market and company databases and 0.33% for analyst and consulting research. Editorial publications and vendor-owned sources both sit near the index median, at 0.88% and 0.84%. Benchmark within your class.
Why does my brand have no published citation rate?
The Index publishes a domain rate only after at least 10 observations across at least 7 distinct run dates. In this release, 503 of 22,213 cited domains clear that floor. A domain marked collecting has been observed but has not yet accumulated enough distinct run dates for a stable rate. See how citation rate is calculated and the citation rate definition.
Who coined Machine Relations?
Jaxon Parrott, founder of AuthorityTech, coined Machine Relations in 2024 to name the discipline of earning AI citations and recommendations through third-party credibility. The framework is documented at machinerelations.ai.
Sources and method
All index figures are read from the public Machine Relations Index, release mri_score_v2.0+2026-09-19+0cad03121f60, retrieved in full from machinerelations.ai/data/machine-relations-index.json on 2026-09-19 at 14:40 UTC. Observation window 2026-05-10 to 2026-09-19, 126 days. Six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity. 912 eligible queries, 15,883 answer runs, 125,115 source events, 22,213 cited source domains, 157 taxonomy strata of which 88 are published and 63 are still collecting.
Segment ranks and totals are read from each domain's relative_signal.category_signals[] field rather than derived by sorting, because the Index uses standard competition ranking — tied domains share the better rank — and re-deriving order by sorting pushes every tied domain after the first one down. Citation rates are read from the same records. Percentiles are linear-interpolated over the full set of published placements, the convention described in the NIST/SEMATECH Engineering Statistics Handbook. Human-readable segment pages carrying the same figures are at machinerelations.ai/index, for example cybersecurity best-tools questions.
The Index measures the market's buyer questions, not queries about any one brand, so it reports which sources engines cite — never whether a specific brand's own pages win. The six observed engines each document their own retrieval and crawling behaviour, and a benchmark built on citation rate should be read against those operator descriptions rather than against a vendor's summary of them: OpenAI bots, Google crawlers, Google AI Mode in Search, Perplexity bots and web search for Claude. Independent studies cited above: arXiv 2512.09483, arXiv 2608.19230, arXiv 2604.25707 and PMLR v318. Related Machine Relations work on the shape of the distribution: how concentrated AI citations are.
Correlation stays correlation. A domain's citation rate describes what engines cited in the window. It does not establish that any particular placement, page change or campaign caused it.