Wikipedia Is Cited in 1.22% of AI Answers. The Question Shape Moves It 15x.
Measured against 15,883 monitored answer runs, wikipedia.org ranks #27 of 22,213 cited domains at a 1.22% citation rate. Its rate ranges from 11.43% to 0.74% depending on the question shape, and it is observed in 17 of 25 released categories.
Wikipedia is cited in 1.22% of monitored AI answers. Reddit is cited in 12.67% of the same runs.
Both numbers come from the same release of the Machine Relations Index, read on the same day, against the same denominator: 15,883 monitored answer runs across six answer engines between May 10 and September 19, 2026. wikipedia.org was cited in 193 of them. reddit.com was cited in 2,012.
That gap is the opposite of what most AI-visibility advice implies. It is also not the interesting part. The interesting part is that Wikipedia's citation rate is not one number. Inside a single category it moves from 11.43% to 0.96% depending only on the shape of the question asked — a 12x swing on the same subject matter, from the same engines, in the same window.
What the measurement is
Release mri_score_v2.0+2026-09-19+0cad03121f60, methodology mri_score_v2.0, generated September 19, 2026. Window May 10 to September 19, 2026 — 126 days, 15,883 observed answer runs, 22,213 cited domains, six engines: ChatGPT, Claude, Gemini, Google AI Mode, Google AI Overviews and Perplexity.
The Index pairs a subject category with a question shape to form a stratum, and a stratum publishes a rate only after it clears an evidence floor of 10 observed runs across 7 distinct run dates. Thin strata stay marked as collecting and publish no rate. Every figure below carries its own denominator, because rates from different strata are not comparable to each other.
Citation rate here means the share of observed runs in which the domain appeared as a cited source. It is not traffic, not recommendation quality, and not proof that the answer's claims came from that source.
Where Wikipedia actually stands
| Domain | Index source role | Citation rate | Cited runs | Standing | Engines | Confidence |
|---|---|---|---|---|---|---|
| Community and social platform | 12.67% | 2,012 of 15,883 | #1 of 22,213 | 4 | A | |
| YouTube | Search or media platform | 9.02% | 1,433 of 15,883 | #2 of 22,213 | 6 | A |
| G2 | Market and company database | 2.82% | 448 of 15,883 | #9 of 22,213 | 6 | A |
| Wikipedia | Academic and government source | 1.22% | 193 of 15,883 | #27 of 22,213 | 6 | B |
Two things are true at once, and most arguments about Wikipedia pick one and drop the other.
Wikipedia is genuinely near the head of the index. #27 of 22,213 observed domains puts it in the top 0.12% of everything the engines cited. It was cited on 69 of the window's 126 days, and it is one of the domains observed across all six engines rather than a subset — Reddit, the #1 domain, was observed on four.
And it is still a 1.22% event. In a monitored buying or research answer, Wikipedia shows up roughly once every 82 runs. A visibility programme that treats a Wikipedia article as the unlock is buying a source the engines reach for in one answer out of eighty-two.
The question shape moves the number more than the subject does
Wikipedia has 34 published strata in this release and 4 more still collecting. Sorted by rate, the top and the bottom are not different industries. They are different question shapes.
| Category | Question shape | Citation rate | Cited runs | Standing in stratum |
|---|---|---|---|---|
| Deep Tech & Hardware | Is it worth it | 11.43% | 12 of 105 | #8 of 181 |
| Deep Tech & Hardware | Top lists | 6.60% | 7 of 106 | #23 of 279 |
| Deep Tech & Hardware | How buyers choose | 6.15% | 8 of 130 | #24 of 195 |
| AI Security & Privacy | Comparisons | 6.14% | 7 of 114 | #19 of 263 |
| Fintech | Top lists | 3.92% | 4 of 102 | #49 of 200 |
| AI Infrastructure | Best tools | 3.90% | 3 of 77 | #72 of 204 |
| … | … | … | … | … |
| Family Software | Comparisons | 0.93% | 1 of 108 | #101 of 101 |
| Cybersecurity | How buyers choose | 0.76% | 1 of 131 | #320 of 325 |
| Consumer Products | Top lists | 0.76% | 1 of 131 | #374 of 380 |
| Consumer Products | Is it worth it | 0.74% | 1 of 136 | #224 of 228 |
Hold Deep Tech & Hardware still and vary only the question. "Is it worth it" cites Wikipedia in 11.43% of runs and ranks it #8 of the 181 domains in that stratum. Problem-first research, same category, same window, cites it in 0.96% of runs and ranks it #155 of 157. Same encyclopedia, same buyer, same engines, twelvefold difference.
Now hold the question shape still and vary the category. "Is it worth it" gives Wikipedia 11.43% in deep tech and 0.74% in consumer products, where it ranks #224 of 228 — fifth from last in its own stratum.
In the news-driven strata, which carry the largest denominators in the release, Wikipedia is steadier and unremarkable: 2.82% in martech and advertising (17 of 602 runs, #55 of 1,244), 2.63% in HR and talent (16 of 608, #54 of 1,249), 2.01% in enterprise software (12 of 596, #72 of 1,375), 1.46% in fintech (9 of 616, #132 of 1,531), 1.44% in cybersecurity (9 of 623, #125 of 1,340).
Across the 25 categories released today, Wikipedia is observed in 17. In the other eight — AI Visibility & GEO, Industrial, Local Services, Logistics & Freight, Physical Consumer, Professional Services, Sales/GTM/CRM, and VC & Private Equity — no Wikipedia citation was observed in this window.
Why the pattern looks like this
The Index classifies every cited domain by source role. wikipedia.org is classified as an Academic and government source — the same role family as a regulator's site or a university department, and a different role from Reddit's community platform, G2's market database, or a vendor's own domain.
That classification predicts the shape pattern better than any authority score does. The strata where Wikipedia climbs are the ones whose answers need a definition, a specification or a neutral description of a technology: "is it worth it" in deep tech, comparisons in AI security and privacy. The strata where it sinks are the ones whose answers need a current shortlist, a price or a verdict on a specific product: top lists and worth-it questions in consumer products, how-buyers-choose in cybersecurity.
An engine reaching for Wikipedia is reaching for reference, not for recommendation. Which means a Wikipedia article about your company does a specific, narrow job: it helps a machine resolve who you are. It does not put you into the answer where the buyer is choosing.
There is a second, structural reason a Wikipedia strategy behaves unlike a content strategy. You do not control the page. Wikipedia's notability standard for organizations requires significant independent coverage before an article can exist at all, its verifiability and reliable sources policies decide what the article may say, and its conflict of interest guideline and paid-contribution disclosure requirement govern who may edit it. Wikidata's notability policy is looser but serves a different function: identity, not prose.
So the earned-media work comes first in both directions. The independent coverage that makes an article possible is the same coverage the engines cite directly, at rates far above 1.22%.
What to do with this
Treat Wikipedia and Wikidata as entity resolution, not citation share. A Wikidata item and a consistent sameAs set across your owned properties is how a machine stops confusing your company with a similarly named one. Google documents the same mechanism for organization structured data. Budget it as identity infrastructure, and judge it by whether engines describe you correctly — not by citation count.
Check your own category and shape before you spend. The swing inside one category is larger than the gap between most categories. Open the stratum leaderboard for the exact category and question shape your buyers use, and read where reference sources sit in that specific list. A ranking that is true for deep-tech "is it worth it" is wrong for consumer-products "top lists" by two orders of position.
Measure against the shapes your buyers actually ask. If your monitored questions skew definitional, a reference-role source like Wikipedia will look powerful in your data. If they skew toward shortlists and verdicts, it will look irrelevant. Both readings are correct about their own sample and wrong as a general claim — which is why the Index keeps each stratum's denominator separate rather than pooling rates.
Sources and method
Every figure above was read on September 19, 2026 from the live public pages of the Machine Relations Index at release mri_score_v2.0+2026-09-19+0cad03121f60, window 2026-05-10 to 2026-09-19, 15,883 observed answer runs, 22,213 cited domains, six engines. The domain profiles read were wikipedia.org, reddit.com, youtube.com and g2.com, each fetched over plain HTTPS and each returning 200. Wikipedia's stratum table on that profile carries 38 rows: 34 published with a rate, 4 still collecting. The 25-category denominator is the released category list on the Index; Wikipedia rows appear under 17 of those 25, and the eight named above carry no observed Wikipedia row in this release.
The release is machine-readable: the public JSON artifact and the release manifest carry the same figures and the same release id, and every domain profile carries a live badge that reflects the current release rather than the date this page was written.
Wikipedia's own policy pages are cited from en.wikipedia.org directly and were each fetched at 200 on September 19, 2026: notability for organizations, verifiability, reliable sources, conflict of interest and paid-contribution disclosure. Wikidata's notability policy and a sample Wikidata item were read the same way. Wikipedia's corpus-level availability, which is why its text reaches models independently of any citation, is documented at Wikimedia's dumps service and its readership at Wikimedia Statistics; open web corpora such as Common Crawl carry it as well.
Limits worth stating plainly. The Index observes root-domain citations in monitored prompts; it does not observe every AI answer, and a citation is not evidence that the answer's claim came from that source. Wikipedia carries Confidence B at the overall level, one grade below the A carried by Reddit, YouTube and G2, so its overall rate is the less settled of the four. Per-stratum confidence is unavailable in this release, so the stratum rates above are reported with their run counts rather than with a grade. Strata marked collecting publish no rate and are excluded from every comparison here. A subsequent release will move these numbers; the release id is printed so the reading can be reproduced or contradicted.