Does Wikipedia Still Decide Your AI Visibility?

Citation share is the percentage of an answer engine's responses or source list in which a given domain appears — the metric every study below is measuring, just with different denominators. Semrush's tracking over several months found that on ChatGPT, Wikipedia appeared in roughly 55% of prompt responses in early August 2025 and fell below 20% by mid-September. Over the same period its share held near 3% on Google's AI Mode and around 0.8% on Perplexity — so the drop was a change at one engine, not a shift across the field.

Other datasets put the level in very different places:

Those findings look irreconcilable. They are not, and understanding why is more useful than picking one.

• Share of responses versus share of citations. "Wikipedia appeared in 55% of responses" and "Wikipedia was 13% of all citations" can both be true, because a single answer cites several sources — one measures presence, the other volume.
• Top-ten share versus total share. Restricting the denominator to the top ten domains inflates every figure inside it. A 26–48% top-ten share is not comparable to a 13% total share.
• Engine and query mix. Only around 11% of domains are cited by both ChatGPT and Perplexity, so a source that dominates one surface can be absent from another. This matters for anyone building an AI visibility programme, since a strategy tuned to one engine can miss the others entirely.

Then there is volatility, which has the most practical consequence. Wikipedia's ChatGPT share halved in weeks, and Reddit's moved from around 60% to around 10% over a similar period — a pattern examined further in how Reddit and forums shape what AI says about a brand. Whatever the true level, it is not a stable asset. Any strategy built on one domain is one platform change away from failing, which argues for diversification rather than chasing whichever source currently leads.

A note on sourcing, since it matters here: most of these studies are published by commercial vendors with a service to sell, methodologies vary, and none are peer reviewed. The direction of travel is consistent across them, which is worth something; the precise percentages are not.

An entity anchor is a stable, corroborated description of who an organisation is that other sources echo, which models treat as reliable background knowledge. Three mechanisms operate independently of citation counts.

Training data weight. Wikipedia is heavily represented in the corpora models learn from, so it shapes what a model believes about an entity before any search happens. That is background knowledge, not a citation, and it appears in no citation tracker — a distinction covered in more depth in entity SEO for AI search.

Description consistency. Models weigh corroboration across sources. A Wikipedia article gives every other publication a canonical description to echo, which produces the consistency that makes an entity recognisable. A brand with a clear, consistently described identity gets summarised accurately; one without gets summarised approximately.

Downstream propagation. Wikipedia content feeds knowledge panels, aggregators and countless derivative pages, so its influence multiplies through sources that are themselves cited.

The practical implication: the goal was never a Wikipedia page. The goal is a stable, verifiable, consistent public description of an organisation. Wikipedia is one route to that — neither the only one, nor an available one for most companies, a point explored further in structured data and entity identity for AEO.

Wikipedia requires notability: demonstrated significance shown through coverage in independent, reliable sources, with undisclosed paid editing and self-promotional articles treated as policy violations. Most B2B companies do not meet that bar, agencies promising a page frequently produce one that is deleted, and a deletion discussion is itself a permanent, public, searchable record — a genuinely bad trade.

What works instead, in rough order of effort to effect:

1. Make your own descriptions identical everywhere. Website boilerplate, LinkedIn, Crunchbase, industry directories, conference bios and press releases should carry the same wording for what an organisation does. Inconsistency is what makes an entity fuzzy to a model, a problem covered in building an AI-ready brand knowledge base.
2. Publish structured data. Organization schema with a clear name, description, founding details and sameAs links to other profiles gives machines an unambiguous identity to attach facts to.
3. Earn third-party mentions, not just links. An unlinked reference in a credible publication still functions as a consensus signal, and coverage across several independent sources does what a Wikipedia article would have done, as set out in why third-party brand mentions matter for GEO.
4. Be present where the engines actually look. The most-cited domains include community and professional platforms, not only publishers, and each engine draws from a different mix.
5. Test in every language sold in. Wikipedia coverage is dramatically thinner outside English, so in most markets the entity-anchor role falls to local sources, local-language directories and translated content — the multilingual dimension of generative engine optimization.

Entity consistency is fundamentally a multilingual problem: organisations described accurately in Bahasa Indonesia, Arabic or Portuguese are the ones that published a consistent description in those languages rather than hoping a translation would appear, an approach detailed in the multilingual content pipeline AI engines cite.