Short answer: Wikipedia AI visibility matters, but mostly in ways that do not show up as citations, and much less than the industry talk suggests for the questions buyers actually ask. Wikipedia shapes what models know about a company through training data and through the entity records that search engines build, and ChatGPT cites it more than any other domain across all topics. But when we asked five AI models 240 buyer questions across 12 categories, Wikipedia was cited 16 times in 1,188 answers, against 504 times for Reddit, and only 2 of those 16 citations were about a specific business. For most brands, a Wikipedia article is a consequence of being notable, not a lever for being recommended. And it is the one channel where trying to pull the lever yourself is explicitly against the rules.

That is our take. The rest of this piece shows the evidence, explains what Wikipedia's own policies allow a company to do, and sets out where the effort usually pays off better.

  • Two different jobs. Wikipedia helps engines know who you are (identity, history, category). It rarely helps them decide whether to recommend you for a purchase.
  • Buyer questions cite other sources. In our benchmark the engines cited Wikipedia for concepts such as "resort fee" or "admission to the bar", not for which hotel or law firm to choose.
  • You cannot buy or write your way in. Wikipedia's notability rules require significant coverage in multiple independent, reliable, secondary sources; its conflict of interest rules strongly discourage you from editing your own article and require paid editors to disclose who pays them.
  • Wikidata and your own entity data are the practical levers. They are smaller, more structured and more within reach, provided you follow their rules too.
  • Measure identity and discovery separately. Wikipedia work, if it works, shows up in how engines describe you, not in how often they recommend you.

Wikipedia AI visibility in numbers: where Wikipedia is cited, and where it is not

The claim that "Wikipedia is the most important source for AI" usually rests on one kind of data: citation shares across every topic engines answer. That data is real. It just answers a different question from the one most marketers are asking.

Across all topics, ChatGPT leans on Wikipedia

Profound analysed 680 million citations collected between August 2024 and June 2025. In ChatGPT, Wikipedia was the most-cited domain, at 7.8% of all citations; Reddit was second at 1.8%. In Google AI Overviews and Perplexity, Reddit led instead, at 2.2% and 6.6% respectively.

ChatGPT: the two most-cited domains (share of all citations)Profound, 680 million citations, Aug 2024 to Jun 2025, all topicsWikipedia7.8%Reddit1.8%In the same data, Reddit led in Google AI Overviews (2.2%) and Perplexity (6.6%); not drawn here.
Figure 1. Across all topics, ChatGPT cites Wikipedia more than any other domain. Share of all citations in ChatGPT. Source: Profound, 680 million citations, August 2024 to June 2025.

Those shares cover everything people ask: history, science, definitions, celebrities, how-to questions. Wikipedia is the obvious reference for most of that. Whether it is also the reference for "which accountant in Leeds" or "best running shoes for flat feet" is a separate question, and the cross-topic share cannot answer it.

On buyer questions, Wikipedia almost disappears

Our own benchmark looked only at buyer questions. On 2026-09-07 we asked 5 AI models 20 questions in each of 12 categories, from dentists in Austin and restaurants in Seattle to e-bikes, skincare and project management software, and recorded every citation. Wikipedia was cited 16 times in 1,188 successful answers. Reddit was cited 504 times in the same answers.

Citations on buyer questions: Wikipedia vs RedditEach category: 5 models × 20 questions; upper grey bar = Reddit, lower bar = WikipediaRestaurants752Fitness gyms720Electric bikes672Furniture580Dentists500HVAC & plumbing370Skincare290Hotels282Real estate agents271Pet supplies253Project mgmt software190Law firms176Wikipedia: 16 in totalReddit: 504 in total
Figure 2. Wikipedia AI visibility on buyer questions: citations to Wikipedia and to Reddit in each of 12 categories. Source: InsightWonder industry benchmark, 5 models × 20 buyer questions per category, collected 2026-09-07.

In half of the twelve categories, Wikipedia was not cited once. Its best showing was law firms, with 6 citations, all for questions about trust and qualifications. By engine, Perplexity accounted for 11 of the 16, ChatGPT and Claude for 2 each and Grok for 1; Gemini cited Wikipedia zero times in our runs.

And when it is cited, it is cited for concepts

The 16 citations are small enough to read one by one, and they tell a consistent story. Fourteen pointed to topic pages: Electric bicycle laws, Step-through frame, Resort fee, Hotel amenity, Personal injury lawyer, several pages on admission to the bar, Pet store, Specialty store, Types of restaurant and a list of Denver neighborhoods. Only two pointed to an article about a specific business: Petco and Nue, a Seattle restaurant.

What the 16 Wikipedia citations pointed toConcept or topic pages14Pages about one business2Concept pages, for exampleElectric bicycle laws · Step-through frameResort fee · Hotel amenity · Types of restaurantAdmission to the bar in the United StatesPersonal injury lawyer · Pet store · Specialty storeBusiness pagesPetco (pet supplies)Nue (Seattle restaurant)
Figure 3. What the 16 Wikipedia citations in the benchmark pointed to: 14 concept or topic pages, 2 pages about a specific business. Source: InsightWonder industry benchmark, collected 2026-09-07.

This is the pattern you would expect from what Wikipedia is: an encyclopedia that explains concepts and documents notable subjects. When a buyer asks "how do I know a lawyer is licensed?", the engine reaches for Wikipedia's explanation of bar admission. When the buyer asks "which personal injury firm in Chicago should I call?", the engine reaches for legal directories, bar associations, firm websites and Reddit. In our law-firm data, Avvo, the American Bar Association and Super Lawyers were each cited more often than Reddit, and none of them is Wikipedia.

How to run the same check for your category

You do not need a twelve-category benchmark to see whether Wikipedia matters for your buyers. A single afternoon is enough for a first read.

  1. Take 20 to 30 buyer questions that do not contain your brand name: recommendations, comparisons, "is it worth it" and "how do I choose" questions.
  2. Run them in the engines your buyers use, and export every cited URL.
  3. Count the Wikipedia citations, and for each one note whether it points to a concept page (a law, a product type, a fee, a technique) or to a page about a specific company.
  4. Run a second, smaller set of identity questions with your name in them ("what is [Company]", "who founded [Company]") and note whether Wikipedia or Wikidata appears there.

If Wikipedia shows up only on concept pages in the first set, and your article (if any) shows up in the second, you are seeing the same split we saw: Wikipedia explains the category and identifies the company, while something else decides the recommendation. If a Wikipedia page about a competitor appears in your buyer answers, read it carefully; it tells you which facts about that competitor the engine treats as settled.

What this data does not show

Citations are the visible part of an answer. They do not show what a model already knew before it searched, and that is where Wikipedia's influence is largest. Our benchmark covers one date, five model versions accessed through their APIs, and buyer questions in English with US sample cities. A brand-name question ("what is [Company]?", "who owns [Company]?") behaves differently, and that is where a Wikipedia article, if you have one, is most likely to be cited.

Why Wikipedia still shapes AI answers: training data and entities

If Wikipedia rarely appears as a citation for buyer questions, why does it matter at all? Because citations are only one of the routes by which a source shapes an answer. Wikipedia's influence runs mainly through two others.

Route one: training data

The Wikimedia Foundation, which runs Wikipedia, put it directly in October 2025: almost all large language models train on Wikipedia datasets, and search engines and social platforms prioritise its information to answer users' questions. Facts that appear in a Wikipedia article about your company (founding year, headquarters, founders, what you make, notable controversies) are likely to be part of what a model "knows" about you without searching. If the article is wrong, a model may repeat the error confidently and without a citation you could trace.

The same post reported a cost of that success: human pageviews on Wikipedia fell roughly 8% compared with the same months a year earlier, which the Foundation attributes to generative AI and social media answering questions directly. Earlier in 2025 it reported that at least 65% of its most resource-consuming traffic came from bots, while bots made up about 35% of overall pageviews. Wikipedia is read less by people and more by machines. For a brand, that means the article matters less as a destination and more as an input.

Route two: entity records

Search engines and AI systems need to know that "Acme", "Acme Inc." and "acme.com" are the same thing, and that it is a company that makes widgets, not the cartoon brand. They resolve this with entity records: Google's Knowledge Graph, Wikidata and similar databases. Wikipedia sits close to the middle of that web. Google's Knowledge Panel Help says panels are generated automatically for entities in the Knowledge Graph, with information drawn from various sources across the web. Wikidata, Wikipedia's structured sister project, links each item to its Wikipedia articles in every language and to identifiers in other databases.

Three routes from Wikipedia to an AI answerWikipediaarticle text, infoboxTraining datawhat a model "already knows"large effect, no visible sourceEntity recordsWikidata, Knowledge Graph, panelsdecides "who you are"Retrieval citationsvisible links in an answerrare on buyer questionsThe first two routes shape identity; only the third shows up as a citation.
Figure 4. Three routes by which Wikipedia reaches an AI answer. Diagram based on Wikimedia Foundation statements on training data and Google's Knowledge Panel Help; no data.

The practical implication: a clean, consistent entity record helps an engine attach the right facts to the right company. That matters most for brands with ambiguous names, brands that changed names, and brands whose category is easy to misread. It matters least for a local business whose name is unique in its city and whose customers ask about services rather than identity.

How a wrong fact travels

The two routes also explain why a small error on Wikipedia can have outsized effects. Consider an illustrative case, not a real one: a company's article gives the wrong founding year. The infobox on Wikipedia draws on the same fact; the Wikidata item may carry it too, and other databases and websites that reuse Wikidata pick it up. A search engine's entity record now has several sources that agree with each other, all ultimately copied from one. A model trained on any of them learns the wrong year, and a live-search engine that checks finds the same wrong year in several places and treats it as confirmed.

The fix runs in the same order. Correct the source, through the talk-page process described below, then correct Wikidata, then make sure your own site states the right fact clearly. Copies elsewhere tend to follow the sources they were copied from, but slowly, and models trained before the correction will keep the old fact until they are updated. This is the strongest argument for checking your Wikipedia and Wikidata records even if you never expect to be cited from them.

Our take

Put the two routes together and the role becomes clear. Wikipedia is upstream infrastructure for identity: who you are, what category you belong to, what is documented about you. It is rarely the source for preference: whether you are a good choice for a particular buyer. Preference comes from reviews, directories, community discussion, comparison articles and your own pages. Spending a quarter chasing a Wikipedia article in the hope of being recommended more often is, for most brands, effort pointed at the wrong layer.

Three claims we hear, and what the evidence supports

  • "ChatGPT reads Wikipedia first, so you need an article." ChatGPT does cite Wikipedia more than any other domain across all topics, according to Profound's data. That is a statement about everything people ask ChatGPT, not about buyer questions, where our benchmark found ChatGPT cited Wikipedia twice in 240 answers.
  • "A Wikipedia article makes engines trust you, so you get recommended." An article makes your identity easier to resolve. Recommendation depends on evidence about whether you are a good choice, which an encyclopedia article is deliberately neutral about. A promotional article that would argue the case is exactly what Wikipedia's rules remove.
  • "More edits mean more visibility." Edit volume has no published role in how engines select sources, and visible editing by a company is more likely to attract scrutiny than citations. Accuracy, not activity, is the goal.

Does your company qualify? Wikipedia's notability rules for companies

Before discussing what to do, it helps to be honest about whether there is anything to do. English Wikipedia's notability guideline for organizations and companies sets the bar, and it is higher than most companies expect.

The core test

A company, organization, product or service is presumed notable if it has been the subject of significant coverage in multiple reliable secondary sources that are independent of the subject. The guideline adds that no company or organization is inherently notable: size, age and importance do not count on their own. Each source has to pass four tests at once: it covers the subject directly and in depth, it is completely independent of the company, it is reliable, and it is secondary. There must be multiple such sources, and at least one must have a regional, national or international audience rather than a purely local one.

Wikipedia notability for companies: every source must pass all fourSignificantcoverage+Fullyindependent+Reliable+SecondaryOne qualifyingsourceMultiple qualifying sources, and at leastone beyond a purely local audiencePresumed notableDoes not count (examples)Routine news: funding, hires, launchesStaff quotes; passing mentionsListings: address, hours, products"Best of", "top 100", "fastest growing" listsDirectories that include everyone
Figure 5. Wikipedia's notability test for companies, summarised from the guideline on organizations and companies (read 2026-10-23). Diagram; no data.

What does not count

The guideline is unusually specific about coverage that looks like evidence but is not. That list is the most useful thing a marketing team can read before asking "should we have a Wikipedia page?", because much of what a PR programme produces falls on it.

Usually counts toward notabilityDoes not count toward notability
An in-depth feature about the company in a national or regional newspaper, written by its own journalistRoutine announcements: funding rounds, acquisitions, hires, product launches, new branches, quarterly results
An independent analysis or investigation of the company's practices or market impactQuotes from your staff in someone else's story, and passing mentions as an example of a type of company
A substantial independent book chapter or academic study about the companyListings: addresses, opening hours, product offerings, statistics, office locations
Detailed independent reviews that discuss the company or product in depth (see the guideline's section on product reviews)Inclusion in "best of", "top 100" or "fastest growing" lists, unless the list itself is notable
Coverage that discusses and evaluates, not just reportsDirectories and databases that try to include every company, sponsorships, awards that are not themselves notable, speeches by your executives

A quick self-assessment

Gather every piece of coverage about your company from the last few years and sort it with the table above. Then ask three questions.

  1. How many pieces are about us, in depth, written independently, by publications with editorial standards? If the answer is fewer than two or three, stop here.
  2. Is at least one of them from a publication with more than a local audience?
  3. Could a volunteer who has never heard of us write a neutral, sourced article from those pieces alone, without our help? If not, the article would depend on material we provide, which is exactly what the rules are built to prevent.

Many good, successful companies fail this test. That is not a judgement on the business. It reflects what an encyclopedia is for. If you fail it, the right response is not to find a way around it, but to accept that your AI visibility will be built elsewhere.

Conflict of interest: what brands must not do on Wikipedia

Wikipedia is the one platform in this space with a detailed, written rulebook specifically about companies editing content about themselves. Ignoring it is the fastest way to turn a neutral article into a story about your company's editing.

The rules in plain language

Wikipedia's conflict of interest guideline covers anyone writing about their employer, client or financial interests. Owners, employees, contractors and investors all count. If that is you, the guideline says you should disclose the conflict, you are strongly discouraged from editing the affected article directly, you may propose changes on the article's talk page using the {{edit COI}} template so other editors can review them, and you should submit new articles through the Articles for Creation process instead of publishing them yourself.

Paid editing has a stricter rule. Anyone paid to edit must disclose who is paying, who the client is and any other relevant affiliation; the guideline describes this as a requirement of the Wikimedia Foundation's terms of use. Paid editing without that declaration has its own name, undisclosed paid editing, and Wikipedia treats it as a serious problem.

What a company may do, and may not do

ActionAllowed?How, if allowed
Rewrite the article to sound more positiveNoNot at all; promotional tone is grounds for removal under "Wikipedia is not a soapbox or means of promotion"
Fix a spelling mistake, a broken link or obvious vandalismYesDirectly, as an "unambiguously uncontroversial" edit, while disclosed; stop if another editor objects
Correct an outdated fact (new CEO, new headquarters, renamed product)Yes, by requestDisclose on your user page, post an {{edit COI}} request on the talk page, cite an independent source
Remove accurate criticism that is reliably sourcedNoYou can point out on the talk page if a criticism is given undue weight, with sources; volunteers decide
Create a new article about your companyDiscouragedOnly through Articles for Creation, disclosed, and only if the notability test above is met
Pay an agency to write or edit the articleOnly with full disclosureThe agency must disclose employer, client and affiliation; most of the same limits on direct editing apply
Ask friends or staff to edit "as volunteers"NoTheir edits are COI edits too; undisclosed, they are the worst case

Why the rules bite harder than on other platforms

Three features make Wikipedia different. Every edit is public and permanent in the page history, with the account or IP address that made it. Experienced volunteers actively look for promotional editing. And, as the guideline itself notes, no one controls articles: once an article exists, anyone may add well-sourced information you would rather not see, and the community may delete or keep the article regardless of what you want.

There is also a legal dimension in some jurisdictions. Wikipedia's guideline records that in 2012 a Munich appeals court (the Oberlandesgericht) ruled that edits made by a company or its agents to influence customers amount to covert advertising under European fair-trading law. Treat that as a reason to talk to counsel before any programme of Wikipedia work, not as a summary of the law in your market.

When someone else edits your article badly

The harder case is not your own editing but other people's: a disgruntled former employee, a competitor, or simply an editor who added an outdated claim. The rules still point the same way, and the temptation to "just fix it" is strongest exactly when it is most dangerous.

  1. Do not revert repeatedly. Undoing another editor's change again and again is an edit war, and a company account in an edit war is the story a reporter will write.
  2. Separate wrong from unwelcome. A claim that is false or unsourced can be challenged on the talk page with evidence. A claim that is accurate and reliably sourced but unflattering is part of the record; your recourse is to argue about weight and balance, with sources, not removal.
  3. Escalate through the right channel. If the talk page is quiet, the guideline points to a related WikiProject or the conflict of interest noticeboard. For articles about people, the biographies of living persons noticeboard exists for exactly this.
  4. Treat defamation as a legal matter. If a claim is defamatory, the guideline allows the subject to remove it quickly and then notify Wikipedia's volunteer response team. Anything beyond that is for counsel.

The AI-specific risk

Because models learn from Wikipedia and engines retrieve from it, a visible controversy about a company's editing can become part of the company's description. A neutral, slightly dull article that you did not write is a better outcome than a polished one with a "conflict of interest" maintenance banner, a public talk-page dispute and a news story about it.

What brands can do legitimately to improve Wikipedia AI visibility

Within the rules, there is a short list of useful, low-risk work. None of it guarantees an article or a particular wording. All of it makes the record more accurate, which is the only goal Wikipedia will support.

Correcting an article when you have a conflict of interest1 Disclosepersonal account,employer named2 Talk page{{edit COI}},one issue each3 Cite sourceswhat to change,independent source4 Volunteer reviewmay accept,may decline5 Log the datethen re-runidentity promptsExceptions: spelling, broken links and obvious vandalism may be fixed directly, while disclosed;if any other editor objects, go back to the talk page.
Figure 6. The disclosed route for correcting an article when you have a conflict of interest, summarised from Wikipedia's COI guideline. Diagram; no data.

If you already have an article

  1. Read it as an engine would. Check the opening paragraph, the infobox and the categories: name, founding date, headquarters, founders, industry, products, website. These are the facts most likely to end up in answers about you.
  2. Create a disclosed account. Use a personal account for a named employee, with a user page that states the employer. Do not use a shared "Acme Marketing" account; Wikipedia's username policy says a username should represent one person and should not be your organisation's name.
  3. Request corrections with sources. On the talk page, post one request per issue with the {{edit COI}} template: what is wrong, what it should say, and an independent reliable source that supports the change. The guideline allows material you published yourself within reason, for example to support an uncontroversial fact such as a new address, but independent sources carry more weight.
  4. Be patient and brief. Requests are reviewed by volunteers, who may decline them. The guideline asks paid and COI editors to respect volunteers' time by keeping discussions concise.
  5. Fix obvious errors directly, sparingly. Spelling, broken links and vandalism are the exceptions where you may edit directly. Anything that changes meaning goes through the talk page.

Writing an edit request volunteers can act on

Requests that get accepted tend to share a shape: one change per request, the exact wording proposed, an independent source, and nothing that reads like marketing. A workable template for the talk page:

{{edit COI}}
I work for [Company] (disclosed on my user page). I'd like to request
one factual correction.

Current text: "[Company] is headquartered in [old city]."
Proposed text: "[Company] is headquartered in [new city]."
Reason: the company moved its headquarters in [month year].
Source: [independent news article, with link and date]
        (company announcement also available: [link])

I won't edit the article myself. Thanks for reviewing. ~~~~

Three habits help. Keep each request short enough to read in a minute. Put the independent source first and your own announcement second, if at all. And never bundle a factual correction with a request to change tone or emphasis: the reviewer will reasonably decline the whole thing. If a request sits unanswered for a long time, the COI guideline points to a related WikiProject's talk page or the conflict of interest noticeboard, rather than editing directly.

If you do not have an article

Do not write one. Work on the thing that makes articles possible: independent, substantial coverage of your company. That means newsworthy work, research others want to report on, products that reviewers want to examine in depth, and openness with journalists. If such coverage accumulates, volunteers may write the article; if you are confident the notability test is met, a disclosed draft through Articles for Creation is the permitted route. It is also the slowest, and the outcome is not yours to decide.

Founders and executives

Articles about people fall under Wikipedia's policy on biographies of living persons, which is stricter about sourcing. The conflict of interest guideline makes one exception to "do not edit your own article": serious errors or defamation may be corrected quickly by the subject, followed by a note to Wikipedia's volunteer response team or the relevant noticeboard. Everything else follows the same talk-page route as company articles.

Wikidata for brands: the smaller, structured lever

Wikidata is the structured database behind Wikipedia's infoboxes and interlanguage links, and it is widely reused. For many brands it is a more realistic target than a Wikipedia article, but it has rules too.

Who can have a Wikidata item

Wikidata's notability policy accepts an item if it meets at least one of three criteria: it has a valid link to a page on a Wikimedia project such as Wikipedia; it refers to a clearly identifiable conceptual or material entity that can be described using serious and publicly available references; or it fulfils a structural need, such as being needed to make statements in other items useful. The second criterion is broader than Wikipedia's notability test, but "serious and publicly available references" still rules out an item built only from your own marketing.

Wikipedia and Wikidata compared

Wikipedia articleWikidata item
Bar for inclusionSignificant coverage in multiple independent, reliable secondary sourcesOne of three criteria, including an identifiable entity described by serious public references
ContentProse written by volunteers, with citationsStructured statements: instance of, country, inception, official website, identifiers
What it does for AI visibilityTraining data and retrieval for identity questionsDisambiguation and linking between databases and languages
What it does not doMake engines recommend you for purchasesMake engines describe you favourably
Main risk for brandsCOI editing, deletion, public disputeLow-quality or self-promotional items being deleted

The statements that matter most for a company

Wikidata describes an item with statements, each made of a property and a value. For identity, a handful of properties do most of the work, and they are the ones to check first:

  • instance of (P31): what kind of thing the item is, for example a business, a public company or a restaurant chain. A wrong value here misfiles the whole entity.
  • official website (P856): the single most useful link for tying the item to your domain.
  • country (P17) and headquarters location (P159): where you are, which matters for local and regional questions.
  • inception (P571) and founded by (P112): the history facts most often repeated in identity answers.
  • industry (P452): the category, which helps an engine decide which questions you are relevant to.

Each statement should carry a reference. A statement without a reference is easier for another editor to challenge and less useful to anyone reusing the data.

Keeping the item accurate

If an item for your company exists, check its core statements: what it is an instance of, country, founding date, official website, and the identifiers that link it to other databases. Errors here propagate quietly. Where the item is wrong, suggest corrections with references and keep the same spirit as on Wikipedia: disclose who you are, add facts, not adjectives. Then make your own site agree with it. Our guide to entity markup and sameAs covers how to point your Organization markup at the Wikidata item and your other official profiles, so that every record describes the same entity.

Knowledge panels and other entity surfaces

Most of the identity work that Wikipedia does for large brands, smaller brands can do through surfaces that are designed for them to manage.

Google's knowledge panel

Google's help pages say knowledge panels are generated automatically and updated as information on the web changes, and that changes are considered from two directions: the entity itself and general user feedback. An official representative can claim the panel by signing in to one of the official sites or profiles Google associates with the entity, then suggest changes. Not every panel can be claimed; when it cannot, the feedback link is the route. For local businesses, Google points to a different product, the Business Profile, which is managed directly.

Your own entity data

The quietest and most controllable entity source is your own site: a clear About page, consistent name and address, Organization markup with sameAs links to your official profiles, and the same facts repeated identically across directories. Engines reconcile entities by agreement between sources. Being consistent is worth more than being on any single prestigious source.

When an engine confuses you with someone else

Name collisions are where entity work pays off fastest. If identity prompts return facts about a different company, a place or a product with a similar name, the engine has attached your name to the wrong entity. Work from the most controllable sources outward: state your full legal name, category and location plainly on your own About page; make your Organization markup point to your official profiles and, if one exists, your Wikidata item; ask for corrections on the Wikidata items of both entities if either carries the other's facts; and use the knowledge panel feedback link if the panel shows the wrong entity. Then re-run the identity prompts monthly. Collisions rarely clear in a single cycle, but the direction is usually visible within a few runs.

Ranking the entity levers by effort and control

LeverWho controls itEffortTypical use
Your site and Organization markupYouLowEvery brand, first
Google Business Profile or claimed knowledge panelYou, within Google's rulesLow to mediumLocal businesses; brands with a panel
Directory and review profilesShared with the platformMediumMost brands; essential in local categories
Wikidata itemCommunity, with your disclosed suggestionsMediumBrands with serious public references, especially with ambiguous names
Wikipedia articleCommunity onlyHigh, uncertainCompanies that already meet notability

Where the effort pays off instead, for buyer questions

If Wikipedia is the identity layer, what is the preference layer? Our benchmark gives a concrete answer for each kind of category, and it is mostly not Wikipedia.

  • Local services and hospitality: review sites and directories (Tripadvisor for hotels and restaurants, Zocdoc for dentists, Avvo for lawyers, Angi for home services), city subreddits and local publications.
  • Consumer products: enthusiast communities, product reviewers, retailers and marketplaces, comparison articles.
  • Business software: vendor sites, comparison publishers and software review platforms.

The common thread is third-party corroboration: independent sources that describe you specifically, in the context of a choice. Our guide to third-party corroboration walks through finding and auditing those sources for your category, and the law-firm benchmark is a useful example of a category where professional directories, not Wikipedia, carried the citations.

A decision framework

Your situationWikipedia priorityDo this first
Large, widely covered company with an articleHigh for accuracyAudit the article and Wikidata item; request corrections through the talk page
Mid-size company, some national coverage, no articleLow to mediumEarn more independent coverage; keep Wikidata and your own entity data accurate
Local businessVery lowBusiness Profile, reviews, directories, local community presence
Product brand in a hobby categoryLowEnthusiast communities, independent reviewers, retailer listings
Brand with an ambiguous or changed nameMedium, via WikidataEntity consistency: Wikidata, sameAs, consistent naming across profiles
Company facing an inaccurate or hostile articleHigh, carefullyDisclosed talk-page requests with independent sources; legal advice for anything defamatory

How to measure Wikipedia AI visibility work

Because Wikipedia mainly affects identity, measure identity. Discovery metrics such as mention rate on "best X" questions will rarely move because of Wikipedia changes, and attributing any movement to them would be guesswork.

Two prompt groups, measured separately

  • Identity prompts: "What is [Company]?", "Who founded [Company]?", "Where is [Company] based?", "Is [Company] the same as [similarly named company]?". Score each answer for factual accuracy against a fixed fact sheet.
  • Discovery prompts: your normal buyer questions without your name. Track mention rate and citations as usual, but do not expect Wikipedia work to move them.

What to record

  1. A fact sheet of 10 to 15 core facts with sources, agreed internally.
  2. For each identity prompt and engine: which facts were right, which wrong, which missing, and whether Wikipedia or Wikidata was cited.
  3. The dates of any accepted Wikipedia or Wikidata corrections, so you can look for changes after them.
  4. A monthly re-run of the same prompts. Engines that search live may pick up a corrected article within weeks; answers drawn from a model's own knowledge change only when the model is updated, which can take much longer.

A simple scoring sheet

Keep the scoring mechanical so that different people score the same answer the same way. One row per prompt, engine and run:

prompt        | "Who founded [Company] and when?"
engine / date | [engine] / [run date]
facts checked | founder: correct | founding year: WRONG (says [year])
              | headquarters: missing
sources cited | wikipedia.org/wiki/[Company]; [company site]
score         | 1 correct / 1 wrong / 1 missing
action        | founding year wrong on Wikipedia too -> talk-page request

Summed across prompts and engines, the "wrong" count is the number to drive down. Track it separately for engines that searched and cited sources and for answers with no citations, because the two change on different timescales.

What success looks like

Success is fewer factual errors in identity answers, fewer confusions with similarly named entities, and a stable description of what you do. It is not a sudden rise in recommendations. If a team reports that Wikipedia work "increased AI visibility" on discovery prompts, ask which prompts, which engines, and what else changed in the same period.

Our own position, then: treat Wikipedia AI visibility as hygiene for companies that are already notable, and as a long-term result of being worth writing about for everyone else. If you want to track identity and discovery prompts side by side on a fixed schedule, our AI visibility tracking keeps the two groups separate so that each kind of work is judged on the metric it can actually move.

Sources