Obility Editorial · · 4 min read
Models learn from Wikipedia and from the pages it cites
“Get a Wikipedia page” is near the top of most AEO checklists, usually with no account of why it would help. The best public evidence is in two model papers. OpenAI’s GPT-3 paper lists English Wikipedia as 3 percent of the training mix, sampled about 3.4 times during training while most of the web crawl was seen less than once. The same paper used Wikipedia, alongside its other curated sets, as the example of high-quality text when it trained a classifier to decide which Common Crawl pages to keep.
Meta’s LLaMA paper goes further. Wikipedia made up 4.5 percent of its training data and was seen about two and a half times. For the Common Crawl portion, which was 67 percent of the mix, Meta trained a model to tell pages used as references in Wikipedia apart from random pages, and discarded the pages that did not look like references. Wikipedia shaped these models twice: through its own text, and through its taste in sources.
The boundary matters. OpenAI, Anthropic and Google do not publish the training mix behind today’s assistants, so these papers describe a common technique, not the recipe for ChatGPT, Claude or Gemini. And a buying question such as “best invoicing software for agencies” is often answered by searching the web, where the cited pages are whatever retrieval finds for that question. An encyclopedia entry about your company rarely compares you with competitors, which is what that question asks.
Wikipedia’s company bar rejects most of what marketing produces
Wikipedia’s notability guideline for organizations and companies presumes a company notable only if it has had significant coverage in multiple reliable secondary sources that are independent of it. Each source has to pass all four tests on its own. The guideline says outright that it sets a higher bar than the general one in order to stop marketing and public relations professionals gaming the rules, and that only unpaid sources count.
Its list of coverage that does not count reads like a typical press page: funding announcements, product launches, hires and departures, appearances at industry events, quotes from your own staff, press releases and articles substantially based on them, “best of” and “fastest growing” lists, and directory listings. A single strong source is almost never enough, and a series of articles by the same journalist counts as one source. The guideline adds that figures such as revenue, headcount or capital raised do not make coverage significant.
Writing your own page is the quickest way to lose it
Wikipedia’s conflict of interest guideline says people connected to a subject are strongly discouraged from editing its article directly. They are expected to propose changes on the article’s talk page for other editors to review, and to send new articles through the Articles for Creation process rather than publishing them. The Wikimedia Foundation’s Terms of Use go further for anyone paid to edit: they must disclose their employer, client and affiliation, and disputes over undisclosed paid editing can go to binding mediation.
If an agency offers to “get you on Wikipedia”, ask how it discloses its edits and which independent sources it will cite. A page built on announcements and sponsored coverage fails the notability test above, and a deleted article helps no one.
Build the coverage the page would need
The useful reading of the Wikipedia advice is the bar itself. The independent, in-depth coverage that would justify an article is the same kind of material those model builders filtered towards and that live answers can retrieve and cite. Count the pieces about your company that pass the four tests: written by someone unaffiliated, discussing you in depth, in a publication with editorial standards, offering their own analysis. If the honest count is one or two, that shortfall matters more than the missing page.
Then look at which outside sources the answers to your buying questions already cite. Obility’s Answer Engine Insights keeps the cited sources next to each answer and the competitors it recommends, so you can see which publishers, review sites and analysts the engines lean on in your category, and where your competitors are covered in depth and you are not. That is a target list for genuine coverage: give those writers something worth analysing, such as product access for real testing or data you actually hold. A review where the product was given free in exchange for coverage is, by Wikipedia’s own definition, not independent.
If you already have a page, keep it accurate
An existing article is worth maintaining, through the talk page and with a reliable source for every requested change. When an AI answer repeats a stale fact about you, check whether your Wikipedia article or one of its sources says the same thing. Obility’s FactCheck & Sentiment workflow records the exact prompt, answer, engine and date, separates a factual error from an opinion, and reviews any cited third-party material, so a correction request starts from evidence rather than annoyance.
Expect no guaranteed effect either way. No engine publishes how much weight it gives a Wikipedia entry, and a page created next month cannot be shown to cause a change in answers. Coverage from people who have no stake in you is the input that holds up under both Wikipedia’s rules and the way models are built.