How to Compare Owned Content and Independent Evidence in AI Answers

When a company checks how AI answers describe it, two kinds of evidence matter: owned content, which the company controls, and independent evidence, published by sources outside its control. This comparison is part of generative engine optimization (GEO), work that helps generative AI products find and accurately describe a business.

Owned pages are direct sources for current product details, policies, and prices. Independent reviews, reporting, research, and customer accounts can provide outside context for comparisons and claims about reputation or performance. To judge an answer, check which sources it cites, who controls them, and whether they support its specific claims. There is no universal share of owned and independent sources that every company should target.

How sources shape an AI answer

An AI product with web search can turn a question into searches, retrieve web pages, and use selected information to write an answer. The exact question matters: a request for current pricing calls for different evidence than a request to compare providers. Studies have found that AI search products differ in which sources they retrieve and in how much their results vary across repeated searches and products, as described in research on generative AI web search.

An answer can also use information stored in the AI model, the part of the product that generates text, rather than information from a page shown beside the answer. A 2026 study of commercial questions found brand mentions without a cited page naming the brand in some ChatGPT responses. The study’s reported rate applies to its own data and method, not to AI answers in general. It does show why a missing citation does not prove that a brand or source had no influence. (The study of AI citations and model knowledge)

When reviewing a response, record three things separately: whether it mentions the company, whether it shows a citation, and whether that citation supports the claim. A cited page may not support the sentence beside it. Earlier human evaluation of generative search products also found citation-support problems, though its findings apply to the products tested in 2023 rather than to current products. (The evaluation of citation support)

What owned and independent sources tell you

Owned pages are best placed to state what a company says about its current products, policies, and services. That makes them useful for questions with verifiable, company-controlled answers, such as what a product includes or how a policy works. Ownership matters because a company’s page states the company’s position; it does not independently confirm that position.

Independent evidence can help readers assess claims that require comparison or outside scrutiny. A review that explains how it tested a product, for example, offers a different kind of evidence from a company’s product page. Readers should consider how the source gathered its evidence and whether that evidence is relevant. They should also consider the source’s connection to the company. A page on another domain is not automatically independent: it could be sponsored, affiliate-driven, syndicated, or controlled by a partner.

Research supports examining the source mix, but not applying a fixed ratio to every business. A 2026 preprint that examined 128 brands across 12 markets and 13 languages reported that 85.7% of its URL-grounded citations pointed to sites the brands did not own. The researchers also noted limits in how they decided who owned each site and which citations to count. That result describes their sample; it does not establish the share a particular business should expect. (The cross-market study of brand sources)

How to audit answers for your own business

Start by asking questions that need different kinds of evidence. Use a current-facts question, such as one about a product feature or policy; a comparison question; and a question about customer experience or reputation. Include versions that name the company and versions that do not. Questions that name the company test how AI describes information about it. Questions that leave out the company show whether AI mentions it without being prompted.

For each test, record the exact question and which AI product you used, along with the date. Also note the search or research mode and the language or market. Save the full answer and its visible citations. Repeat some questions later and, where practical, run them more than once. A single response is an observation, not a stable measure of visibility.

Then review each cited page. Note who controls it and whether it appears to have a commercial relationship. Check what evidence it provides and whether the information is current. Open the page and compare its actual wording or evidence with the answer’s claim. For example, if an answer says one product is easier to set up than another, check whether the cited page actually compared setup, or whether it only describes one product.

Group the results by question type and AI product. Look for recurring patterns: official pages missing from answers about current product facts, outdated pages appearing repeatedly, or comparisons resting on sources that do not test the stated criteria. Do not turn a small sample into a universal score. Differences across products, markets, and runs make a fixed threshold misleading.

Improve the evidence, not just the source count

Keep company-owned pages accurate and make their information easy to check. Put important product, pricing, policy, and support details in clear text, and update them when the underlying facts change. Google says its Search AI features rely on established Search eligibility and does not require special AI files or special schema. Its guidance focuses on making pages accessible to crawlers, linking to them from other pages on the site, and writing content that people and search systems can find and understand. (Google’s guidance for AI features)

For claims about comparisons or reputation, seek evidence that can genuinely test or contextualize the claim. That might mean a review with a disclosed method, relevant public records, transparent research, or firsthand reporting. More mentions do not, by themselves, make evidence stronger. Do not present a sponsored or company-controlled page as independent.

Check technical access separately from answer visibility. As of November 11, 2026, OpenAI documents OAI-SearchBot as the crawler used to surface websites in ChatGPT search and says site owners can manage its access separately from the crawler used for training-related purposes. When a site lets a crawler fetch a page, that supports the page’s eligibility for search. Crawler access does not establish whether the search product indexed the page or retrieved it for a particular question. It also does not show whether the product cited the page or used it to make a business decision. (OpenAI’s crawler documentation)

Where comparisons go wrong

The most common mistake is treating a citation count as a measure of accuracy. A citation may not support the claim beside it, and an answer may use information without showing a source that names the company. Compare each claim with the cited page. Record company mentions, citations, and whether the citation supports the claim as separate observations.

Do not label every external page as independent. Check who publishes or controls it and whether a commercial relationship affects its role. A company’s guest post or an affiliate review may appear on a separate domain, as may an independent investigation; each offers different evidence.

Finally, do not infer business impact from visibility alone. A crawler request, a citation, or a brand mention does not show that the answer generated a visit, lead, or sale. Where your analytics allow, track traffic you can identify and business outcomes separately. Do not claim that a citation caused a result unless you have evidence to support that conclusion.

What I'm building

Delegate tasks. Get software.

Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.

Take a look at vroni.com