Contradictions on your own site: an overlooked part of fixing AI search misinformation

Start by checking your own articles against your company pages

When teams work on AI search misinformation, they tend to look at what other websites say about them. Yet on a company's own site, articles can also say something different from its company overview, pricing or service pages. In September 2026, Rootpublish compared the public pages of six sites, including its own, with each site's company pages. Three of the six sites had contradictions.

This audit compared public pages with each other. It did not examine answers from AI search itself, so it does not show whether any contradiction actually led to a wrong description. Even so, when a company's own sources contradict each other, it becomes harder to expect an accurate description in return.

Google says that the fundamentals of SEO still apply to AI Overviews and AI Mode, and that no special text file for AI or dedicated structured data is required. It also says that meeting the requirements does not guarantee that content will be shown. When content is prepared with generative AI, Google asks for a focus on accuracy, quality and relevance. Before looking for special tactics, the starting point is to confirm that your own statements are accurate and do not contradict each other.

Overstatements outnumbered wrong figures

The contradictions we found included the following:

• Describing a free assessment as covering more than it actually does
• Referring to a service by a different name
• Stating a three-month minimum contract, while the pricing page says there is no minimum contract period
• Quoting a price in a customer example that is lower than the cheapest plan on the pricing page
• Promising operation "on autopilot" in a headline, although the product requires a person's approval before publishing

Statements that overstate the scope or terms of the company's own services were more common than simple numerical errors. If you only look for typos and outdated figures, these statements are easy to miss. When reviewing, pay particular attention to words about scope, contract terms, prices and how much is automated.

Results from six sites

The results by site were as follows. On a company blog that mass-produces AI-written articles, we found four contradictions in 12 articles (one of them open to interpretation). On one company that publishes columns about AI, we found two in 12 articles. On Rootpublish's own site, we found two in 322 passages.

On a hand-written personal blog, and on two companies whose articles ended with a standard block describing their services, we found no clear contradictions. With only six sites, however, this does not establish a general relationship between how articles are written and how many contradictions they contain.

How the audit worked

We followed these steps. First, we split each public page into passages. Next, we selected passages that mention the company or service name, or first-person words such as "we" and "our company". We then compared the selected passages with the statements on the company overview, pricing and service pages.

Claude Opus judged whether a passage contradicted the company pages, and Claude read each flagged passage to confirm it. Only part of the results has been checked by a person. For each site, we covered up to its 12 newest articles.

Results depend on the AI model that does the judging

We also compared AI models for judging contradictions. On 102 synthetic cases with known answers, accuracy was:

• Claude Opus 5.5 and Fable 5.1: 100%
• Sonnet 5: 99%
• GPT via Codex, and Jev, a model built only for judging: 97%
• Haiku 4.5: 88%

However, for the four overstatements of a company's own services found in real AI-written articles, Jev and Sonnet caught none of them, while Opus caught them. Scores on synthetic data did not guarantee results on real articles. When you choose a model, do not decide on test-set accuracy alone. Check whether it catches contradictions in your own real articles before you rely on it.

Estimated costs

Checking all 1,212 passages of 12 articles with Jev, the judging-only model, cost about $0.25 at the published rate. That approach, however, did not catch the actual contradictions.

We therefore examined a method that selects only the passages about the company itself (about a tenth of the total) and has Opus judge them. We estimate this method at about $0.05 per article. This is an estimate, and we did not measure the tokens used for reasoning. In this verification, narrowing the scope and prioritizing accuracy found the real contradictions, whereas cheaply checking everything did not.

Fixing a headline on our own site

Rootpublish audited 322 passages across eight public pages of rootpublish.com. The audit found that the WordPress headline, 「自動運営」 in Japanese and "On autopilot" in English, contradicted the product specification, which requires a person's approval before publishing.

We changed the headline to 「一次情報で育てる」 and "Grown from what you know" and published it. A re-audit of the same pages after publication, covering 129 passages, found no contradictions.

This is a single fix on a single site. It does not show that the same steps will remove contradictions on every site.

Limits of this audit

This audit is a small sample of six sites and up to 12 articles each. Many judgments rest only on Opus and a check by Claude, and review by a person is not complete.

The way passages are selected also has a weakness. If an article and the company pages use different names for a product, the relevant passage can be missed. And if the company page itself is out of date, it is the company page, not the article, that needs correcting. Finding a contradiction does not necessarily mean the article is wrong.

Steps to start checking your own site

Based on these records, here are steps for starting the same check on your own site. They are a suggested process derived from this verification, not measured results showing better search rankings or more AI citations.

1. Choose the pages to check against. Treat your company overview, pricing and service pages as the reference, and first confirm that they are up to date. If they are outdated, correct them before the articles.
2. Split published articles into passages, and select those that mention your company or service names, or words such as "we" and "our company". Adding variations such as former service names and abbreviations reduces missed passages.
3. Compare the selected passages with the reference pages. Look beyond wrong figures and focus on overstatements of scope, contract period, price and the degree of automation.
4. Use AI judgments to surface candidates, and have a person make the final decision by reading the article and the reference page. Choose the judging model only after testing whether it catches contradictions in your own real articles.
5. Review and publish each correction one at a time, then check the same page again after publication. Rootpublish is designed so that, for search improvements, one approved and published change is observed on the same page for seven days and judged on measured results. Search rankings and AI citations are not guaranteed.

Contradictions can also be prevented while articles are being written. Rootpublish records the ID and version of the company information each draft paragraph relies on, and at approval it checks whether that information has changed. This version check does not guarantee that the text means the same as its source, so a person still reads the text and its sources before publication.