Does AI state your prices correctly? Findings from a study of 30 B2B companies that publish their pricing

Different values for 12 of 30 companies

Even when a B2B company publishes its prices, AI does not necessarily state its prices and contract terms correctly. When Rootpublish checked 30 companies on September 28, 2026, we found answers for 12 of them that stated, without qualification, values for prices, plans, contract terms or services that differ from each company's pricing and service pages (hereafter, the canonical pages). That was 4 of 15 companies in Japan and 8 of 15 in English-speaking countries.

Separately, for 3 companies (2 in Japan, 1 in an English-speaking country), the AI answered that prices or terms written on the canonical pages were "not published". Conversely, for 3 companies we did not count discrepancies that the automated judgment had marked as clear, after checking them on the canonical pages. The reasons were that the scope of the features covered might be different, that the statement could be correct depending on how the page is read, and that the page itself noted that terms "vary by page".

Before the study, we decided that discrepancies at 9 or more of the 30 companies would mean the problem is widespread, and that fewer than 3 would mean dropping the hypothesis. The result of 12 companies is above that line.

How we checked

The subjects were 30 B2B companies that publish their prices, in fields such as AI adoption support, marketing support, map search optimization (MEO), fractional CMO services and SaaS. We asked an AI with web search (Claude Sonnet, searching from Japan) the same questions a prospective buyer would ask, once per company. We asked about pricing plans and amounts, the minimum contract period and cancellation terms, the main services, and the target customers. To questions in English, we added "Please answer in English."

We compared the answers with what the canonical pages state. Rootpublish opened the canonical pages and checked each clear discrepancy one by one. Statements that could be correct depending on how the page is read were not counted.

We used one AI and asked about each company once. This study cannot tell whether other AIs, such as ChatGPT, would show the same proportion. We have not published the names of the companies we checked.

Examples of actual errors

The errors were not limited to wrong amounts. All of the following are real examples, with the company names withheld.

For one company, the AI answered "Basic plan: from ¥330,000 per month (tax included)". But the pricing page has three plans and no plan called "Basic". ¥330,000 a month was the price of a higher plan with a different name. The same amount also appeared in a blog article by the same company.

In another example, the AI answered that the initial fee was "zero". On the actual pricing page, every plan carries an initial fee of ¥100,000 (excluding tax) at the start.

For companies in English-speaking countries, there were errors in contract periods, refunds and how pricing works. The AI answered "a minimum of 6 months, typically 12–24 month contracts", but the pricing page says no long-term contract is required and starts with a 3-month trial period. The AI also answered that there is "no refund guarantee", but the home page offers refunds for 14 days to clients who are not satisfied with the quality. On how pricing works, the AI answered with plans: "Pro at $79 a month and Team at $399 a month". The actual pricing page has no plans; it quotes an annual fee tailored to the work.

Where the wrong figures came from

Of the 12 companies with wrong values, the same figure appeared on a page the AI cited for 4. For 3 of those, the same figure appeared on another page or in a blog of the company itself: an amount for a plan name that does not exist, an old price and contract period, and terms for a custom quote.

For the other 8 companies, the cited pages did not contain the same figures, and the answers appear to have been filled in from outdated information or guesswork. In other words, a company's own pages are only part of the cause.

Note that this check of sources tracked a limited set of figures: amounts of 100 or more, numbers with a currency or %, and periods in months or years. The canonical pages themselves were not counted as sources.

Patterns in which a company's own articles are the source

On October 6, 2026, we additionally checked 9 Japanese companies that publish their prices and publish articles mentioning their own plans. For 4 of them, we confirmed errors in prices or contract periods. The sources of the errors fell into the following patterns.

First, the AI gave an amount from the company's own article explaining market rates (such as "from ¥300,000 per day") as that company's price. Second, after a company changed its prices, its old articles still carried the previous annual-payment amount. Third, the site stated the minimum contract period in two ways, "3 months" and "six months", and the AI gave one of them. Fourth, the company's article still described its previous plans and contract period, and the AI gave them as the current pricing.

For the remaining 5 companies, we found no clear errors with a source on the company's pages.

An audit of company sites carried out at a different time found a similar statement: the pricing page said "no minimum contract period", while an article gave the minimum contract period as "3 months".

A single question can easily miss errors

On September 29, 2026, we also tried splitting the questions into three and asking each of them to a separate AI. The three questions covered the services as a whole, the cheapest way to start, and comparison with competitors and who the service suits.

For one company, the 3 discrepancies found with a single question rose to 9 with three questions, 7 of them clear. The "Basic plan at ¥330,000 a month" that does not exist appeared in all three answers.

When we asked the same questions about the same company twice, the content of the answers changed each time. Both times, however, they were wrong. Getting a correct answer once does not mean the AI will always answer correctly.

How to check your own company

To check by hand, ask an AI with web search the questions a prospective buyer would actually ask, as they are: for example, pricing plans and amounts, the minimum contract period and cancellation terms, and the cheapest way to start. Then compare each answer with your own pricing and service pages. If a figure differs, open the pages the AI cited and look for which of your own articles carries the same figure.

Rootpublish's AI facts check puts these steps together. It asks the three questions a prospective buyer would ask, each to a separate AI that has only web search. Until the answers are in, the content of the company's pages is kept out of the conversation, because an AI that has seen the right answer first leans toward it. The answers are compared with what the pricing and service pages say, and only differences with quotes that actually exist in both the answer and the page are kept. For wrong figures, it looks for the source among the pages the answer cited and puts it in a report. Measuring the same site again shows a comparison with the previous run.

There are three ways to use it. In Claude Code, it is free as an MIT-licensed plugin. The AI processing runs on the user's own Claude, needs no API key and sends nothing to Rootpublish. One run uses about $1.50 worth of subscription usage (October 2026, version 0.3.0). From ChatGPT and others, you can try it with the MCP server offered on a trial basis (https://rootpublish.com/mcp). In that case, however, accuracy is lower, because the AI that answered checks its own answers. If you ask us by email, we check for free.

How to fix the errors you find

If the source is on your own pages, the first remedy is to bring that statement in line with the canonical pages. Judging from the patterns in the study, check these places. First, whether articles explaining market rates are written so that their amounts can be told apart from your own prices. Next, whether old articles still carry previous amounts or plans after a price change. Finally, whether terms such as the minimum contract period are written in two different ways on the site.

When you are unsure which is correct, also check whether the canonical page itself is outdated. If the canonical page is the older one, it is the canonical page, not the article, that needs fixing.

However, for 8 of the 12 companies, the cited pages did not contain the same figures. Fixing your own pages does not remove every source of errors. We have also not yet checked, in this study, whether the AI's answers changed after the fixes.

Google explains that the fundamentals of conventional SEO also apply to AI Overviews and AI Mode, and that there is no need to add special text files for AI or dedicated structured data. It also says that meeting the requirements to be shown does not guarantee crawling, indexing or display.

What we found when we measured Rootpublish itself

On October 7, 2026, we used the same method to ask about Rootpublish itself, in Japanese and English. The AI could not find Rootpublish and named other companies with similar names (such as Rootstock Publishing). Before any question of price errors, it did not recognize that Rootpublish exists.

Of 12 questions about checking AI answers and about pricing, only one answer mentioned Rootpublish. The page cited then was a listing on a directory site of MCP servers.

So we did three things. We made the AI facts check the focus of the home page. We put up a page that answers who runs Rootpublish, what it offers and what it costs (/about), along with llms.txt. And we stated that Rootpublish is separate from the companies with similar names. We have not yet been able to measure the effect. We remeasure every week with the same questions.

What this record shows

What we can say is that, even when companies publish their prices, it was not unusual for AI to state, without qualification, prices or terms that differ from the published ones. Some of the errors came from a company's own old articles or articles about market rates, and those a company can check and fix itself.

This record also has limits. The September study asked one AI about each company once. Answers to the same question changed from run to run. And whether the AI's answers become correct after pages are fixed has not yet been measured, including for our own site. We cannot guarantee that AI answers will improve or that a site will be cited in AI search. That is exactly why we recommend measuring again regularly with the same questions, rather than checking once.