Questions to Ask Before Your Customers Ask an AI Assistant About Your Products

If your customers ask an AI assistant such as ChatGPT or Gemini about your products, the answer may draw on pages that are out of date or that your company did not write. Few companies have checked. The six questions below are the ones a leadership team should be able to answer, and each comes with the reason it matters and a description of a good answer.

Answering them takes no new software. It takes a list of the facts that matter, a set of real customer questions and someone whose job it is to compare the answers with the approved information.

What the evidence shows, from one sector

The published evidence used here comes from medicine, where researchers have compared AI answers with reference sources. It says nothing direct about how your customers behave, so read it as proof that the problem can be measured. The American Medical Association surveyed 1,692 US physicians in early 2026. Its report, dated March 2026, says 72% use AI for at least one task in their practice, up from 48% in 2024. The most common task, at 39%, was summarising medical research and standards of care. The figures are self-reported.

Researchers have also tested what the tools say. A study in BMJ Quality and Safety, published online on 1 October 2024, put patient questions about the 50 most prescribed US drugs to Bing Copilot in April 2023. Of the answers that could be checked against a reference source, 26.0% were partly inconsistent with it and 3.3% were fully inconsistent. The site the assistant cited most often was a consumer drug site, drugs.com, with 25.8% of all citations. The tool tested is now more than three years old, so read the rates as a measurement from that date.

Results depend on the task. A later study with the same lead author, in Naunyn-Schmiedeberg's Archives of Pharmacology, published online on 12 June 2026, found that Copilot and Gemini were accurate on 20 oral cancer drugs, with a median accuracy of 100%. The answers were incomplete, with median completeness of 61.1% for Copilot and 73.8% for Gemini. A citation is no guarantee either. A study in Nature Communications, published on 16 April 2025, checked medical answers from seven models against the sources each answer cited and found that "between 50% and 90% of LLM responses are not fully supported" by those sources.

Which of our facts must be right, and where is the reference version of each?

Start with the facts where a wrong answer could cost a sale or create a legal problem. Prices are one type. Others include specifications, compatibility, availability, warranty terms and safety instructions. Profound, a vendor in this field, says in its launch post of 14 July 2026 that accuracy checking works best on questions with "a deterministically correct answer: pricing, specs, features, availability, policies." That is a fair test for your own list. Include a fact only if an approved document settles it.

The exercise can show that no single reference version exists. A specification sheet may disagree with the warranty page, for example, and nobody has decided which one governs. A good answer to this question is a register that links each fact to its approved source and names the person who maintains it. Where a fact differs by market or product version, the register says so, and it records the date of the last change.

What do AI assistants say today, and which pages do they cite?

An impression is easy to form here, and a record takes work. Use support tickets and sales calls to find the questions customers ask most. Put each question to the assistants your customers are likely to use, and save a dated copy of every answer with the pages it cites. Note which assistant and which mode you used. Ask each question more than once. In a study of drug interactions in intensive care patients, reported in the Journal of the American Pharmacists Association and published online on 6 January 2026, ChatGPT-5 gave the same answer in only 68% of repeated runs.

The cited pages matter as much as the answers. Sort them into your own pages, pages run by partners and resellers, review and comparison sites, and everything else. In the BMJ study above the most cited source was a third-party site, and the same may be true for your products. When you compare an answer with the reference version, keep a wrong answer separate from an incomplete one, because the two call for different fixes.

Are we measuring how often we are mentioned, whether what is said is correct, or both?

These are two measurements, and a team can do well on one while failing the other. The first is visibility, meaning how often and where a brand appears. Profound's product page offers to show "how often your brand appears in AI answers" and which websites those answers draw on. Semrush describes its AI Visibility Toolkit as reporting a brand's share of voice and sentiment in AI answers. Both companies address marketing teams.

The second measurement is correctness, and it needs a reference to compare against. Profound added this in 2026. Its FactCheck announcement of 14 July 2026 says each claim in an answer is scored as accurate or inaccurate against a knowledge base that the brand supplies from its own site and documents. Semrush said in a release dated 6 March 2025 that its enterprise product can "Monitor AI-driven content for inaccuracies". The pages cited here do not say what reference that monitoring uses. Any accuracy score is only as good as the reference behind it, which returns you to the first question.

In some regulated fields the company does not choose the reference. For medicines in the European Union, Article 87(2) of Directive 2001/83/EC says all advertising "must comply with the particulars listed in the summary of product characteristics", a document the regulator approves. A good answer to this question names both measurements and reports them separately, because a high visibility figure says nothing about whether the answers are right.

When a fact changes, how long until our pages, our partners' pages and the answers catch up?

A changed fact has to reach more places than the main product page. Superseded PDFs can stay online, and resellers may keep an earlier specification sheet. An assistant that searches the web can find any of them. The model behind an assistant is also trained on text up to a cutoff date, so for anything newer the answer depends on what the assistant retrieves when the question is asked.

One historical measurement shows how long the lag can be on a site a company does not control. A study in the New England Journal of Medicine looked at Wikipedia pages for drugs that received US safety warnings in 2011 and 2012. As reported by the Regulatory Affairs Professionals Society on 27 June 2014, 41% of the relevant pages were updated within two weeks and 36% were still unchanged more than a year later. The study is old and covers a single site, so take it as an illustration and measure your own lag.

A good answer to this question has a figure in it, measured on your own last change. It also has a distribution list. When a fact in the register changes, the people who run the website, the documentation, the partner portal and the support scripts are told promptly, and someone asks the assistants the same questions again a few weeks later.

What can we change, and what can we not?

You can change your own pages. Correct the published fact and show the date it took effect, then decide whether each superseded document is removed or redirected. You can also ask the owners of cited third-party pages to correct theirs. Both vendors above describe this route. Profound's FactCheck announcement describes drafting "an update request to the author" of a third-party source, and Semrush's page on brand sentiment in AI answers points to updating owned content and earning mentions on the sources that AI models cite.

You cannot directly edit what a public assistant run by another company says. You can only influence later answers through the sources it reads. The Nature Communications study above found answers that cited a source and still departed from it, so it is reasonable to expect that a corrected source will not always produce a corrected answer. That is our inference, since the study did not test corrections. Plan to fix sources, ask the same questions again and keep the dated records of both.

Who owns this: marketing, product, legal or IT?

Each of these teams holds a part. Marketing sees the answers and the effect on the brand. Product owns most of the facts. Legal decides what may be claimed. IT runs the website and the content systems where old versions sit. The risk with shared ownership and no lead is that nobody checks again after a change.

Our view is that a single named person should own the checking and the register, with the authority to ask other teams for corrections, while each fact keeps its own owner in product or legal. A sensible default is the digital or marketing team that already runs the public product pages. A regulated company is the exception. There the owner should sit in the function that answers to the regulator, with marketing as a contributor.

What to do this month

1. Write the register of facts that must be right, with the approved source and the maintainer of each.
2. Collect real customer questions from support and sales, and ask them of the assistants your customers are likely to use.
3. Save each answer and its cited pages with the date, and mark every answer that disagrees with the approved source.
4. Correct your own pages first, then write to the owners of the third-party pages that were cited with wrong information.
5. Name the owner of the checking and set the date of the next round.

Where QueryNow fits

QueryNow builds the foundation enterprise AI runs on, and keeping a current reference version of company documents is part of that work. You can talk to our team about your own case. For pharma teams QueryNow makes Juncture, a product that compares what a company and AI tools say about a medicine with its official prescribing information. Mahesh Tatiparthy, the founder of QueryNow, will discuss this subject on a panel webinar later this year.

Update, 11 November 2026: the panel, Winning the AI Battle for HCP Attention, is confirmed for Thursday 17 December 2026, 16:00 to 17:00 CET. The registration link will be added here when it opens.