What Happens When AI Has Conflicting Information About Your Business?

When an AI system finds conflicting information about a business, the result is not necessarily a wrong answer. The result is uncertainty: about which account the system will use, and about whether the description a buyer receives is accurate enough for the decision they are making.

That uncertainty can exist even when a company’s own website is accurate, because the website is only one of the sources a system may draw on.

The Same Company, Several Descriptions

Consider an environmental consultancy. Its website accurately describes what it does today: it specializes in contaminated land assessment for residential developers, under a managing director appointed eighteen months ago.

Elsewhere, the picture is less tidy. An industry directory still lists the ecology surveys and planning support the firm stopped offering three years ago. A trade publication’s company profile names the former managing director. Several smaller sites repeat a description from an old press release, written when the firm was a generalist.

None of these sources is malicious. Each was accurate when it was written.

Now a developer asks an AI assistant Which environmental consultancies specialize in contaminated land assessments for residential developments, and how do their capabilities compare?

Depending on what the system retrieves, it may encounter two versions of the same firm: the specialist it is today and the generalist it used to be. Which account the system relies on can influence how the consultancy is classified and which competitors it is measured against. The developer sees the answer. They may never see the disagreement behind it.

Two problems are worth separating here. Information can be wrong somewhere on the web, which an evidence review can find. Or an AI system can actually repeat it, which only observing real answers can show. The two are related, but one doesn’t prove the other.

Why AI Encounters Conflicting Accounts

Information about a business is spread across its website, directories, trade and local press, association listings, review sites, partner pages and public records. The business controls only some of them.

Each is updated on its own schedule, if at all, and businesses usually change faster than the record of them does. A new service or a new managing director reaches the website first. A directory entry written four years ago doesn’t update itself.

Some conflicts come from confusion rather than age. Two firms with similar names, a parent company and its subsidiary, or a professional who has moved between firms can leave information attached to the wrong subject.

And some apparent conflicts aren’t conflicts at all. A firm that had 40 staff in 2022 and has 65 today has two accurate headcounts. A consultant who led a practice at one firm and now leads another has held two true job titles.

Statements like these disagree only when the dates and context are lost. A system that treats two time-bound facts as a contradiction can produce a wrong answer even though every source is correct.

What Happens When the Sources Disagree

No AI provider publishes a universal rule for which source wins when sources disagree, and there’s no reason to assume different systems resolve conflicts the same way.

Research does show that this is a hard problem. In 2025, researchers at Google Research proposed a taxonomy of conflicts between retrieved sources, with a different desired response for each type. Complementary information should be combined. Genuine disagreement should be presented as disagreement. Where sources differ because one is out of date, the current information should take priority. Sources that are likely false should be set aside. The researchers found that language models “often struggle to appropriately resolve conflicts between sources.”

A benchmark presented at ACL 2026 examined 1,814 real-world questions paired with retrieved web passages. In 57.2% of the questions, those passages contained explicit contradictions. That figure describes the benchmark, not how often conflicts occur in everyday searches. The authors report “substantial performance gaps” when the models they tested, including GPT-4.1 and Claude 3.7 Sonnet, had to reason across the conflicting references. Both models have since been superseded.

These studies test how models handle conflicting evidence under research conditions. They do not establish how today’s AI assistants will respond to a particular business.

Where conflicting accounts of a company exist, several outcomes are plausible:

• The system presents one account with confidence, current or not.
• It combines details from different periods into a description that was never true at any single point.
• It acknowledges the disagreement, or hedges.
• It resolves the conflict correctly and uses the current information.
• It never retrieves the conflicting material, so the conflict never reaches the answer.

Which of these happens can vary with the system, the question being asked and what was retrieved for it. The same business can be described accurately in one answer and inaccurately in the next.

Conflicting evidence doesn’t guarantee a wrong answer. It makes the answer less predictable.

Why It Matters Commercially

A single wrong detail can decide a shortlist if it touches the buyer’s selection criteria.

The damage can run in either direction. A business can be included for a service it stopped offering, potentially undermining its credibility when the mismatch surfaces. Or it can be excluded for a capability it recently built, because the record still describes the business before it changed.

The second is harder to detect. Nothing reaches a company when it is left off a list it should have been on.

This is where conflicting information becomes a commercial AI visibility problem. A business can be recognized, correctly identified and mentioned, and still be misrepresented on the one attribute the buyer cares about. Being named is not the same as being understood.

For professional services firms, B2B suppliers and luxury real estate professionals, those attributes are often the ones most likely to drift: credentials, years of experience, specialties, leadership, geographic coverage, notable clients and transactions.

We have seen this kind of drift in our own work. In one luxury real estate market we monitored, two biographies on the same brokerage website gave different years of experience for the same broker. That establishes a conflict in the published evidence. It doesn’t establish that an AI system repeated the discrepancy, or that a buyer acted on it.

An inconsistency in the evidence is a risk. An error in an AI response is an observation. A lost opportunity is a consequence, and the hardest of the three to establish.

How to Recognize the Problem

Start with the questions that matter commercially. Asking an assistant to compare the business with competitors, or to suggest firms for a particular kind of work, tells you more than asking it to describe the company by name.

Then check what the answers claim against what is currently true. Do responses describe the services, leadership, coverage and credentials the business has today? Where a system shows its sources, which source supports each claim, and is it current?

Agreement between answers isn’t proof of accuracy. A system can repeat the same outdated claim every time it is asked, and a consistent error is more significant than a one-off, especially when it can be traced to a source that still says it.

This kind of diagnosis sits within a wider question: what AI systems understand about a business, and how that understanding shows up in their answers. The aim is to separate a conflict in the evidence from an error in an answer, and an isolated error from a pattern.

What a Business Can Control

The first-party record is the part a business fully controls. Its website should state current services, leadership, locations and credentials clearly, without the same figure appearing differently on two pages. Because buyers increasingly check the website after reading an AI answer, it is also where an outdated claim can be contradicted.

Beyond the website, the third-party records that matter most are the ones describing what the business does and who leads it.

Google’s guidance points the same way for its own products. Its structured data guidelines require markup to be “a true representation of the page content”, and its Business Profile guidance warns that if a business’s information isn’t accurate, its profile “might not show up for relevant searches”. That guidance concerns Google Search. It supports good maintenance practice, but it doesn’t tell us how independent AI assistants choose between contradictory sources.

No business can control every page that mentions it, or decide which source a system will believe. Correcting what an AI system already gets wrong is a separate question, with its own limits.

Accurate Isn’t the Same as Consistent

A company may know exactly what has changed about its capabilities, leadership or expertise. The wider record may still describe the business it used to be.

When that outdated record enters an AI-assisted buying decision, the company can be evaluated against a version of itself that no longer exists.