AI Can Cite Your Business. That Doesn't Mean It Checked the Facts.

Put "We've served 10,000 customers" on a company's website and everyone understands what it is: a claim made by the company. Put the same sentence in an AI-generated recommendation, add a citation to that website, and it begins to look like an independently researched finding. The number hasn't changed. The underlying evidence hasn't changed. But the presentation has, and with it, the authority the statement appears to carry.

This is one of the more consequential problems emerging in AI-assisted business discovery. A cited answer feels more trustworthy than an uncited one, often for good reason. The link gives you somewhere to look. But a citation establishes where information was found, not necessarily whether anyone verified it. The difference matters when that information is helping someone choose a contractor, a law firm, a dentist, or a senior living community.

We've spent a lot of time discussing whether AI can find the right businesses. We need to spend more time asking what the systems can actually establish about those businesses.

Where the answers come from

In September, BrightLocal published an unusually useful study of 200,085 localized AI searches across ChatGPT, Google AI Mode, and Google AI Overviews. The researchers tracked recommendations across different prompts, locations, and repeated runs, and analyzed 1,967,762 citations appearing in the results.

Business websites accounted for 42% of those citations. Google Business Profiles, review platforms, directories, and other sources contributed to the rest. The study also found substantial variation in the businesses mentioned across platforms and repeated searches. It's a large observational study of generated responses, not a controlled experiment proving why any particular business was selected.

The website finding is important. Despite predictions that AI will make company websites irrelevant, those sites remain prominent among the sources surfaced in AI-assisted local discovery. They're often where the richest descriptions of a company's capabilities and history live.

They're also written, for the most part, by the companies themselves.

That's normal. A service-business website is supposed to explain what the company does, where it works, and why someone might hire it. The trouble comes when information created to market a business is presented in a recommendation without a clear distinction between the claim and the evidence behind it.

To be precise, BrightLocal's research doesn't show that AI systems never check such claims. Nor does the appearance of a citation prove that the cited page was the only information considered. It tells us which links appeared in the answers. That's a meaningful observation, but a narrower one than many marketers will be tempted to make.

A link answers one question. Hiring someone involves several more.

Consider a hypothetical family law firm whose website says its attorneys have more than 50 years of combined experience. The claim could be completely accurate. An AI assistant might reproduce it and link to the firm's biography page. What has been established?

We know where the statement appeared. We may still need to know how the figure was calculated, which attorneys it includes, and whether that experience is relevant to the particular matter a prospective client needs help with. Fifty combined years of practice isn't the same as fifty years handling custody disputes, or fifty years of experience for the attorney who would actually take the case.

Now take an HVAC contractor that says it completed 3,000 jobs last year. There may be excellent records behind that number. But was it counted from completed invoices, dispatched appointments, or work orders? Were canceled jobs removed? Are repeat visits counted separately? A number calculated from a system of record can be useful evidence, provided the calculation and period are clear. A number placed on a homepage without that context is harder to interpret.

Or consider a senior living community advertising around-the-clock nursing. A family may reasonably want to know whether that means a licensed nurse is physically present at the particular location at all hours, available on call, or part of a broader clinical network. The difference could be central to the decision, yet all three arrangements might be described in reassuring language on a marketing page.

In each example, there are really three questions: Who made the claim? What supports it? What does it actually prove?

Those questions aren't interchangeable. A citation can help answer the first. A registry, professional credential, documented policy, or operating record may help answer the second. The third requires interpretation and context. Often, a business can satisfy one question without satisfying the others.

This is why the distinction between accurate, verifiable, and decision-useful matters. A fact can be accurate but hard for an outsider to verify. It can be independently verified and still say very little about whether a particular provider is right for a particular customer.

The facts buyers care about aren't all the same kind of facts

Business information gets flattened when everything is reduced to a website description or a star rating. In reality, it falls into several categories with different evidentiary standards.

Some facts can be checked against an authoritative public source: a professional license, a registered business name, or a credential with a current issuing body. Even here, the source, jurisdiction, status, and date matter. A license that was active three years ago doesn't establish current good standing.

Other facts originate with the business but can be documented: the services it offers, the communities it serves, the warranties it provides, or the kinds of clients it accepts. Those statements may be entirely legitimate without having been independently verified. Labeling them as business-attested makes their origin clear; it doesn't make them false or somehow less useful.

Then there are operating measures, such as completed jobs, distinct customers served, or the share of customers who return. They may be computable from authorized business systems, but they still require definitions. A repeat-customer rate for a dental practice can't be interpreted the same way as one for a roofing company. The underlying purchase cycles are different. Even a properly calculated metric is not a universal measure of service quality.

Finally, there are judgments that records cannot settle: whether a lawyer is the right advocate, whether a physician communicates well, whether a contractor's team will handle a difficult installation carefully. Reviews, credentials, and operating history may inform those decisions. None eliminates the human judgment involved.

An AI system that treats all four categories as equally established risks creating a persuasive description without a sound basis for the parts that matter most.

Consumers aren't finished checking

There is an encouraging reality check in the consumer research. In a July 2026 BrightLocal survey of 1,227 U.S. consumers who had recently searched online for a local business, 75% reported using more than one channel during their most recent search.

Among respondents who were already using AI for local searches, only 18% said they felt ready to contact a business recommended by AI without doing additional research. Most wanted to check elsewhere or learn more first.

That doesn't prove consumers are specifically worried about unsupported claims. The survey doesn't establish their motives in that level of detail. But it does challenge the notion that an AI-generated shortlist automatically settles the decision. For many people, the answer is a starting point.

Now think about what happens as AI assistants get better at handling follow-up questions themselves. Someone may no longer need to leave the conversation to ask whether a contractor is licensed, whether a practice performs a particular procedure, or whether a law firm takes a certain kind of case. The assistant may attempt to resolve those questions on the spot.

That's useful, provided it knows when the available information supports an answer and when it doesn't. The more convenient the experience becomes, the easier it may be to overlook the gap between a sourced assertion and a substantiated fact.

Machine-readable doesn't mean independently verified

There's another version of this confusion playing out in AI-search marketing. Businesses are being told to restructure their websites, add special AI files, or publish more metadata so assistants will understand and recommend them.

Some technical hygiene is valuable. Clear pages, accessible content, accurate business details, and structured data all have legitimate purposes. But Google's guidance for generative search explicitly says that no special Schema.org markup or llms.txt file is required for its AI search features. Google's position is that traditional search fundamentals still apply.

More importantly for this discussion, organizing a statement into a machine-readable field doesn't improve the evidence behind it. "10,000 customers served" can be written in a paragraph or encoded as structured data. If it was simply entered by the business with no supporting definition or source, its status hasn't changed.

What would change the situation is a clearer record of how the claim was established: who attested to it, which registry confirmed it, or which operating records supported the calculation, along with an as-of date and any important limits. That information would also be useful to a person reviewing the business. Whether a particular AI system retrieves or uses it is a separate question that needs evidence of its own.

This is the thinking behind TrustRecord, which we're building at TrueSignal. The free record makes business-provided facts public with their attested status clear. Where supported systems are connected, additional operating metrics can be computed from authorized source records and published with their provenance and dates. The aim is to make the evidence legible and attributable, without pretending that publishing it guarantees a citation or recommendation.

A better standard for a convincing answer

I don't think the solution is to demand an audit for every sentence an AI produces. Most ordinary business facts don't require that. If a company says it installs heat pumps, the useful first step is to make that service clearly discoverable and correctly attributed to the company.

But when an answer moves from description to evaluation, the standard ought to rise. There is a meaningful difference between saying a company offers a service and saying it is experienced at that service; between listing a credential and confirming it is current; between reporting a customer count and inferring customer satisfaction from it.

A trustworthy answer should preserve those distinctions. It should be comfortable saying "according to the company" when that's the source, citing a regulator when a credential has actually been checked, and identifying an operating figure as computed from specified records when that's what supports it. It should also be willing to say the evidence isn't available.

We're still learning how often today's AI systems do any of this well. The BrightLocal findings tell us a great deal about the links that appear alongside local recommendations. They don't tell us whether the underlying claims were tested against independent records, or whether the recommended businesses were the most qualified choices. That remains an important research gap.

The industry has spent years making business information easier to find. AI has made it dramatically easier to assemble that information into a convincing answer. The next step is to make the basis for those answers easier to examine.

A citation tells you where a statement came from. A good recommendation should help you understand how much that statement is worth.