What AI understands and says about your business, and how to check it
A model does not read only your website. It cross-checks what your site says against your Business Profile, your professional profiles, the press, directories and your job postings. When those sources contradict each other, the model's understanding gets blurry and the recommendation goes to the wrong customers. The diagnostic takes one hour, and it is judged on repetition rather than on a single answer.
- A recommendation aimed at the wrong type of customer comes more often from a contradiction between your sources than from a lack of visibility.
- The diagnostic is done by hand, on two models, by asking the same questions on three separate occasions.
- A single answer proves nothing: models vary from one run to the next, so you judge on what keeps coming back.
- The most contradicted sources are the least watched: job postings, professional profiles, old directory listings.
- The correction follows a precise order, because some sources feed others.
On this page
Entity consistency
Entity consistency is the degree of agreement between all the public sources that describe the same company: its website, its Business Profile, its professional profiles, industry directories, the press, public data and its job postings. It covers what the company does, for whom, where and at what scale. A language model builds its understanding by cross-checking these sources, and a contradiction between them shows up as an imprecise description rather than as a missing answer.
Why the recommendation goes to the wrong place
The most common complaint about assistants is not being absent. It is being put in the wrong category: a company recommended to customers that are too small, in a specialty it has dropped, or in a region it no longer serves.
The mechanism is simple. A model that comes across three different descriptions of the same company does not pick the most recent one: it builds a synthesis, and that synthesis leans toward whatever is repeated most often. An old specialty mentioned in four directories weighs more than a new specialty announced on a single page of your site.
The practical consequence is that the useful move is not to publish more. It is to bring what already exists into line, which costs time and almost nothing else.
This topic is separate from correcting information that is factually false (which has its own procedure) and from the way an assistant shortlists vendors. Here, the subject is consistency: no source is wrong, they simply do not say the same thing.
How a model arbitrates between two sources that contradict each other
Understanding the arbitration keeps you from correcting in the wrong place. Four elements carry weight, and none of them is your preference.
- Repetition: a statement found in six documents wins over a statement found in only one, even if that one is your site.
- The apparent date: a recently dated page weighs more than an undated page, which penalizes sites whose service pages carry no update date.
- The independence of the source: a directory or a media outlet is treated as outside confirmation, your site as a declaration.
- The sharpness of the wording: a short, explicit sentence gets picked up as is, while vague wording gets replaced by whatever the model finds elsewhere.
The consequence is counterintuitive, and it settles a lot of cases. Correcting your site without correcting the outside sources is not enough, because the weight of repetition works against you. Correcting five directories while keeping a vague page on your site is not enough either, because nothing sharp is available to pick up.
What to do about the sources you cannot correct
Some of the documents that describe you are out of your control: a press article from 2019, an archived job posting, a directory listing nobody has access to anymore. Deleting them is neither possible nor desirable in most cases.
The answer is frequency rather than deletion. Recent, dated and sharp sources end up weighing more than old ones, provided there are more of them. A dated service page, an up-to-date Business Profile, two aligned professional profiles and a recent mention in an active directory are generally enough to overturn an outdated description.
The particular case of information that is factually false calls for a different approach with the platform, and that is another topic.
The diagnostic, in one hour
The review is done by hand. It needs no paid tool, and its value lies in the method rather than the technology.
Prepare six questions, not one
Who you are, what you do, for what type of customer, where, how much you cost and who you are compared to. Those are the six questions asked by a person who is evaluating you.
Ask them to two models, with no history
A fresh session, without a logged-in account if possible, so that your own past conversations do not influence the answer.
Repeat on three separate occasions
Two days apart is enough. What comes back three times describes the model's understanding. What appears once describes a single run.
Note the sources cited under each answer
They give you the list of documents to correct, in the order of their real influence rather than their assumed influence.
Record the gaps, not the errors
Three columns: what is accurate, what is outdated, what comes from somewhere else. The middle column is the longest, and it is the one that gets corrected fastest.
One hour is enough for a first pass. The result is a gap table that names each gap, with the source of each one. That document becomes your work plan.
Two runs of the same test give different answers, and the gap can be wide. Presenting a single screenshot as proof of what an AI thinks of your company gives a fragile conclusion, in one direction as in the other.
The sources to check, one by one
The list below is ordered by how often a contradiction is observed, not by how well known the source is.
| Source | What most often contradicts itself there | Who can correct it |
|---|---|---|
| Job postings, current and archived | The services being developed, the size of the team, the tools used | You, but the archives remain |
| Professional profiles of the partners | The positioning, the industries served, the titles | Each person, individually |
| Old directory listings | Dropped services, the address, the service area | You, once you have recovered access |
| Google Business Profile | The services listed, the description, the service area | You, directly |
| Press and news releases | A dated specialty, a partnership that has ended | Nobody, but the frequency of recent sources makes up for it |
| Public data and registries | The legal name, the declared activities | You, through the administrative channel |
The first two rows usually come as a surprise, and they pay off the most. A job posting describes a company in precise, recent terms, it is well indexed, and nobody rereads it once the position is filled.
Two clarifications missing everywhere else
The first is about what you announce. A patent filed is not a product deployed, a certification in progress is not a certification obtained, and a partnership announced is not an active partnership. Models repeat these mentions in the present tense, without the conditional that appeared in the news release. The rule that protects you is to write in the past or the future tense whatever is not in service, everywhere, including in the documents you consider minor.
The second is about reading the tests. Answers vary from one run to the next, sometimes a lot. A gap observed once justifies no decision. A gap observed three times out of three describes a real problem. The rule also holds when the answer is flattering.
The correction plan, in order
The order has a reason: some sources feed others, and correcting downstream while the upstream is still wrong produces a new contradiction rather than convergence.
Settle the reference statement first
One sentence that says what you do, for whom and where. It will then be copied as is everywhere, with no creative variation from one platform to the next.
Correct your site
The about page, the service pages and the footer. It is the source the others will cite, so it goes first.
Align the Business Profile
Services, description, service area. It has become a cited source in its own right, which is covered in our note on the Google Business Profile.
Go back over the professional profiles
The company's and the partners', with the same reference statement. It is the correction most often forgotten.
Clean up the directories
The ones the diagnostic brought up as priorities, and the old ones you still have access to. An outdated listing you cannot edit gets reported to the directory.
Redo the diagnostic after eight weeks
The delay matches the time it takes for sources to be read again. A test redone the following week measures nothing.
What deserves a correction and what can wait
Not every contradiction costs the same. Three criteria are enough to sort them.
- Is the source cited in your records: a contradiction in a document the models do not consult can wait.
- Is the gap about who you serve: that is the type of error that brings in the wrong requests, so it is the most costly.
- Is the gap under your control: a listing you can edit gets corrected today, an old press article gets offset by the frequency of recent sources.
A vendor who presents a screenshot of a single answer as a diagnostic has not accounted for the variation between runs. One who hands you a gap table with the source of each gap has done the work.
Running this diagnostic and building the correction plan is part of what we cover in a 90-minute consultation.
What stays in-house is the reference statement: deciding how your company describes itself today is a leadership decision, not an execution task, and everything else follows from it. What can be delegated is recording the gaps, taking inventory of the sources and following up on corrections with the directories, because it is long, repetitive and teaches you nothing. The point that most often gets stuck in practice is access: finding out who created a listing six years ago takes longer than correcting it. Starting with an inventory of your accesses keeps you from discovering the problem in the middle of the project.
Checking your consistency
The procedure for information that is factually false is different, and it is detailed in our note on false information from an assistant. The way an assistant builds a list of vendors is covered in AI-assisted vendor shortlisting. The work on mentions that follows once your sources are consistent is in getting cited by AI without buying links.
Taking back control of what the sources say about you is at the heart of the Strengthen your visibility in AI answers goal.
Already running a marketing team? See how we plug in as reinforcement on answer engine optimization.
Frequently asked questions about what an AI understands about your business
How do I find out what ChatGPT says about my business?
By asking six questions in a fresh session, on two models, on three occasions two days apart, then noting the sources cited under each answer. What comes back three times describes the model's understanding. What appears only once describes a single run.
Why does AI recommend me to the wrong customers?
Most often because your public sources do not say the same thing about you. A model leans toward whatever is repeated most often, so an old specialty found in several directories weighs more than a new one announced on a single page of your site.
Which source should be corrected first?
Your site, once a reference statement is settled, because it is the source the others will cite. Next come the Business Profile, the partners' professional profiles, then the directories the diagnostic brought up.
How long before the correction shows?
Allow about eight weeks before redoing the diagnostic. That is the time it takes for sources to be read again and picked up in the answers. A test redone the following week only measures the usual variation between runs.
Does a single wrong answer justify stepping in?
No. Models vary from one run to the next, including on identical questions. A gap observed once justifies no decision. The same gap observed three times out of three describes a real problem to correct.
Should you mention a patent or a certification in progress?
Yes, provided you write it in the future tense or the conditional everywhere. Models repeat these mentions in the present tense and turn a filing into a deployed product. The rule also applies to the documents you consider minor, such as old news releases and professional profiles.
- Semrush and Kevin Indig, Only 25% of cited sources overlap between ChatGPT's different reasoning modes, June 30, 2026, on how sources vary between two runs.
- Profound, Google AI Mode's shift to citing itself, July 2026, on the weight of Business Profiles in citations.
- Zyppy, AI citation ranking factors, May 7, 2026, on the relative weight of brand mentions.
Geneviève puts the strategy for your engagement into action. She leads all our web development projects: Shopify, WordPress and the new ways of building a site with AI. She manages our team of developers and translates your business needs into technical language. She runs your organic search (SEO), your visibility in AI answers (GEO) and your site's conversion rate optimization (CRO). Her work is at the heart of three goals: Attract customers with SEO and AI, Improve your site's conversion, and Strengthen your visibility in AI answers. With Gabriel, she also builds the landing pages for your advertising campaigns. She writes mainly about SEO, AI visibility and web design.
About Falia →
SEO
Paid advertising
Conversion
AI visibility
Web design