The fastest way to find out what AI models believe about your business is to ask them, then go and find the sources that made them believe it. This is a one-day method. It needs a spreadsheet and no tools, and it will produce more actionable findings than most paid audits.
This is the starting point for the ecosystem validation cluster. Do this before you fix anything, because you cannot correct contradictions you have not found.
Step 1: interrogate the models. One hour
Five questions, five platforms, answers recorded verbatim.
The questions:
- “What is [business name]?”
- “What does [business name] do?”
- “Where is [business name] based?”
- “When was [business name] founded and who runs it?”
- “Who are the best [your category] in [your area]?”
The platforms: ChatGPT, Perplexity, Google AI Overviews, Gemini and Copilot. Claude too if your buyers use it.
Record the answers word for word in a document, dated. Do not paraphrase. The exact wording is the evidence, and you will want to compare it against the same test in six months.
Then read them for four things.
Contradictions between platforms. Two different founding years, two addresses, two descriptions. Each contradiction points at a real source somewhere saying each version.
Confident errors. The most useful finding. If a model states something wrong with confidence, a source it trusts is saying it. That source is findable.
Entity confusion. Details from a similarly named company, or from a dissolved predecessor. Very common where a business has restructured. See public records and governance surfaces.
Absence. No usable answer, or a redirect to a directory. That is not a citation problem, it is an identity problem, and it means the foundation layer is not done.
Question 5 does a different job. It tells you who is being named instead of you, which is a competitor list assembled by the systems you are trying to influence rather than by your own assumptions.
The three signals, and what each one tells you
Before step 2, know what you are reading. The answers you have just recorded contain three different signals, they mean different things, and treating them as one number is how people misdiagnose the problem.
Recognition. Does the model know you exist and can it describe you correctly? This is a yes-or-no about identity, and it is the answer to questions 1 to 4. Failure here is an entity problem and nothing downstream will work until it is fixed.
Mention. Are you named anywhere in an answer, including in passing, in a list, or as one of several options? This tells you that you are in the source pool but not necessarily preferred.
Citation. Are you named as the source of a claim, linked, or recommended with a reason attached? This is the commercial signal. A citation carries the model’s own endorsement in a way a mention does not.
The three do not move together, and the gap between them is the diagnosis.
| Pattern | What it means | What to work on |
|---|---|---|
| No recognition | The entity does not resolve. You are not competing yet | Registers, canonical description, schema. Steps 2 to 5 below |
| Recognition, no mentions | The model knows you but will not put you in an answer | Corroboration. Independent sources, reviews, coverage |
| Mentions, no citations | You are in the pool but not the preferred source | Extractability. Your pages answer vaguely where rivals answer precisely |
| Citations on one platform only | Your footprint suits one retrieval pattern | Find what that platform is reading and widen it |
| Confident wrong answers | A trusted source is publishing something false | Find the source. This is the highest-value finding available |
That last row is worth dwelling on. A confident error is better news than absence, because it is traceable. Something a model trusts is saying it, that something is findable, and correcting one authoritative source frequently corrects every downstream repetition of it.
Read the reasons, not just the names. When an assistant recommends a competitor, it usually says why: they specialise in something, they serve a sector, they published a figure. That clause is the criterion the model applied. It tells you what would have to be true of your own footprint for you to be named instead, and it is the single most useful sentence in the whole exercise.
Watch for blending. Details from a similarly named company, a dissolved predecessor or a different industry appearing in your description is entity confusion rather than absence, and it needs the register work in public records and governance surfaces.
One caution on method. Answers vary between runs, so a single test proves very little. Use the same prompts, the same wording, the same month, and read the trend rather than any one reading. Two runs a few days apart before you conclude anything is a reasonable discipline.
Step 2: find every source. Three to four hours
The bulk of the work, and the part people cut short. Be thorough, because the source doing the damage is almost always one nobody remembered existed.
Search for all of these, and go at least three pages deep:
- Your exact business name, in quotes
- Your business name without the legal suffix
- Any former trading name
- Your telephone number
- Your registered address
- Your founder’s name plus the business name
- Your business name plus “reviews”
- Your business name plus each town you serve
Then check directly:
| Source type | Specifics |
|---|---|
| Public registers | Companies House, your regulator, ICO, trade bodies, accreditations |
| Map and place | Google Business Profile, Apple Business Connect, Bing Places |
| Reviews | Google, Trustpilot, your sector platform |
| Social | Every platform, including accounts nobody uses |
| Directories | Sector directories, chambers, aggregators, marketplace profiles |
| Company data sites | The aggregators that mirror Companies House |
| Media | Anything that has written about you |
| Employee profiles | LinkedIn, for at least your leadership team |
Step 3: record what each one says. Two hours
One spreadsheet. One row per source. These columns:
Source · URL · Business name as written · Address · Phone · Website URL · Founded · Description · Named people · Last updated · Do we control it
Fill it in exactly as written. Do not tidy as you go. “Ltd” versus “Limited” is a finding, and if you normalise it while typing you will lose it.
Then sort by each column in turn. Every column where the rows disagree is a finding, and the disagreements will be obvious the moment the data is sorted.
Step 4: prioritise. Contradictions before gaps
The order that matters:
1. Contradictions on authoritative sources. A wrong registered office on Companies House is the single most damaging finding available, because it is authoritative, machine-readable and mirrored across dozens of company data sites.
2. Contradictions on sources you control. Your own social profiles, directory listings you claimed, your website footer. Fast, free, entirely within your gift.
3. Contradictions on sources you do not control. Correction requests, claim processes. Slower, still worth it.
4. Stale or abandoned profiles. Correct them or delete them. Do not leave them. Covered in social platform consistency.
5. Gaps. Sources where you should exist and do not. Genuinely the lowest priority, and the one most audits lead with because adding listings is easier to sell than removing contradictions.
That ordering is the main thing this page is for. Most businesses given an audit start by adding listings, which increases the number of sources describing an entity that still does not reconcile.
Step 5: fix, and record what you changed
Work down the list. Keep a dated log of every change, because the next step depends on it.
Two things worth knowing before you start.
Some sources will not change. Dissolved companies stay on the register. Some directories will not respond. Where a contradiction cannot be removed, the mitigation is to strengthen every other signal so the weight of agreement is overwhelming, and to state the position publicly. MarGen does this on its own entity map, where unconfirmed profiles are listed as outstanding rather than quietly claimed.
Propagation is not instant. Register corrections take weeks to reach the aggregators that mirror them, and some re-sync on their own schedule. Expect the model answers to lag your corrections by four to eight weeks, and do not conclude the work failed at week three.
Step 6: re-test, and diarise
Re-run step 1 at 6 weeks, 12 weeks and 6 months. Same questions, same platforms, same wording, recorded the same way. You are looking for the confident errors to disappear and the descriptions to converge.
Then diarise the whole audit for six months, and add it to the process for four trigger events: a move, a rename, a leadership change and a domain change. Those four create almost every new contradiction, and the six-month repeat catches whatever they missed.
What good looks like
The finish line is not a score. It is this: ask five assistants who you are and get the same answer five times, and it is the answer you would have given.
When that is true, corroboration is doing its job and the rest of the AI visibility work has something solid to build on. When it is not true, everything downstream is being spent on a brand the systems are not sure about, which is the argument made at length in AI visibility strategy.
A note on why this is not a data page
Rank4AI publishes what it learned auditing 1,400 UK businesses. MarGen does not yet publish an equivalent dataset, and this page deliberately does not invent one. Original research is a separate piece of work with its own standards, and a page of unsourced percentages would fail the same test this site applies to everybody else.
What is here instead is the method, which is the part you can use today. When the dataset exists and the sample is large enough to be non-identifying, it will be published with the method attached and this page will link to it.
Off-Site Ecosystem Audit: Common Questions
How do I audit what AI models know about my business?
Ask five assistants the same five questions about your business and record every answer verbatim. Who is this business, what does it do, where is it based, when was it founded, who runs it. Contradictions between the answers point straight at contradicting sources, which is a faster and cheaper diagnosis than any audit tool will give you.
How long does an off-site ecosystem audit take?
About a day for a small business and two to three for one with a long history, multiple trading names or several locations. The interrogation is an hour. Finding every public source is the bulk of it. Recording what each one says is mechanical, and the second run six months later takes a fraction of the time.
What should I look for in an ecosystem audit?
Disagreement, not absence. Record the business name, address, phone number, founding year, description and named leadership from every public source you can find, put them in one spreadsheet and look for columns where the rows differ. Every difference is a finding. Missing listings are a lower priority than wrong ones.
Do I need a tool to audit my off-site presence?
No, and doing the first one by hand is better. You read the actual sources rather than a score, which is where the useful findings are. Citation and listing management tools are worth considering once the corrections are done and you need to keep a large footprint aligned, but they will not tell you what a model currently believes.
What is the most common finding in an ecosystem audit?
Address inconsistency, usually from a move that was never propagated to Companies House, old directory listings and abandoned social profiles. Second is a founding year that differs between the website, Companies House and LinkedIn. Third is a former trading name still live on sources nobody has logged into for years.
How often should I repeat the audit?
Every six months, plus immediately after a move, a rename, a leadership change or a domain change. The repeat is much faster because you already have the spreadsheet. The trigger events matter more than the calendar, because they are what creates new contradictions.
Should I audit competitors too?
It is worth an hour. Run the same five questions about the two competitors who appear in AI answers when you do not, and look at where they are listed that you are not. It tells you which surfaces the models are actually consulting in your sector, which is more useful than a general best practice list.
What do I fix first after an audit?
Contradictions before gaps, and authoritative sources before minor ones. A wrong address on Companies House outranks a missing listing on a directory nobody reads. Work down from public registers, through social profile fields, to directories, and treat missing listings as a later, optional phase.
Where to Go Next
Fix what the audit found: public records and governance · social platform consistency · review platforms and trust signals · media distribution and entity reinforcement
Related diagnostics: AI giving wrong information about my company · why am I not in AI search results · how to audit your AI search visibility
Have it delivered: MarGen is a UK GEO agency with published pricing, a published method and documented results.