Somebody is looking your company up right now, and there is a decent chance it is not a person. It is an AI assistant, assembling an answer for a person who asked "is this company legit" or "what does this company do" and expects a verdict in seconds.

The verdict does not come from your homepage. It comes from the agreement, or disagreement, between everything public about you.

Machines do not read your website. They reconcile it

When an assistant is asked about a company, it does something closer to due diligence than to reading. It retrieves your website, your business registry entry, your LinkedIn company page, your domain records, directory listings, press mentions if any exist. Then it does the part most companies never think about. It compares them.

Same legal name everywhere, or three variants of it. Same founding year, or a website that says one thing and a registry that says another. An address on the site that matches the registered address, or does not. A LinkedIn description that sounds like the same company, or reads like it was written by a different one two rebrands ago.

Where the sources agree, the assistant answers with confidence. Where they conflict, it hedges, guesses, or quietly leaves you out of the answer. There is no error message. You just stop appearing, or appear wrong, and you will not know unless you go ask.

This is entity resolution, the same problem search engines have worked on for years, deciding whether mentions across the web refer to one real-world thing. What has changed is who consumes the result. It used to feed a ranking. Now it feeds a sentence a human takes at face value.

The inconsistency tax is paid twice

Here is why this matters more than it did two years ago. The people who used to look you up manually now ask an assistant first. Partnership managers screening applicants. Procurement teams checking a vendor. Compliance analysts deciding whether to onboard you. A buyer asking whether you are worth a call.

These are exactly the people trained to treat inconsistency as a warning sign. A human reviewer who finds your website saying one founding year and your registry saying another does not think "typo." They think "what else is off," and they have fifty other companies in the queue.

So inconsistency costs you twice. The machine downgrades its confidence and gives a vaguer, more hedged answer about you. Then the human, primed by that hedged answer, reads any remaining mismatch as confirmation. An AI assistant is a due diligence reviewer that never gets tired and never gives you the benefit of the doubt.

The uncomfortable part is that most inconsistency is not dishonesty. It is drift. The site got rewritten, the LinkedIn page did not. Someone shortened the legal name because it looked cleaner. The old address survived in a footer. Nobody was lying. The machine cannot tell the difference.

One description, repeated verbatim

The single highest-value fix is boring. Write one paragraph that says what your company is, in plain language, with the legal name, the location, the founding year, and what you actually do. Then use it everywhere, word for word. Homepage metadata, About page, LinkedIn, every directory and profile you control.

Verbatim matters. Not "roughly the same," not "adapted for each channel." Word for word. When a machine finds identical language across independent surfaces, that is a strong signal the surfaces describe the same entity and the description is deliberate. Paraphrases force the machine to judge whether two descriptions mean the same thing, and every judgment call is a chance to lose.

We do this for our own studio. One canonical paragraph, reused on every profile we hold, and it feels almost too simple to count as work. It counts. Consistency is not a style choice. It is the signal.

The instinct to freshen the copy per platform comes from writing for humans, who get bored. Machines do not get bored. They get confused, and confused machines describe you badly.

Structured data, or saying it so machines cannot misread it

Your website's visible text is written for people, and machines extract facts from prose imperfectly. Structured data fixes that. In practice this means a small block of code on your site, usually Organization schema in a format called JSON-LD, which is a label maker for facts. Invisible to visitors, unambiguous to machines.

You do not need to write it yourself, but you should know what belongs in it, because most implementations are half empty.

The block should state your legal name exactly as registered, not your brand name with the suffix dropped. Your founding date. Your registered address, matching the registry character for character. Your official email and logo. And the field almost everyone skips, called sameAs, which lists the URLs of your official profiles. Your LinkedIn page. Your entry in the public business registry. sameAs is you telling machines, explicitly, which pages on the internet are actually you, so they never have to guess whether that similarly named profile belongs to your company or someone else's.

Ten minutes of markup removes an entire category of misreading. Prose can be extracted wrong. A labeled field cannot.

Registries anchor everything else

Every source discussed so far is one you can edit this afternoon, which is exactly why, on its own, none of it proves anything. A public business registry entry is different. Companies House in the UK, the Secretary of State filings in US states, and their equivalents elsewhere are third-party records with real registration dates, filed names, and filed addresses. You cannot backdate one and you cannot casually rewrite one.

That is what makes the registry the anchor. Machines, and the reviewers behind them, weigh sources by how hard they are to fake. Your website says what you want it to say. The registry says what you filed and when. When your site, your schema, and your profiles all match the registry, every softer source inherits its credibility. When they contradict it, the registry wins and everything you wrote loses.

The practical rule, treat your registry entry as the master record. Every name, address, and date you publish anywhere should be checked against it, not against your own homepage.

What a small company can do in a week

None of this requires a budget. It requires an afternoon of honesty and a few days of cleanup.

Day one. Write the canonical paragraph. Legal name, plain description, location, founding year. Pull the legal name, address, and incorporation date directly from your registry entry, not from memory.

Day two. Audit yourself the way a machine would. Search your company name. Open every surface you find, website, LinkedIn, directories, old profiles you forgot existed. List every place the name, address, description, or founding year diverges. There will be more than you expect.

Days three and four. Fix the divergences. Replace every description with the canonical paragraph, verbatim. Kill or update stale profiles. Make the address in your footer match the registered address exactly.

Day five. Add Organization schema to your site with legal name, founding date, address, email, and sameAs links to your LinkedIn page and your registry entry. Link back the other way too, your LinkedIn profile should point at your domain. Check that your robots.txt file, the small file telling crawlers what they may read, is not blocking AI crawlers from the site you just cleaned up.

Then test it. Ask ChatGPT, Perplexity, and Google's AI assistants what your company is. Read the answers as a skeptical stranger would. That output is your first impression now, for machines and for the growing number of humans who ask machines before they ask you.

You cannot control what AI assistants say about your company. You can control whether every source they check tells the same story. Right now, that is the whole game.