How Do I Find Out What ChatGPT Says About My Company?

Ask the engines yourself, from a browser with no account signed in, using the questions a buyer would type rather than your company name. Run each question three times and open every link it cites.

The whole thing takes about half an hour. Most brands never get a real answer out of it, because the check they run is built to flatter them, and the rest of this is about running one that does not.

The check most people run tells them nothing

Four things turn a check into a compliment, and most people manage at least two of them on the first try.

You are signed in. The engine personalises around who you are and hands you back your own brand. We walked into this one ourselves, and the mechanism is explained in how AI engines decide which brands to recommend. Use a private window with no account.

You asked by name. Typing your company name tells you whether the model holds a description of you. It does not tell you whether you would be named to somebody who only knows what they need. Those are different questions, and only the second one is worth money.

You asked a leading question. “Is Alkeme a good packaging studio?” gets a yes from almost any assistant, because the question already contains the premise. Assistants agree with the shape of the question put to them. Ask who, not whether.

You ran it once. The same question asked three times can produce three different shortlists. One run is a screenshot, not a measurement, and a single lucky run is the most common reason a founder believes they are visible.

What a reliable check looks like

Ten buyer questions, three engines, three runs each, and five columns written down. Ninety answers sounds like a lot and takes half an hour, because you are scanning for your own name, not reading.

Write the ten questions once and never reword them. They are the questions somebody with your problem types, not the ones you wish they typed: which packaging design studio to use for a coffee brand launching in the EU, who designs brand identities for dental clinics, who builds a product configurator into a website, which studio works founder-direct with no account manager.

For each answer, record five things:

Did you appear. Yes or no, with no interpretation.

Where in the list. First of three names is not the same result as third of eight.

Which sources were cited. Everyone skips this column and it is the most useful one, because it tells you which pages the answer was assembled from, and therefore where you need to exist.

Whether the engine searched or answered from memory. An answer with no sources at all came out of training data and is months or years old. An answer with live sources was built just now. Those two need different repairs.

The date. Without it, next month’s run has nothing to compare against.

Then repeat monthly with the same questions. The value is in the comparison, and if you reword the questions each time, you will never know whether the world changed or only your phrasing did.

How to tell a real citation from an invented one

Open every cited link. That single habit separates a measurement from a story, and there are exactly three outcomes.

The address does not resolve. The model produced a plausible URL that was never real, which happens most often with deep links into directories and listicles.

The address resolves, but the sentence attributed to it is not on the page. The source exists and the claim was assembled anyway. This is the one that fools people, because the citation looks legitimate until you read the page.

The address resolves and the sentence is there. That is a genuine citation, and it is the only kind worth counting.

A second useful probe: ask the engine to quote the exact sentence it is relying on. A real citation produces a sentence you can find with a search on the page. An invented one produces a paraphrase that exists nowhere.

The one measurement that does not depend on asking

Your own server logs tell you when AI engines actually opened your pages, and they cannot be flattered. Every engine identifies itself with a user agent string, and the strings fall into two groups that mean completely different things.

Training and index crawlers visit constantly whether anybody asked about you or not. Their volume tells you that your site is readable, nothing more.

The on demand fetchers are the ones worth watching, because they only run when a real person asked a question and the engine decided to open your page to answer it. On our own site, over thirty days, those accounted for 310 page opens: 171 from ChatGPT, 83 from Perplexity, 56 from Claude. No public tool reports that number for you and Search Console does not show it. You read it off your own logs or your CDN analytics.

Two cautions. A fetch is not a recommendation: it means the engine reached your page while composing an answer, which is the step before being named in one. And a user agent string is a claim, not an identity, so a name in your logs can be someone imitating a crawler. Google-Extended is the clearest example, since it is a control token you put in robots.txt rather than a crawler that visits, so a log line carrying that name is not Google.

Reading the result you got

Five patterns cover almost every outcome, and each one points at a different repair.

Absent from every engine, on every question. The problem is foundational rather than editorial. The most common cause is a site whose text only appears after JavaScript runs, so there is nothing for the engines to read.

Present only when you type your company name. The model knows you exist but does not connect you to what you sell. What is missing is facts: what you do, who for, in which markets, since when.

Present on one engine and absent on the others. Normal, because the indexes behind them differ. It becomes a signal only if the same engine keeps you out three months running.

Present, with the facts wrong. The most irritating result and the one with the longest repair time. Next section.

Cited as a source, while somebody else gets recommended. Not an error. If your page lists ten studios, the engine can use it to recommend any of them.

What to do when the answer about you is wrong

Correct it at the source, because nothing you type into a conversation changes what the next person is told. The model is repeating a consensus it assembled, so you change the consensus.

Work out first which kind of wrong you are looking at. A fact that was true once and is now stale needs the old version removed or dated wherever it still lives. A fact that belongs to a company with a name close to yours is an attribution problem, and it gets fixed by making your own details unambiguous: full legal name, location, founding year, category, identical in every place you appear. A fact nobody ever published is the model filling a gap, and the repair is to fill that gap yourself with a clear sentence in a place built to be quoted.

Then make the correct version identical everywhere: your site, your business listing, every profile, the same words rather than inspired synonyms. Then wait, because this takes months rather than weeks.

What not to do with a bad result: write ten articles next month to fix it. If the site cannot be read without JavaScript, those ten articles cannot be read either. And do not buy a placement, because nobody holds sellable positions inside an answer. The work that actually moves it is in how to get ChatGPT to recommend your brand.

How we think about it

The check above is the first part of an online presence audit, and the quick part. The rest is the half a spreadsheet cannot give you: why the answers look the way they do, and the order in which the causes get repaired. Making an engine understand and attribute you correctly as a company is its own piece of work.

If your market is not asking AI assistants about your category yet, we will tell you that instead of selling you a monthly report about it.

FAQ

Where does ChatGPT find facts about a UK company? From your own site, Google Business Profile, Companies House, UK directories and press, so those facts need to agree.

How do I find out what ChatGPT says about my company? Open ChatGPT in a browser with no account signed in, ask the ten questions a buyer would type rather than your company name, and run each one three times. Record whether you appeared, where in the list you sat, and which sources were cited. Repeat monthly with the same questions written down.

Why do I need to ask the same question more than once? Because the answers vary between runs. The same question asked three times can produce three different shortlists, so one run tells you almost nothing. Three runs tell you whether you are reliably in the answer, occasionally in it, or absent.

How can I tell whether a citation is real or invented? Open the link. There are three outcomes: the address does not resolve, which means it was fabricated; the page loads but the sentence attributed to it is not there, which means the attribution was invented; or the page loads and the sentence is on it, which is a real citation. Only the third one counts.

Can I see AI engines visiting my site? Yes, in your server or CDN logs. Training crawlers visit constantly and tell you nothing about demand. The on demand fetchers are the useful ones, because they only run when a real person asked something that made the engine open your page. On our own site those accounted for 310 page opens in thirty days.

What do I do if ChatGPT says something false about my company? Correct it at the source, not in the chat. Nothing you type into a conversation changes what the model tells the next person. Work out whether the fact was once true, belongs to a company with a similar name, or was invented, then make the correct version appear identically on your site, your business listing and every profile where you exist.