AI visibility is the topic on everyone's mind, and with good reason. In October 2026 we ran full audits for two firms: a law firm, and a lender whose press coverage we track with our in-house Rostrum.AI media monitoring. What the assistants said about each matched neither its Google rankings nor its coverage, and neither organisation's marketing team had a clear picture of their brand's performance on AI.
Neither Google rank nor press coverage predicted the AI answers. In our two audits in October 2026, a law firm outside Google's top 30 for all nine of its buyer questions was named in 62% of AI answers; a lender with more than 250 news articles in a year was named in 18%. Of the 793 links the answers cited, about half were firms' own websites, 21% directories and rankings, and under 2% national and trade press.
This is the checklist we now use to determine AI visibility, what we've learned by developing it, and the numbers you need to read the results honestly. If you want to get a clear picture of how your brand is performing on AI, this is the kind of approach you need to take.
The checklist
Write 25 to 30 questions your buyers actually ask, from first calls and tenders, in their words. Add two or three questions about your firm and score those separately.
Choose the three or four assistants your buyers use most.
Ask every question three times in each assistant, in a fresh chat each time, in the same week each quarter.
For each answer, record the firms it named, in order, what it said about you and the links it cited.
Score each firm as a share of answers, and check any change against the margin before you report it.
Sort every cited link by type, and open the most cited pages to see where their facts came from.
Read every answer that describes you, and note what is old or wrong and the page it came from.
Turn what you find into four or five jobs, each with an owner.
Give the board three lines: your share against competitors with its margin, what the assistants get wrong about you, and the sources you most need to be in.
Run the same questions again next quarter, and each subsequent quarter.
The rest of this article explains what each step in the process does, and why it's included.
What did we find when we ran the audit?
Mostly that the answers change every time you ask. Our last piece, which PR coverage matters for SEO and AI search, ended by recommending a quarterly check with fixed questions, repeated tests and competitor comparisons, so we ran one properly for two firms.
Ask the same assistant the same question twice and, across 237 such pairs, it gave the same list of firms only 21% of the time. Two answers to one question typically shared under half their firms. SparkToro and Gumshoe found the same thing at scale: across nearly 3,000 prompts run between November and December 2025, the same list of brands came back fewer than once in every 100 times. Their prompts asked for longer lists of brands in bigger consumer markets; ours were narrow questions in b2b professional markets, where the lists are short and an exact repeat comes easier. Either way, one answer tells you very little. The chart below shows what that looks like for a single question.
One question, asked 14 times. In one of our October 2026 audits, 14 answers to the same buyer question named 38 different firms between them, and 5 different firms came first. Our client was named in 10 of the 14, anywhere from 2nd to 17th.
The assistants also disagreed with each other, by as much as 37 points on the same firm:
Assistant | Law firm | Lender |
|---|---|---|
ChatGPT | 81% | 24% |
Google AI Overviews | 52% | 28% |
Google AI Mode | 74% | 10% |
Gemini | 44% | 19% |
Perplexity | 56% | 14% |
Claude | 56% | not once |
Overall | 62% | 18% |
The assistants named 89 firms in one audit and 90 in the other. Of those, 85 and 84 were not on the client's own competitor list, so you do need to record every firm an answer names, including the ones you have never thought of as rivals.
Google told us almost nothing. The law firm ranked outside the top 30 for every one of its nine buyer questions, and topped the results only where its own name was in the query. The lender sat outside Google's top 10 for all of its buyer questions.
B2B buyers are moving this way. In a G2 survey of 1,076 B2B software buyers in March 2026, 51% said they start research in an AI chatbot more often than in Google, up from 29% eleven months earlier. Those are b2b software buyers, so read it as a direction of travel.
How do you write the questions your buyers actually ask?
From real conversations, in the words a buyer would use. People do not type keywords into an assistant; they describe a situation and ask for names.
Start with first calls and tenders. What did new clients ask, and what did the last tender want?
Write them as a buyer speaks. Compare "best employment lawyers" with "Which employment lawyers in London should an HR director shortlist for a tribunal claim?"
Mix three kinds: open questions, questions with a situation in them, and sceptical ones about whether the big names are worth the fees.
Keep the questions about you out of the score, because they will always mention you.
Ask the same set every quarter, so the numbers stay comparable.
If you only appear in the open answers, that is general reputation. If you appear in the situation answers, someone has written about you doing that specific thing, which is harder to win and worth more.
How many questions do you need?
More than most firms use. We asked nine questions, and that was too few to track change.
Asking the same question again mostly gets you the same kind of answer. The firms an assistant favours for one question tend to come back for that question, in other assistants and on other days. So a repeat teaches you less than a new question would. In our audits, nine questions asked in six assistants three times each gave 162 answers, and they carried about as much information as 35 fully independent answers would.
With nine questions, a quarterly change of less than about 20 points could be chance. If you treat all 162 answers as independent, the margin looks like nine points, which is how firms end up reporting noise as progress.
To catch a change reliably, it has to clear the noise with room to spare, so the figures below are larger than the margins above. The table shows the smallest change you could reliably detect between two quarters, asking each question three times, for a firm named in about a fifth of answers:
Questions | Assistants | Answers a quarter | Smallest change you would reliably catch |
|---|---|---|---|
9 | 6 | 162 | about 28 points |
15 | 3 | 135 | about 23 points |
30 | 3 | 270 | about 16 points |
30 | 6 | 540 | about 15 points |
Two things follow. Questions matter far more than assistants: 30 questions in three assistants catches nearly as much as 30 in six, for half the work. And repeats beyond three add little; spend the effort on more questions.
So write 25 to 30 questions and ask them in the three or four assistants your buyers use. By hand that is around 300 answers a quarter, or roughly a day of someone's time. Even then, a change of less than about 12 points could still be chance, so treat it as a reason to watch the next quarter, and keep it out of the board report.
Those margins are the cautious reading, which pools every answer. A sharper one compares each question with itself from one quarter to the next, because whatever makes a question favour a firm tends to persist; a statistician can run that for you from the same grid.
Where do the answers really come from?
From the links the assistants cite, and from the sources behind those links. List every cited link, sort it by type, then open the most cited pages and ask where they got their facts.
In our two audits the answers cited 793 links, and roughly half pointed to firms' own websites. Directories and rankings accounted for 21%, comparison lists 2%, and national and trade press under 2%.
The two markets behaved differently, which is why you need your own map. In the legal market, Legal 500 and Chambers were the two most cited sources, ahead of every firm's own website, and the law firm was named in most answers with little independent press coverage. In business finance, the four most cited pages were competitors' pages written for the city in the buyer's question, and the lender had no such page. Its 250 articles barely registered, because across both audits the assistants cited national and trade press in under 2% of links.
That fits our last piece, which argued that the type of coverage matters more than the amount.
What is the assistant saying about you?
Read every answer that describes you, because no score will show you what they say. For the law firm, all 105 descriptions were accurate. For the lender, 26 of 59 mentioned the name it traded under until a few years ago, and one of the pages the assistants cited was a broker's listing still filed under the old name.
None of that is strictly wrong, which is why an automated check is not enough. Ours flagged 19 of the lender's answers as out of date, and on reading them most were simply mentions of the old name. What matters is the pattern: the assistants describe the firm the internet wrote about, and every part of that description has a page behind it that you can update, replace or outweigh.
What do you do with what you find?
Turn the source map into a short list of jobs, each with an owner:
Get the directories and rankings right where they appear in your answers. Check the entries, make the submissions and line up the referees.
Make your own service pages say plainly what you do, who for and with what evidence, such as a named example or a figure. Roughly half the citations were firms' own websites, and assistants tend to quote what is written plainly. If your buyers ask about a city, write a page for that city: in our lender's market, those pages were the most cited of all.
Aim coverage at the outlets and lists that actually appear in the answers to your buyers' questions. They may not be the titles you would have chosen first.
Fix old or wrong facts at their source, so you change the page the assistant reads rather than the answer in front of you.
Next quarter's grid will show which aspects of your work have had an effect.
What should the board see?
Three lines, with the detail kept in the working file:
Your share of answers against competitors, with the margin shown honestly.
What the assistants are getting wrong about you, and where it comes from.
The three or four sources you most need to be in next quarter.
Can a tool do the counting?
Yes, and at 300 answers a quarter the maths is the difficult part. Rostrum has built a free AI visibility audit tool to do it, and the first part of it runs here.
How does AI read your website? Put in any website address on our free AI visibility audit tool and you get a free check of how AI reads it, with fixes ready to paste.
Firms in financial, legal, professional and B2B services and property can go further, benefiting from the full audit across the six assistants we check: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude. You also get your Google rank from your buyers' city, who is named in your place, the sources behind the answers and what to fix first. The report lands by email once you confirm your work email (we never share your email).
If you would rather we ran the audit for you, get in touch. And thank you to everyone who read and shared our last piece.
Frequently asked questions
How often should we check our AI visibility?
Once a quarter suits most firms. The answers move week to week, so checking more often tells you about noise rather than progress. Put the same week in the diary each quarter and keep the questions fixed.
How many times should we ask each question?
Three times in each assistant, in a fresh chat each time. After that, add questions rather than repeats: in our audits, answers to the same question moved together, so a new question told us far more than another repeat of an old one.
How many questions do we need?
About 25 to 30 if you want to track change from quarter to quarter. With nine, as we found, only a move of more than about 20 points is clearly more than chance.
Does a good Google ranking mean AI assistants will recommend us?
No. In our audits one firm ranked outside Google's top 30 for every buyer question yet appeared in 62% of AI answers. Ahrefs found much the same across 15,000 prompts in a study published on 11 August 2025, with only 12% of cited links in Google's top 10 for the same prompt. A strong Google position helps, but it no longer predicts what the assistants say.
