TL;DR: When you check what AI search engines say about your business, it’s important to view the results as only one data point. AI answers can vary widely, so use a systematic approach by trying multiple prompts and engines. This repeatable habit, made scalable with Aurora, can help maintain an accurate presence in AI-driven search results.
Table of Contents
Most business owners have never once opened an AI assistant and asked it the question their customers ask. They have typed in their own company name, read the answer, and felt either reassured or rattled. That is not a check. This is how to check what AI search engines say about your business properly: which prompts to run, on which engines, what to write down, and what to do with what comes back.
Why this is worth checking now
Not because the sky is falling. Because the habit is spreading faster than most marketing plans have noticed, and because the answers are easier to influence early than late.
Ofcom found that more than half of UK adults, 54%, now use AI tools such as ChatGPT, Copilot or Gemini, rising to 79% of 16 to 24 year olds. That is a survey of 7,533 UK adults, published in April 2026. Google’s AI Mode, the conversational tab in Search, launched in the UK in July 2025 and passed a billion monthly users globally by the middle of 2026.
Buying through those tools is a smaller story, and it is worth being honest about that. Ofcom’s September 2026 tracker found 3% of online adults had used an AI tool to make a price, product or service comparison in the telecoms market. Retail data tells the same shape of story from the other end: Adobe Analytics measured AI-referred traffic to US retail sites up 393% year on year in the first quarter of 2026, while noting in its own holiday report that the base of users remains modest. Fast growth from a small base is still a small base.
So the case for checking is not that you are losing customers to AI today. It is that a small, fast-growing share of people are forming shortlists somewhere you cannot see, the answers they get are built from sources you can influence, and the businesses that fix their facts now will be the ones described accurately when the numbers are no longer small.
What it means to check what AI search engines say about your business
An AI visibility check is simple to describe and easy to do badly. You ask AI assistants the questions your customers actually ask, on purpose, in a fixed set, and you record exactly what comes back.
That is different from typing your company name into ChatGPT once, reading a flattering paragraph and deciding you are fine. It is also different from typing it once, getting nothing, and deciding the sky is falling. One prompt on one engine on one day is one data point from one session. Answers differ run to run and per user, so the check only becomes useful when you repeat it the same way and watch the pattern.
It is worth separating this from the search rankings you already track. Ranking tells you where your page sits on a results page when someone searches a phrase. An AI visibility check tells you whether an assistant names you at all when someone describes a need, how it describes you, and which sources it leaned on to decide. The two overlap, because both reward clear, accurate, well corroborated information, but they are not the same measurement and a good position in one does not buy you the other. We have written separately on what GEO and SEO actually have in common.
Why the same question gives different answers
This is the part that catches people out, so it is worth settling before you run anything.
Microsoft’s own developer documentation puts it plainly: “By default if you ask an Azure OpenAI Chat Completion model the same question multiple times you’re likely to get a different response.” Even with the settings designed to make output repeatable, it adds, “Determinism isn’t guaranteed.”
Google says something similar about its two search surfaces: “AI Mode and AI Overviews may use different models and techniques, so the set of responses and links they show will vary.” Google also notes that AI Overviews “often don’t trigger” at all, because they only appear when its systems judge them useful.
Location and context move the answer too. Anthropic’s developer documentation for Claude’s web search includes a parameter for the user’s location, precisely so results can be localised.
None of that makes the check pointless. It makes a single check pointless. It is also why the answer you get is not a scorecard on your business: Deloitte’s 2026 UK consumer research found that 27% of AI users “still do not realise that Gen AI is a probabilistic tool that may generate inaccuracies, and whose outputs should be checked”. One run is one data point from one session. Ten dated runs across six engines is a measurement.
The prompts to actually run
Most people run one prompt: “tell me about [my company]”. It is the least useful question you can ask, because you have handed the assistant the answer. Real buyers rarely start with your name. Run four kinds of prompt instead.
Branded prompts
“Tell me about [business].” “Is [business] any good?” “What do people say about [business]?”
These tell you what the engines believe about you as a described entity: what you sell, where you are, who you serve, and whether any of it is out of date. Watch for quietly wrong details, an old address, a service you dropped, a founder who left, a price that was never yours.
Category prompts, the ones that matter most
“Best [what you sell] in [your town].” “Who should I use for [the problem you solve]?” “Recommend a [category] for [specific situation].”
Open a fresh chat, ask the question a buyer would ask, and see whether you appear at all. This is the [“best [service] near me” question we have written about before](https://thegeoagency.io/best-service-near-me/), asked of an assistant instead of a search box. This is the prompt family that decides whether you are on the shortlist, and it is where most businesses find out they are not. Ask it without your brand name anywhere in the conversation.
Comparison prompts
“[Business] vs [competitor].” “How does [business] compare with other [category] in [place]?”
These surface who the engines consider your peer set, which is not always who you think. They also expose the criteria the assistant reaches for, price, speed, specialism, reviews, and whether your strengths appear at all.
Trust prompts
“Is [business] legitimate?” “Is [business] safe to use?” “Any complaints about [business]?”
These reveal what an assistant does with thin or conflicting information. Where the facts about you are inconsistent, an assistant tends to hedge, and hedging reads to a buyer as doubt.
Which engines to check, and why one isn’t enough
Check six surfaces, because they behave differently and your buyers do not all use the same one.
- ChatGPT, the default for most people, and usually the first place a business discovers it is invisible.
- Google AI Overviews, the summary above the blue links, which reaches people who never intended to use an assistant at all.
- Google AI Mode, the fuller conversational surface, which behaves less like a results page and more like an assistant.
- Gemini, Google’s assistant proper.
- Perplexity, which leans hard on citations and shows its sources, making it the most useful engine for working out which pages shaped an answer.
- Claude, common in professional and B2B settings.
The same prompt can produce different names on each, because they draw on different sources and weigh them differently. Checking one engine and calling it done is like surveying one customer and calling it research.
Two practical notes. Sign out, or use a private window, so your own history does not flatter the result. And run each prompt more than once: assistants are not deterministic, so a name that appears in one run can vanish in the next, and that variation is itself worth recording.
What to record every time
A check you cannot compare to last month is a feeling, not a measurement. Record six things for every prompt you run.
1. The prompt, word for word, exactly as you typed it.
2. The engine, named, and the date you ran it.
3. The full answer, copied verbatim, not summarised. The wording is the evidence.
4. Whether you were named at all, and how high up.
5. Who else was named, in what order, and what the assistant said about them.
6. The sources cited, where the engine shows them.
Keep the runs where you were absent. A missing mention is not an embarrassment to be deleted, it is the most useful row in the sheet, because it tells you which question you have not answered anywhere the engines can read.
A worked example
Take a fictional but ordinary case: an independent accountancy practice in Leeds, three partners, strong local reputation, a website that reads well to humans.
Branded prompt: every engine describes the firm roughly correctly, though two still list a service line the firm stopped offering last year.
Category prompt, “best accountants in Leeds for small businesses”: the firm appears in one engine out of six. The other five name two national chains, a directory listing and a competitor down the road whose site answers that exact question on a dedicated page.
Trust prompt: answers hedge, because the firm’s opening hours and address differ between its website, its Google Business Profile and two directories.
Nothing here is a scandal, and none of it means the firm is bad at accountancy. It means the question buyers actually ask is being answered by somebody else, and the inconsistencies give the assistants a reason to be vague. Both are fixable, and you only know about them because the check was run properly and written down.
Common mistakes when you check
- Asking only about your own name. It confirms the engine can read your website. It tells you nothing about whether you win the question a buyer actually asks.
- Checking once. Answers move. One run is a snapshot, three months of runs is evidence.
- Saving a screenshot instead of the text. You cannot search, compare or quote a screenshot next quarter.
- Checking one engine. They read different sources and answer differently, so one engine is a sample of one.
- Reading one bad answer as a verdict on the business. It means one prompt, one engine, one moment did not surface you. That is a content and corroboration problem, not a character reference.
- Changing the prompts every time. If the questions move, the results are not comparable. Fix the list, then leave it alone.
Turning findings into action
A check is only worth running if something changes afterwards. Three things move the needle, in this order.
Fix the facts about you, everywhere. Assistants assemble an answer from whatever they can find, and where sources disagree they hedge or guess. Your own site, your Google Business Profile, your directory listings and your social profiles should agree on name, address, services, hours and who you serve. This is unglamorous and it is the highest-yield work most businesses can do.
Answer the questions that came back empty. The prompts where nobody was named, or where a competitor was, are a content brief written by your buyers. Write the page that answers that exact question plainly, with specifics an assistant can lift: who it is for, what it costs, what is included, what happens next.
One myth to retire while you are in there. Google’s own guidance for AI features says: “You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search”, and “Structured data isn’t required for generative AI search”. It is equally blunt about guarantees: “Indexing and serving aren’t guaranteed.” Nobody can sell you a placement.
Earn the corroboration. Assistants trust what other credible sources say about you more than what you say about yourself. Reviews that describe specifics, coverage in trade press and local media, accurate listings in the directories your sector actually uses. If you want the longer version, see why your business is not showing up in AI search.
Then run the same check again next month. The point is not the score on any given day, it is the direction over a quarter.
From a manual check to a repeatable process
Doing this by hand works, and I would rather you did it by hand than not at all. It stops working at about the point you have twenty questions, six engines and a month-on-month comparison to keep, which is roughly two hours of copy and paste every time, and the temptation to skip the runs where you were not mentioned.
That is the job Aurora does: the same questions, the same engines, on a schedule, with every answer saved word for word and dated, including the ones where you do not appear. You can start for free and run your first check today.
Whichever way you do it, the discipline is the same. Fixed questions. Named engines. Verbatim answers. Dates on everything. Gaps kept alongside the wins.
What is an AI visibility check?
It is asking AI assistants the questions your customers actually ask, in a fixed set, on named engines, and recording exactly what comes back: whether you were named, how you were described, who else appeared, and which sources the answer leaned on.
How often should I check what AI search engines say about my business?
Monthly is enough for most businesses. The value is in the trend, not in any single run, so keep the questions and the engines identical each time and compare month to month.
Does one bad AI answer mean something is wrong with my business?
No. It means one prompt, on one engine, on one day, did not surface you. Answers vary between runs, users and sessions by design, so treat any single result as one data point from one session.
Which AI engines should I check?
Check six: ChatGPT, Google AI Overviews, Google AI Mode, Gemini, Perplexity and Claude. They draw on different sources and answer differently, so checking one engine is a sample of one.
Do I need special markup or an llms.txt file to appear in AI answers?
Not for Google. Its guidance says you do not need new machine readable files, AI text files or Markdown to appear in Google Search, and that structured data is not required for generative AI search. It also says indexing and serving are never guaranteed.
What should I record for each prompt?
The prompt word for word, the engine, the date, the full verbatim answer, whether you were named and how high up, who else was named, and the sources cited. Keep the runs where you did not appear; those are the most useful rows.
AI Visibility Expert and Founder of The GEO Agency, helping travel, hospitality and property companies win visibility in AI search. 28 years of digital marketing, brand strategy and enterprise digital transformation experience with brands including Virgin Media, O2, Samsung, LEGO, RBS/NatWest and ITN. Author of the ChatGPT Blueprint, Anthropic Claude Partner Network member and Claude Certified Architect.
