Ask the machine what it thinks of you
Before you read any further, open a temporary chat in ChatGPT and ask the question your best prospect would ask. Not your brand name, the problem they have.
"Best supplier for X in the UK."
"Alternatives to [your biggest competitor]."
"Which companies can help us solve X?"
Look at what comes back. Who is named? Why are they recommended? Which sources support the answer? And are you there at all?
That five-minute exercise tells you something your website analytics cannot: whether you are making a shortlist you never knew existed.
AI visibility is more than being mentioned
Brand visibility in AI search is really three separate things. You will also see this work called AI SEO or generative engine optimisation, and the three names describe the same job.
- Mentioned: your name appears.
- Recommended: you appear as an option with a reason to choose you.
- Cited: your content is used as evidence in an answer.
Don't confuse them. You can be cited repeatedly as an expert source without ever being recommended as a supplier. You can also appear on buying shortlists while being described using positioning you abandoned two years ago.
So a useful audit doesn't simply ask "are we visible?" It asks where you are visible, what you are visible for, what is being said about you, who appears instead, and what evidence is shaping the answer.
Why this matters now
AI is already part of business-to-business vendor research. Semrush surveyed 643 US professionals in spring 2026 and found that, among those using AI at work, 66% regularly use it to research products, vendors or solutions. It appears throughout the decision: 72% use it during early research, 62% to compare vendors, 48% to narrow a shortlist and 45% to support the final decision.
Most strikingly, 92% said AI had shaped their vendor shortlist, with 45% saying the impact was significant.
Much of that decision forms before sales knows a buying process exists. 6sense's 2025 Buyer Experience Report, covering nearly 4,000 buyers, found that buyers contacted sellers around 61% of the way through the journey, and that 94% of buying groups had already ranked their shortlist before engaging anyone. The supplier ranked first before seller contact won roughly eight out of ten deals.
That is the commercial risk. A buyer can ask AI to understand the category, identify suppliers and compare them. If you don't make that consideration set, there is no website visit, no form fill and no lost opportunity in the CRM. You don't know you lost, because as far as your funnel is concerned the opportunity never existed.
SEO still matters, but it isn't the same thing
One tempting assumption is that if you rank well organically, AI systems will find and cite you. The evidence suggests otherwise.
Ahrefs analysed 15,000 long-tail queries in August 2025, comparing citations from ChatGPT, Gemini, Copilot and Perplexity with conventional results. On average, only around 12% of URLs cited by AI assistants also appeared in Google's top 10 for the same query. Perplexity showed more overlap, at roughly one in three, but across the assistants the relationship was weak. Semrush found higher overlap for Google AI Overviews specifically, and a looser relationship for Google AI Mode.
The conclusion isn't that search engine optimisation no longer matters. It is that Google rank and AI visibility are not interchangeable metrics. A page can perform well in conventional search and fail to become a cited source. An AI system may surface a source that doesn't rank prominently for the exact prompt.
So measure both.
Run the audit yourself before buying a platform
There is now an expanding category of software dedicated to AI visibility. Don't start there. Start by understanding the problem manually.
Build a prompt library around how people buy
Create a fixed set of questions representing the buying journey rather than your keyword list. Five groups are useful: category ("best [category] for [customer type]"), comparison ("alternatives to [competitor]"), problem ("why are our leads disappearing between marketing and sales?"), brand ("what does [company] do?", "who is [company] best suited to?") and trust ("who are the credible experts on [subject]?").
Thirty to fifty prompts is enough to expose meaningful patterns without creating an audit nobody wants to repeat.
Run them under repeatable conditions
AI answers vary. The same question can produce different sources, brands and wording between runs, so don't treat one screenshot as data.
Run important prompts several times. Use temporary or non-personalised sessions where possible. Test the assistants your audience actually uses: Semrush's 2026 research found 71% of AI-using respondents used ChatGPT for product research, 61% Gemini and 45% Microsoft Copilot, against 18% Perplexity and 14% Claude. Your audience may differ, but that is a better basis for prioritisation than testing every model equally.
Record the date, assistant and prompt exactly. You are building a repeatable benchmark, not collecting interesting screenshots.
Measure what could actually change a buying decision
For every response, record whether you were named, whether you were presented as an appropriate choice, roughly where you appeared, who appeared instead, what reason the assistant gave for each recommendation, which sources supported the answer, and whether anything said about you was wrong or outdated.
That final point is often the fastest win. Wrong pricing. An old service. Outdated positioning. The wrong target market. A company description that hasn't kept pace with the business.
Visibility with the wrong message can be worse than absence.
Then look at the sources
This is where the audit becomes a strategy. Don't simply count how often your website was cited. Look at what kinds of sources repeatedly influence the answer: competitors' websites, trade publications, review platforms, analysts, Reddit, comparison sites, independent experts.
The pattern tells you what evidence the system relies on for your category, and that tells you what to do next. If review platforms dominate, another 20 blog posts won't solve your problem. If independent expert commentary dominates, this is a public relations and authority problem. If competitors' detailed product pages dominate, your content may simply not answer the questions precisely enough.
Citation analysis turns "we aren't appearing" into a remediation brief.
What a manual audit can and cannot tell you
A structured manual audit answers the first questions that matter: are we appearing, are we recommended, are we described accurately, who consistently appears instead, which sources keep being cited, and where are the obvious gaps.
What it cannot reliably give you is exhaustive share of voice across hundreds of prompts, markets and answer variations. It won't give you historical data you didn't collect. And running a large prompt library across multiple assistants, markets and time periods manually becomes operationally unrealistic.
There is no credible independent evidence that a manual audit captures a fixed percentage of what a commercial monitoring platform would reveal. Don't pretend there is. The better distinction is simpler: a manual audit diagnoses the problem, a platform industrialises the measurement. Run the first before deciding whether you need the second.
Check whether AI systems can actually access you
Before rewriting your content strategy, check the technical basics. Be precise, because training crawlers and search crawlers are not the same thing.
OpenAI. OAI-SearchBot is the search crawler, and OpenAI says sites that opt out will not be surfaced in ChatGPT search answers. GPTBot crawls material that may be used to train generative models. ChatGPT-User handles certain user-initiated requests and does not determine inclusion in ChatGPT Search. Blocking GPTBot while allowing OAI-SearchBot is a legitimate configuration if you want search visibility without training access.
Anthropic draws the same distinction: ClaudeBot for model development, Claude-SearchBot for search discovery and indexing, Claude-User for retrieving content in response to a request. Blocking Claude-SearchBot may reduce visibility in Claude search results.
Google. Google-Extended is a robots.txt control token rather than a separate crawler. It governs whether crawled content can be used for Gemini training and grounding, and Google states it does not affect inclusion or ranking in Google Search.
So "unblock the AI crawlers" is no longer useful advice. Decide which forms of access you want, configure the appropriate agents deliberately, and re-check the vendor documentation before making changes. These policies move quickly, and the detail above reflects September 2026.
Four things an absence usually tells you
Once technical access is ruled out, most problems fall into four categories.
Nobody independent describes you. If almost everything authoritative about your business comes from your own website, the evidence base is thin. The answer isn't more self-description. It's corroboration.
You aren't using the language of the problem. Your website says "revenue orchestration". Your buyer asks how to stop good leads disappearing between marketing and sales. If your content doesn't connect those two ideas, discoverability gets harder for humans and machines alike.
Your category is unclear. Businesses spanning several disciplines often struggle to appear in category recommendations because there is no consensus about where they belong. That is a positioning problem before it is an AI problem.
The evidence exists but isn't retrievable. Important explanations buried in PDFs, content behind forms, thin pages, answers several clicks deep, or blocked retrieval agents.
Only one of the four is primarily solved by technical search work.
When do you need a platform?
Once the manual audit becomes too large to run reliably, monitoring technology starts making sense. Several products now cover this, and the category is consolidating quickly: tools that began as monitoring dashboards have added content workflows, and tools that began as content platforms have added visibility tracking. Published pricing and engine coverage both change often enough that any comparison needs dating rather than treating as settled.
Choose based on the job you need doing rather than the category label, and check current coverage and pricing directly before committing.
You may not need either. If you have 30 prompts, operate in one market and have just discovered that nobody outside your company credibly describes what you do, software isn't your constraint. You already know what needs fixing.
A platform starts earning its place when your prompt set is too large to run consistently, you need multiple markets or languages, several product lines need separate monitoring, competitive share of voice needs tracking over time, or AI visibility has become a board-level metric rather than an exploratory exercise.
Until then, spend the money fixing the evidence rather than measuring its absence more precisely.
Fix things in the right order
First, correct what is wrong. Inaccurate descriptions, old products, outdated pricing and obsolete positioning are usually easier to fix than absence.
Second, clarify what you do. Can a human, or a machine, understand from one page what you do, who you do it for, what problem you solve and why you are different?
mhance shows why this matters. The company had built a strong reputation delivering Microsoft integration projects for not-for-profit organisations, and its website still reflected that history. But around 40% of its customer base was by then commercial, and the business increasingly operated as an ongoing consultancy across Dynamics 365, cloud and data. Anything relying on the public website would have described the business it used to be. That is a positioning problem with an AI consequence.
Third, build corroboration. Independent evidence matters: original research, expert commentary, earned media, credible reviews, relevant industry coverage, third-party profiles.
Ripple Energy illustrates building that external footprint. As a relatively new entrant, it generated 74 pieces of media coverage during 2023, including ten national titles, with reported reach exceeding 522 million. That isn't proof of AI visibility. It is proof of the kind of independent evidence base from which visibility can be built.
Finally, fix retrievability. Make important information available in HTML. Answer questions directly. Use clear headings and specific entity names. Check crawler configuration. Don't hide your strongest explanations behind a form.
Don't turn the audit into another dashboard
The number that ultimately matters isn't how often your company appears. It is whether an intelligent stranger, reading what the internet currently says about you, could answer three questions:
What do you do? Why should I trust you? Why might I choose you over the alternatives?
If the internet cannot answer those clearly, an AI assistant won't either. And if the internet can answer them but AI still isn't surfacing you, you have a much more specific problem to solve.
Run the audit. Find out which problem you have. Fix that before you buy another dashboard.
More on how we approach AI visibility, or, if the wider issue is that buyers don't know you well enough yet, how we help you win attention and trust.
We offer a free AI visibility audit.
Submit the contact form and we will assess how your brand appears across the AI assistants your buyers actually use, and where the most important visibility gaps sit.
