What 20 AI Visibility Audits Revealed About How AI Recommends Businesses

A company can rank well in search, have a clear website and be accurately described by AI systems, yet still be absent when a potential customer asks for a recommendation.

That’s one of the clearest patterns in our AI visibility audits. Between June and September 2026, we audited 20 businesses across software, professional services, finance and hardware. We tested buyer questions across five AI engines and eight markets, then reviewed the answers, citations and sources behind them.

Across 1,474 scored answers, the audited company received a linked citation 182 times: an overall citation rate of 12%. Five of the 20 companies weren’t cited once. Yet every engine, in every audit, returned a written recommendation naming somebody.

The results show why AI visibility can’t be measured simply by checking whether a company gets mentioned in a chatbot response. The more useful questions are: does AI recommend the business, does it cite evidence for that recommendation, and does it do so for the questions buyers actually ask?

How we measured AI visibility

For each audit, we tested unbranded buyer questions: queries someone might ask when they have a problem but don’t yet know which company sells the solution. Examples included questions such as “Who is the best broker for this?” and “How do I automate that?”

A standard audit tested 60 answers across five engines:

• Google AI Mode: 24 questions
• ChatGPT: 9 questions
• Claude: 9 questions
• Gemini: 9 questions
• Perplexity: 9 questions

Some multi-market audits used larger question sets, ranging from 68 to 168 answers.

We counted an answer as a citation only when it included a link to the audited company. If the answer named the company but didn’t link to it, we counted it as absent. That makes the overall 12% rate a measure of linked citations, rather than every unlinked mention.

We also ran two controls: whether an engine could find the company when asked about it by name, and what it already appeared to know about the company.

The 20-audit results

The citation rate varied considerably by company and sector. The table shows the linked citations recorded in each audit.

Half of the audited companies scored 3.3% or less, and five weren’t cited at all. Four scored above 25%, but each still had at least one question on which it received no citation.

These figures aren’t a market-wide benchmark. We selected the 20 companies ourselves, and the audit results are a snapshot of defined questions and configurations. They do, however, show the range of outcomes, and help reveal the different issues that can sit behind a low score.

1. AI can know a company without recommending it

In all 20 audits, brand-specific searches returned the company. When asked directly, the engines often described it accurately, including details such as its founding year, regulation, executives and published customer results.

But that knowledge didn’t reliably carry over to unbranded questions about the company’s category or the problems it solves.

In other words, an AI system might know what a business does but fail to connect it to the answer when a buyer asks for help. That’s a different problem from not being known at all, and the fixes may be different too.

A useful audit should test both situations: what AI says when asked about a company by name, and who it recommends when a buyer describes a need without naming a provider.

2. Strong search rankings don’t guarantee AI citations

Several audited companies ranked well in conventional search for relevant terms, but weren’t cited in corresponding AI answers.

One regional accountancy firm ranked eighth nationally for a term with 5,400 monthly searches. It wasn’t cited in any of the four AI answers tested on that topic. A consumer finance business ranked ninth for a furniture-finance term but didn’t appear in an AI answer about buying a sofa on finance. A penetration testing platform ranked second nationally for its main category term, yet half its US citations came from a single blog post.

Those rankings were real, but they didn’t show whether AI engines were using the company’s pages as evidence in answers. A ranking report can’t tell you which competitor an AI engine recommended or which sources it used to support that recommendation.

3. Specific, answer-led pages earned citations

The audits found examples where clear, specific pages attracted citations across multiple engines.

One page on a regional accountancy firm’s website was cited by all five engines. It named an industry body, referred to a real client and stated a specialism. One page on an infrastructure software company’s site was cited repeatedly across two markets. It was around 700 words long and clearly explained the application and its constraints.

The reverse showed up too. Two blog posts on a penetration testing platform’s website received seven of the domain’s nine blog citations. Nine comparison pages, built for the questions being asked, received none. The blog posts had full article and question markup; the comparison pages had less heading structure and thinner markup.

On a mortgage broker’s site, the homepage had 23 headings, three written as questions, and received ten citations. A page for foreign nationals had six headings, none phrased as a question, and received no citations across 12 chances.

These examples don’t prove that a particular heading count or page length will produce citations. They do show that content needs to make its subject, audience and answer clear, and that intended purpose alone doesn’t guarantee an AI engine will select a page.

4. AI answers draw on sources beyond a company’s website

A company’s own domain was often only one part of the source material behind a cited answer. Other sources included review platforms, directories, roundups, comparison pages, trade publications and video.

In one audit, YouTube accounted for 45% of all English-language citations, ahead of every vendor and review site. Video was the largest single source in four of the 20 audits, and none of those four companies had a presence there.

The report also found gaps on other third-party sources:

• On a payments audit, none of 12 roundups reviewed included the company.
• On a consumer finance audit, none of 12 third-party guides reviewed included it.
• On an AI research software audit, none of 21 category pages reviewed included the company, while a direct rival had published a “top ten alternatives” list and was cited for it.
• On a consumer insights audit, none of ten listicles reviewed included the company; two had been written by competitors that ranked themselves in their own lists.

Review profiles were also a recurring issue. Two companies that performed well on Google’s answers but poorly on standalone assistants didn’t have listings on G2, Capterra or TrustRadius. One company had three review profiles live but empty, including one that was still unclaimed. Across the 20 audits, only one company had a working review presence, and the engines cited it in all four of its markets.

The sources that matter will vary by sector. The point is to find out which sources AI is already drawing on, then check whether they describe the business accurately and include relevant evidence.

5. Missing company facts leave room for contradictions

In nine of the 20 audits, the engines returned contradictory facts about the company.

In one audit, four engines gave four different founding years, and one suggested the company might be a rebrand of a competitor. Another produced three founding years, three funding totals and four headcounts, and surfaced an old corporate identity the company had abandoned. For one regional accountancy firm, engines returned staff figures of 150 and 220 and listed five offices, while the firm’s own careers page showed a higher staff number. Another engine named the wrong professional network. A UK lender was credited with a product sold by a similarly named US business.

In these examples, the companies’ About pages didn’t provide much checkable information about their founding date, size, structure or regulatory position. When a company doesn’t clearly state those facts, engines may find them elsewhere, and the sources may conflict.

Structured data was another gap: in 19 of the 20 audits, the site either lacked the structured data the engines could read or used it in a form that parsers couldn’t use. The single company with usable structured data was absent for other reasons.

Clear, consistent facts on a company’s site and relevant profiles give AI systems stronger material to work with. Structured data can help machines interpret those facts, although it doesn’t guarantee a recommendation.

6. Overall scores can hide gaps in valuable questions

A high aggregate citation rate doesn’t mean a company appears everywhere buyers need it.

The logistics hardware company had the highest rate in the audit set at 48%, but it was absent from all five “best system” questions and both pricing questions. A workforce software company led the vendor field in Google’s answers but received no citations across four questions about its own published outcomes. A penetration testing platform scored 35% overall but received no citations across eight UK questions.

In a set of 96 Google AI Mode answers for a mortgage broker, the broker was cited in 13 of 20 questions that named the product category, and six of the other 76 questions. A regional accountancy firm received three citations across 24 Google answers; two came when the question named a town where the firm had an office.

The pattern is a reminder to examine performance by question and buying stage, not just by overall rate. A company may appear for broad category questions yet be absent from pricing, product comparison or outcome-related questions.

7. A local or international result may look very different

Visibility changed across markets, sometimes on the same questions.

One company’s citation rate in the UAE was more than twice its rate in the UK. A penetration testing platform had six times as many ranking keywords in the US as in the UK, but no UK service page. A consumer insights business scored zero in Spain; the answers went to two companies, neither of them Spanish.

The separate UK professional services screen showed a similar local dimension. Among uncited firms, the median result was that a local peer received 14 of the 18 opportunities tested.

Market performance can depend on local directories, comparison sites, competitors and service pages. A business that is visible in one country or town shouldn’t assume its AI visibility carries over elsewhere.

8. Technical issues can make useful content inaccessible

Some visibility problems weren’t caused by content or reputation. Technical issues stopped crawlers from reaching or interpreting pages.

One company’s CDN returned a 403 response to eight AI crawlers, even though the site served Google, Bing and ordinary browsers normally. Its robots.txt file allowed access. Two of the blocked crawlers were live fetchers that open pages while a buyer is asking about a company.

Other examples included:

• A website relaunch that left 115 of 117 previously ranking URLs returning 404 errors, with no redirects. The report recorded 687 AI answers that had cited those URLs.
• A robots.txt file that blocked every PDF the company published, while an engine cited a competitor’s PDF report in the same answer.
• A machine-readable file written to explain what the company sells that pointed to a server which no longer resolved.

These defects were invisible during normal use of the websites. Checking crawler access, redirects, PDFs, structured data and machine-readable files can reveal problems that a standard browser review misses.

A separate screen of 231 UK professional services firms

Alongside the 20 full audits, we screened 231 UK professional services firms on 3 August 2026. Each firm was tested against 18 buyer questions in its own town, using Google AI Mode and Perplexity.

The results were:

• 180 of 231 firms (78%) weren’t cited once.
• The firms’ own websites accounted for 1.35% of the citations generated by their buyer questions.
• Questions naming a specialism produced citations 2.5 times as often as generic category questions: 12.8% versus 5.1%.
• 24 firms appeared for a specialism question but not for the obvious category question.

This was a shorter screen with a different method from the 20 full audits. It didn’t include ChatGPT, which was consistently the weakest of the five engines in the full audits. The 78% figure is therefore more likely to flatter than understate the result. It shouldn’t be combined with the full-audit figures or treated as a representative market average.

Still, it highlights a practical opportunity: a firm may be absent from a broad “best accountant in my town” answer and appear when a buyer asks about a more specific need.

The competitor gap can be large, and sometimes reachable

The number of citations received by competing businesses varied from modest gaps to very large ones.

In one audit, an AI research software company received five cited answers, while a direct rival received 357. The largest incumbent in the category received 62,950. Elsewhere, a regional accountancy firm was losing to a competitor that held ten of the 305 references across the full run, supported by a handful of well-built pages.

These figures show why identifying the competing sources matters. A business may not need to displace the largest company in its category to improve visibility. In some cases, a smaller competitor is capturing relevant answers with a limited number of useful pages.

What the findings mean for businesses

The audits produced 11 recurring diagnoses, including “ranks but is not quoted”, “described but not linked”, “answered from the registry”, “strong overall, zero in one place” and “the crawler is locked out”. The same score can therefore point to different causes, and different priorities.

A practical review should:

1. Test brand and buyer questions separately. Check what AI says when given the company name, then test the questions buyers ask without naming a provider.
2. Record citations and misses by question. Note who was recommended instead, which page or source was cited, and whether the company was named without a link.
3. Review third-party coverage. Check the review sites, directories, roundups, comparison pages and video sources relevant to the category.
4. Make company facts easy to verify. Clarify details such as founding date, structure, scale, regulation, named people and customer evidence.
5. Check technical access. Review crawler responses, redirects, PDFs, structured data and machine-readable files.
6. Prioritise questions with a real buyer opportunity. A high overall score can still hide a complete absence from pricing, product or outcome questions.

The report also found that proof businesses already own is often difficult for AI systems to read. Published results, named clients, accreditations and commendations may sit in press releases, gated PDFs, testimonial carousels or internal project notes. Making existing evidence easier to access and interpret can be a different task from having to create new evidence.

A baseline, not a permanent score

AI answers vary between runs, so a single audit is a snapshot. The report recommends repeating the same question set after 90 days and treating the movement between runs as the number to trust.

One audited company was remeasured without acting on the findings and returned the same score. The report’s examples of improvement came from separate client programmes, not from the 20 audits:

• A healthcare compliance platform’s pipeline increased from £240,000 to £645,000 over six months; its largest contract in that period came through ChatGPT.
• A digital verification provider’s decision-stage AI answers increased from 2 of 8 to 15 of 24 across four assistants in one month. The programme also recorded a seven-figure contract attributed directly to the AI search work.

Those results are separate from the audit sample and shouldn’t be read as a guaranteed outcome. They do show why measuring changes over time can be more useful than treating one citation rate as a verdict.

The category answer is still being written

Across the audits, three businesses competed in categories where no company was cited at all. In other cases, one rival was starting to lead, but its margin was narrow enough for another business to challenge.

That means AI visibility work isn’t always about catching a dominant incumbent. Sometimes the question is unclaimed. Sometimes the business is known but not connected to buyer intent. Sometimes the answer is being built from directories, third-party guides or a competitor’s page.

The first step is finding out which situation applies.

If you’d like to see which situation applies to your business, you can book a free AI visibility audit.

Read the full report: AI answer visibility: what twenty audits found

The report includes the full audit methodology, the 20 company-level citation rates, the separate 231-firm screen, 11 diagnoses and further examples from the audits.