
A 27-observation Ontario pilot found that Google AI Mode, Gemini and Perplexity often produced different shortlists for equivalent local business recommendation prompts. Of seven fully comparable tests, only one contained any business common to all three platforms.
Why one AI recommendation is not a ranking
AI search answers often sound final. A customer asks for a nearby clinic, law firm or contractor and receives a short list with reasons, ratings and source links. The presentation can make the answer feel like a settled ranking, even though another AI platform may produce a substantially different list from almost the same request.
A single check can confirm that a company appeared at one moment, in one product and under one set of conditions. It cannot show that the company is consistently visible across the systems customers use. Treating one result as a general ranking can mislead the business and the people making decisions from its reports.
How the Ontario pilot was run
On September 16, 2026, I used the same recommendation prompt structure in Google AI Mode, Gemini and Perplexity. Each test used a fresh request and told the system not to rely on earlier chat context.
The pilot covered physiotherapy clinics, family law firms and HVAC contractors in Vaughan, Toronto and Mississauga. This created 27 planned observations across three platforms, three service categories and three cities. Twenty-five answers were returned, or 92.6 percent of the planned set. Google AI Mode completed seven of its nine tests, while Gemini and Perplexity completed all nine.
Each prompt was run once, and location permission was not shared. Account state also differed by product: Perplexity was signed out, Gemini was mainly tested through its public signed-out interface, and Google AI Mode alternated between signed-out and an existing Google session. These conditions make the work a dated pilot rather than a controlled market study.
Recommend up to five [category] in [city], Ontario. Explain why and cite sources. Do not use prior chat context.
What the platforms returned
Across its seven completed answers, Google AI Mode returned 35 businesses in 35 recommendation slots. Gemini returned 44 businesses in 44 slots across nine answers. Perplexity returned 43 businesses in 45 slots; two slots contained directories rather than individual businesses.
The two non-business results matter because the prompt asked for local businesses, not lists of businesses. They show that candidate validity should be measured alongside brand mentions and source citations.
The shortlists rarely agreed
Seven category-and-city combinations produced usable answers from all three platforms. Only one of those seven included any business that appeared in all three answers. That is 14.3 percent of the fully comparable tests.
The shared case was physiotherapy in Vaughan. My Physio Sports and Rehab Centre and Vaughan Physiotherapy Clinic appeared in the Google AI Mode, Gemini and Perplexity answers. None of the other fully completed comparisons contained a business common to all three platforms. Pairwise overlap was also limited, with most platform pairs sharing no more than one or two businesses.
Low agreement does not tell us that one platform was right and another was wrong. The answers may reflect different indexes, retrieval methods, source preferences, product modes and freshness. The useful finding is that local visibility was fragmented: a company could appear prominent in one answer while being absent from another.
Different source patterns shaped the answers
The supporting sources also differed. Google AI Mode and Gemini frequently surfaced local business facts such as addresses, categories, ratings, review counts and information from official websites. Clear service pages and accurate local profiles were common elements around the businesses they recommended.
Perplexity showed a stronger pattern of relying on third-party material. Every Perplexity answer in the pilot cited at least one directory, ranking, comparison, roundup or public listing. Two recommendation slots returned a directory rather than an individual business.
These observations do not prove that a specific directory or website caused a recommendation. They do show why a business cannot rely only on its own website. AI systems can use independent pages to confirm what the business does, where it operates and whether other sources recognize it in that context.
AI search visibility affects the customer journey
A customer may use more than one AI product while researching a local service. Each interaction can create a different first impression. If a business is represented accurately in one answer but omitted or described incorrectly in another, the customer journey changes before the person reaches the company website.
This makes AI visibility a measurement issue as much as a content issue. Teams need to know where the business appears, which facts are shown, which pages support the answer and where the gaps occur. Without that baseline, it is easy to publish more content without knowing whether the content addresses the actual problem.
A four step AI search visibility audit
- Build a fixed prompt panel from the questions customers ask when choosing the service. Keep the wording stable enough to compare results over time.
- Standardize the test conditions. Record the date, product, mode, account state, location setting and exact prompt, and use fresh sessions where possible.
- Record the complete answer. Capture whether the business appeared, whether its facts were correct, which pages were cited, whether every recommendation was a valid business and which competitors were named.
- Retest on a schedule and after meaningful website, profile or media changes. Judge patterns across several checks instead of treating one favourable screenshot as proof of progress.
How businesses can improve AI search visibility
First-party information should be consistent. The business name, category, address, phone number, service area and main services should match across the website and major public profiles. Service and location pages should explain the offer in plain language and identify who provides it.
Independent confirmation also matters. Relevant professional associations, credible directories, genuine customer reviews, local reporting, original research and expert quotations can connect the company with a topic or location. Mass directory submissions and paid links on sites created only to sell placements provide little useful evidence for a reader or an AI system.
The aim is accurate, verifiable coverage. When the company website and independent sources describe the same entity consistently, an AI system has clearer evidence to work with. The monitoring panel can then show whether those changes are followed by broader and more accurate visibility.
No business can guarantee inclusion in an AI answer. The practical goal is to make the company easier to understand and verify, then measure whether its visibility and factual accuracy improve across several systems and dates.
Limits and responsible reporting
This was a small pilot, and AI answers can change with time, model updates, location, account state and the sources available during retrieval. The results should not be used to rank the tested businesses or estimate platform market share. They also cannot establish that one source caused a company to appear.
The result supports a narrower conclusion. Businesses should evaluate AI visibility across several systems and dates. A repeatable record of recommendations and sources provides a stronger basis for decisions than a single screenshot, especially when those results are reported to clients or leadership.



