Who AI Actually Recommends: Inside a 65,000-Answer Benchmark of Commercial PromptsArticle · digest for November 17
Infographic blueprint comparing AI search outputs with branching paths for live search indexes and isolated models.
Article5 sections≈ 8 min

Who AI Actually Recommends: Inside a 65,000-Answer Benchmark of Commercial Prompts

Between July and November 2026, analysts at GUSAROV logged and scored 65,827 commercial answers across nine AI interfaces covering 1,271 projects. When the researchers compared 11,697 direct overlaps—the exact same buyer prompt ("who do you recommend in my city") fed simultaneously to at least three models—a sobering fact jumped out: every AI agreed on the same company in just 1.8% of queries. In 69.8% of cases, models named no brand at all, and in 28.4%, their shortlists diverged.

That number snapped me out of any complacency. When a business owner checks whether AI recommends their company, they usually open one familiar chatbot, type a couple of prompts, and draw sweeping conclusions. So why does one model put a business in first place while the model next door acts like you don't exist—and what rules actually govern who gets recommended?

The "Single AI" Illusion: Why One Quick Check in ChatGPT Lies to Business Owners

A comforting myth floats around the market that "AI visibility" is a single score: if ChatGPT praises your business, you've supposedly locked down generative search. GUSAROV's slice of 799 projects (each tested with at least five answers across four or more models) proves otherwise. All tested interfaces named the brand simultaneously for only 8.1% of projects. For 44.9% of companies, only a subset of models recommended them (nearly half the time, just one model out of five), and 46.9% of projects weren't named by a single AI. In an anonymized auto-repair case on the same prompt set, Perplexity hit 80% visibility while DeepSeek scored a flat 0%. Checking your company in ChatGPT alone tests ChatGPT—not your market presence.

The spread comes down to where a model pulls its facts the second a user asks. All nine interfaces split into two camps:

  • Models answering from old memory (frozen training weights). Classic example: Claude without web search, which named a company in only 5.0% of commercial answers, and the lightweight ChatGPT model at 6.2%. Without live web access, an algorithm only recalls giants with years of footprint in its training data.
  • Systems wired to a live search index (RAG—Retrieval-Augmented Generation). The moment a prompt lands, the system queries a live index, scans fresh web pages, maps, and directories, and assembles an answer on the fly. In a 10-project control sample, toggling web search on in Claude lifted brand mentions from 6.2% to 19.6% and website link share from zero to 15.9%.

Commercial Citation Leaders: How Perplexity and Google AI Overviews Pick Winners

In Andrey Gusarov's ranking by how often a brand is named directly in the answer body, search-native engines topped the board: Perplexity (20.1%) and Google AI Overviews (19.2%) name a business in roughly one in five commercial answers—four times more often than Claude without search (5.0%). Mid-table sat Search with Alice (11.8%), Gemini (11.6%), Claude with search (10.0%), DeepSeek (9.6%), and ChatGPT with search (8.4%).

You have to separate being recommended by name in the text from appearing as a source link. In Google AI Overviews, a company's name appears in 19.2% of answers, while a link to its domain shows up in the source cards underneath 26.1% of the time (at Perplexity, 24.7%). A citation link gives you a shot at a click, but it doesn't mean the model actually endorsed your business in prose.

Brands named by ChatGPT are already recommended by Google AI Overviews 70.1% of the time, whereas overlap in the reverse direction is only 20.4%. Under Google Search Central guidelines, AI Overviews require no secret hack: a page simply needs to be indexed and eligible to show with a snippet. We've already looked at how a technical AI audit of metadata and page structure stops you from losing those mentions.

Market size plays a role too. In Russia, AI models name brands 1.7× less often than in Belarus: 7.5% on average versus 12.6%. The reason is strict shelf space: an AI shortlist averages 6.2 companies (median 5 slots), and far more contenders fight for those slots in the larger market. On the bright side, once a chat model does name your brand, it lands in #1 position 54.2% of the time, and negative sentiment makes up just 0.8% versus 62.8% positive framing.

The Yandex Window of Opportunity: Why Alice Still Stays Quiet on 60.9% of Commercial Queries

Yandex's Search with Alice behaves differently: it names a brand in the answer text 11.8% of the time, yet combining text mentions with site URLs and direct links pushes total visibility to 26.3% (links appear in 23.2% of answers). Alice is stingier with verbal praise, but eager to act as a navigator and drop your URL straight in front of the user.

The biggest takeaway for local markets, though, is the share of results with no AI block at all. Across 60.9% of commercial queries (and 66.8% for Russian projects), Alice's generative block doesn't even trigger yet, whereas Google AI Overviews leaves only 19.4% of commercial queries without an AI header.

On two-thirds of commercial searches in Yandex, buyers still pick vendors from classic organic results, Yandex Maps, and directories. While that window is open, traditional SEO and geo-listings still carry full weight—and when Yandex rolls AI answers out to the remaining demand, Alice will pull its recommendations from the pages and map profiles ranking at the top today.

Industry Filters: Why Healthcare Hits Guardrails While Ingredients Beat Luxury in Beauty

Across industries in GUSAROV's benchmark, AI most frequently names IT companies (15.8%), online stores (14.8%), and construction firms (14.6%), followed by education (12.2%) and marketing agencies (10.9%). Healthcare gets named least often—just 7.2% by name. That's YMYL ("Your Money or Your Life") safety filtering in action: models avoid endorsing specific private clinics and fall back on general medical protocols.

Where guardrails don't block recommendations, AI is scrambling old brand hierarchies. A 5W PR study of more than 80 commercial beauty prompts found that AI recommends the active ingredient first and the brand second. Shoppers ask for "retinol for beginners" or "vitamin C serum," and the model picks clear formulations and clinical backing:

  • Four brands built around transparent "pharmacy-grade" positioning captured 22% of all skincare mentions, beating players with massive ad budgets.
  • Luxury giants Chanel, Estée Lauder, and Lancôme failed to crack the top 10 even once on ingredient-specific queries.
  • Brands speaking the language of skin concerns and stocked across multiple retailers win: Charlotte Tilbury made top 5 in both skincare and makeup, Rare Beauty took #2 in makeup, and carrying stock in both Sephora and Ulta delivers 1.2× more mentions. On how an external footprint shields your brand from AI distortions, see our piece on managing company reputation in the AI era.

A Monthly Measurement Protocol: Tracking AI Visibility Yourself

Start with Google AI Overviews, Perplexity, and Search with Alice—they generate three times as many commercial touchpoints and read the live web index. Track numbers over time: in an e-commerce case from GUSAROV's database, raw visibility held at 28–30%, while position- and link-weighted Share of Voice (AI SOV) swung between 14% and 21%. You don't need pricey software to keep tabs on this—run a monthly manual protocol.

Monthly Checklist for Measuring Your Business Visibility in AI

  • Build a core list of 20–30 non-branded commercial prompts ("recommend a service company in city").
  • Run the list across search-linked AI blocks (AI Overviews, Alice, Perplexity) and one chat model (ChatGPT or Claude).
  • Log direct brand mentions in the text separately from website links in source cards.
  • Note the share of queries with no AI block, where standard SEO and maps can still capture traffic.
  • Write down 3–5 competitors appearing in answers and 5–10 third-party source sites the AI pulls from.
  • Check site indexing and repeat the audit monthly, queuing 2–3 updates to your pages and directory profiles.

Drop those numbers into a spreadsheet and your exact situation jumps off the page: total invisibility, over-reliance on a single model, footnotes without a real endorsement, or stuck at the bottom of the shortlist.

Watching classic organic space shrink in Google makes the next 6–12 months easy to read. That 60.9% AI-free window in Yandex will start closing, and chatbots will keep tightening their links to live search indexes. The winners in this race won't be people chasing "secret prompts"—they'll be the businesses translating service pages into crisp specs, keeping map profiles spotless, and earning spots in independent rankings today so algorithms recommend them first tomorrow.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 3