Who AI Actually Cites: A Study of 360 AI Answers on Small Business MarketingArticle · digest for November 12
Infographic blueprint of generative search architecture with a diagram showing citation distribution between platforms and agencies.
Article4 sections≈ 8 min

Who AI Actually Cites: A Study of 360 AI Answers on Small Business Marketing

When you ask a neural network which advertising channel to choose for your business, in 49 out of 66 answers with links (74% of cases!) the AI relies on articles written by the very companies selling you that channel. In real life, we switch on a mental filter automatically: if a store clerk praises a coffee machine from their own shelf, you discount for their bias. A chatbot, however, speaks in the tone of an independent consultant and never warns you whose blog it is paraphrasing. At the same time, there is no monopoly in the results: out of 502 unique domains cited in a new study on small business promotion, 232 sites appeared in the output exactly once.

Those two numbers caught my attention right away because they change how you look at promotion in generative search. As a business owner, you need to know who stands behind the advice in the chat and where blind trust in an algorithm can cost you your budget. And if you run a company blog or work at an agency, you need to see whose articles AI picks as sources and how to get into those recommendations without a massive newsroom.

Eight Niches and 360 Answers: How AI Advisors for Small Businesses Were Tested

On November 10, 2026, Daria Zavyalova, Head of Marketing at TexTerra, and her team tested eight types of small businesses: a coffee shop, a beauty salon, an online clothing store, an auto repair shop, an online school, an apartment renovation crew, a law or accounting firm, and a small B2B manufacturer.

For each niche, they prepared five questions: where to start promotion, which channel to choose, how much to budget for ads, what to know about the law, and how to tell if advertising is paying off. All 40 questions were asked to Alisa (Moscow), ChatGPT, and Google AI (Russia) three times each in fresh chats. The 360 answers contained 1,850 links to 502 domains.

The sources were tagged by author and commercial interest. Company and user posts on vc.ru, Habr, RBC Companies, Klerk, and Sostav blogs were classified as user-generated and corporate content, while the media category included only editorial outlets and bank brand media (SberBusiness Live, T-Business Secrets).

The good news: all three systems tailor promotion channels to the specifics of the business. In repeated answers for the same niche, about 61% of the named channels overlap, compared to only 26% across different niches. A coffee shop is advised to use Maps, VKontakte, and neighborhood communities while holding off on Direct; a renovation crew is steered toward Avito and Profi; and a manufacturer is pointed to trade shows and industry platforms.

Optimizing content for AI answers is called GEO (Generative Engine Optimization) or AEO (Answer Engine Optimization). Put simply, you used to fight for a click on a blue link, whereas now you package facts so a neural network uses your page as the backbone of its answer. We previously covered the shift from classic SEO to GEO optimization, and TexTerra's breakdown showed how the 1,850 links are split:

  • 34% (630 links) point to blogs of ad platforms and business services;
  • 22% (407 links) go to blogs of agencies and independent specialists;
  • 14% are taken by user and corporate posts on open publishing platforms;
  • 11% go to editorial media, and 8% to legal databases and government websites.

Every major service earns more citations than the most visible agency: Yandex Ads and Yandex Direct help centers got 37 and 33 mentions, Bitrix24 got 34, eLama got 33, and Kontur got 29, while the top-performing agency received 12 citations. Platforms have the budgets and a direct incentive to spend years writing guides for small businesses. Meanwhile, the 407 agency links were spread across 191 domains (138 of them got only one or two citations). In 141 of the 293 answers with links, at least one source sells the very tool being recommended (62 of 98 for Alisa, 58 of 112 for Google, and 21 of 83 for ChatGPT; out of 268 such links, 192 lead to services and 76 to agencies).

So where is the entry point for smaller teams? 56% of all agency citations go to narrow-niche articles whose URL and topic explicitly name the industry ("auto repair shop promotion," "beauty salon advertising," "where to find apartment renovation orders"), whereas such pages make up only 39% of the overall sample.

Look at the running example from the report. An auto shop owner asks: "Should an auto repair shop list on Avito or invest in Yandex Direct?" The AI systems recommend both channels and cite the Avito blog and Direct help docs, but right alongside the giants they cite a small studio with a detailed article on promoting an auto repair shop on Avito. Trying to outcompete a platform with a generic "how to promote a business" guide is pointless, whereas specific industry details and case studies put practicing specialists into AI answers. The second springboard is open platforms: vc.ru became the most cited domain in the study (82 links), Sostav blogs delivered 45 citations, Dzen brought 23, and RBC Companies brought 13.

Three Different Internets: Why Alisa, ChatGPT, and Google Barely Overlap in Sources

Asked the exact same question, Alisa, ChatGPT, and Google AI return almost completely non-overlapping lists of websites (only 2–9% shared domains): Alisa and Google share 9% of domains, Alisa and ChatGPT share 4%, and Google and ChatGPT share 2%. Across the entire study, only 12 websites were cited by all three systems (including vc.ru, ConsultantPlus, eLama, Direct help docs, and two bank media outlets).

Each model has its own personality: Alisa leans on service blogs (43% of its links) and agencies, giving Yandex's own sites 8%. Google AI favors open platforms with company and user posts (23% of links, including 23 citations from Instagram*, owned by Meta, which is designated extremist and banned in the Russian Federation). ChatGPT cites narrowly: its top 10 domains account for 58% of its links (compared to a quarter for Alisa and Google), nearly 40% go to ConsultantPlus, Garant, and the FAS website, and 139 of 194 links appear only as a domain badge without a page URL.

Even within the same system, repeating a question in a new chat yields only 30–40% domain overlap. Google's documentation explains that AI Overviews and AI Mode fire off a series of related search subqueries. This was also confirmed by Rand Fishkin's test at SparkToro and Gumshoe.ai across 3,000 runs of 12 prompts: an identical list of recommendations repeated in fewer than 1% of cases, and the exact brand order once in 1,000 runs. That is why you need to apply site optimization rules for AI search alongside regular visibility-share tracking across multiple runs (TexTerra checks up to 20 prompts for a GEO audit).

Conversion Fantasies and Fines Up to 500 Thousand Rubles: Where AI Lets Entrepreneurs Down

Out of 360 answers, 67 (19%) had no links at all, and half of those were questions about evaluating ad ROI: 36 out of 72 answers lacked sources (19 of 24 in ChatGPT, 12 of 24 in Alisa, 5 of 24 in Google). Even so, in 66 of 72 answers (92%), the models confidently pulled numbers from memory—giving ROI and ROMI formulas and calling "2–3% a normal conversion rate," even though a neighborhood coffee shop and an online school have completely different unit economics.

On direct legal questions, the AI systems accurately describe ad labeling, ORD, erid tokens, and the ban on advertising on Instagram* and Facebook* (owned by Meta, designated extremist and banned in the Russian Federation) in effect since September 1, 2025. Yet dangerous slip-ups surface in details and budget questions:

  • Fine risk for a coffee shop: when asked about a coffee shop's budget, Google suggested allocating 20–30% of funds to "Reels / VK Clips / banned social networks," including targeted ads and barter, which exposes a legal entity to a fine of up to 500,000 rubles.
  • Confusion over the 3% fee (Law No. 479-FZ): in two answers, the fee was described as an advertiser payment, even though it is paid by ad distributors, system operators, and intermediaries (the advertiser pays it only when ordering from a foreign entity).
  • Shifted deadlines and blocked platforms: one answer pushed the tax ban on deducting unlabeled advertising expenses to "2026" instead of 2025, and two answers recommended LinkedIn—blocked for a decade—to an online school and a manufacturer.

Treat AI advice like a tip from a well-read friend: it can help you sketch a starting plan for a coffee shop or auto repair shop, but before paying any invoices, check the numbers and laws against your cash register, CRM, and primary sources. For authors and entrepreneurs, these findings also hold an encouraging truth: you don't need to turn into a faceless content factory repeating generic truisms after big corporations. When you honestly and thoroughly break down the real-world practice of your narrow niche, you save time, keep your peace of mind, and attract the exact clients who need your firsthand experience.

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