In 94% of B2B shortlists, AI summaries barely show the companies' own websites
A contractor's site ranks in Google's top 10, but when someone asks ChatGPT "who can do this in my city," it names directories, not that contractor. Across 200 B2B queries, an AI summary appeared in 94 of 100 shortlists, and company websites almost never made the list. The model cross-checks a company by its tax ID in external registries and does not trust self-description. So you need a listing in an external registry that links back to your site.
AI-processed from Habr AI; edited by Hamidun NewsThe problem is that a company's site can sit confidently in the top 10 of regular search, yet a neural network still won't name it. Imagine a customer asking ChatGPT, Perplexity or Google's AI summary "who can do this in my city." Instead of the contractors themselves, the model returns a list drawn from external directories and reference sites.
The author of an analysis on Habr ran 200 real B2B queries, meaning questions about finding a contractor. In 94 of 100 industry shortlists, an AI summary did appear, but corporate sites almost never made it in. The headline's "94% of answers" is better read this way: in 94 of 100 shortlists, the companies' own sites barely made the list.
If you are responsible for a website or marketing, this suggests which company data you need to publish outside your own domain.
How the model vets a contractor
The author then manually examined how Gemini 3 Flash and gpt-5-mini behave with web search (runs on 31 October and 1 November 2026). Both followed a similar pattern:
- first, a search by company name, city and Latin transliteration;
- then the model extracts the 9-digit tax number (UNP) from the page and sends a separate query using only those digits;
- after that, it checks whether the legal entity exists and whether its services match the description on the site.
The logic is simple: to the model, text on a company's own domain remains self-promotion. Claims like "twenty years of experience" and any prices don't confirm themselves. If search results contain no independent confirmation by UNP, the model either discards the site or takes data from old registry files. These often hold outdated phone numbers and vague activity codes such as "wholesale trade of a wide range of goods."
Why llms.txt doesn't help, and bots read JSON and Markdown
On 15 June 2026, Google's Search guidance stated explicitly that llms.txt is used neither in ranking nor in AI overviews. Logs from an open directory of 30,900 listings back this up. Over 30 days, the directory recorded 773,704 requests. Of these, 705,323 were verified by official IPs and reverse DNS, meaning they came from real bots rather than traffic impersonating them.
The breakdown: industry shortlists got 616,242 requests, JSON listings 38,139, and Markdown listings 20,110. The /llms.txt and /llms-full.txt files were opened just 14 times. In the daily log for 16 November, external clients made 7,087 requests to the open REST API, filtering companies by city, type of service and whether they have a website.
By bot: GPTBot made 383,898 requests, PerplexityBot 308,764, ClaudeBot 40,614. Among traditional search engines, YandexBot made 5,534, Applebot 2,587, Googlebot 1,956, bingbot 975. Real-time user-triggered requests are noticeably smaller: ChatGPT-User 82, OAI-SearchBot 48. The author notes that older directories often return 403 or 429 errors to parsers, have no clean Markdown and are overloaded with ads. Clean machine-readable data wins simply because it is easy to collect.
A company's "digital passport" and caveats
In the author's view, optimizing a landing page alone is not enough. You need an external listing in a neutral registry that meets four requirements:
- Tie to a state registry (UNP or INN): the legal entity's status, registration date and region are pulled from official databases and can't be edited by hand.
- Specific services and geography instead of generic OKED codes, plus price benchmarks and answers to frequent questions.
- Delivery without wrappers: one listing available as server-side HTML, Markdown, JSON and Schema.org markup with a taxID field.
- Two-way linking with the website: a claim by the director is not enough; you need a link from the site to the listing or a sameAs property in Organization markup.
The author connects this to the merger of Google AI Overviews and AI Mode (19 May 2026) and the development of the WebMCP specification at the W3C: browser agents increasingly read structured data rather than parse page layout.
There are caveats. The data comes from the author's own directory, so he has a commercial interest, and in places the piece reads like a product presentation. The sample is small (200 queries), and bot traffic to the directory does not yet prove that a model will later cite a specific company.
The practical takeaway is useful regardless of the directory. Check that your UNP or INN is easy to find on your site and matches the data in registries. Make sure independent pages about the company exist with current services and phone numbers. Add Organization markup with sameAs and taxID. And don't count on llms.txt: in this data, almost no one reads it.
- Nov 18, 2026MoEngage Plugs AI Agents Into Its Existing APIs: Glance's Push CTR Up 38%
- Nov 18, 2026Tencent Hy4 flash scores 2.79 out of 4 vs. 2.81 for Haiku 5.5 and costs 18 times less per task
- Nov 18, 2026Five AI Models Tried to Fix the Same Bug: Two Succeeded, Three Broke More Than They Fixed
- Nov 18, 2026Developer gives a web AI chat hands: an MCP server with 14 tools and 21,000 lines of code
Key Questions Answered
How the model vets a contractor?
The author then manually examined how Gemini 3 Flash and gpt-5-mini behave with web search (runs on 31 October and 1 November 2026). Both followed a similar pattern: first, a search by company name, city and Latin transliteration; then the model extracts the 9-digit tax number (UNP) from the page and sends a separate query using only those digits; after that, it checks whether the legal entity exists and whether its services match the description on the site.
Why llms.txt doesn't help, and bots read JSON and Markdown?
On 15 June 2026, Google's Search guidance stated explicitly that llms.txt is used neither in ranking nor in AI overviews. Logs from an open directory of 30,900 listings back this up. Over 30 days, the directory recorded 773,704 requests. Of these, 705,323 were verified by official IPs and reverse DNS, meaning they came from real bots rather than traffic impersonating them. The breakdown: industry shortlists got 616,242 requests, JSON listings 38,139, and Markdown listings 20,110.
What are the key details on A company's "digital passport" and caveats?
In the author's view, optimizing a landing page alone is not enough. You need an external listing in a neutral registry that meets four requirements: Tie to a state registry (UNP or INN): the legal entity's status, registration date and region are pulled from official databases and can't be edited by hand. Specific services and geography instead of generic OKED codes, plus price benchmarks and answers to frequent questions.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.