How AI Assistants Decide Which Business to Recommend

ChatGPT decides which business to recommend by searching the web for the buyer’s question, reading the pages that rank, and naming the businesses those pages support. In LeadsNow’s own polling (15 August to 15 October 2026), a page ranked top 3 in search was cited in 21% of 298 ChatGPT polls; outside the top 10, 0% of 154.

Rules of thumb (each is sourced on this page or is arithmetic from the worked example below):

• No rank, no recommendation. In LeadsNow’s polling, pages outside the search top 10 for the question were cited in 0% of 154 ChatGPT polls.
• The assistant runs its own searches. Google says AI Mode may issue “multiple related searches” (query fan-out), and Microsoft says Microsoft 365 Copilot sends a generated query to the Bing search service. Rank for the sub-questions too.
• Shortlists beat articles. Top-3 “best X” listicles were cited in 35% of polls (62 of 179) against 1% (1 of 80) for top-3 how-to articles.
• Hiring questions are where businesses get named. “Who should I hire” questions were cited 27% of the time against 4% for informational questions.
• Two strong pages outweigh eight weak ones. In the worked example, two top-3 listicles earn 7 of 8.3 expected citations.
• Do not block the search crawlers. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.

What are the four gates an AI assistant puts a business through?

An AI assistant recommends a business only after it passes four gates in order: it must be found in search, read cleanly, listed on a shortlist page, and then survive the buyer’s own check. LeadsNow calls this the Four-Gate model: observed behaviour, not a published algorithm. Vendors document the first gate, where most businesses drop out.

FAQ: how AI assistants decide which business to recommend

How does ChatGPT decide which business to recommend?

ChatGPT searches the web for the question, reads the pages that rank, and names the businesses those pages support. OpenAI documents the crawler (OAI-SearchBot) but not a ranking formula. In LeadsNow’s polling, pages outside the search top 10 were cited in 0% of 154 polls.

Does Google AI Mode use different rules from ordinary Google Search?

No extra technical rules apply: Google Search Central’s AI features guide says a page must be indexed and eligible to show in Google Search with a snippet, with no additional technical requirements. AI Mode may run several related searches for one prompt, so sub-questions matter.

Does Claude decide the same way as ChatGPT?

Claude follows the same broad pattern of searching, fetching and citing, with its own crawlers: Anthropic says Claude-SearchBot indexes content for search and Claude-User fetches pages when a user asks. LeadsNow sees Claude’s live fetches lean towards cost, benchmark and pricing pages.

Do AI recommendations turn into paying clients?

They can, but the evidence is thin everywhere. LeadsNow’s booking records show three clients in October 2026 who said Claude, Google AI Mode or Microsoft Copilot sent them. That is a small sample, so treat AI referral as an emerging channel and track it per booker.

Should I block AI crawlers to protect my content?

Blocking the search crawlers removes you from those answers. OpenAI says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, and Anthropic says blocking Claude-SearchBot may reduce your visibility in Claude’s search results.

How does each AI assistant find businesses to recommend?

ChatGPT, Google AI Mode, Microsoft Copilot and Claude each retrieve live web pages before naming a business, but each vendor documents a different part of that retrieval. The table below states only what each vendor’s own documentation said when checked on 29 October 2026, plus LeadsNow’s own server-log counts.

None of the four vendors publishes how it chooses which named business goes first, so any list of AI “ranking factors” is someone’s observation. PerplexityBot made 5,776 fetches of leadsnow.ai in the same window, more than any of the other three AI agents LeadsNow counted: the four engines above are not the whole market.

Why does search rank decide most AI recommendations?

Search rank decides most AI recommendations because an assistant can only name businesses that appear on the pages its searches return. In LeadsNow’s ChatGPT polling between 15 August and 15 October 2026, the share of polls citing a LeadsNow page fell from 21% (top 3) to 5% (rank 4-10) to 0% (outside the top 10), using DuckDuckGo and Bing rank as the visibility proxy.

That proxy is a measurement choice, not a claim about which index ChatGPT uses. For Microsoft 365 Copilot the link is documented directly, because Microsoft says it sends a generated query to the Bing search service. For Google AI Mode, Google says pages must be indexed and snippet-eligible in ordinary Search. If you are already ranking and still not being named, the diagnosis lives on why you are not showing up in ChatGPT answers, and why most retrieved pages still go uncited is covered in how to be in the 15% of pages ChatGPT actually cites.

Which pages make it onto the AI shortlist?

Ranked shortlist pages make it onto the AI shortlist far more often than articles: in LeadsNow’s polling, top-3 “best X” listicles were cited in 35% of polls (62 of 179) and top-3 how-to articles in 1% (1 of 80). Both are subsets of the 15 August to 15 October 2026 top-3 polls.

The question shape matters as much as the page shape. Across LeadsNow’s ChatGPT polling from July to 15 October 2026 (1,069 polls in total), “who should I hire” and “best X agency” questions produced a citation 27% of the time, against 4% for informational questions.

So publish a fair ranked comparison for each hiring question your buyers ask, keep your business details consistent everywhere (entity consistency for AI search covers the method), and put the answer and table in the first third of the page.

Worked example: how many recommendations does a 10-question set earn?

A business with 10 buyer questions and a mixed set of pages can estimate its expected AI citations by multiplying asks per question by the cite rate for each page’s rank band and type. The example below uses LeadsNow’s polling rates as inputs and 10 asks per question, so each row is 10 asks.

The arithmetic is 20 x 0.35 = 7.0, 30 x 0.01 = 0.3, 20 x 0.05 = 1.0 and 30 x 0 = 0, for 8.3 citations per 100 asks. Caveats: the listicle and how-to rates are overlapping subsets of the top-3 polls, and a citation is not a client. Substitute your own question count and ask volume.

The fastest improvement here is not a new page: moving the two rank 4-10 pages into the top 3 as listicles lifts them from 1.0 to 7.0, taking the set to 14.3 per 100 asks.

What happens after an AI assistant names your business?

After an AI assistant names your business, the buyer runs gate 4 themselves: they open your site, check proof and pricing, and either book or move to the next name. LeadsNow’s own booking records for October 2026 show three clients who said an AI assistant sent them: one from Claude and one from Google AI Mode (both confirmed 15 October 2026), and one from Microsoft Copilot (confirmed 20 October 2026).

The Copilot booker, a financial-adviser buyer, had asked Copilot for the best lead generation for financial advisors and saw LeadsNow in its top three. That is the full sample: three bookers, too small to rank engines, and none came from ChatGPT, the engine LeadsNow polls most.

So the pages that win gate 4 are the ones a sceptical buyer checks: the homepage, pricing, proof and the booking page. LeadsNow’s live fetch logs show Claude’s fetches lean towards cost, benchmark and pricing pages, which means the same facts must be stated plainly where both a buyer and a bot will look. Converting that click is covered in how to capture leads from ChatGPT referrals; proving where the client came from is covered in measuring AI referral traffic and attribution. The simplest attribution step is to ask every new booker how they found you: LeadsNow knows about all three of its AI-referred bookers only because they said so.

Which stage takes the longest, and what makes it slip?

The search gate takes the longest of the four stages, because a new page must be crawled, indexed and then earn a top-3 rank before any assistant can read it. The read, shortlist and verify gates happen within a single conversation once the page ranks.

Three documented things make the search gate slip:

• Blocking the wrong crawler. OpenAI says opted-out sites will not appear in ChatGPT search answers and that a robots.txt change can take about 24 hours to take effect in its systems.
• Not being indexed. Google says indexing and serving are not guaranteed even when a page meets every requirement, and AI Mode only links pages eligible for a Search snippet.
• Ranking only for the head term. Because AI Mode may fan one prompt out into several related searches, a page that ranks only for the broad term can miss the sub-question that actually gets cited.

How long the whole sequence takes for a typical site is a separate question, answered on how long AEO takes to work. Doing it yourself costs mostly hours: picking buyer questions, writing fair comparison pages, and re-checking rank monthly.

Sources

1. Google Search Central: AI features and your website, checked 29 October 2026.
2. OpenAI: Overview of OpenAI crawlers, checked 29 October 2026.
3. Microsoft: Transparency Note for Microsoft Copilot, checked 29 October 2026.
4. Microsoft Learn: web search in Microsoft Copilot and Copilot Chat, checked 29 October 2026.
5. Anthropic: ClaudeBot, Claude-User and Claude-SearchBot, checked 29 October 2026.
6. LeadsNow AI-answer polling (ChatGPT with browsing; DuckDuckGo and Bing rank as proxy): rank bands and page types 15 August to 15 October 2026; question-shape split July to 15 October 2026 (1,069 polls). LeadsNow server logs, 1-27 October 2026. LeadsNow booking records, October 2026.