How ChatGPT, Gemini and Perplexity Choose Which Suppliers to Recommend

Most AI answer engines work in two steps. First, retrieval: the engine runs something close to an ordinary web search — Google AI Mode draws on Google Search, while OpenAI documents its own search crawler, OAI-SearchBot, but not which index ChatGPT search queries — and pulls back a short list of pages that look relevant to the question. Second, generation: the underlying language model reads those pages and writes an answer, choosing which sources to name and which to leave out. This two-step process is sometimes called "grounding" (grounding means tying an answer to real retrieved pages instead of the model guessing from memory).

That matters for GEO (generative engine optimization, explained in GEO vs SEO) because it means an engine can only cite a supplier it actually found in that first retrieval step. A company that is invisible to ordinary web search — blocked crawlers, no indexed pages, nothing current — is invisible to the AI engine too, no matter how good its product is.

Some engines add a further step called query fan-out: instead of searching once for the literal prompt, the engine expands it into several related sub-questions, runs retrieval for each, then merges the evidence before writing one answer. A supplier that only answers the exact phrase a buyer typed, and nothing about the related sub-questions an engine fans out to, can still be missed even if the literal page is found.

For a buyer question like 'best contract manufacturer for [product]' or '[part] supplier with short lead time,' AI engines lean on the same kinds of sources a human researcher would check first: industry directories (Thomasnet/Thomas, Kompass and similar), trade associations, review and rating sites, and discussion threads on sites like Reddit where buyers compare real experiences. A supplier's own website matters too, but usually as a secondary confirmation once the engine already has a candidate name from one of those other sources.

That is a structural reason a listing in an industry directory or a trade association can matter as much for AI visibility as it does for a human buyer doing the same research — the engine is reading the same public web that buyer would read, not a private database of manufacturers. Selling to consumers through an online store instead? The retail version, built on product facts and reviews, is how AI search chooses which stores to cite.

No, and nobody honestly can — not Leadee, not an agency, not a directory. AI engines run the same prompt differently from one run to the next, change their retrieval sources over time, and weigh signals no outside party fully controls. What is realistic is improving the odds: the checklist above removes the reasons an engine cannot find or trust a supplier, and measuring repeatedly over time (see the methodology) shows whether those changes are actually moving mention and citation rates, rather than guessing from one chat answer.

This is one real run, not a benchmark: in a Netbase JSC internal measurement run on 2026-10-23, Leadee tracked 3 web-to-print buyer prompts across OpenAI, Perplexity and Google AI Mode for printcart.com (a Netbase-owned product of Netbase JSC) against named market players, with n=1 (one run per engine, not yet repeated for a confidence interval).

The domains the engines cited most often that week were apps.shopify.com, customerscanvas.com, designnbuy.com, woocommerce.com, reddit.com, teeinblue.com and help.customily.com — marketplaces, platform documentation and a community forum, not any single market player's own marketing page. That single run illustrates the pattern above (directories, platforms and forums over brand pages) on one real brand; it is an example, not a claim about every industry.

A buyer who does not know any vendor yet asks a shortlist question: "best web-to-print software for commercial print shops", "Zakeke alternatives for product customization", "best Vietnam software development companies for US startups". Later, once they have names, they ask trust and comparison questions with the vendor's name in them. The first kind decides whether a supplier is considered at all. How ChatGPT in particular finds and cites brands is covered on the ChatGPT SEO overview.

On 2026-10-25 we asked those questions for the 7 sites of Netbase JSC (the company behind Leadee): 8 US questions each, 3 runs per engine. On the questions without the brand name, only 2 of the 7 were ever named: CMSmart in 13 of 40 runs and Printcart (Netbase-owned) in 7 of 40. NetbaseJSC, Web2Print, SoftWSP, OutsourcingVN and AdminKit were named 0 times in 37-42 runs each. Asked by name, all 7 were named in at least some runs.

ChatGPT itself (OpenAI gpt-5 with web search) answered 25 runs before the account ran out of credit, all on shortlist questions. It named a Netbase brand in 3 of them (12%, 95% interval 4.2-30.0%), all CMSmart, on the WooCommerce product-designer questions where CMSmart also did best on the other engines. ChatGPT vs Gemini vs Perplexity compares the three engines on the same questions.

Mostly from third-party pages, not from the suppliers' own sites. The sources that stood out per sector were clutch.co for outsourcing and software services, apps.shopify.com for product customizers, designnbuy.com and software directories such as G2 and Capterra for web-to-print, mxtoolbox.com for DNS and email tools, and binarly.io and finitestate.io for firmware security. The two Netbase sites that made shortlists are also the two whose own domains were cited most in their niche: cmsmart.net (96 citations) and printcart.com (Netbase-owned) with solution.printcart.com (87).

That is the practical lesson for a supplier: be present on the pages the engine reads for your category, and make your own site a source worth citing. Which directories matter depends on the sector, so check the cited domains for your own questions first. How to get cited when ChatGPT prefers other websites shows how, and ChatGPT SEO for professional services firms looks closer at the two service firms. These counts pool three engines; our analyses do not split cited domains by engine.