# How AI decides what to recommend How ChatGPT, Claude, Gemini and Perplexity pick the products they name: live search, third-party pages, search rank, familiarity and reviews. With sources. By Dasha. Updated November 1 2026. https://heardofyou.io/guides/how-ai-decides-what-to-recommend Ask an AI assistant for "the best invoicing app for freelancers" and it usually does three things: 1. It runs a few web searches of its own. 2. It reads what other people wrote about the options: lists, reviews, forum threads. Far more than vendor sites. 3. It writes an answer that leans toward brands it already knows. Small review and rating edges decide the close calls. Your product gets named when it shows up in those searches and on those third-party pages. Below is each step with the evidence, and what it means for a small product. ## Step 1: it usually searches first Modern assistants do not answer product questions from memory alone. In one data set, ChatGPT searched the web in 31% of prompts. That was about 2.17 searches per prompt, averaging 5.48 words each. "Reviews", "comparison" and the current year were common words (Nectiv, via Search Engine Land). For software questions, 76% of ChatGPT's searches added a year, as in "best invoicing app 2026" (StatWharf). One finding matters more than any other. In Tannenbaum's 2026 study, a brand missing from the assistant's live search results got named by GPT 2.8% of the time. When the brand's own site was cited and branded searches happened too, 91.4%. Being in the material the assistant pulls in is most of the game. **2.8% vs 91.4% : how often GPT named a brand that was absent from its live search results, against one whose site was cited : Tannenbaum, 2026 (pre-print)** Ordinary search rank still counts inside that. Across 815,484 query-page pairs, 58% of pages at position 1 got cited, against 14% at position 10 (AirOps and Kevin Indig). The same study found that covering lots of related sub-topics ("fan-out coverage") barely mattered. Rank mattered. Breadth did not. ## Step 2: it reads what other people say about you AI answers lean hard on pages the brand does not control. For US software questions, 72.7% of cited sources were earned media: reviews, articles, lists by third parties. A normal Google results page: 45.4% (Chen et al., University of Toronto). Across 207 million citations, Muck Rack found over 84% came from earned or third-party sources. Paid content: 0.3%. Which third parties? - Reddit and YouTube are in every engine's top five cited domains. - ChatGPT leans editorial: Wikipedia, Forbes, TechRadar. - G2 is in Perplexity's top five but not the others (Peec, a vendor study). Across 75,000 brands, YouTube mentions had the strongest correlation with AI visibility (about 0.737). Branded web mentions came next (0.66 to 0.71). Domain authority was weak (0.27) (Ahrefs, a vendor study). OpenAI describes what ChatGPT uses for shopping: the query plus memory, structured product data, third-party content, and price, availability and reviews. It says product results are not ads. ## Step 3: ties go to the brand it already knows With identical specs, the well-known brand was recommended 100% of the time (Chu and Hou, 2026). That is the incumbent's advantage, and it is real. The same study found what breaks it. A rating edge of under 0.1 stars for the smaller competitor flipped the pick. A Columbia study of shopping agents found the same lever: - A rating up by 0.1 took a product from a 10% choice rate to 15 to 20%. - An "Overall Pick" style badge raised it to 20 to 43%. - A "Sponsored" tag pushed it down to about 8 to 9%. Agents also favoured products listed first. One trusted source can override the model's favourite. In about 26,000 Wharton tests, one Wirecutter or Reddit source changed the pick. The order of the sources mattered too. ## Step 4: wording and dates move things too Small things in the question change the answer. Adding "I love hiking" redirected Claude Opus in three-quarters of runs (Wharton GAIL). Asking for "the best" option nudged models toward AI-flavoured picks by 8 points in one study. Models also favour newer dates. A more recent publication date moved passages up in all seven models tested, by as much as 95 places (Fang et al.). Where a candidate sits in a list changes whether it gets picked (Bito et al.). The honest reading: recency and position are biases, not strategies. A real update date on a page you really updated is fine. Faking dates is manipulation, and this site does not do it. ## Step 5: the answer changes every time you ask Two runs of the same prompt gave the same brand list less than 1 time in 100. The same list in the same order: about 1 time in 1,000 (SparkToro, 2,961 responses). Category leaders still showed up in 55 to 100% of answers. The wobble is at the edges of the list, which is exactly where a small product lives. Across 536,301 answers, ChatGPT, Claude and Gemini all named the same company only about 3% of the time (BusySeed, via IPS). So one screenshot of ChatGPT naming you (or not) means very little. That is why we never report a rank. We run each question three times on each assistant and report how often you are named. See [how we measure](/how-we-measure). ## What this means if you are small 1. Get into the assistant's live search results. Your own site for your category's queries, and the third-party pages that already rank for them. 2. Earn mentions on the sites each assistant cites for your category: "best X" lists, review sites, Reddit threads, a YouTube walkthrough. 3. Collect honest reviews. A fraction of a star can beat a bigger name in a close call. 4. Make your homepage say what it is and who it is for on the first screen. The assistant and the buyer both check there. 5. Measure with repeated runs, not one answer. ## Limits of this evidence Several of the strongest results (Tannenbaum, Chu and Hou) are 2026 pre-prints, not yet peer reviewed. Ahrefs, Peec and Muck Rack sell AI visibility tools or PR software, so their studies are labelled vendor studies here. They are large but correlational. Models change often, and any figure can shift with the next update. Nothing on this page lets anyone promise what an assistant will say. ## What to do next Run the check. Our [do-it-yourself guide](/guides/check-if-chatgpt-recommends-you) shows the exact method. Or we run it for you free: 5 buyer questions, 3 runs each, on ChatGPT, Gemini, Claude, Perplexity and Muse. Results by email within 2 working days. Already know you are missing? Read [why AI doesn't recommend your product](/guides/why-ai-doesnt-recommend-you). ## Questions **Does ChatGPT read my website when it recommends products?** Sometimes, and usually not first. When it searches, it mostly reads pages other people wrote about you: lists, reviews, Reddit threads. Your own site is where it checks facts like price and audience. **Is there a ranking I can climb?** No. The same question asked twice gives the same list of brands less than 1 time in 100. The useful measure is how often you are named across many runs. We call it a mention rate. **Do the four assistants agree?** Rarely. In one study of 536,301 answers, ChatGPT, Claude and Gemini all named the same company only about 3% of the time. Check each one separately. **Can I pay to be recommended?** Not in the answer itself. OpenAI says product results are not ads, and paid content is about 0.3% of citations. ChatGPT shows ads under answers to free users. That is a different thing. ## Sources - Tannenbaum, A fitted stage model of brand mention in AI search, 2026 (pre-print) (https://arxiv.org/abs/2609.23162) - Nectiv, ChatGPT search prompt data, via Search Engine Land (https://searchengineland.com/chatgpt-search-prompts-data-463407) - StatWharf, ChatGPT query fan-out research (https://statwharf.com/research/chatgpt-query-fan-out/) - AirOps and Kevin Indig, 815,484 query-page pairs (https://www.getpassionfruit.com/research/is-query-fan-out-a-marketing-buzzword-or-something-relevant) - Chen et al., University of Toronto, earned media in AI search, 2025 (https://arxiv.org/html/2509.08919) - Muck Rack, Generative Pulse, 207 million citations, October 2026 (https://www.globenewswire.com/news-release/2026/10/28/3390134/0/en/generative-pulse-earned-media-still-drives-more-than-84-of-ai-citations.html) - Chu and Hou, brand bias and rating edges, 2026 (pre-print) (https://arxiv.org/abs/2606.17443) - Columbia, What Is Your AI Agent Buying? (https://ar5iv.labs.arxiv.org/html/2508.02630) - Wharton GAIL, agentic shopping technical report, 2026 (https://gail.wharton.upenn.edu/research-and-insights/technical-report-agentic-shopping/) - SparkToro, AIs are highly inconsistent when recommending brands (https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/) - BusySeed via IPS, three assistants agree on 3 in 100 recommendations (https://ipsnews.net/business/2026/10/23/three-ai-assistants-agree-on-only-about-3-in-100-company-recommendations-study-finds/) - Peec, most-cited domains by engine (vendor) (https://docs.peec.ai/research/most-cited-domains-by-engine) - Ahrefs, AI brand visibility correlations, 75,000 brands (vendor) (https://ahrefs.com/blog/ai-brand-visibility-correlations/) - OpenAI Help, Shopping with ChatGPT search (https://help.openai.com/en/articles/11128490-shopping-with-chatgpt-search) - Fang et al., recency bias in LLM rankers (https://arxiv.org/abs/2509.11353) - Bito et al., position bias in LLM recommendations (https://arxiv.org/abs/2508.02020)