How ChatGPT Decides Which Software Products to Recommend

ChatGPT picks software from two inputs: what the model absorbed about your category during training, and, when it searches, the pages it retrieves for that prompt. Both reward the same thing, a product the wider web names consistently for a specific job. If you're a founder or growth lead at an AI startup and ChatGPT keeps naming three rivals and never you, this is the machinery behind that answer. Perplexity, Claude, Gemini, Google AI Overviews, Google AI Mode and Copilot each build their shortlists differently. What follows is ChatGPT's machinery only.

How ChatGPT decides which software products to recommend, in three stages

ChatGPT builds a software shortlist in three stages: it reads the prompt against what the model already knows, it may search the web, and it writes an answer that blends both. At no point does it consult a vendor ranking you can submit to. Each stage has different inputs, and you can influence most of them only through what other sites say about you.

The three stages behind a ChatGPT software answer

What does ChatGPT know about your product before it searches?

Before any search, ChatGPT knows only what it learned in training. OpenAI's help article on how its models are developed says they are trained on publicly available internet content, information from third-party partners and information provided by users, trainers and researchers, and that it uses only internet content that is freely and openly accessible. If the open web rarely described your product before the training cutoff, the model has little to go on.

The same article explains the drift founders notice. The model predicts the most likely next word, several continuations are usually plausible, and so the same question can produce different answers on different occasions. A tool the web mentions constantly next to a use case wins that prediction often. A tool mentioned once, on its own homepage, rarely does.

Training is also where blocking choices start to matter. OpenAI says site owners control training access through the GPTBot user agent. Blocking GPTBot keeps future content out of training, which is a reasonable choice for some companies, but it does nothing for search answers, which run on a different crawler.

When does ChatGPT search the web for software recommendations?

ChatGPT searches when it judges that a question would benefit from current information, and a user can force a search from the tools menu. OpenAI's help page on searching the web with ChatGPT describes both routes. Software questions about current pricing, recent launches or this year's options fit that description; a request for a general explanation may be answered from memory alone.

When it searches, ChatGPT typically rewrites the question into one or more targeted queries for its search partners, and can send more specific follow-up queries after reading the first results. OpenAI's own example turns a long research question into "CCR8 immunotherapy drug development 2025". If memory is on, saved memories can shape the rewritten query, and approximate location from the IP address is used too. The partner privacy policies the page links are Microsoft's and Shopify's.

That rewrite is why the page ranking for a buyer's exact sentence is often not the page ChatGPT cites. Ahrefs found in August 2025, across 15,000 long-tail queries, that only 8% of ChatGPT's in-text citations ranked in Google's top 10 for the original prompt, and 8.1% in Bing's top 10. Shorter, simpler queries tell a different story.

ChatGPT search draws heavily on Bing, which Microsoft announced as ChatGPT's default search experience in May 2023. One access rule sits under all of it. OpenAI's crawler documentation says sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though they can still appear as navigational links. Why the Seer and Ahrefs figures differ so much, and whether a Bing listing alone gets you in, is covered in does adding your site to Bing Webmaster Tools get you into ChatGPT.

Which sources does ChatGPT lean on for software questions?

For shortlist prompts, ChatGPT search leans on third-party best-of lists, reference pages and reviews. Most of the evidence about your product therefore reaches ChatGPT through pages you do not own, written by people deciding which tools to include.

The same Ahrefs study found that when a software vendor was recommended, its own self-promotional "best" list showed up in more than a third of those ChatGPT responses, and that 35% of cited lists were published on low-authority domains. A list does not need a famous publisher to be used. It needs to exist, be recent and match the query ChatGPT actually ran.

Reference pages carry weight too. In Profound's analysis of 680 million citations from August 2024 to June 2025, Wikipedia accounted for 7.8% of all ChatGPT citations and 47.9% of citations among its top 10 sources. Review platforms carry the most weight with buyers checking an AI answer: in G2's survey of 1,076 B2B software buyers, published in April 2026, 85% said they think more highly of vendors that AI includes in its answers, and citations from software review sites stood out as their top trust signal. Building that presence is covered in how review sites like G2 influence AI recommendations.

Swipe sideways to see every column

Signal	How it reaches ChatGPT	What the evidence says	What you control
Category association	Public web text absorbed in training	OpenAI trains on public web content, partner data and feedback	The same description of what you do, repeated on many sites over time
Best-of lists	Retrieved by search	43.8% of cited page types in Ahrefs' December 2025 study	Inclusion in recent roundups for your use case
Review platforms	Retrieved by search and absorbed in training	Buyers rank review-site citations as their top trust signal (G2's survey of 1,076 B2B software buyers, published in April 2026)	Reviews written in buyer language on the sites your category uses
Brand mentions	Both	Branded web mentions had the strongest correlation (0.664) with brand visibility in Google AI Overviews (Ahrefs); we have not seen an equivalent study for ChatGPT	Mentions in articles, threads and lists that name your use case
Your own pages	Retrieved by search if OAI-SearchBot can crawl them	Sites that opt out of OAI-SearchBot are left out of search answers (OpenAI)	Crawlable product, pricing and comparison pages that answer first
Entity consistency	Both	Bing's guidelines say clear, consistent entity naming improves grounding visibility	One product name, one category and one description everywhere

The brand mentions row comes from Ahrefs' study of 75,000 brands, published in May 2025, which measured Google AI Overviews only. We include it because it is the largest public test of mentions against links, and mentions beat links there by a wide margin. The entity row comes from Bing's Webmaster Guidelines, which ask for consistent naming of products and organizations.

Why does ChatGPT recommend different software to different people?

Each answer is generated fresh, shaped by the exact wording and sometimes by the person asking. SparkToro and Gumshoe had 600 volunteers run 12 prompts through ChatGPT, Claude and Google's AI 2,961 times, published in January 2026. There was less than a 1 in 100 chance that ChatGPT would return the same list of brands twice, and about 1 in 1,000 that two lists would share the same order.

Wording varies even more than founders expect. Across 142 human-written prompts about the same headphone purchase, average semantic similarity between prompts was 0.081, yet the top brands still appeared in 55% to 77% of the 994 answers. Presence across many runs is stable enough to measure. Position in one answer is noise.

Personal context adds another layer: memory can shape the rewritten search query, and location is inferred from the IP address. Your own account is the worst test bench you have, so run checks in a Temporary Chat or logged out, as our guide to tracking brand mentions in ChatGPT explains. Paying does not change the list either. OpenAI's ads FAQ says ads run on separate systems from the chat model, appear below responses labeled as sponsored, and give advertisers no ability to shape, rank or alter answers.

What can an AI startup change to get recommended by ChatGPT?

You can change three things: whether ChatGPT can reach your pages, whether those pages state plainly what you do, and how many independent sources repeat it. The order below runs from fastest to slowest.

Five changes, fastest first

When the first two steps are done and you are still missing, the cause is usually on the third-party side, and why your brand does not show up in ChatGPT walks through each test. This is the work a ChatGPT SEO agency is hired for, and most of it can also be done in-house with patience.

How long does it take for ChatGPT to start recommending a product?

Search answers can change within weeks, while model answers change only when OpenAI ships a model trained on newer data, on a schedule it does not publish. SISTRIX's study of 82,619 prompts over 17 weeks, published in May 2026, found ChatGPT search replaced 74% of its cited sources each week. That churn cuts both ways: a new list placement can surface quickly, and it can disappear just as fast.

We have not seen reliable data on how long a new brand takes to enter ChatGPT's model answers. Plan on search answers first, and treat model answers as the slow result of months of consistent mentions.