How to get cited and recommended by ChatGPT

ChatGPT names brands from its training or, when it searches the web (opens in a new tab), from pages it retrieves. Household brands appeared in 73% of first answers to generic prompts and small brands in 11% (opens in a new tab), so a newer brand must be found through retrieval: pages ChatGPT can fetch, specific facts it can compare, and third-party sources that name you. Answers change from run to run (opens in a new tab), so measure the share of repeated runs that name you, not a rank.

This guide covers getting a brand, product, or business named in ChatGPT's answers: how the model forms recommendations, what helps a smaller brand, and how to measure the result. Getting a specific page retrieved and cited is covered in how to rank in ChatGPT search. For other assistants, see how ChatGPT, Gemini, Claude, Perplexity, and Copilot find and cite sources.

Scope and definitions

• Mention. ChatGPT names your brand anywhere in an answer.
• Recommendation. ChatGPT names your brand as an option for the user's need, often in a shortlist. It also tends to pick one: in an audit of 117 product queries, ChatGPT stated a first-person preference in 79% of the answers that recommended products (opens in a new tab) (preprint; September 2026).
• Citation. ChatGPT links a page as a source. The page is usually not yours: only 2.9% of 149,912 citations pointed to the tracked brand's own website (opens in a new tab) (preprint; the author co-founded the tracking platform studied). A third-party list that names you produces a mention with no link to your site.
• Unbranded prompt. A question that describes a need without naming a brand, such as "Which payroll software suits a 50-person company?" Being named on unbranded prompts is what "being recommended" means here.
• Share of runs. The percentage of repeated runs of a prompt in which ChatGPT names you. Share of voice divides your mentions by all mentions of a fixed set of competing brands.
• Entity. A distinct thing, such as a company, person, or product, that models and search engines try to recognize and keep apart from others.

How ChatGPT forms recommendations

Two routes: training data and retrieval

Parametric knowledge is what a model absorbed in training, and it favors well-known names. GPT-4's accuracy fell from well-known to lesser-known entities and averaged 31% on an 18,000-question benchmark (opens in a new tab) (peer-reviewed). In controlled experiments, a fact could be reliably extracted from a model only when it had appeared in varied restatements during training (opens in a new tab) (peer-reviewed). A brand first written about after a model's training data was collected is not in that knowledge at all.

Retrieval is the second route, and ChatGPT does not always use it. ChatGPT ran a web search on 34.5% of queries in February 2026 (opens in a new tab) (vendor study; US clickstream panel), and in a German-language panel of commercial prompts, 57.8% of ChatGPT runs returned no citations (opens in a new tab) (preprint). The two routes reward different things. Among 112 startups from the 2025 Product Hunt leaderboard, a model without web access named 5.4% of them at least once in discovery-style questions, while Perplexity's search-based model named 27.7% (opens in a new tab) (preprint; master's thesis; December 2025).

Recommendations are distributions, not rankings

• When the same prompt was repeated, ChatGPT and Google's AI returned the same list of brands less than once in 100 runs, and Claude only slightly more often (opens in a new tab) (vendor study; 2,961 runs on ChatGPT, Claude, and Google's AI features, November and December 2025; a co-investigator works for a tracking vendor).
• ChatGPT's recommended products had a mean Jaccard similarity of 0.178 across three repeats of the same query (opens in a new tab), so most picks changed between repeats (preprint).
• Rates are steadier than lists. 77.5% of brand, prompt, and engine combinations were either always or never mentioned, and sentiment was 6.7 times noisier than whether the brand was mentioned at all (opens in a new tab) (preprint; 102,025 API responses, March to May 2026).

One answer is one sample. The share of runs that name you is the quantity you can measure and improve.

Big brands start ahead

• On a first answer to a generic prompt, household brands appeared 73% of the time, mid-market brands 44%, and small brands 11% (opens in a new tab) (preprint; 102 brands across five assistants' APIs).
• LLM-based recommenders over-recommend popular items (opens in a new tab) (peer-reviewed; conversational recommendation).
• The same 2025 startups appeared in 3.32% of discovery questions to the model without web access and 8.29% to the search-based model (opens in a new tab) (preprint; 784 discovery questions per model).
• With equal specifications, a real brand was chosen in 100% of trials against nine fictional ones (opens in a new tab) (preprint).

The head start is not absolute. When the user's country was stated, GPT-4o recommended local brands 75% of the time (opens in a new tab) (peer-reviewed). Context in the question changes who wins.

Specific, visible facts can flip a choice

• A fictional brand beat a real one half the time with only a 0.075-star rating edge, 1.6 times the reviews, or a 7.3% lower price (opens in a new tab) (preprint; GPT-4o-mini, Claude Sonnet, and Gemini 3 Flash through their APIs).
• In a pre-registered study of 12 models and 61,459 calls, a 4.7 rating instead of 3.9 raised the chance of being chosen by 31.65 percentage points, a $249 price instead of $129 lowered it by 30.03 points, and 2,100 reviews instead of 45 raised it by 8.31 points (opens in a new tab) (preprint; hotel choices).
• In a simple retrieval setup with one optimized challenger, the well-known brand survived in only 5.0% to 6.7% of runs, because the retriever ranked it low on text similarity (opens in a new tab). The authors do not claim this holds for commercial systems.

These are laboratory tests of consumer products and hotels, not B2B services, and ChatGPT's interface and its API shared only 12.0% of cited domains for the same queries (opens in a new tab) (preprint). Buyers check the facts as well: 27% of US business professionals who use AI at work said AI vendor recommendations don't reflect real pricing or contract structures (opens in a new tab) (vendor survey; 519 respondents, March to April 2026).

Third-party corroboration: lists, reviews, and communities

• In API tests of ChatGPT's search model, 95.1% of cited sources for niche brands were earned media such as reviews and editorial coverage (opens in a new tab) (preprint; mid-2025).
• Ranked "best of" listicles made up about 21% of all citations (opens in a new tab) (preprint; five assistants' APIs). Across 750 ChatGPT prompts about software, products, and agencies, "best X" blog lists were 43.8% of cited page types (opens in a new tab) (vendor study, December 2025).
• Across 75,000 brands, YouTube mentions (about 0.74 across ChatGPT and Google's AI features) and branded web mentions (0.664 for ChatGPT) correlated with AI visibility, while link metrics such as the number of backlinks showed "very weak correlations" (opens in a new tab) (vendor study; correlation, not cause). The guide to brand mentions covers how to earn them.
• OpenAI accesses Reddit's Data API to "better understand and showcase Reddit content, especially on recent topics" (opens in a new tab) (company announcement, May 2024).

In the answer engine optimization (AEO) agency category, a published test of 48 answers from ChatGPT, Gemini, and Google's AI features (August 20 to September 2, 2026) found that the most-named agency appeared in 58% of answers, almost every frequently cited source was an agency's own "best of" list ranking itself first, and one agency's site was the source for 14 of the 48 answers (opens in a new tab) (agency-published; small sample). A second test of six runs concluded that "most of the pages that decide the answer are owned by the agencies competing for the answer" (opens in a new tab) (agency-published, September 2026).

Entity consistency

• Models "often yield ambiguous answers or incorrectly merge information belonging to different entities" that share a name (opens in a new tab) (peer-reviewed; 36,098 examples).
• Consistent rather than contradictory claims raised the odds of being cited first in at least four of six models (opens in a new tab) (peer-reviewed; 252,000 laboratory trials).
• Google asks for author markup that links to a page that uniquely identifies the author (opens in a new tab) (official documentation). But adding schema did not measurably raise ChatGPT citations (opens in a new tab) (vendor quasi-experiment), so structured data and sameAs links are hygiene, not a proven lever.

Where a recommendation leads

After ChatGPT switched to clickable brand names in answers in May 2026, homepage visits rose from 26–29% of its referrals to 62–63% (opens in a new tab) (vendor panel data; panel size not disclosed; July 2026). So a buyer who clicks your name usually lands on your homepage and can check it against what ChatGPT said.

Steps

The findings above come from the cited studies. The steps below are AEO HQ's recommendations based on them.

1. Make your pages retrievable. Allow OAI-SearchBot and OpenAI's IP addresses, get indexed by Bing, and server-render your facts. The ChatGPT search guide linked above has the details.
2. Start with specific prompts, not "best X". Specific queries gave more stable results than broad ones (opens in a new tab). Target questions where your facts decide the answer, such as a price limit, a buyer type, or a delivery model.
3. Publish the facts a model can compare on the page that answers each prompt: price, scope, deliverables, timeline, who it is for, and ratings with their source. Keep them in the HTML text, not in images or scripts.
4. Make the homepage confirm the answer. State the offer, the price, and who it is for near the top.
5. Use one name, one description, and the same facts everywhere: your site, profiles, directories, and author pages. If a founder shares a common name, use one distinctive descriptor every time.
6. Earn corroboration where ChatGPT looks: lists that publish their method, reviews on third-party profiles, disclosed participation in communities, and trade coverage of original data.
7. Publish original, citable evidence. Across ChatGPT, Google, and Perplexity, the cited pages that most shaped answers were rich in definitions, numbers, comparisons, and procedural steps (opens in a new tab) (preprint; descriptive; 602 prompts). The guide on how to write content for answer engines covers the format.
8. Measure the share of runs, as described in the next section.

How to measure recommendations

• Build a fixed prompt panel. Use unbranded buyer questions with several paraphrases each. People word the same need very differently: 142 human-written prompts for one intent had a mean semantic similarity of 0.081, yet produced similar brand sets (opens in a new tab).
• Repeat each prompt. The standard error of a brand's detection rate fell from 0.370 with one run to 0.081 with seven and 0.062 with eight, and to 0.033 when aggregated over 28 days (opens in a new tab) (preprint). A review of 45 studies suggests 7–8 repetitions and 3–5 paraphrases as a starting point, with the product, mode, model, date, locale, account, and search setting recorded for each run (opens in a new tab) (preprint).
• Prefer more prompts to more runs. Re-running a prompt shrinks only the variation within that prompt, and clustered standard errors can be over three times larger than naive ones (opens in a new tab) (preprint).
• Report intervals. Normal-approximation intervals are too narrow below a few hundred data points; nominal 95% intervals covered the true value only 92.5% of the time at 100 data points (opens in a new tab) (peer-reviewed), so use Wilson or Bayesian intervals. By simple arithmetic, zero mentions in n independent runs puts the true rate below about 3/n with 95% confidence: 0 of 100 suggests under 3%.
• Use clean sessions. ChatGPT may use saved memories when it rewrites search queries (opens in a new tab). In a published test, a logged-in account that had discussed a brand put it first, while a temporary chat with memory off did not name it (opens in a new tab).
• Test the product buyers use. API runs are a cheap proxy, but the interface and the API share few sources (see above).
• Connect to outcomes. ChatGPT adds utm_source=chatgpt.com to referral URLs (opens in a new tab); see how to track AI referral traffic in GA4. Also ask buyers how they found you: in one agency's records, first-touch attribution credited AI for 28 of the 189 leads who named an AI tool (15%) (opens in a new tab) (single firm, which sells AEO services). The wider method is in how to measure AI visibility.

What the evidence shows and does not show

Antipatterns

These antipatterns can move answers in lab tests or for a while, but they are deceptive, fragile, or both.

1. Fabricated claims or credentials. Authority-style language, including fabricated clinical claims, broke the leading brand's hold in 55–99% of trials, but when every challenger used it, the first mover's payoff fell from +0.802 to +0.007 (opens in a new tab). The gain is deceptive and vanishes once others copy it.
2. Hidden text or instructions aimed at AI. Hidden instructions made a product 2.5 times more likely to be recommended by Bing Copilot in 2024 tests (opens in a new tab) (peer-reviewed), and hidden text overrode negative reviews in ChatGPT search (opens in a new tab). OpenAI runs automated monitors to block prompt injection (opens in a new tab).
3. Fake reviews, astroturfed posts, or planted pages. One polluted page among the retrieved results fooled LLM recommenders up to 27% of the time (opens in a new tab) (peer-reviewed; accepted, not yet published). That is manipulation, and it puts the trust the brand needs at risk.
4. Self-ranking "best of" lists and paid list placements. ChatGPT flagged the self-ranking in several answers (opens in a new tab), and a practitioner linked self-promotional listicles to visibility drops in January 2026 (opens in a new tab).
5. Rewriting your site for generative engine optimization (GEO). Much of this advice fails in testing: a 2025 benchmark found significant gains in 3 of 54 cases (opens in a new tab), and rewriting page text hurt retrieval (opens in a new tab).
6. Reporting "#1 in ChatGPT." Lists change on almost every run (opens in a new tab), so a rank is not a stable quantity.
7. Judging visibility from your own logged-in account. Saved memories can shape ChatGPT's searches (opens in a new tab), so your own account is not a neutral test.
8. Paying for a guarantee. OpenAI says placement in ChatGPT search is not guaranteed (opens in a new tab).

Checklist

Scope: one brand's readiness to be recommended by ChatGPT. Prerequisites: server logs, analytics, and a list of 20 or more buyer questions. Checklist version 1, September 27, 2026.

Frequently asked questions

Why does ChatGPT recommend my company to me but not to my customers?

Your account shapes your answers. ChatGPT may use saved memories when it rewrites search queries (opens in a new tab), and revealed user identity significantly changed chatbot recommendations (opens in a new tab) (peer-reviewed). In the published test above, a logged-in account that had discussed a brand put it first, and a temporary chat did not name it (opens in a new tab). Check in a temporary chat or a signed-out session.

How do I get my product recommended by ChatGPT?

Make the comparison facts visible and consistent: price, specifications, ratings, and review counts. Small differences in these flipped choices in lab tests (opens in a new tab). For product questions, ChatGPT's most frequent sources were review publishers, led by techradar.com in 13.8% of answers with sources (opens in a new tab) (preprint), so independent reviews matter.

How do I get my brand mentioned by ChatGPT?

Most mentions come from other sites, not your own: only 2.9% of citations in one large study pointed to the brand's own domain (opens in a new tab). Earn places in independent lists, reviews, and coverage, and keep your facts identical everywhere.

How long does it take to get recommended by ChatGPT?

No study has measured it. The fastest-moving small brands gained 10 to 20 percentage points of visibility between March and May 2026, and the author notes that a brand with no prior web presence can take longer to surface (opens in a new tab) (preprint). Until later models are trained, a new brand depends on retrieval.

Next steps

The Instant AEO Audit ($499) includes how AI assistants answer your buyers' questions, a side-by-side analysis of up to three competitors, a live crawl of your key pages, and AI crawler access checks.

Change log

• September 27, 2026: First published.