How AI assistants decide which brands to recommend

AI assistants like ChatGPT, Gemini, Claude, Google AI Overviews and Google AI Mode decide which brands to recommend in two ways: from patterns learned in their training data, and from live web sources retrieved at the moment the question is asked. There is no fixed ranking behind those recommendations: SparkToro research found less than a 1-in-100 chance that ChatGPT gives the same list of brands in any two responses to the same query (Ahrefs). Brand selection is probabilistic, not positional: your brand is either part of the answer often enough to matter, or it isn't. This guide explains both selection mechanisms, what the citation evidence shows about retrieval, how the engines differ, and how to check your own AI visibility, how often AI assistants name your brand when buyers ask for recommendations.

How does ChatGPT decide which brands to recommend?

ChatGPT decides which brands to recommend by combining what it learned during training with what it finds when it searches the web mid-conversation. For questions with buying intent, the live search is involved more often than not: an October 2025 analysis by Nectiv found that commercial-intent prompts trigger web searches in ChatGPT 53.5% of the time, against 18.7% for informational queries (HubSpot). So when someone asks "best CRM for a 10-person agency that bills hourly", the assistant is more often reading live pages and synthesising an answer from them than reciting memory.

Gemini, Claude and Google AI Overviews are built the same broad way: a trained model paired with live search or grounding. Grounding means the engine is handed live web results and writes its answer from them, rather than answering from memory alone. Most published citation research covers ChatGPT and Google AI Overviews, so the evidence below is strongest for those two and thinner for the rest.

Do AI assistants pick brands from training data or from live search?

Both, and the two work differently. Training data determines what an AI assistant already believes about your category; live retrieval determines what it reads about your category today. They reward different things on different timescales, and most brands can only meaningfully influence the second.

The practical consequence: you cannot rewrite a model's training data, but you can change what retrieval finds, which is why the rest of this piece focuses on retrieval-time selection.

What do AI engines look for when they choose sources to cite?

The published evidence points to three properties of the pages AI engines cite: recency, extractable structure, and agreement across independent sources. No engine publishes its selection rules, so these are patterns observed in large citation datasets rather than confirmed mechanisms, but they replicate across several independent studies.

Fresh pages get cited far more often than stale ones

AI engines cite recently published or updated content disproportionately. An AirOps study found that 95% of ChatGPT citations come from content published or updated within the last 10 months, and that pages with a visible "last updated" timestamp receive 1.8x more citations (Semrush). Ahrefs' 17-million-citation study across seven AI platforms found AI assistants prefer content roughly 25.7% fresher than Google's organic results, with ChatGPT citing URLs 393 days newer in-text (Ahrefs). On the AirOps figures, content older than about ten months accounts for roughly one ChatGPT citation in twenty.

Pages engines can extract from get cited more

Engines cite pages they can lift complete answers from. Kevin Indig's analysis of 1.2 million ChatGPT citations found that cited text is nearly twice as likely to contain definitive phrases such as "is defined as" (36.2% vs 20.2%), and that question marks in headings roughly double citation rate (18% vs 8.9%) (Ahrefs). Research compiled by ZipTie points the same way: self-contained 50–150-word chunks earn 2.3x more AI citations (Ekamoira), and the Princeton/Georgia Tech GEO paper presented at ACM KDD 2024 found that adding statistics lifts AI-answer visibility 15–40% (ZipTie). A page that buries its answer is harder to quote, and a page that is not quoted cannot carry your brand into an answer.

Brands named across independent sources appear more often

Brands mentioned across several independent sources turn up in AI answers more often than brands mentioned in one place. A Search Engine Land study of more than 800 websites across 11 industries found multi-platform brand mentions correlate with AI visibility at r=0.87, a strong statistical association, though a correlation of that kind does not by itself show the mentions cause the visibility (ZipTie). Query fan-out points the same way: engines decompose one question into several sub-queries, and a Search Engine Land study of 10,000 keywords found pages ranking for those sub-queries are 161% more likely to be cited, with fan-out accounting for 51% of all AI citations. A brand present across many of those sub-answers has more chances to be named than one present in a single source.

Consistency is measurable, and worth measuring before you act on it. Run a free AI visibility check to see how often your brand survives into an answer across a realistic set of buyer questions, rather than inferring it from a handful of prompts you tried by hand.

Which sources do AI assistants actually cite?

Video, reference, community and professional-network platforms take the largest share of citations in the published studies, and the mix is both highly concentrated and unstable. A Surfer SEO study found the top five domains cited in Google AI Overviews were youtube.com at 23.29%, wikipedia.org at 18.41%, google.com at 16.38%, reddit.com at 9.37% and linkedin.com at 8.80% (Search Engine Land). Concentration runs deep: industry analysis compiled by ZipTie found the top 1% of domains capture 64% of all AI citations and the top 10% capture 84% (ZipTie).

Three qualifications matter before you act on any league table of cited domains.

The mix shifts sharply over time. Semrush's three-month study (230,000 prompts across three engines, 13 weekly snapshots, 100 million-plus citations) recorded ChatGPT citing Reddit in roughly 60% of responses in early August 2025, collapsing to around 10% by mid-September, while Wikipedia dropped from about 55% to under 20% (Semrush). The Semrush AI Visibility Index separately found 40–60% of AI-cited sources change from month to month (Search Engine Land).

The mix changes by industry. In health queries, Surfer SEO and Amsive found the top AI Overview citations flip to nih.gov at 38.89%, youtube.com at 27.91% and healthline.com at 15.06% (Search Engine Land): the sources that matter in your category are not the sources that matter on average.

Format matters as well as domain. For commercial queries specifically, analysis by Glen Allsopp found "best of" list posts account for around 43.83% of ChatGPT citations (Ahrefs). If the comparison round-ups in your category omit you, the engines reading them will too.

The only reliable way to know which sources feed answers in your category is to look at the citations in real answers to real buyer questions. For the full breakdown of the source ecosystem, see where AI engines get their answers.

Why does consistency of mentions beat any single placement?

Consistency correlates with visibility because AI assistants draw on many sources per answer rather than one. Evertune's study of 75,000 brands found that brands present on four or more non-affiliated forums are 2.8x more likely to appear in ChatGPT responses (ZipTie). Ahrefs' own 75,000-brand study found that the number of pages on a website has extremely low correlation with ChatGPT visibility, while YouTube mentions and mention impressions showed the strongest correlation (Ahrefs). On those figures, how often independent sources mention you tracks AI visibility more closely than how much you publish on your own site.

This is also why AI visibility and Google rankings have come apart. In a 15,000-prompt study, Ahrefs found only 12% of URLs cited by ChatGPT, Gemini and Copilot appear in Google's top 10 for the same query (ZipTie). Ahrefs separately reports that around 80% of citations by large language models (the models behind ChatGPT, Gemini and Claude) come from pages that do not rank in Google's top 100 at all. Seer Interactive found 28% of ChatGPT's most-cited pages have zero Google organic visibility (Onely). Ranking well is not proof AI recommends you, and ranking badly is not proof it doesn't.

Because you cannot infer AI visibility from your search rankings, the only way to know where you stand is to measure it directly: run a free AI visibility check and see whether AI assistants name your brand in answers to non-branded questions (buyer questions that never mention your brand, the ones asked by people who do not yet know who to ask for) and which brands get named instead.

Do ChatGPT, Gemini and AI Overviews choose brands the same way?

No, the engines share the same broad mechanism but weight their sources differently, so the same brand can be visible on one engine and absent on another. The clearest measured divide is how closely each engine tracks Google's index. Authoritas and Optimizely studies found 40–76% of AI Overview citations also appear in the top 10 organic results (Search Engine Land). AI Overviews stay comparatively close to Google's rankings. ChatGPT sits further away: per the Ahrefs figures above, only 12% of the URLs cited by ChatGPT, Gemini and Copilot come from Google's top 10 for the same query.

Engine by engine, and with the evidence gaps stated plainly:

• Google AI Overviews draws heavily on Google's own index: a Rich Sanger / Authoritas study found links in organic position 1 had a 53% chance of appearing in AI Overviews, against 36.9% at position 10 (Search Engine Land).
• ChatGPT is the engine with the most published research behind it, and that research shows a weak link to Google rankings alongside a measurable association with third-party mentions: the Evertune forum finding and the Ahrefs YouTube-mention correlation above were both measured on ChatGPT alone.
• Gemini answers from live, Google-grounded retrieval. It appears inside the Ahrefs 15,000-prompt figure above, but we have not found a published study isolating Gemini's own source selection, so we make no engine-specific claim about it.
• Claude also retrieves and synthesises when it searches the web. We are not aware of a published citation study measuring Claude's source mix, so how it differs from the others is not something we can evidence.

One honest limitation applies to every engine: answers vary run to run. The SparkToro finding at the top of this piece (less than a 1-in-100 chance of the same brand list twice) means any single answer is a sample, not a verdict. Measuring visibility properly means asking many questions and looking at coverage rates, not screenshotting one good answer. It also means nobody, Skavora included, can promise to put you in an AI answer; what can be measured is how often you appear, and what the engines were reading when you did.

If you want the strategy side, what to change on and off your site once you know where you stand, see how generative engine optimisation (GEO) differs from SEO.

How can you improve your chances of being recommended?

You cannot control which brands AI assistants recommend, but you can improve your odds, because the retrieval-side signals above are observable and most of them are earnable. The working loop:

• Measure first. Run realistic, non-branded buyer questions through AI assistants and record which brands get named. Skavora's free check does this with 25 buyer questions generated from live research on your brand, run live with web grounding on one engine, Google's Gemini, and returns a scored report including the cited-source breakdown ranked by influence. The £9.99 five-engine report runs the same questions across Gemini, ChatGPT, Claude, Google AI Overviews and Google AI Mode.
• Keep your own coverage fresh and extractable. Direct answers under question-phrased headings, self-contained paragraphs, a real updated date.
• Earn mentions where the engines already look. Your scan's source list shows which review sites, communities, video channels and publications feed answers in your category: that list is your outreach plan.
• Re-scan periodically. With 40–60% of cited sources changing month to month per the Semrush AI Visibility Index, a snapshot ages quickly. Skavora runs scans on demand rather than monitoring continuously, so re-running the same questions is how you see movement.

Skavora runs 25 realistic buyer questions live through AI assistants and shows whether they recommend your brand, and who they recommend instead. The free check takes a few minutes, needs no signup, and covers one engine: Gemini with live web grounding.