How AI assistants choose which brands to recommend

Ask ChatGPT to recommend an accountant in your city, a project management tool, or a good local roofer, and it will answer with names. Confidently, in complete sentences, usually three to five of them. If your business isn't on the list, you've just watched a customer conversation happen without you, and unlike a search results page, there's no page two.

The consoling news is that these recommendations aren't mysterious. Over the past two years, researchers have analyzed hundreds of millions of AI citations to work out where assistants get their answers, and the mechanics turn out to be knowable, consistent per platform, and, for a small business willing to do specific work, influenceable. Here's how the sausage is made.

The two memories an AI assistant draws from

Every recommendation comes from one of two places, and the distinction decides what you can do about it.

Training data is the model's long-term memory: the internet as it existed when the model was trained, compressed into patterns. When a model answers without searching, it names the brands that appeared most often, most positively, in the most contexts, across that snapshot. You can't edit a trained model, and there's a lag of months to years. What you published and what others said about you in the past is what lives here.

Retrieval is the short-term memory: modern assistants with search, meaning ChatGPT's search mode, Perplexity, and Google's AI Overviews, fetch live web pages at answer time and compose their recommendation from what those pages say. This half works on a much faster clock, and it behaves like a strange new kind of search engine: it cares about which pages it can fetch, trust, and quote.

Both memories feed on the same underlying thing, which the citation data makes visible: what the written web says about your category, and whether you're part of that record.

Where the answers actually come from

The biggest mapping to date comes from Profound, which analyzed 680 million citations across ChatGPT, Google AI Overviews, and Perplexity over ten months to mid-2025. The headline finding is that each platform has a house taste.

ChatGPT leans on Wikipedia more than any other single source: 7.8% of all its citations, and nearly half of the citations among its top ten sources. Perplexity's favorite is Reddit, at 6.6% of its citations. Google's AI Overviews also cite Reddit more than anything else. Community discussion and reference material, in other words, carry startling weight: when an assistant recommends "what people say is good", it is often literally summarizing forum threads and encyclopedia entries.

Ahrefs' analysis of ChatGPT's top 1,000 cited pages adds the sobering structural detail: about two-thirds of those pages are types a marketer can't directly influence, Wikipedia at 29.7%, companies' own homepages at 23.8%, app stores and the like beyond that. The cited pages skew heavily toward trusted, established domains, with 65.3% sitting on domains in the top fifth of Ahrefs' authority scale. And Semrush's controlled study of 1,000 domains found a strong correlation, 0.65, between a site's authority score and how often AI platforms mention it. The rich get cited.

But three findings crack the door open for everyone else. First, freshness matters enormously: nearly 90% of ChatGPT's dated citations had been updated in the current year. Second, Ahrefs found 28.3% of ChatGPT's most-cited pages have zero visibility in classic Google rankings, so the citation game demonstrably isn't just the ranking game restated. Third, Yext's study of 17.2 million citations found that verified, structured, directly distributed business data, the kind that lives in listings and machine-readable formats, made up more than half of distinct citation sources. Assistants love clean facts from managed sources, and managing those is open to any business at any size.

Why Reddit keeps appearing, and what it signals

Reddit's dominance isn't accidental. In February 2024, Google announced a partnership giving it access to Reddit's data feed, described as "real-time, structured, unique content", for use across its products. The platforms decided that authentic human discussion is the scarce ingredient for answering "what's actually good" questions, and they piped it in at the source.

The signal for you: AI assistants are trying to reconstruct word-of-mouth. That's why review threads, comparison discussions, and community mentions punch above their SEO weight in AI answers. It's also why faking it is a dead end, both because platforms police it and because your entry in the machine's model of word-of-mouth is built from many small, uncoordinated mentions over time, which is exactly what can't be manufactured in a weekend.

What earns a recommendation, practically

Pulling the research into a to-do list, roughly in order of leverage for a small business:

• Be consistently describable. A model can only recommend an entity it can pin down. Same business name everywhere, a homepage that states plainly what you do, for whom, and where, and structured data marking it up. Remember homepages were nearly a quarter of ChatGPT's top citations: when an assistant checks who you are, your own site is the document it reads. (Making your business a clear entity to machines is its own discipline, and this batch's companion draft on entity SEO covers it in depth.)
• Show up in the managed-data layer. Yext's 54.53% figure is the case for boring hygiene: your Google Business Profile, the major directories for your industry, review platforms, all accurate and consistent. This is unglamorous, cheap, and directly upstream of what assistants cite.
• Exist where the discussions happen. Genuine participation where your customers already talk, answering questions in your specialty, being the business people name when someone asks for recommendations, feeds precisely the sources Perplexity and AI Overviews weight most. Earned mentions, not manufactured ones.
• Publish citable facts. The academic GEO study found pages with statistics, quotations, and citations gained up to 40% more visibility in generative answers. An assistant composing an answer needs concrete material to quote, and original data only you have is the most defensible version of it. Extractable, direct answers on your pages serve the same function.
• Keep it fresh. The near-90% updated-this-year figure makes content refreshing, long an SEO nicety, into an AI-visibility requirement. A yearly pass over your key pages is now table stakes.
• Don't abandon classic authority. The 0.65 correlation means the old game, being linked, referenced, and established, remains the strongest single predictor of the new game. AI search optimization is additive to SEO, not a replacement for it.

A thirty-day starter plan

The list above is a strategy; here's the first month of it, sized for one owner's spare hours.

Week one: hear what the machines currently say. Ask ChatGPT, Perplexity, and Google's AI mode the ten questions your customers would ask, "who's good at X in [city]", "what is [your business]", "best [service] for [customer type]", and write down every answer: mentioned or not, cited or not, facts right or wrong. This baseline costs an hour and turns everything that follows from faith into measurement.

Week two: fix what they got wrong. Every wrong fact traces to a source: an outdated directory entry, an inconsistent name, a stale page on your own site. Correct your homepage's plain-language description of what you do and where, complete your Google Business Profile, and sweep your top few directory listings into agreement. This is the highest-certainty work in the whole program, because you're not competing with anyone, just repairing your own record.

Week three: publish one citable page. Pick the question from week one's panel that assistants answered most vaguely, and publish the page that answers it with specifics no one else has: your real prices, your real timelines, a direct extractable answer up top. One genuinely quotable page in a thin answer-space does more than five generic posts in a crowded one.

Week four: earn one mention where assistants look. One genuine appearance in the discussion layer, a helpful, non-promotional answer in the forum or community where your customers actually ask, or a review request to your five happiest recent customers. Small, real, and aimed exactly at the sources the citation studies say carry weight.

Then repeat the week-one panel monthly and watch the baseline move. It moves slowly, these systems retrain and re-retrieve on their own schedules, but it moves, and the businesses measuring it are the ones that notice which efforts moved it.

Set expectations by query type

One calibration saves a lot of frustration. For head-to-head queries against giants, "best CRM", "best running shoes", the authority skew in the data means small brands are largely shut out today, the same way page one of Google was. But recommendations get dramatically more winnable as queries get specific: your service in your city, your specialty for a particular kind of customer, your product category with a qualifier. Those answers are assembled from thinner evidence, which means a few good sources mentioning you can be decisive. Specific and local is where a small business should check what the assistants currently say and aim first. Note also that answers vary between runs and phrasings: an assistant that skips you today may name you tomorrow, so judge progress on repeated checks, not one query.

Quick answers to the follow-ups

Can I pay to be recommended? Not organically, and no legitimate vendor can sell placement inside model answers. Money helps only indirectly, by funding the real inputs: content, data quality, and presence in the places assistants read. Anyone selling "guaranteed AI mentions" is selling weather.

Do the assistants know my website exists? Checkable in minutes: ask one directly "what is [yourbusiness].com?" and see whether the answer reflects your actual site or a guess. Then confirm the retrieval side, whether AI crawlers can and do fetch your pages, since a blocked or unreadable site is out of the citation game regardless of merit.

A chatbot said something false about my business. What can I do? Fix the sources first, stale directories, old pages, inconsistent facts, because answers regenerate from evidence and corrected evidence corrects future answers. For persistent, damaging errors, the platforms have feedback mechanisms, but the source-repair route is the one that reliably works.

Does being recommended actually bring customers? The click-through is only part of it: recommended brands see the payoff partly as direct visits and branded searches later, which is why measuring AI visibility means watching those shadows too, not just referral counts.

The one-line version

AI assistants recommend the brands that the written web describes most clearly, most credibly, and most recently: reference sites, community discussion, managed business data, and authoritative pages, blended per platform. You influence that not by tricking a model but by becoming well-documented, well-mentioned in the places assistants actually read, and quotable when they get there, starting with the specific and local queries where the contest is still open.