How AI Assistants Decide Which Brands to Recommend

When someone asks ChatGPT "what's the best project management tool for a small agency," a name comes out the other end in two seconds. The interesting question isn't what it answered — it's how. Was that name baked into the model months ago during training? Pulled from a web search that ran while you waited? Both? Neither?

Getting this wrong is expensive. Teams pour effort into tactics that target the wrong mechanism — optimizing for a retrieval layer the assistant didn't even use, or chasing a training-data effect they can't influence on a useful timescale. So let's open the box. Not all the way — nobody outside the labs can — but far enough that you can make decisions instead of guesses.

Two completely different machines wear the same face

Every modern assistant blends two sources of knowledge, and they behave nothing alike.

Training data is the model's long-term memory. During pretraining, the model ingested a huge slice of the web, books, forums, documentation, and code, then compressed all of it into weights. It doesn't store sentences; it stores statistical associations. So when you ask about CRMs, the model has a learned sense that "Salesforce," "HubSpot," and "Pipedrive" are strongly co-associated with that category — because they appeared near those concepts thousands of times across the corpus. This is frozen. It reflects the web as it looked at the data cutoff, not today.

Retrieval is short-term, live memory. The assistant runs a search (its own index, Bing, Google, a partner API), pulls a handful of pages, and feeds the text into the prompt alongside your question. The model then reads those pages right now and answers partly from them. This is what powers Perplexity end-to-end, what ChatGPT does when it decides to "search the web," and what Google's AI Overviews lean on heavily.

Here's the distinction that actually matters for your work:

If a brand is recommended from training data, you influenced that 6 to 18 months ago and you can't change it this quarter. If it's recommended from retrieval, you can influence it this week — by changing what the open web says about you right now. Almost all the leverage you have lives in the retrieval layer. Optimize there first.

How a shortlist actually forms

The model doesn't "look up the best tool." It generates a list, and that list is shaped before a single word is produced. Roughly, three things happen in sequence.

1. Candidate generation. The model assembles a pool of names that are plausibly relevant. From training data, this is whatever has the strongest associative pull for the query's concepts. From retrieval, it's whoever shows up in the fetched pages. A brand absent from both simply cannot be recommended — there is no path. This is the single most common reason a good product never gets named: it never entered the candidate pool.

2. Filtering and ranking. The model down-weights candidates that look risky, contradicted, or thinly supported, and up-weights ones that are corroborated across multiple sources, recently mentioned, and described in concrete terms. A brand named once in one blog post is fragile. A brand named across ten independent sources, with consistent positioning, is hard to dislodge.

3. Framing. The survivors get ordered and described. Whether you're listed first, called "best for enterprise," or buried as "another option" depends on the language the sources used about you, not just whether they mentioned you.

The uncomfortable implication: you can be in the candidate pool and still lose at step two or three. Being mentioned isn't the goal. Being mentioned consistently, recently, and concretely is.

Authority and consensus: the real ranking signals

Classic SEO trained everyone to think about backlinks and domain authority. Those still matter, but for answer engines the dominant signal is consensus — does the open web broadly agree that you belong in this category?

LLMs are, at their core, agreement-detection machines. When twelve different sources independently say "Notion is good for team wikis," that repetition becomes a strong statistical signal the model trusts. One source making a bold claim about itself is noise. Many independent sources converging is signal. This is why a single brilliant landing page rarely moves the needle and why being named in roundups, comparison articles, Reddit threads, and review sites does.

A few specific authority signals that punch above their weight:

• Third-party comparison and "best X for Y" listicles. These are retrieval gold because they're literally structured as candidate lists. If you're absent from the listicles that rank for your category, you're absent from the shortlist.
• Community and forum content — Reddit, Hacker News, Stack Overflow, niche Discords that get indexed. Assistants weight these heavily because they read as unbiased peer opinion. They're also where consensus visibly forms.
• Structured, citable facts — Wikipedia, Crunchbase, G2, Capterra, well-maintained docs. These give the model unambiguous, low-risk statements to anchor on.
• Your own site, but only the parts that read like reference material. Specs, pricing, comparison pages, FAQs. Marketing prose gets paraphrased away; concrete facts get quoted.

Notice what's missing: clever copywriting, brand voice, hero animations. The model can't see your design. It reads text and weighs how often, how consistently, and how recently that text appears across the corpus and the live web.

Why recency and citations carry so much weight

In retrieval mode, the assistant is reading fresh pages, so recency is a tiebreaker that often decides everything. A comparison article updated last month outranks a definitive one from 2023, because the model — and the search index feeding it — both treat freshness as a proxy for correctness. Categories move fast; the assistant assumes the newest source knows about the latest entrant or price change. If your category's top sources haven't mentioned you since your last rebrand, the assistant is describing a version of you that no longer exists.

Citations matter for a structural reason people miss. When an assistant cites sources (Perplexity always, ChatGPT and Gemini often), it is constraining itself to say things those sources support. That makes citable, well-sourced claims about your brand disproportionately powerful: they're the raw material the assistant is allowed to repeat. A claim no source backs up is a claim the cited-mode assistant won't make, no matter how true it is.

There's a second-order effect. Once an assistant cites a page that positions you well, users click it, other writers find it, and it gets re-cited. Citations compound. The brands that win in answer engines tend to win progressively — each mention makes the next one likelier.

What this means you should actually do

Stop optimizing for one monolithic "AI." Different assistants weight the two machines differently, and your tactics should follow.

For retrieval-heavy assistants (Perplexity, AI Overviews, ChatGPT-with-search): Win the open web now. Get into the comparison listicles. Earn genuine community mentions. Keep your reference pages — pricing, specs, comparisons, FAQs — accurate, dated, and crawlable. Make sure third parties describe you in the exact language you want repeated, because that language is what gets repeated.

For training-heavy answers (ChatGPT with search off, base-model responses): Play the long game. Consistent, broad presence across the web over months is what eventually gets compressed into the weights. There's no shortcut, and anyone selling you one is selling you nothing.

A practical sequence I'd run:

• Audit what assistants currently say about you and your category. Ask the same buying question across ChatGPT, Perplexity, Gemini and Claude. Note who gets named, in what order, with what framing. This is your real baseline — not your search rankings. (This is exactly what a free AEOeye audit automates: it queries the major assistants for your category and shows where you appear, where you're invisible, and who's eating your shortlist slots.)
• Find the sources the assistants cite. Those pages are your highest-leverage targets. Getting added to, or favorably updated in, a frequently-cited comparison article moves you faster than ten new blog posts on your own domain.
• Fix the consensus gap. If the web's language about you is stale, inconsistent, or thin, that's your work. Align how third parties describe you, and refresh the facts that matter.
• Re-measure monthly. Retrieval-driven answers shift week to week. Treat AI visibility as a metric you track, not a project you finish.

The honest limits

Be skeptical of anyone claiming a deterministic playbook. These systems are partly stochastic — ask the same question twice and the wording, sometimes the names, shift. The labs change retrieval logic and ranking without announcement. No one outside Anthropic, OpenAI, or Google knows the exact weighting between training and retrieval for a given query.

But the mechanism is knowable enough to act on. Candidate pool, consensus, recency, citability — those four levers are real, observable, and influenceable. You don't need to crack the black box. You need to make sure that when the assistant assembles its shortlist, you're in the pool, you're corroborated, you're current, and you're described in words you'd be happy to have repeated to a buyer. Do that consistently and you stop hoping to get recommended. You start engineering it.