Two ways AI knows your brand: training data vs live retrieval

When ChatGPT names your brand, or fails to, the words on screen can come from two entirely different machines wearing the same interface. The first machine is the language model itself: billions of parameters that absorbed a snapshot of the web, books and other text during pretraining. Whatever that corpus said about your brand, accurate or stale or absent, is compressed into the model's weights and stays there until the vendor trains a new model. The industry calls this parametric knowledge. You can think of it as the assistant's long-term memory, formed at a moment in the past you did not choose.

The second machine is a search pipeline bolted to the model. When a question looks like it needs current facts, the assistant issues search queries, fetches indexed pages, and hands passages to the model to summarize, usually with citations attached. This is live retrieval, often called RAG, and it changes the physics of the problem completely: the answer now depends on what is indexed and citable this week rather than on what was true at training time.

Marketers who treat AI answers as one system end up with one blurry strategy. Treating the channels separately produces two sharp ones, with different levers, different timelines and different owners. The rest of this article takes the channels one at a time, shows how to tell which one produced any given answer, and lays out what each channel rewards.

A brand enters training data by being written about. Pretraining corpora sweep up encyclopedias, forums, review sites, news, documentation and the broad crawlable web, so the models tend to know brands roughly in proportion to their independent written footprint. A company discussed across Wikipedia, Reddit threads, review platforms and trade coverage gets encoded richly. A company whose entire written presence is its own website gets encoded thinly or missed, which is one reason obscure but well-documented open source tools often outshine funded startups inside model memory.

Two properties of this channel shape everything you can do about it. First, it is frozen. A model trained before your repricing will describe your old prices with total confidence, and no amount of updating your website reaches into its weights. Stale answers about pricing, discontinued features and old positioning are the signature failure of this channel, and the correction paths for them are covered in when AI gets your brand wrong. Second, it is unscheduled. Vendors retrain when they retrain; no roadmap tells you when the next model absorbs the last two years of the web, so any strategy aimed at this channel has to be comfortable paying off at an unknown date.

The channel still matters enormously, for a quiet reason: it sets the assistant's priors. Baked-in knowledge shapes which brands a model treats as the natural defaults of a category, how it frames retrieved snippets, and what it says on the many prompts where search never fires at all. Retrieval decorates the answer; the model's memory decides its center of gravity.

The retrieval channel runs on infrastructure you can watch move. ChatGPT Search launched on Bing's index, and OpenAI now operates its own crawler, OAI-SearchBot, with a crawl that has roughly tripled since August 2025 according to Botify's analysis. Perplexity retrieves from its own index, reported at more than 50 billion pages, and cites generously: about 8.2 sources per answer, roughly 3.4 times ChatGPT's citation density. Google's AI features cite from Google's organic index, which makes being indexed and rankable in classic search a precondition for appearing in them. The full pipeline for OpenAI's side is traced in how ChatGPT Search works.

What the channel cites is measured too, and the diet is lopsided. 5W Research found Wikipedia at 13.15 percent and Reddit at 11.97 percent of ChatGPT's U.S. citations, together over a quarter of the total, while the Wall Street Journal, New York Times and Bloomberg miss the top 20 entirely. Community and reference content beats prestige media in this channel, which should recalibrate where you spend earned-media effort.

The property that makes this channel strategically precious is its clock speed. A page you publish today can be crawled, indexed and cited inside a couple of weeks. The same gap that is unfixable in the training channel until an unannounced retrain is a two-week content task in the retrieval channel. That asymmetry is the entire reason AI visibility is a workable discipline rather than a waiting game.

The diagnostic is usually one glance: citations. An answer that links sources, quotes current numbers and names recent events came through retrieval, and the cited pages tell you exactly which documents to influence. An answer with no citations, generic phrasing and hedges about recency is the model speaking from memory. Between those poles sit blended answers, where memory frames the narrative and retrieval fills in specifics, which is the common case for buying questions on modern assistants.

A few probes sharpen the diagnosis. Ask the same question with web search enabled and disabled where the interface allows it, and diff the brands named. Ask for current pricing: a memory answer confidently states stale numbers, while a retrieval answer either cites a source or declines. Ask about something that changed in the last six months; only retrieval can know. Run each probe more than once, because answers vary run to run even with settings held constant. SparkToro measured under a 1 percent chance that two identical ChatGPT runs return the same brand list, a problem dissected in why AI gives a different answer every time you ask.

The channel split is why Reachroller stores the full answer text for every tracked question rather than a bare mention count. Reading the stored answers tells you whether your losses are retrieval losses, where rivals' pages are being cited and yours are absent, or memory losses, where the model simply does not consider you part of the category. The two diagnoses trigger different work, as the next sections show.

The channel mix is an engine-level personality trait, and the usage numbers hint at how much of each assistant's output never touches the web at all. ChatGPT reached roughly 900 million weekly active users in early 2026, while ChatGPT Search processes an estimated 250 to 500 million weekly queries. Those are different units and rough estimates, but the gap suggests a large share of ChatGPT conversations, including plenty of brand and category questions, are answered from model memory without a single retrieval. For those answers, your website is irrelevant and your training-data footprint is everything.

Perplexity sits at the opposite pole. It is retrieval-first by design, runs every answer through its own index of more than 50 billion pages, and cites about 8.2 sources per answer. Winning Perplexity is therefore almost purely a retrieval game, and the fastest engine to test fixes on. Google's AI Overviews and AI Mode are retrieval-native too, but with a specific dependency: they cite from Google's organic index, so classic indexing and rankability remain the entry ticket. Claude and Grok invoke search on demand, which makes their mix question-dependent in ways worth testing rather than assuming.

The practical consequence: your channel priorities should follow your engine priorities. A brand whose buyers live in Perplexity and Google surfaces can win mostly through the fast retrieval loop. A brand whose buyers ask ChatGPT conversational questions with search disabled needs the slow footprint work more urgently, because memory is answering. This is also why Reachroller reports scores per engine rather than blended: the same brand routinely holds strong retrieval-backed visibility on one engine while remaining invisible to another engine's memory.

Retrieval strategy is publishing strategy with a deadline you can actually see. The loop has four steps. Identify the specific unbranded questions where assistants cite rivals and omit you. Publish a page that answers each question directly and citably. Get the page indexed by Google and Bing, since the engines read search indexes and a page outside them does not exist for this channel. Then recheck the answer after one to two weeks and see whether it flipped.

What makes a page citable is measured rather than mystical. The Princeton-led GEO study, published at KDD 2024, tested nine optimization methods and found that adding quotations, statistics and cited sources boosted a page's visibility in generative engine responses by up to roughly 40 percent, while keyword stuffing performed near the bottom. Structure matters too, with an honest caveat: SE Ranking found about 71 percent of pages cited by ChatGPT carry structured data, a correlation rather than proven causation, while an Ahrefs study of 1,885 already-cited pages in May 2026 found adding schema produced no measurable citation lift on those pages. The section-by-section anatomy of a citable page is in how to write content AI engines actually cite.

This loop is precisely what Reachroller automates. The tracking finds the lost questions, each generated fix ships publish-ready with the URL slug, title tag, meta description, schema markup and indexing steps, and a scheduled recheck shows whether the stored answer now names you. One credit is one AI answer, a generated fix costs ten, and plans start at $29 per month; the whole mechanism is on the how it works page. The deliberate scope note: Reachroller works this channel through official engine APIs only, with ChatGPT live today and Claude, Gemini, Perplexity and Grok built and rolling out.

You cannot edit a model's weights, but you can shape what the next training run reads. The lever is the independent written footprint: reference sites, community discussion, review platforms, comparisons and coverage that mention your brand in useful, factual contexts. The same sources that dominate citation studies are strong candidates for training corpora, so work that earns third-party mentions pays into both channels at once. That double payoff is the case made in digital PR for AI visibility.

Community presence deserves particular weight. Reddit's share of AI citations grew roughly 73 percent in commercial categories across 2025 and 2026, and community threads about tools and vendors are exactly the kind of text that teaches models which brands belong to which problems. Earned, authentic participation compounds here; astroturfing gets brands banned and poisons the record permanently.

Set expectations accordingly. This channel has no recheck date, no two-week loop, and no dashboard that shows the needle move next month. It is closer to brand equity than to campaign work: unglamorous, cumulative, and decisive in the long run because it determines what every future model believes your category looks like the day it ships. Fund it steadily, and measure it by the footprint itself, counting the independent pages that describe you accurately, rather than by short-term answer changes it cannot promise.

In practice you are never choosing one channel. Assistant answers to buying questions are blends: memory proposes the frame and the familiar names, retrieval supplies the fresh specifics and the citations, and the mix varies by engine, by question and by run. A sane brand strategy therefore runs both plays in parallel, sized by their clocks: the retrieval loop as the quarterly operating cadence, because it produces measurable flips in weeks, and the footprint work as the always-on background investment, because it compounds into every future model.

Measurement has to respect both clocks too. Retrieval progress shows up as flipped answers on rechecked questions, and it is the number a monthly report can honestly claim. Training-data progress shows up as slow drift in uncited answers over quarters, and pretending to control its timing is how this industry loses credibility. Track both, promise only the first.

The practical starting point is a baseline: run your buyer questions across the engines, store the answers, and sort your losses into retrieval losses and memory losses by reading what got cited. That sorted list is your strategy, written by the engines themselves. Reachroller produces it in an afternoon with evidence attached, and the manual version is a well-spent day that will convince you the two-channel split is real before you automate it.