01 Two ChatGPTs answer every question
The ChatGPT that answers from model memory draws on training data with a cutoff, describing your brand the way the public web described it months or years ago. The ChatGPT that answers in search mode retrieves live pages and cites them with links, and parts of that pipeline have historically drawn on Bing's index, the same infrastructure that feeds Microsoft Copilot's web answers. Users do not deliberately choose between the two, and the same question can route differently on different days.
The practical consequence: a brand can look strong in one channel and be invisible in the other. New companies tend to exist only in search mode. Older brands can coast on stale model memory that misstates what they now sell. Any serious ChatGPT visibility program begins by testing both channels separately and treating them as separate problems. The test is straightforward: ask your buyers' questions with browsing on and with it off, and record the difference. The gap between those two answers is your work list.
02 What makes ChatGPT name a brand
When someone asks for the best expense platform for a small agency, ChatGPT assembles candidates from comparison articles, review coverage, industry press, and its trained sense of who matters in the category. Brands that appear repeatedly across independent best-of and versus coverage, described consistently and with specific claims attached, get named. Brands that exist mainly on their own websites get summarized generically or skipped. The pattern holds across the categories we track: recommendation answers are assembled from the sources a diligent human researcher would read, then compressed. Win those sources and you win the compression.
So the inputs are unglamorous and compounding: press placements with backlinks in the outlets that cover your category, consistent entity facts everywhere your brand is described, and pages that state plainly who you are for. Those links do double duty, feeding both live retrieval and the training data behind model memory, which is the mechanism covered on our LLM SEO page.
03 Measuring share of voice in ChatGPT
One screenshot of a good answer proves nothing, because answers vary by prompt phrasing, user context, and model version. We measure instead with a fixed panel: the same buyer-intent prompts, re-run monthly, with results recorded for who gets named, in what order, with what description, and which sources search mode cites. Across enough prompts, a share of voice number emerges that is stable enough to steer by. We also log referral traffic from ChatGPT where analytics allows it, since citation clicks are a small but clean signal that an answer named you.
That panel is also the honesty mechanism. When your share rises, you see it. When a model update shuffles the category, you see that too, labeled as what it is rather than spun. Our full method is documented in measuring AI visibility.
04 The program that moves ChatGPT visibility
Monthly retainers combine press placements aimed at the publications ChatGPT's search mode actually cites for your category, entity cleanup so every source agrees on your facts, and quotable page structures for the claims only you can make. Model memory moves slowly, on training cycles. Search citations can move within weeks of coverage landing. We report the two channels separately so you always know which lever is working. The same placements tend to move Claude as well, since Anthropic's models learn from overlapping corpora and cite overlapping sources.
Where a category is crowded, we narrow the panel to the segments you can plausibly own first, because a strong share of a specific question beats a rounding error on a broad one. We never promise that ChatGPT will recommend you, because nobody controls that output. We promise the inputs it demonstrably favors and measurement you can audit. Retainers start at $3,500 per month with a minimum of five monthly placements, detailed on pricing.