# Why We're Launching Claude Lite: Most of Claude's Buying Answers Never Touch the Web Claude answers most buying questions from memory, without searching the web. Claude Lite tracks those answers and shows how Claude describes your brand. **Published:** October 5, 2026 **Author:** Ivan Slobodin & JP Garbaccio --- Ask Claude which CRM a 20-person sales team should use and you'll get three or four names back, stated with confidence. It's natural to assume Claude ran a search first, read a few pages, and picked the brands those pages recommended. Most of the time, it didn't. On buying questions, Claude usually answers from what it already knows, a picture of the category it built during training, without reading a single page. That matters because AI search tracking has been built around the search path. It follows which pages an engine reads and cites, and how those citations move when your content changes. The search path is real and worth measuring, but it isn't the one Claude takes for most commercial questions. When Claude answers from memory, what counts is whether your brand is in that memory, and citation tracking has no way to see it. Below, we share what we found across 1,374 prompts, why it leaves a gap in how Claude gets tracked (by us too), and how Claude Lite and Grok Lite measure the answers that have gone untracked. ## How Often Claude Searches Before Answering In July 2026, we ran the same 1,374 prompts through the consumer Claude app and ChatGPT. For each prompt, we captured the network traffic to see whether the engine searched and, when it did, which queries it ran and which pages it fetched and cited. ChatGPT searched on 89% of prompts, so it goes to the web for almost everything. Claude searched on 48%, which means more than half of its answers came from memory alone. The gap widens as questions get closer to a purchase. The table below shows how often Claude answered without searching for each type of question. | Question type | Answered without searching | | :---- | :---- | | Advice ("what should I do about...") | 97.5% | | Comparisons ("X or Y?") | 88.7% | | Reputation ("is this brand any good?") | 86.7% | | Software and SaaS | 61% | | Shopping | 29% | | Local ("near me") | 2% | The questions that do send Claude to the web have answers that change week to week, like what's nearby, what's in stock, or what's in the news. The questions where a buyer chooses between you and a competitor mostly don't, so Claude answers those from memory. We also checked that this wasn't down to how the prompts were worded. We rewrote 16 of the buying questions 10 different ways each. ChatGPT searched on every version, while Claude searched on one version at most for each of the 16 questions. ## Claude Gives 2 Kinds of Answers to Buying Questions Claude answers a buying question in one of two ways, and each one rewards something different. ### When Claude searches the web When Claude searches, it reads about seven pages and cites about three. Its results line up with Brave's index far more closely than Google's. Of the domains Claude fetched, 63% were in Brave's top 10 for the same query, compared with 34% in Google's top 10. For marketers, winning this answer works much like SEO, just against an index most teams don't watch. You need to be on the pages Claude reads and be the source it credits, and content changes can move your results within weeks. ### When Claude answers from memory When Claude doesn't search, there's no page to rank and no citation to win. The answer reflects what the model learned about your brand and your category before its training cutoff. That picture builds up over years, from how widely and how credibly your brand has been written about, and it mostly shifts when a new model version comes out. Most brands, and most tracking tools, only watch the searched answer. On Claude, that's the smaller share of commercial answers. ## The Gap in How Claude Gets Tracked Tracking tools, ours included, measure Claude through the API with web search switched on for every prompt. That's the right setup for the searched answer, because it shows which pages get cited and whether your content work is landing. However, it also means tracking only ever sees the searched version of Claude. A buyer who asks a comparison question in the Claude app gets the memory version almost nine times out of 10. If you only track the searched path, you measure one of Claude's two answers well and miss the other one entirely. One more finding shaped how we built the fix. AI answers vary from one run to the next, and in our stability analysis, 98.7% of brand appearances changed between repeat runs. That's why [AI visibility tracking](https://www.searchable.com/blog/ai-visibility-tracking) depends on running the same prompts again and again, since a single check of what Claude says about you tells you very little on its own. Anything that measures the memory answer has to work the same way and report the trend. ## How Claude Lite Measures the Memory Answer We're launching Claude Lite to measure the memory answer directly. It answers from the model's training data only, and nothing is looked up when the prompt runs. The results show the picture of your brand that Claude already holds, so you can see which brands it mentions in your category and how it describes yours, mistakes included. We built Lite for the other half of Claude's answers, and it differs from full tracking in four ways. - **It measures the memory answer directly.** That's the path Claude takes for most comparison, advice, and reputation questions. - **It shows your brand as the model learned it,** including outdated pricing or discontinued products. Knowing what the model picked up about you is where you start if you want to [change what AI says about your brand](https://www.searchable.com/blog/ai-brand-sentiment). - **It runs on your project's normal schedule,** so you read a trend across repeated runs instead of relying on a one-off snapshot. - **It separates your two sources of visibility.** With full and Lite side by side, you can see how much of your visibility comes from your content and how much comes from your reputation. Lite also costs less to run, because there's no web search behind each prompt. It's priced at $0.50 per prompt per month, compared with $2.00 for full tracking. The lower price comes from how it works, but the reason to track it is the research above. ### How to read full Claude and Claude Lite results together Each engine is useful on its own, but the clearest signal is the gap between them. | Full Claude | Claude Lite | What it means | | :---- | :---- | :---- | | Strong | Weak | Your content performs, but your brand isn't embedded in the model yet. You show up when Claude searches and go missing when it doesn't. | | Weak | Strong | You have an established reputation that your current content isn't making the most of. | | Strong | Strong | You're covered whichever path Claude takes. | | Weak | Weak | You're missing from the shortlist on both paths. | ### What Claude Lite won't show you Because Lite measures memory, some of its data is thinner or less reliable than what a full engine gives you. - **Source data is sparse.** Sources, Query Fanout, and Gap Analysis will look thin for Lite engines, and that's expected. - **Links are unverified.** Any links Lite returns come from the model's memory, so they may not point to a live page. - **Established brands show up more.** Lite favors brands that have been written about widely over a long period. Newer brands and recent launches may appear less, and that gap is part of what it measures. Visibility, share of voice, and sentiment work the same way as for any other engine, and Lite uses your existing prompt allowance. ## Grok Lite Uses the Same Approach The logic isn't specific to Claude. Every model holds a trained picture of your brand, and when it doesn't search, that picture is the answer. Grok has the same split, so we built Grok Lite the same way. There's a full Grok engine that searches the web for each prompt and a Lite engine that answers from training data, priced at the same $2.00 and $0.50. Our capture study only covered Claude and ChatGPT, so we don't have a search rate to share for Grok. The method still applies, though. If you track the memory answer and the searched answer separately, the gap between them shows where your visibility on Grok comes from. ## How to Add Claude Lite and Grok Lite Claude Lite and Grok Lite are available now. You can add them from the engine list in your project, or talk to your account manager to set them up. New to Searchable? You can add Claude Lite or Grok Lite when you choose a plan on our [pricing page](https://www.searchable.com/pricing). --- [Back to Blog](https://www.searchable.com/blog) | [Searchable Homepage](https://www.searchable.com)