Blog · 25 September 2026 · By HeardOf

Training data or live search: why ChatGPT can know your brand in one mode and not the other

Does ChatGPT use live search?

Sometimes, and ChatGPT decides. OpenAI's help page "Searching the web with ChatGPT", in the Internet Archive's capture of 19 September 2026, says "ChatGPT may search the web automatically when your question would benefit from current information", and that you can start a search yourself from View all tools, or by entering "/" and selecting Search. So if you typed "does chatgpt use live search" into a search box, the answer on OpenAI's own page is: when it judges the question needs it, or when you ask.

The same page says that when ChatGPT search partners with other search providers it "typically rewrites your query into one or more targeted queries" and sends them to those providers, and it points to Microsoft's and Shopify's privacy policies; that "Responses that use web search may include citations"; that "People who are not signed in can also use web search"; and that a site should "allow OAI-Searchbot to crawl the site" to be eligible.

What the page does not say is what an answer without a search is built from. It offers a way to add one afterwards: "To regenerate an existing response with web results, select the refresh control and choose Try again or Search the web, when available." We read that as an existing response that did not use the web. That is our reading, not a sentence of OpenAI's.

We read the archive's copy because the live page answered our request on 25 September 2026 with HTTP 403 and a Cloudflare challenge script, and we do not read past a bot check. The newest copy we could read is the archive's capture of 19 September 2026, at the address the article now carries under its current title, "Searching the web with ChatGPT"; captures at its old address after 6 August 2026 are recorded as 403.

What does ChatGPT answer from when it does not search?

From the model's training, by Evertune's description, since OpenAI's archived page does not say. Evertune's help page "AI Models Available", dated 13 May 2026, historical, says that in the ChatGPT app "the AI model either generates the answer from its base model knowledge … or it retrieves the answer from live search", and lists August 2025 as the training cutoff for ChatGPT-5.4-mini. That cutoff is Evertune's figure; OpenAI's page, as captured on 19 September 2026, prints no cutoff.

Anthropic states the mechanism for its own model. Its web search tool page, read 25 September 2026, says the tool gives Claude "up-to-date information beyond its knowledge cutoff", that with the tool added "Claude determines when to search based on the prompt", and that it searches for "Information about specific organizations, people, or products that might have changed", while it "answers directly without searching when the request draws on stable knowledge". Anthropic's models page, the same day, prints a training data cutoff for each of its current models: July 2025 for Claude Haiku 4.5, January 2026 for Claude Sonnet 5, June 2026 for Claude Opus 5.5 and Claude Fable 5.1.

That is how one brand gets two answers, and this paragraph is our reasoning from those pages, not a measurement. A brand that launched, renamed itself or changed what it sells after a model's cutoff is not in that model's training as it is now. A search the same day can find its current pages, if the search reaches them. Which of the two you get depends on whether ChatGPT searched, which it decides unless you ask it to.

Do Perplexity and Google AI Overviews search every time?

Perplexity says it does; Google says AI Overviews often don't show at all, and that when they do they may run several searches and link only to indexed pages. Perplexity's help page "How does Perplexity work?", last modified 3 September 2026 per the page, says "When you ask Perplexity a question, it uses advanced AI to search the internet in real-time", and that "Each answer includes numbered citations". It names GPT-5 and "Claude 4.6 Sonnet" among the models it uses. The live page answered us on 25 September 2026 with a Cloudflare "Just a moment" page, so we read the Internet Archive's capture of 14 September 2026.

Google's page "AI Features and Your Website", last updated 10 December 2025, historical, says AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources", and that to be a supporting link a page "must be indexed and eligible to be shown in Google Search with a snippet". It also says AI Overviews "often don't trigger". The same page keeps training apart from Search: Googlebot rules are the control for Search, AI features included, and Google-Extended is where it points to limit "AI training and grounding in some of Google's other systems".

What does a free brand check actually measure?

The two we read say, on their own pages, that they measure training data. Our comparison with Otterly, reading its pages on 23 September 2026, recorded that its free check, an AI Brand Authority Check, gives a 0–100 score "that its page says reflects training data rather than live answers" (HeardOf vs Otterly). We re-read that page on 25 September 2026. Its footer says "Results reflect AI training data, not real-time web". Its header says "Live · Claude + GPT-5.4 + Google KG", and its tiles say "Real API" and "Live, Not Cached". The page says "We query four sources in real time" in one place and "3 Sources Checked" in another, and lists three: Claude Sonnet "via Anthropic's training data", GPT-5.4 "via OpenAI's corpus", and Google's Knowledge Graph Search API. We read the "Live" labels as describing the call and the footer as describing what the answer is drawn from; that is our reading, not Otterly's.

HubSpot's AI Search Grader, read the same day, is described on its page as "a free, one-time check that reveals what ChatGPT, Perplexity, and Gemini say about you based on their training data", and names GPT-5.4 mini as the OpenAI model it asks.

Both say so on their own pages. In our reading, a score built this way can tell you whether a model's training holds your brand, and cannot tell you whether a search run today would find you.

Which model answered, and do the tools say?

Six of nineteen AI visibility tools name a model version on their own pages, in our census of how nineteen tools get their answers, read 24 September 2026: Peec, Rankscale, Evertune, Goodie, Ahrefs and HubSpot. We re-read the six vendors' version pages on 25 September 2026, and each still carries its version string. The other thirteen named none on the pages we read on 24 September; we did not re-read them.

In our reading, for an answer given without a search, the model is the date. Anthropic's current models alone span training data cutoffs from July 2025 to June 2026, on the page cited above. Naming the mode is a different thing from naming the model: Semrush says its toolkit reads "the latest model of ChatGPT in search mode", which is a mode and a product, not a version. Peec says that for most platforms "the platform decides which model to use and whether it runs a web search", because it reads the app as a logged-out user would. Ahrefs asks Claude "with web search" and the rest "using the default model".

What does HeardOf ask, and do our calls search the web?

We ask ChatGPT through OpenAI's API on gpt-5.6-luna, Perplexity through its API on sonar and Gemini through Google's API on gemini-2.5-flash, with model names and routes as read on 19 September 2026, and we read Google AI Overviews from a Google results page fetched for us by SerpAPI, as of 23 September 2026 — all per our post on which engines we check, as updated 23 September 2026; whether those calls search, that post does not say.

That post does not say whether the three API calls have web search switched on, and so this post does not say either way. What it does print is the model behind each of the three API engines' answers and the date it was read, which is the half a reader needs to place an answer that came from training. The AI Overviews answer is a block on a Google results page, which Google's page above says may use a "query fan-out" of related searches, and whose supporting links must be pages in its index.

The free check on our homepage asks one engine, Gemini, three prompts, about one domain, as read on 19 September 2026, per the same post, and the post does not say whether that call searches either.

How do you tell which route a tool is reading?

Ask it three things, each answered for itself by at least one page above. Which route: search on, search off, or whatever the app decides, as Peec and Semrush state. Which model version, as six of nineteen tools did in our read of 24 September 2026. And which day, because a search reads the web as it stood that day and a model's training stops at its cutoff. A tool that answers all three lets you tell whether a missing brand is missing from the model or from the search. A tool that answers none leaves you with a number and no way to tell which.

HeardOf vs Peec AI: what each one measures, and what Peec does better