Know where ChatGPT's answer comes from before you track it
ChatGPT draws on four different sources, and only some of them respond to anything you do. Here is how to sort that out and measure AI traffic in your own analytics.
Comments
No comments yet. If something here matches or contradicts what you have seen on your own site, say so.
ChatGPT answers draw on training data, live search, licensed publishers and user context, and you can only influence some of them directly.Share
Blocking GPTBot does not remove you from ChatGPT search, which uses a separate crawler called OAI-SearchBot.Share
AI traffic benchmarks range from 1.26x to 4.4x organic, so compare conversions and revenue in your own analytics.Share
Ahrefs published an explainer this month laying out the four places ChatGPT gets its information: data it was trained on before release, live web searches it runs when it needs current facts, content from publishers OpenAI pays to license, and the user's own context such as the prompt, files, memory and connected apps. It reads like background, but it changes how you should read any AI visibility report, because a tracked prompt where your brand appears is a blend of those sources, and you can only influence some of them.
Where a ChatGPT answer comes from
The table in the Ahrefs piece is the useful part, because it says which sources a site owner can touch. Training data is the model's base knowledge, and your influence there is indirect and slow. If your site is public and doesn't block OpenAI's training crawler, your pages may end up in what future models learn from, but that only shows up in the next model, months later, and you can't see what it learned. Ahrefs also notes that OpenAI no longer publishes a breakdown of its training data. The last one was for GPT-3 in 2020, so nobody outside the company knows what is in the mix today.
Live web search is the source you can work on. Ahrefs says you influence it by ranking for the topic in the results ChatGPT pulls from and by letting OpenAI's search crawler, OAI-SearchBot, access your site. Licensed publisher content only responds to coverage in those publications, and user context only matters if the person brings you in with a pasted link, an uploaded file or a connected app.
The crawler detail trips up a lot of owners. GPTBot, which collects training data, and OAI-SearchBot, which serves search, are separate settings. Ahrefs says blocking GPTBot doesn't remove you from ChatGPT's search results, and that GPTBot is the most-blocked AI crawler in its study of about 140 million websites. If you blocked it a while ago to protect your content, check that you didn't also shut out the search crawler by accident.
What prompt tracking can and cannot tell you
Prompt tracking is worth doing, but it measures a mixture, and the tools can't tell you which ingredient produced a given mention. Our research notes put it plainly: clickable citations attached to a claim suggest live retrieval, a brand mention with no visible source may come from training data, and the two can mix, so you cannot prove the origin from the answer alone. A retrieval system also may not cite every fact it used, so an uncited mention can still involve search.
That should shape how you read the numbers. Moz argues that owners fixate on citations when they should care about mentions, and that the share of responses mentioning you across a large set of prompts is a better starting gauge than traditional ranking position. Moz also says these answers are not very likely to generate a click, so this is closer to awareness than to the old search funnel. I agree with that framing, and it adds a wrinkle: if a mention comes from training data, you can't act on it directly, whereas a mention that comes from a roundup on another site gives you something to do, namely getting into that roundup.
Coverage is the other trap. Semrush's methodology guide points out that a dashboard tracking hundreds or a thousand prompts can look comprehensive while most of those prompts represent the same few buyer situations. It also notes that the same prompt can produce different responses over time, so a mention that keeps showing up is stronger evidence than a single result. A smaller set that covers different buyer decisions beats a long list of near-duplicates.
Measure the return in your own analytics
Benchmarks for AI traffic quality disagree with each other, which is the best argument for looking at your own data. Semrush's traffic quality guide lists a Semrush study that found AI visitors worth 4.4 times traditional organic visitors, and a Siege Media analysis with a median AI-to-organic conversion rate of 1.26x. Shopify's Q2 2026 analysis found AI-referred shoppers about 80% more likely to convert, but watches converted at 2.4 times the organic rate while apparel was up 60%. Adobe's July 2026 data had retail AI visitors converting 60% better and travel 1% worse.
Siege Media also found that AI traffic matched or exceeded organic conversion rates on 72% of the sites studied, yet among sites with reliable revenue tracking, AI produced higher revenue per session on fewer than half. Conversion rate and revenue can point in different directions, so a benchmark from someone else's site tells you very little about yours.
What to check this week
Start with your robots.txt and see what you have said about GPTBot and OAI-SearchBot. Decide on each one separately, since one governs training and the other governs search. If you want ChatGPT to know your brand, Ahrefs suggests you probably want to leave training open, but that trade-off is yours to make.
Next, run a handful of prompts your buyers would plausibly use and look at whether the answer shows linked sources. Where it does, that points to live retrieval, and your work is ranking for that topic and appearing in the pages being cited. Where your brand is named with no source, treat it as something you can observe but not directly change.
Then build your prompt list around distinct buyer situations rather than rewordings of one question, and read mention share across the set instead of position on any one answer.
Finally, set up AI referral reporting in your analytics. Group visits from sources like chatgpt.com separately from direct and organic search, tag any links you control with UTMs, and compare conversions and revenue per session against your other channels. Do this before paying for more tracking, because it tells you whether the channel is worth the investment.
What we do not know yet
We don't have a reliable figure for how often ChatGPT uses live search versus its trained knowledge on commercial queries, and the one number in our notes is secondhand, so we're not repeating it. We also can't say how much any single source contributes to a given answer, and OpenAI no longer publishes what its models were trained on.
“Conversion rate and revenue can point in different directions, so a benchmark from someone else's site tells you very little about yours.”