New: ChatGPT now sends real traffic to the websites it recommends. See why it matters →

How we estimate ChatGPT traffic

ChatGPT publishes no data on what people ask or which websites it recommends. So every number on this site is an estimate, built by anchoring to something that is measured: real search demand on Google. This page explains exactly how, and links the studies behind each assumption so you can check our work.

The idea in one line

Real demand for a topic is already measured by Google. ChatGPT's product-recommendation traffic is a roughly fixed fraction of that demand, so we can estimate it from Google's numbers instead of guessing how many prompts ChatGPT receives.

How the estimate works

Five steps, from measured Google demand to an estimated dollar value per website.

  1. 1

    Start with real Google demand

    For each commercial category, we take the monthly Google search volume for the matching "best X" query (for example, "best car insurance"), measured by Keywords Everywhere. This is real, observed demand, not a model output. Google handles more than 5 trillion searches a year, which is the bedrock the whole estimate stands on.

  2. 2

    Scale it to ChatGPT demand

    People ask ChatGPT for recommendations far less often than they search Google, but in a stable proportion. We divide Google demand by 8. That ratio lines up with Ahrefs' finding that ChatGPT handles about 12% of Google's search volume (1 ÷ 0.12 ≈ 8). So "best car insurance", searched on Google by hundreds of thousands of people a month, implies tens of thousands asking ChatGPT the same thing.

  3. 3

    Ask ChatGPT, five times

    We put each category to ChatGPT and record which websites it recommends and in what order. We run every category five times, because AI recommendations vary noticeably from run to run. A site has to show up in at least two of the five runs to count, and we score it by how often, and how highly, it appears.

  4. 4

    Split the demand across the recommendations

    Not everyone who asks clicks through, and the top recommendation gets far more attention than the tenth. We divide the estimated demand across the recommended sites using a click-through-rate-by-rank curve: higher positions get a larger share, tapering down the list. We keep the absolute click rates deliberately conservative, because academic studies of AI search find users click well under one link per session. (In May 2026 ChatGPT began showing clickable links inside its answers, which is making this outbound traffic larger and more measurable over time.)

  5. 5

    Turn visits into value

    Finally we multiply each website's estimated visits by the category's cost-per-click (the advertising price of that audience, again from Keywords Everywhere) to get an estimated monthly dollar value. It answers a simple question: what would this ChatGPT visibility cost to buy as ads?

A worked example

Illustrative figures, to show the arithmetic end to end.

Prompt category
best car insurance
Google searches for “car insurance” / month (measured) 450,000
÷ 8 → people asking ChatGPT / month (estimated) ≈ 56,000
ChatGPT's recommended sites, across 5 runs, split by rank:
1geico.com largest share → est. visits
2progressive.com smaller share
3statefarm.com smaller still
… tapering down the ranked list
Each site's visits × cost-per-click (estimated) = est. monthly value

Why anchor to Google?

Other approaches guess at ChatGPT prompt volumes from small user panels or by inventing synthetic prompts, neither of which gives a reliable count. Google search volume, by contrast, is a real, decades-old measurement of how much demand exists for a topic. Tying our estimate to it means the relative picture, which websites win a category and by how much, rests on solid ground even where the absolute numbers carry more uncertainty.

It is the most rigorous proxy available today, but it is still a proxy. The next section is honest about where the uncertainty lives.

Assumptions, and honest caveats

The 8× ratio is a midpoint

It matches the most-cited like-for-like estimate (Ahrefs' 12%), but the honest range is wide: roughly if you count all ChatGPT activity, up to 18–26× on the strictest "search-only" definition. We use 8× as a defensible middle, and ChatGPT is trending more search-like over time, which keeps it reasonable going forward. It is also a global average: the real ChatGPT-to-Google ratio varies by topic, since ChatGPT skews informational while Google still owns most commercial and local intent. Peer-reviewed clickstream research finds LLM traffic share runs several times higher for complex, research-heavy categories than for simple ones. We apply one ratio across every category today; per-category calibration is a planned refinement.

Click-through rates are estimated, not measured

No one has published a per-position click curve for ChatGPT. We base the shape on measured search click-through behavior and keep the totals conservative, in line with the limited ChatGPT click data that has leaked and with academic findings on AI search. Treat absolute visit counts as order-of-magnitude; treat the rankings as robust. And since May 2026, ChatGPT places clickable links inside its answers, which roughly doubled referral traffic to websites (Similarweb, Profound) and makes AI recommendations matter more than ever. Our curve is still anchored to pre-change measured data, so today's estimates likely understate current post-May-2026 traffic. We would rather under-show than overstate.

We measure recommendation, not logged clicks

This site tracks which websites ChatGPT recommends, and estimates the traffic that implies. We do not have access to anyone's actual ChatGPT conversations or click logs, and we never will. A category here represents the real demand behind the many ways people ask about a topic, not a single logged prompt.

We strip out brand and navigational noise

We focus on non-branded discovery, the categories where ChatGPT is genuinely choosing who to recommend. Queries that just name a brand, or that ChatGPT answers by pointing back to Google, are filtered out so the leaderboard reflects earned visibility, not brand search.

Sources

The Google–ChatGPT volume ratio

ChatGPT scale and usage

Click-through behavior

Run-to-run variation

In short: we anchor to real Google demand, scale it to ChatGPT, ask ChatGPT what it recommends, and split the result by rank and value. Every figure is an estimate, the relative rankings are the strong signal, and the absolute numbers are best read as order-of-magnitude.

AI Brand Tracker is independent and not affiliated with OpenAI. We currently track 354,994 websites across 1,011,972 prompt categories.