AI search is becoming an operating surface for brand teams. The practical question is not whether a single model response changed—it is whether the change reveals something your team can verify and improve.
What the new dataset says
BrightEdge’s September 24 AI Market Pulse reports that ChatGPT generated 95.1% of the AI-platform referral traffic in its tracked dataset during August. The company places Gemini at 2.4%, Perplexity at 1.1%, and Claude at 0.8%. In the same release, BrightEdge says ChatGPT referral traffic grew 101% from January through August and 36% in August alone. The Next Web and ICTMagazine independently reported the release and its headline figures. This is timely evidence that, within one vendor’s referral dataset, outbound web traffic from AI assistants was heavily concentrated in ChatGPT last month.
Referral share is not usage share
The headline measures a share of referral traffic sent to websites by the AI platforms in BrightEdge’s index. It does not measure each assistant’s total users, prompt volume, answer quality, time spent in a product, market share, or the number of brands mentioned in answers. A person can use an assistant extensively without following a link, and a platform can have broad reach while sending relatively few outbound referrals. The reported 95.1% therefore describes one part of the discovery path: tracked visits that left an AI interface for the web. It should not be restated as “95.1% of AI search” or “95.1% of all AI answers.”
The concentration is striking, but the denominator is narrow
BrightEdge says ChatGPT sent roughly 22 times the referrals of Gemini, Perplexity, and Claude combined in August. The company also says total AI referral traffic grew about 73% in less than a year. Those figures make the channel worth watching, particularly for teams seeing traffic from several answer systems. But BrightEdge’s own Market Pulse puts AI search at 1.9% of total search referrals, compared with Google at 86.9%. That broader comparison is important: a platform can dominate a small, growing slice of referred traffic without making traditional search irrelevant to a site’s acquisition mix.
The published caveat belongs beside the chart
BrightEdge’s current Market Pulse includes an explicit note that analytics platforms often adjust how they report AI referrals. It says BrightEdge updated historical data in July 2026 to account for those changes, and that OpenAI/ChatGPT data—its largest AI-referral source—was affected. That does not invalidate the August snapshot. It does mean that a year-to-date comparison, a market-share line, or an apparent platform jump may include measurement reclassification as well as observed traffic change. The public page does not provide a full methodology, site panel, geography mix, raw counts, or the size of the historical revision. Those limits should travel with the numbers.
What the result can tell a brand team
The practical signal is not that every team should move its search budget to ChatGPT. It is that referral analysis needs an assistant-level view. Keep a dedicated, documented channel grouping for identifiable AI referrers, then compare sessions, engaged visits, key events, assisted conversions, landing pages, markets, and dates with the same discipline used for any acquisition source. Preserve the raw referrer and your grouping rules, because analytics classification can change. A sudden rise in one channel may be commercially useful, a reporting change, or both; the record should make later interpretation possible.
What it cannot tell a content or GEO team
A referral visit does not reveal the full answer a user saw, whether a brand was named or recommended, which sources were displayed, how a competitor was framed, or why a model selected a particular link. Conversely, a useful brand mention may create no referral at all. Treat referral data, answer captures, displayed citations, brand and competitor language, and conversion data as separate observations that can be compared but not substituted for one another. This distinction matters most when a team is evaluating visibility work: an answer-level improvement is not automatically a traffic result, and a traffic result is not proof of an answer-level recommendation change.
A restrained response for the next reporting cycle
First, review whether your analytics property identifies ChatGPT, Gemini, Perplexity, Claude, Copilot, and other relevant referrers consistently; document exclusions such as in-app traffic or privacy-limited sessions. Next, select a small set of high-value landing pages and inspect the associated AI referrals alongside organic search and direct traffic, keeping the market and date window fixed. Finally, maintain a controlled prompt record for the buyer questions those pages serve: full response, links, sources, brands named, and collection conditions. That combined evidence can show whether a referral pattern coincides with an observable answer change without claiming one caused the other.
What remains unknown
BrightEdge’s public release does not establish why ChatGPT’s referral share moved in August, whether the result holds across industries or countries, or whether another independent dataset using the same definitions would produce the same platform mix. The two independent news reports corroborate the company’s published figures, but they rely on BrightEdge as the underlying measurement source. The release also does not show the distribution of referrals across sites, the query types involved, or the quality and downstream value of those visits. Those are questions for first-party analytics and reproducible, platform-specific research.
Bottom line
BrightEdge’s latest snapshot is a useful reminder that AI referrals are not evenly distributed: its August dataset assigns 95.1% of tracked AI-platform referrals to ChatGPT. The number is relevant to traffic measurement, but it is not a universal platform-ranking statistic or a proxy for brand visibility. Use it as a prompt to segment and preserve your own referral evidence, while keeping answer content, citations, and commercial outcomes in their proper, separate records.