Brands shown by ChatGPT got more visits, mostly through search. What that does to attribution

What the study measured

Similarweb, which sells web traffic and AI visibility data, published "The Downstream Impact of AI Visibility" in June 2026.¹ It asks a question most analytics setups cannot answer: when ChatGPT recommends a brand, does the person later visit that brand's site, and by which route?

The method is a panel study. Similarweb followed opted-in US desktop users in its panel from July to December 2025, across three brand pairs in finance, travel and beauty.² A user qualified when ChatGPT recommended one of the brands in an answer, and prompts that already named the brand were excluded. Visits were then counted for seven days, and only users with no visit to that brand's site in the previous four weeks were included.²

That design matters. Excluding prompts that named the brand removes buyers who had already decided. Restricting to new visitors removes loyal customers. What is left is close to the question a marketer wants answered: did the answer send someone new?

Do ChatGPT recommendations lead to site visits?

Yes, in this panel, and the detail that matters is the route they took. Brands recommended by ChatGPT were 2.5 times more likely to be visited within seven days than brands that were not.² The recommended brand drew more visits than its rival in all six scenarios, though by different margins in each pair.²

The channel split is what changes attribution. Of the AI-influenced visits, 55.9% arrived through search, against 40.4% of other visits. Direct visits made up 19.9% of AI-influenced visits against 38.8% of others, and only 8.8% arrived as an AI referral, against 5.0%.² Most people who acted on the recommendation looked the brand up rather than clicking through.

The direct-traffic gap has a simple reading. New visitors who have just heard of a brand do not know its address, so they search for it. Returning customers type it or use a bookmark. AI exposure therefore looks like branded search demand from people who were not customers before.

What the study does not show

It does not show that the same holds for your brand, your engine mix or your buyers. The limits are worth stating plainly before anyone quotes the multiplier in a meeting.

• One engine. Only ChatGPT recommendations were tracked, not Gemini, Perplexity, Claude, Copilot or Google's AI features.
• Desktop only, in one country. Search Engine Journal's report of the study notes that mobile and other markets were excluded, and that Similarweb plans to extend it.³
• Six consumer brands, all well known. A lesser-known B2B vendor may see a weaker effect, because the buyer has less reason to remember the name.
• A vendor study. Similarweb sells the data and tools that measure this effect, and the full report is gated behind a form.

It is also a comparison within a panel, not a controlled experiment. The exclusions make it stronger than a simple correlation, but people who receive a recommendation for a brand may differ from those who do not in ways the design cannot remove. Our view: treat the direction as solid and the size as specific to these brands.

How it fits the referral data

It explains why AI referral lines look so small. Across 101,574 websites in 250 countries, SE Ranking found AI search sent 0.32% of all visits in January–April 2026.⁴ If most of the effect of a recommendation arrives through search, that small share describes the click-through path only, not the influence.

We made the same argument in our guide to which AI search numbers belong in a board report, where AI-influenced visits are one of five measures that get confused. The Similarweb study is the best evidence so far for where that fifth measure shows up: in branded search.

What to change in last-click reporting

Stop giving branded search all the credit for the demand it captures. In a last-click model, a buyer who read a ChatGPT recommendation and then searched for your name is a branded search conversion. The model is not wrong about the last step. It is silent about the step before.

1. Report branded search clicks weekly next to your share-of-answer series, so a rise in one can be read against the other.
2. Split branded search into new and returning visitors where your analytics allow it. The Similarweb effect is about new visitors.
3. Ask buyers. A free-text "how did you hear about us" field, plus an AI assistant option, catches the cases where the route is lost.
4. Protect your branded results. If AI exposure sends people to search for your name, a reseller, comparison page or competitor ad on that results page can take the visit.
5. Keep the AI channel, but label it a floor.

Our view: do not reallocate branded search budget on the strength of one panel study. Do change how the branded search line is described. Part of what it counts is demand that an AI answer created, and our guide to proving that AI visibility reaches pipeline shows how to capture the fields that make that visible in the CRM.