--- language: en title: "AI mentions and website visits: what the Profound study allows us to conclude" description: "More than two million conversations help investigate visits after brand mentions. Understand the comparison, its causal limits, and what the absence of a referral does not prove." canonical: "https://tropk.ai/en/resources/research/ai-mentions-and-website-visits-what-the-profound-study-allows-us-to-conclude" last_modified: "2026-10-12T16:14:20.158Z" --- # AI mentions and website visits: what the Profound study allows us to conclude Reviewed on 2026-10-06. ## What is being investigated This is an editorial review by Tropk, revised on October 6, 2026, of research by Profound published on July 1. The data and statistical analysis belong to Profound. The question is whether users who encounter a brand in an [AI response](/en/resources/glossary/ai-response) visit its website afterward more often than an estimated baseline. The study finds a positive association; that does not turn a mention into a purchase or rule out all alternative explanations. The distinction matters because marketing teams may see a brand recommended and find no identifiable click in analytics. A user may search for the brand later, type in the address, or continue through another channel. At the same time, a visit without a referral does not reveal what prompted it. The analysis should keep both possibilities open rather than attributing all direct traffic to AI. ## Key points - Profound analyzes more than two million conversations from a consented U.S. panel, between January and June 2026. - The study compares visits after a brand mention with a forecast based on previous windows. - Adjusted effects are presented in percentage points and measure visits, not sales or revenue. - The observed association does not rule out prior interest or other explanations for visits. - Separate brand representation, navigation, and business outcomes; traffic without a referral should not automatically be attributed to AI. ## How Profound built the comparison The [original research](https://www.tryprofound.com/blog/the-ai-mention-effect)uses more than two million conversations from January to June 2026, from a consented U.S. panel with browsing and interactions on [ChatGPT](/en/resources/glossary/chatgpt), [Gemini](/en/resources/glossary/gemini), and [AI Overviews](/en/resources/glossary/ai-overviews). Exposure requires the brand to appear in the response, but not in the [prompt](/en/resources/glossary/prompt). Brand searches or visits in the previous week are excluded. Visits in the following seven days are compared with a linear forecast based on three previous seven-day windows. Uncertainty is estimated using user-clustered bootstrap. The study takes care to reduce observable prior interest, but exposure is not randomly assigned. Users may be closer to buying for reasons the previous window does not capture. The counterfactual is a forecast, not a group of identical people randomly selected not to see the brand. This distinction should be clear before making any commercial promise. ## How to interpret the reported effects Profound reports adjusted effects of 2.07 percentage points on ChatGPT, 3.21 on Gemini, and 2.96 on AI Overviews. The respective 95% confidence intervals are 1.67–2.48, 2.34–4.08, and 2.76–3.17 percentage points. Among subsequent first visits, 42% occurred within the first 24 hours. These denominators and results appear in the [original report](https://www.tryprofound.com/blog/the-ai-mention-effect). The study measures visits, not contracts or revenue. Percentage points are not relative percentages. And a result among exposed people, with specific filters, cannot be multiplied by all website visits. Temporal proximity helps formulate hypotheses about influence, but does not by itself identify the origin of each visit. An executive presentation should show the result and the group definition on the same screen. ## Three layers the marketing report should distinguish The first layer is representation: did the brand appear, was it described accurately, and was it relevant to the need? The second is navigation: was there a visit identified by referral, an increase in brand searches, or another observable signal? The third is business outcome: did the visitor take a meaningful action and, where follow-up is available, progress in the commercial relationship? None of these layers replaces the others. A favorable mention may not generate a visit. A visit may lead to an unsuitable page. A conversion may occur after multiple exposures across different channels. Working with these layers helps identify the problem before choosing an action: correct the information, improve the destination, or review how it is measured. A single visibility metric cannot answer all three questions. ## What referral data reveals—and what it leaves open An identifiable source referral is evidence of navigation, but it does not reconstruct the entire journey. Parameters may be missing, the visit may happen later, and the user may switch contexts. The absence of an identifier should not be used to prove an absence of influence. Nor does it authorize automatically reclassifying direct or organic sessions as visits from AI. The team can combine aggregated referral data, brand queries, and commercial actions, documenting changes in advertising, press, events, and offers. If all these signals grow at the same time, there is a hypothesis to investigate. It is important to preserve campaign history: external publicity may increase both the likelihood of a mention and the likelihood of a visit, without one causing the other. Voluntary source questions in customer surveys can add context, provided people have room to report more than one touchpoint and the team accepts imprecise answers. A buyer's recollection is not an exact record of exposure. It works better as complementary evidence than as a way to fill every [attribution gap](/en/resources/glossary/attribution-gap) with artificial certainty. ## How to test the hypothesis in Brazil The dataset is U.S.-based, and the article does not report a Brazilian subset. Search habits, categories, languages, and contact channels may differ. The number of conversations also does not equal the number of independent people. So the study supports investigating the local customer journey, but does not provide a ready-made coefficient for predicting visits for a Brazilian brand. Start with a stable set of discovery and comparison questions. Record mentions, relevance, and sources by engine. In parallel, track aggregated website metrics and brand queries over comparable periods. Decide in advance which website actions indicate interest: reading documentation, requesting a demo, or initiating contact, depending on the business. Do not change the definition after seeing which number increased. If the team corrects a recurring error in sources, document the intervention and continue tracking. Compare with topics where no intervention was made and note simultaneous campaigns. This is a proposal for operational learning, not a replication of Profound's panel. Without observing individual exposure and a suitable counterfactual, the conclusion should remain that the signals are consistent with one another, not that the action caused a specific number of visits. ## Questions for a monthly review - Did the brand appear in questions that reflect real needs, with accurate attributes and relevant caveats? - Does the destination a buyer finds explain what the response promised? Is there continuity between the information, the page, and the next step? - Does the analysis distinguish identified referrals, hypotheses about influence, and commercial outcomes? What campaigns or changes could also explain the period? ## Related reading Check [AI mentions](/en/resources/glossary/ai-mention), [attribution](/en/resources/glossary/attribution) and [AI referral traffic](/en/resources/glossary/ai-referral-traffic) before comparing metrics. The [research prompt and benchmark](/en/resources/prompts/prompt-for-writing-a-research-study-or-benchmark) help make criteria and limitations explicit. [Monitoring](/en/product/monitoring)organizes the observation of responses; [Insights](/en/product/insights)helps turn differences into priorities. Evidence should accompany the recommendation, including when it does not yet allow the origin of a visit to be determined. ## Sources and references - [The AI Mention Effect]() — Profound — Nikolas Laskaris ## Frequently asked questions ### Does the Profound study prove that an AI mention causes a visit to the brand's website? No. Profound's research identifies an association between exposure to a brand in AI responses and subsequent visits, using filters and a forecast based on prior periods. Exposure was not randomly assigned, and unobserved factors may influence visits. The study measures navigation; it does not automatically demonstrate purchases or revenue caused by the mention. ### What population and period did Profound analyze in its study of AI mentions and visits? Profound used more than two million conversations from January to June 2026, from a consented U.S. panel with interactions on ChatGPT, Gemini, and AI Overviews. The analysis tracks visits in the seven days after brand mentions and applies prior-interest filters. The report does not present a representative sample of Brazil. ### How should we interpret the effects of 2.07, 3.21, and 2.96 points in Profound's study of visits after mentions? The reported adjusted effects are 2.07 percentage points for ChatGPT, 3.21 for Gemini, and 2.96 for AI Overviews, within the study's groups and method. Percentage points are not relative percentages. These results measure subsequent visits and cannot be multiplied by all website visits or directly converted into sales. ### Does the absence of an AI referral in analytics mean a visit was not influenced by AI? No. A person may learn about a brand in an AI response and later visit the website via search, a typed-in address, or another channel. The absence of a referral also does not prove AI influence. Combine aggregated metrics, brand queries, and campaign context, keeping identified traffic, hypotheses about influence, and commercial outcomes separate. ## Continue a leitura - [Attribution]() - [AI mention]() - [Prompt for writing a research study or benchmark]() - [AI recommendation]() - [AI referral traffic]() - [AI response monitoring]() - [Insights and opportunities]() ## About the author [Luiz Gardelin]() Luiz Gardelin is the founder of Tropk.ai and works in GEO/AEO, product vision, strategy, and growth, helping brands understand and improve their presence in AI responses. --- [All pages in Markdown]()