{ "@context": "https://schema.org", "@type": "Article", "schema_version": "1.1", "content_item_id": "guide-ai-search-publisher-referrals-experiment.en", "translation_group_id": "guide-ai-search-publisher-referrals-experiment", "locale": "en", "type": "guide", "section": "knowledge", "slug": "ai-search-publisher-referrals-experiment", "title": "How AI search changes publisher referrals: a Google field experiment", "description": "How to read a randomized Google Search experiment and measure brand appearance, visible links, external clicks, site visits, and target actions separately.", "direct_answer": "In a randomized field experiment with 1,100 participants, forcing Google's generative answer mode reduced external clicks per search by 18.8 percentage points relative to current Google Search. This does not imply the same traffic loss for every site: the study measured outbound clicks from Google, not appearances of a particular brand, visits to a particular page, target actions, or revenue.", "sections": [ { "heading": "1. The practical question this study can answer", "paragraphs": [ "The paper addresses a narrow question: does behavior change when current Google Search is replaced by a mode that produces a ready-made answer, or when generative summaries are removed from the ordinary results page? It does not test how often an individual brand is named, how one specific page becomes a source, or how much money a particular publisher earns.", "Its best use is therefore as a test of a measurement design. A brand appearing in an answer, or its page appearing among visible sources, is not yet an outbound click. A visit recorded by a site is not yet a target action. The experiment provides a causal estimate for one stage—an outbound Google click—under a particular March 2026 intervention." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "2. How the field experiment was designed", "paragraphs": [ "The researchers recruited US adults who used Chrome as their primary browser and Google as their primary search engine. From March 17 to 19, 2026, 1,444 people completed an initial survey and installed the research extension. The behavioral analysis included 1,100 participants who made at least one search during the following seven days; 956 completed the final survey.", "The extension first observed three days of current search. Participants were then randomly assigned to one of three conditions: unchanged Google Search; an attempt to hide generative summaries and block the separate AI experience; or forced redirection of every search into `AI Mode`. Random assignment supports a careful causal claim inside this sample, but it does not turn the result into a forecast for every market or later Google version." ], "bullets": [ "Three days of current search before the intervention.", "Seven days in the assigned condition.", "Three conditions: current search, an attempt to remove generative features, and forced generative answers." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "3. What the external click measure means", "paragraphs": [ "The authors defined click-through rate (CTR) as external clicks originating from Google divided by Google searches. The count of clicks excluded movement within google.com, such as a follow-up search. The total count of searches excluded Google's separate image, news, and shopping pages. The primary value was calculated for each participant.", "In the conversational experience, every new turn counted as another search. The authors also tested a version that counted the whole conversation as one search: the direction remained the same, but the estimate changed. This boundary matters. The figure of 18.8 is a percentage-point difference between experimental conditions, not an 18.8% relative decline and not the share of visits lost by a particular site." ], "bullets": [ "Top count: external clicks that started on Google.", "Total checked: Google searches excluding image, news, and shopping pages.", "Primary unit: a participant, not an individual publisher page." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "4. What changed under forced generative answers", "paragraphs": [ "The comparison between current search and forced `AI Mode` included 714 participants. Assignment to the generative-answer group reduced CTR by 18.8 percentage points. The 95% confidence interval ranged from minus 22.2 to minus 15.3 points. The extension successfully routed 94.7% of this group's searches into the assigned experience, so the intervention substantially changed what participants used.", "The estimate is a group difference, not a site-level result. The paper also reports fewer people clicking news sites, Reddit, and Wikipedia, but those categories overlap and do not form a universal list of affected publishers. The narrower practical meaning is stronger: a ready-made answer can keep an action inside Google, so greater presence in an answer cannot automatically be recorded as greater site traffic." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "5. Why the no-generative-summary result needs its own reading", "paragraphs": [ "For participants whose extension actually hid generative summaries, the authors estimated an 8.8 percentage-point increase in CTR relative to current search. This calculation included 615 participants, and its 95% confidence interval ranged from plus 2.3 to plus 15.3 points. It estimates the effect of receiving the intervention rather than the simple difference between everyone initially assigned to the groups.", "Google changed its page structure during the study and broke part of the extension's hiding logic. About 90% of detected summaries were hidden on the first day, but success later fell close to zero; overall, 51.1% were hidden. The result should not be presented as a clean seven-day comparison of Google with and without summaries. It is a separate estimate with a stronger assumption and a material technical limitation." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "6. More time inside search does not mean more site visits", "paragraphs": [ "Forced generative answers reduced average search sessions by 0.92 per day while increasing the length of a session by 0.43 minutes. Both effects came from the 714-participant comparison. A person could spend more time inside one Google session and still leave for an external site less often. Those outcomes are compatible rather than contradictory.", "A brand report therefore needs more than one visibility score. A saved answer can show whether a name or link appeared. A click event can show that a link was selected. A server record or web analytics can show a visit. Only a separately defined action can show a business outcome. Neither time spent in search nor the presence of a link substitutes for the next stage." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "7. Five events that should not be collapsed into one metric", "paragraphs": [ "The experiment measured movement from search to an external site, but it did not code mentions of individual brands or connect every answer to a purchase. The next step is not to transfer its average estimate to a private funnel. Instead, preserve five different observations, each with its own record, period, and total number checked.", "If a brand appears in 20 of 100 archived answers, that is 20/100 answer appearances. If its official address is visible eight times, that is 8/100 visible links. If analytics records three visits from a marked channel, that comes from another record and another total. Target actions are counted only among visits that can actually be observed under a rule defined in advance." ], "bullets": [ "The name appears in an archived answer.", "A visible link points to the intended page.", "The user selects an outbound link.", "The site records a visit.", "A predefined target action is completed." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "8. How to run a local check without false causality", "paragraphs": [ "Fix one set of search requests, country, language, Google experience, dates, and account state. Archive the full response and every visible address. Mark brand naming, recommendation wording, and exact links separately. For site traffic, use marked addresses, Search Console, and web analytics, but define in advance which visits and target actions enter the calculation.", "Compare like-for-like periods and show raw counts next to rates. A change in external visits should not be assigned automatically to a generative feature: demand, ordinary search position, request mix, devices, seasonality, or measurement rules may also have changed. A causal statement needs its own controlled comparison; the Wang and colleagues experiment supports the mechanism only within its assigned conditions." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "9. Limitations and alternative explanations", "paragraphs": [ "The sample consisted of US users who preferred Chrome and Google Search; it was younger, more highly educated, and more politically left-leaning than the overall US population. People were forced to use their assigned experience for seven days, although only 0.6% of their searches used `AI Mode` before the experiment. Interface unfamiliarity may have influenced behavior apart from lasting preferences.", "The forced generative experience did not show ads during the study, and Google changed the structure of its summary pages while data collection was running. Interface design, monetization, and user habits may have changed since March 2026. The paper concerns Google and does not establish effects for ChatGPT, Perplexity, Yandex, or any other service. It also does not measure revenue for individual publishers or prove that one displayed source caused or prevented one visit." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] }, { "heading": "10. Practical decision for editorial and analytics teams", "paragraphs": [ "Use the experiment as a reason to separate goals, not as a traffic-loss forecast. Editorial work owns clear, checkable material. External communications own independent support. Search specialists own an accessible page and the correct address. Analysts own separate records for answer appearances, visible links, clicks, visits, and target actions.", "Act at the break point. If the brand is absent, examine the answer and its support. If a link appears but visits do not, inspect the interface, address marking, and visit measurement. If visits arrive but actions do not, improve the landing page or offer. Do not declare success from answer presence when the goal was site visits, and do not declare failure from visits when the study measured only answer presence." ], "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ] } ], "published_at": "2026-10-08", "modified_at": "2026-10-10", "data_through": "2026-03-29", "next_review_at": "2026-11-10", "author": "GeoAeoAle Editorial", "origin": "editorial", "topics": [ "ai-search", "publisher-referrals", "measurement", "field-experiment" ], "publisher": "GeoAeoAle Editorial", "license": "https://creativecommons.org/licenses/by/4.0/", "canonical_url": "https://geoaeoale.com/en/knowledge/ai-search-publisher-referrals-experiment/", "claims": [ { "claim_id": "experiment-scope", "text": "The authors preregistered a US field experiment: 1,444 people installed the extension, 1,100 made at least one search during the seven-day comparison period, and 956 completed the final survey; three ordinary-search days preceded random assignment.", "status": "observed", "confidence": "high", "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ], "publication_status": "public", "confidentiality": "public" }, { "claim_id": "external-click-rate-definition", "text": "The primary referral measure was external clicks originating from Google divided by Google searches; clicks within google.com and searches on image, news, and shopping pages were excluded, while each turn in the generative answer mode counted as a separate search.", "status": "observed", "confidence": "high", "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ], "publication_status": "public", "confidentiality": "public" }, { "claim_id": "ai-mode-click-effect", "text": "Across the 714 participants in the relevant comparison, assignment to Google's forced generative answer mode reduced external clicks per search by 18.8 percentage points; the 95% confidence interval ranged from minus 22.2 to minus 15.3 points.", "status": "observed", "confidence": "high", "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ], "publication_status": "public", "confidentiality": "public" }, { "claim_id": "no-ai-click-effect", "text": "For 615 participants, the estimated effect of successfully hiding generative summaries was an 8.8 percentage-point increase in external clicks per search, with a 95% interval from plus 2.3 to plus 15.3 points; only 51.1% of detected summaries were successfully hidden overall.", "status": "observed", "confidence": "high", "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ], "publication_status": "public", "confidentiality": "public" }, { "claim_id": "session-pattern", "text": "In the same 714-participant comparison, forced generative answers reduced search sessions by 0.92 per day while increasing average session length by 0.43 minutes.", "status": "observed", "confidence": "high", "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ], "publication_status": "public", "confidentiality": "public" }, { "claim_id": "measurement-boundary", "text": "Editorial inference: brand appearance, a visible source link, an outbound click, a recorded site visit, and a target action should remain separate measurements because the experiment directly tested search behavior and outbound clicks only.", "status": "inference", "confidence": "high", "source_ids": [ "arxiv-ai-search-referrals-2026-08-18" ], "publication_status": "public", "confidentiality": "public" } ], "sources": [ { "source_id": "arxiv-ai-search-referrals-2026-08-18", "canonical_url": "https://arxiv.org/html/2608.18352", "title": "AI in Search Reduces Publisher Referrals Without Improving User Experience: Experimental Evidence", "publisher": "Stephanie T. Wang et al. / arXiv", "source_type": "primary-research", "locale": "en", "published_at": null, "checked_at": "2026-10-10", "sha256": "f7625a03fb5fec715266f9411634f23c20b7ddd7b4d9384f0965c41d9bb8b594", "license": "CC BY-NC-ND 4.0", "visibility": "public" } ], "related_slugs": [ "ai-visibility", "ai-citation", "source-exposure-brand-mention-diagnostic", "visible-links-search-record" ], "distribution_links": [ { "kind": "telegram", "url": "https://t.me/geoaeoale/32", "label": "Short version on Telegram" } ], "limitations": [ "The experiment ran in the United States with Chrome and Google Search users for seven assigned-condition days after a three-day ordinary-search period.", "The paper does not name one final collection date. March 29, 2026 in the metadata is a conservative upper bound derived from the last recruitment wave on March 19 and the ten-day protocol.", "The generative answer experience was forced; participants used it for only 0.6% of searches before the intervention.", "The attempt to hide generative summaries succeeded for only 51.1% of detected cases because Google's interface changed.", "The estimates apply to Google's March 2026 implementation and do not transfer automatically to other countries, dates, or AI services.", "The study measured outbound clicks and perceptions of search, not visits, target actions, or revenue for a particular brand." ], "corrections": [] }