🤖 AI

Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia

This study provides causal evidence that Google's AI Overviews reduce daily traffic to English Wikipedia articles by approximately 15%, with the substitution effect being most pronounced for cultural topics compared to STEM subjects.

Original authors: Mehrzad Khosravi, Hema Yoganarasimhan

Published 2026-03-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mehrzad Khosravi, Hema Yoganarasimhan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: The "Magic Answer" vs. The "Library"

Imagine you are hungry and you ask a friend, "Where can I get the best pizza in town?"

The Old Way (Traditional Search): Your friend gives you a list of five pizza places, their phone numbers, and their addresses. You have to pick one, walk over there, and look at the menu yourself.
The New Way (Google's AI Overview): Your friend pulls out a magic tablet, instantly reads the menus of all five places, and says, "The best pepperoni is at Tony's, and the best veggie is at Maria's. Here is the price and the exact address." You get your answer right there. You don't need to visit the pizza shops' websites anymore.

The Problem: The pizza shops (publishers like Wikipedia) are worried. They say, "If you give me the answer right here, why would I ever visit my own website? I lose customers, and I lose money from ads."
The Counter-Argument: Google says, "No, no! We are just helping you find the right shop faster. We are sending you more people because we made searching easier."

This paper asks: Who is right? Does the "Magic Answer" (AI) steal traffic from the "Library" (Wikipedia), or does it help?

How They Tested It: The "Time Travel" Experiment

The researchers couldn't just ask Google or Wikipedia what happened; they needed hard proof. They used a clever trick called a "Staggered Rollout."

Think of it like a new flavor of ice cream being released in a grocery store chain:

The Test: In March 2024, Google started showing these "Magic Answers" (AI Overviews) to people in the United States (English speakers).
The Control: At the same time, people in India, Japan, Indonesia, and Brazil (speaking Hindi, Japanese, Indonesian, and Portuguese) were still using the old "List of Links" search. They didn't get the Magic Answers yet.

The researchers looked at the same Wikipedia articles (e.g., "Who is Taylor Swift?" or "What is photosynthesis?") in both groups.

Group A (Treatment): English Wikipedia (US users got the AI).
Group B (Control): Hindi/Japanese/etc. Wikipedia (Users did not get the AI yet).

Since the articles are about the same topics, any difference in traffic between the two groups after March 2024 must be because of the AI.

The Findings: The "Traffic Cliff"

The results were clear and surprising.

1. The Traffic Drop:
Once the US started seeing the AI summaries, traffic to English Wikipedia dropped by about 15%.

The Analogy: Imagine a busy highway leading to a popular tourist destination. Suddenly, a helicopter starts dropping flyers with the exact map and ticket prices right onto the highway. People stop driving to the destination because they already have the info. The destination gets 15% fewer cars.

2. The "Substitution" Effect:
The AI didn't just send people to the website; it replaced the need to visit. If you ask "Who is the president of France?" and the AI gives you the name and a photo, you don't click the link to read the full biography. You got what you needed.

3. It Depends on the Topic:
The drop wasn't the same for everything.

Culture & Pop Culture (Big Drop): Questions like "Who dated whom?" or "What is the plot of this movie?" are easy for AI to summarize in one sentence. Traffic to these Wikipedia pages crashed the hardest.
STEM & Science (Smaller Drop): Questions about complex math or science often need deep context, formulas, or verification. People still felt the need to click through to the "Library" to get the full details. The AI couldn't fully replace the deep dive.

The Economic Impact: The "Lost Rent"

Wikipedia doesn't run ads, so they didn't lose money directly in this study. But the researchers calculated what this would mean for a normal news website.

The Math: The drop in traffic meant roughly 11.5 million fewer visits per day for English Wikipedia.
The Money: If this were a regular news site, that traffic drop would equal a loss of $35 million to $100 million per year in advertising revenue.

The Metaphor: Imagine a mall. Google is the giant billboard outside the mall that now gives you a coupon and the store's address right on the billboard. You don't walk into the mall to find the store. The mall (the publisher) loses the foot traffic, and the billboard owner (Google) keeps all the ad money from the people standing outside looking at the billboard.

Why This Matters

This study provides the first "smoking gun" evidence that Generative AI in search is stealing traffic from content creators.

For Publishers: If AI answers your questions, you might go out of business. This is bad for the internet because if no one pays writers or editors, the quality of information on the web might drop.
For Google: They are becoming the "answer engine," but they are doing it by potentially starving the sources they rely on.
For Policymakers: The paper suggests we might need new rules. Maybe Google should pay a "rent" to Wikipedia and other publishers for using their content to build these AI answers, similar to how you pay a musician when you stream their song.

The Bottom Line

The "Magic Answer" is convenient for us, the users. But for the "Librarians" of the internet, it's a threat. The study shows that when the answer is given to you for free on the search page, you stop visiting the source. And if the source stops getting visitors, it might stop producing answers for the AI to read in the first place.

1. Problem Statement

The paper addresses a critical economic and structural shift in the digital information ecosystem: the transition of search engines from "link lists" to "answer engines" powered by Large Language Models (LLMs). Specifically, it investigates the causal impact of Google's AI Overviews (AIO) on the inbound traffic of third-party publishers.

The Conflict: Publishers argue that AIOs act as substitutes, satisfying user queries directly on the Search Engine Results Page (SERP) and cannibalizing clicks to source pages. Search platforms argue they are complements, lowering search costs and directing users to relevant sources via cited links.
The Gap: While prior literature on news aggregators suggests complementarity, AIOs differ fundamentally by providing synthesized answers before organic links, potentially satisfying informational intent without a click. The paper seeks to provide the first causal evidence on whether this specific mechanism displaces publisher traffic.

2. Methodology

The authors employ a Difference-in-Differences (DiD) design, leveraging the staggered rollout of Google AIO and Wikipedia's multilingual structure to create a quasi-experimental setting.

A. Data Construction

Source: Wikimedia Analytics API (publicly available daily page view data).
Sample: 161,382 matched article–language pairs covering 52,262 unique English articles and their corresponding versions in Hindi, Indonesian, Japanese, and Portuguese.
Timeframe: October 28, 2023, to August 14, 2024.
Selection Criteria: Articles were selected based on pre-treatment popularity (appearing in the top 1,000 most-viewed pages in at least one language between July 1, 2023, and March 21, 2024) to avoid conditioning on post-treatment outcomes.

B. Identification Strategy

Treatment Group: English Wikipedia articles. The English edition was exposed to AIO starting March 22, 2024 (US testing phase).
Control Group: The same underlying articles in Hindi, Indonesian, Japanese, and Portuguese. These editions were not exposed to AIO during the observation period (AIO rollout in these regions occurred later, starting August 2024).
Proxy for Exposure: Since the API does not filter by user geography, the authors use the language edition as a proxy for AIO exposure, assuming English traffic is heavily concentrated in the US (where AIO launched first).

C. Econometric Model

The authors estimate a two-way fixed effects model at the article-language-day level:
$views_{i,l,t} = \alpha_{i,l} + \delta_t + \beta (English_l \times Post_t) + \epsilon_{i,l,t}$

Outcome Variable: Daily page views ( $views_{i,l,t}$ ). The authors explicitly use levels (not log-transformed) to align with revenue implications (ad impressions scale with absolute volume) and to avoid bias identified in recent literature regarding log-transformed DiD with heterogeneous group sizes.
Fixed Effects:
- $\alpha_{i,l}$ : Article-Language fixed effects (absorbs time-invariant popularity and language-specific audience size).
- $\delta_t$ : Date fixed effects (absorbs common time shocks, seasonality, and platform-wide algorithm updates).
Coefficient of Interest ( $\beta$ ): Captures the average change in English page views relative to control languages after the AIO introduction.
Inference: Standard errors are two-way clustered by article and date to account for serial correlation and common daily shocks.

3. Key Results

A. Main Effect: Traffic Displacement

The study finds a statistically significant and economically large reduction in traffic for the treated group.

Magnitude: Exposure to AIO reduces daily traffic to English Wikipedia articles by approximately 15% relative to the control group.
Absolute Impact: Scaling this to the sample of 52,262 English articles implies a loss of roughly 11.5 million daily page views, or 4.21 billion fewer visits annually.
Robustness: Results hold across alternative specifications (PPML, weighted log), different time windows (excluding the initial test phase), and aggregated weekly data.

B. Heterogeneous Effects by Topic

The substitution effect is not uniform; it varies significantly by content category, supporting the hypothesis that AIO substitutes for queries where short, synthesized answers suffice.

Culture: Largest decline (-19.6%). These queries often require factual summaries easily generated by LLMs.
Geography: -16.6%.
History & Society: -9.9%.
STEM: Smallest decline (-7.4%). The authors suggest STEM queries may require deeper nuance, verification, or complex context that users still seek via source links.

C. Economic Implications

The authors translate the traffic loss into revenue equivalents for ad-supported publishers:

Assuming a standard Cost Per Mille (CPM) range of \ $3.28–\$ 9.37 and a publisher retention rate of 70%, the traffic loss translates to an estimated annual revenue loss of \ $35.7 million to \$ 102.1 million for a publisher of Wikipedia's scale.
This represents a direct transfer of advertising value from content creators to the search platform.

4. Key Contributions

Causal Evidence on Generative AI: Unlike prior studies on aggregators, this paper isolates the impact of generative summaries that sit above organic links, providing the first causal estimate of traffic cannibalization by LLM-based search features.
Methodological Rigor: The paper addresses common pitfalls in DiD analysis (e.g., the "log-transform bias" in heterogeneous groups) by using levels-based estimation and rigorous parallel trend validation (event studies and pre-trend tests).
Granular Heterogeneity: It demonstrates that the impact of AI search is content-dependent, with "factual/summary" content suffering the most, challenging the notion that AI search is universally beneficial for all publishers.
Policy Relevance: The findings provide empirical backing for concerns regarding market power concentration, the sustainability of the content supply chain, and the need for new attribution mechanisms (e.g., Shapley value-based revenue sharing).

5. Significance and Limitations

Significance: The results suggest that the "answer-first" paradigm fundamentally alters the economics of the open web. If search engines can satisfy user intent without clicks, the traditional ad-revenue model for informational publishers is threatened. This supports arguments for regulatory intervention or new business models (e.g., licensing content for AI training/generation).
Limitations:
- Generalizability: The study focuses on Wikipedia, which is unique as both a primary answer repository and a highly cited source. The effect on smaller publishers or transactional/e-commerce sites might differ.
- Measurement: The use of language as a proxy for geography is imperfect (English has a global readership), meaning the estimated 15% decline is likely a conservative lower bound of the true impact.
- Evolution: AIO features are rapidly evolving; long-term effects on content quality and contributor incentives remain to be seen.

In conclusion, the paper provides robust evidence that Google's AI Overviews act as a substitute for publisher traffic, particularly for informational content, leading to a material reallocation of attention and revenue from publishers to search platforms.