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

Mehrzad Khosravi; Hema Yoganarasimhan

cs.CYFeb 5, 2026

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

Authors: Mehrzad Khosravi, Hema Yoganarasimhan

Organizations: University of Washington

Abstract

Search engines increasingly display AI-generated answers above organic links, potentially displacing traffic to upstream publishers. We estimate the impact of Google's AI Overviews (AIO) on Wikipedia's search traffic using AIO's staggered geographic rollout and Wikipedia's multilingual structure. Our difference-in-differences design compares monthly external-search referrals to English Wikipedia articles with referrals to the same articles in German and French, and finds that default AIO availability reduced English search traffic by 5.45% and 4.82%, respectively. Our results suggest that answer-producing digital intermediaries can materially reallocate attention away from informational publishers, with implications for content monetization, search platform design, and policy.

Figures & tables

Panel of Figure 1 — Figure 1: Demeaned log total monthly search traffic through December 2024 for the English–German and English–French samples. Each series is demeaned using December 2023–April 2024; the dashed line marks May 2024. The English–German and English–French panels contain 499,927 and 530,873 matched article pairs, respectively, with 13 monthly observations in each plotted series.

	English–German	English–French
English $\times$ post	$-0.0561^{***}$	$-0.0494^{***}$
	(0.0117)	(0.0136)
No. observations	12,998,102	13,802,698
Articles (QIDs)	499,927	530,873
Article $\times$ language FE	Yes	Yes
Month FE	Yes	Yes

Table 2: PPML DiD estimates

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1 : A Google search result page without AIO for the query “e=mc2.”

Figure A2 : A Google search result page with AIO for the query “e=mc2.”

	English–German	English–French
English $\times t_{m}$	$-0.0014$	$0.0028$
	(0.0035)	(0.0076)
No. observations	4,999,270	5,308,730
Articles (QIDs)	499,927	530,873
Article $\times$ language FE	Yes	Yes
Month FE	Yes	Yes

Appendix

Table A1: PPML linear pre-treatment trend tests

Figure A3 : Monthly PPML event-study estimates through December 2024 for the English–German and English–French samples. April 2024 is the omitted reference month and May 2024 is event month zero. Pointwise 95% confidence intervals use the direct two-way-clustered standard errors and asymptotic-normal inference from Stata.

Figure A4 : Honest DiD sensitivity intervals under relative-magnitude (

\Delta^{RM}

) restrictions for German and French controls. The target is the equally weighted average of May–December 2024, using the original PPML event-study covariance matrices exported by Stata.

Figure A5 : OLS linear extrapolations of the pre-treatment event-study coefficients. Black points and blue whiskers report the monthly PPML estimates and 95% confidence intervals. The solid red line is fitted to the four nonreference pre-treatment coefficients and anchored at the omitted April 2024 coefficient; the dashed red line is its May–December counterfactual continuation. The dotted gray line marks May 2024.

Figure A6 : WLS linear extrapolations of the pre-treatment event-study coefficients. Black points and blue whiskers report the monthly PPML estimates and 95% confidence intervals. The solid red line uses inverse marginal-variance weights for the four nonreference pre-treatment coefficients and is anchored at the omitted April 2024 coefficient; the dashed red line is its May–December counterfactual continuation. The dotted gray line marks May 2024.

Control	Method	Pre-trend slope	Factual post mean	Counterfactual post mean	Difference
German	OLS	$-0.0009$	$-0.0552^{***}$	$-0.0042$	$-0.0510^{***}$
		(0.0013)	(0.0039)	(0.0057)	(0.0079)
	WLS	$-0.0010$	$-0.0552^{***}$	$-0.0046$	$-0.0505^{***}$
		(0.0013)	(0.0039)	(0.0057)	(0.0078)
French	OLS	$0.0092^{***}$	$-0.0748^{***}$	$0.0414^{***}$	$-0.1163^{***}$
		(0.0012)	(0.0040)	(0.0052)	(0.0072)

Appendix

Table A2 : Linear extrapolations of pre-treatment event-study coefficients

Main assignment: unpublished English count 10; control count 0
	German	French
English $\times$ post	$-0.0561^{***}$	$-0.0494^{***}$
	(0.0117)	(0.0136)
Opposite assignment: unpublished English count 0; control count 10
English $\times$ post	$-0.0566^{***}$	$-0.0502^{***}$
	(0.0117)	(0.0136)

Appendix

Table A3: PPML DiD estimates under alternative post-treatment censoring assignments

	German	French
English $\times$ placebo post	$-0.0026$	$0.0162$
	(0.0094)	(0.0179)
No. observations	4,999,270	5,308,730
Articles (QIDs)	499,927	530,873
Article $\times$ language FE	Yes	Yes
Month FE	Yes	Yes

Appendix

Table A4: Pre-treatment timing-placebo PPML DiD estimates

Figure A7: Demeaned log total monthly search referrals through July 2024 for the English–Japanese sample. Each series is demeaned using December 2023–April 2024; the dashed line marks May 2024.

	Japanese
English $\times$ post	$-0.1806^{***}$
	(0.0110)
No. observations	5,040,864
Articles (QIDs)	315,054
Article $\times$ language FE	Yes
Month FE	Yes

Appendix

Table A6: English–Japanese PPML DiD estimate through July 2024

	Japanese
English $\times t_{m}$	$-0.0020$
	(0.0049)
No. observations	3,150,540
Articles (QIDs)	315,054
Article $\times$ language FE	Yes
Month FE	Yes

Appendix

Table A7: English–Japanese PPML linear pre-treatment trend test

Figure A8: Monthly PPML event-study estimates through July 2024 for the English–Japanese sample. April 2024 is the omitted reference month and May 2024 is event month zero. Pointwise 95% confidence intervals use the direct two-way-clustered standard errors and asymptotic-normal inference from Stata.

Figure A9: Honest DiD sensitivity intervals under relative-magnitude (

\Delta^{RM}

) restrictions for the Japanese control. The target is the equally weighted average of May–July 2024.

Panel of Figure 18 — Figure A10 : Linear extrapolations for the English–Japanese comparison through July 2024. Black points and blue whiskers report the monthly PPML estimates and 95% confidence intervals. The solid red line is fitted to the four nonreference pre-treatment coefficients and anchored at the omitted April 2024 coefficient, using equal weights in panel (a) and inverse marginal-variance weights in panel (b); the dashed red line is its May–July counterfactual continuation. The dotted gray line marks May 2024.

Control	Method	Pre-trend slope	Factual post mean	Counterfactual post mean	Difference
Japanese	OLS	$-0.0049^{*}$	$-0.1675^{***}$	$-0.0098^{*}$	$-0.1577^{***}$
		(0.0022)	(0.0059)	(0.0044)	(0.0085)
	WLS	$-0.0067^{***}$	$-0.1675^{***}$	$-0.0134^{***}$	$-0.1541^{***}$
		(0.0018)	(0.0059)	(0.0036)	(0.0073)

Appendix

Table A8 : Linear extrapolations of pre-treatment event-study coefficients

Information Retrieval

Information Retrieval studies how to find, rank, and organize relevant information in large collections of data.

See all 476 papers

Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain
Jul 8, 2026Qiaoni Shi, Kai Zhu, Kai GuInformation Retrieval
Bocconi University, Milan, Italy
How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews
Apr 30, 2026Riley Grossman, Songjiang Liu, Michael K. Chen +3Generative Engine Optimization Generative Retrieval
New Jersey Institute of Technology Newark, NJ, USA · Nanyang Technological University Singapore, Singapore · Indiana University Bloomington Bloomington, IN, USA
The Impact of AI Search on the Online Content Ecosystem: Evidence from Google and Reddit
May 14, 2026Peibo Zhang, Ruomeng Cui, Dennis J. Zhang
Goizueta Business School, Emory University · Olin Business School, Washington University in St. Louis

Jul 8, 2026J/K move · Enter open · S save

Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain

Qiaoni Shi, Kai Zhu, Kai Gu

Bocconi University, Milan, Italy

Information Retrieval

Open paper AskPDF arXiv

Search engines have long allocated attention on the web by routing users from queries to websites. AI search changes this arrangement because information needs can be resolved inside the intermediary. Using URL-level Comscore U.S. desktop clickstream, we compare ChatGPT and Google information-seeking occasions and exploit ChatGPT Search access expansions to estimate traditional search displacement. ChatGPT produces outbound clicks in only 5.2% of conversation sessions, far below Google's referral ratio. The remaining clicks are not a scaled-down Google stream: they skew toward specialized destinations and away from ad-supported sites. Wider access cuts search use by 9.4%, with search-referral losses largest for informational categories. Our findings identify a central economic shift in digital intermediation: AI search might satisfy information needs inside the intermediary while weakening the referral bargain that has linked search, traffic, and content production on the open web.

Abstract

Information Retrieval

Explore similar work

Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain

How Generative AI Disrupts Search: An Empirical Study of Google Search, Gemini, and AI Overviews

The Impact of AI Search on the Online Content Ecosystem: Evidence from Google and Reddit