The practical brief
The finding
Monthly English Wikipedia search referrals declined an estimated 5.45% relative to German and 4.82% relative to French after Google’s default U.S. AI Overviews rollout. [c1] [c2] [c3] [c11]
Why it matters
Informational content can contribute to answering questions without receiving the visits on which its business case depends.
| Control and sample window | Estimated monthly referral change |
|---|---|
| German; December 2023–December 2024 | −5.45% in monthly English external-search referrals relative to German; 499,927 matched article pairs [c1] [c2] |
| French; December 2023–December 2024 | −4.82% in monthly English external-search referrals relative to French; 530,873 matched article pairs [c1] [c2] |
| Japanese; December 2023–July 2024 | −16.53% in monthly English external-search referrals relative to Japanese; 315,054 matched article pairs [c9] |
What to try
BLURSOR’s practical interpretation
Our interpretation: review dependence on search referrals separately from answer visibility and business outcomes, rather than applying Wikipedia’s decline to your forecast.
Study boundary
The study measures default availability, not individual answer exposure, conversions or proven recovery tactics.
Review content plans that depend on search visits
If your content budget assumes useful information will bring search visitors, this study identifies a risk worth reviewing. Google’s AI Overviews—generated summaries above search links—were followed by fewer direct search referrals to English Wikipedia relative to comparison editions.
That is evidence about an informational publisher, not a forecast for your business. Answer visibility, website visits and sales are different outcomes; this paper measures referrals.
Matched language editions provide the comparison
Mehrzad Khosravi and Hema Yoganarasimhan analyzed monthly Wikimedia clickstream data from December 2023 through December 2024, matching English articles with German or French versions covering the same subjects.
English readership represented greater exposure to the May 2024 U.S. rollout. German and French readership was concentrated in markets without default AI Overviews during the sample. Language therefore proxies geography, not individual exposure.
Their difference-in-differences method compares traffic changes between exposed and comparison groups while accounting for persistent article-language differences and shared monthly changes. A causal interpretation requires proportional traffic trends to have remained comparable without the rollout.
Declines survive checks, but trends require qualification
Monthly English search referrals fell an estimated 5.45% relative to German and 4.82% relative to French. Alternative assignments for suppressed low-traffic counts produced similar declines. An artificial pre-rollout treatment date produced no statistically significant shift.
Trend diagnostics were mixed. Separate linear pre-trend tests were insignificant, but some month-by-month pre-rollout confidence intervals excluded zero. German deviations fluctuated around zero; French deviations showed a positive trend. Anchored extrapolations using ordinary and weighted least squares found significant positive French slopes but insignificant German slopes. Post-rollout traffic remained significantly below both extrapolations for both controls.
Honest DiD, a sensitivity analysis allowing departures from comparable trends, kept May–December effect intervals negative through bounds of 0.40 for German and 0.35 for French. These restrict adjacent counterfactual post-rollout trend changes relative to the largest adjacent pre-rollout change—not proof of parallel trends.
A separate Japanese comparison estimated a 16.53% decline through July. Despite mixed pre-trend diagnostics and significant negative extrapolation slopes, post-rollout means remained below both extrapolations and all displayed sensitivity intervals stayed below zero. Its different audience and shorter window make it directional corroboration, not a pooled estimate.
Evidence [c2] [c4] [c5] [c8] [c9] [c14] [c16] [c17] [c19] [c21]
The economic scale is an illustrative benchmark
Averaging the separately estimated German and French coefficients implies approximately 5.14% lower traffic and 100.27 million fewer direct search-originated visits monthly in the comparison-sample calculation. The 1.20 billion annual figure assumes persistence; it is neither an observed twelve-month effect nor a pooled causal estimate.
For an otherwise comparable ad-supported publisher, the hypothetical advertising equivalent is $0.901–$3.090 million monthly, or $10.82–$37.08 million annually if sustained. The calculation assumes one landing-page view per lost referral, other publishers’ advertising benchmarks, 5.22 ad opportunities per page, 75%–90% fill and a 70% publisher share.
Its $3.28/$9.37 CPM inputs—cost per thousand advertising impressions—span delivery environments including connected television, so they are not general web-display bounds. Wikipedia carries no advertising: these are not its revenue losses or revenue transferred to Google. They exclude downstream navigation, subscriptions, leads and commerce.
Separate referral exposure from business performance
Our interpretation is to review informational search as a changing distribution channel, not assume a particular AI-citation technique will recover visits. The paper tests no such tactic.
Referrals combine search engines, and query-level AI Overview displays are unobserved. The study cannot establish effects on total reading, purchasing, answer quality or every publisher.
- Identify content whose business case depends on search visits; compare referral changes with leads or sales.
- Report answer appearances separately from visits; the study supplies no conversion rate between them.
- Treat the advertising equivalents as scenarios, not your revenue forecast.
The source and evidence
Impact of AI Search Summaries on Website Traffic: Evidence from Google AI Overviews and Wikipedia
Study limitations and disclosures
- Language editions proxy geography, not individual exposure. Referrals include multiple search engines, and query-level AI Overview displays are unobserved.
- Parallel trends are not established. Some pre-rollout confidence intervals exclude zero; anchored French slopes are significantly positive, while German slopes are insignificant. Both controls remain below extrapolated counterfactuals. Sensitivity intervals remain negative through bounds of 0.40 and 0.35, respectively, conditional on restrictions on counterfactual trend changes.
- The Japanese comparison ends before Japan’s August rollout and is not interchangeable with the main comparisons. Mixed pre-trend diagnostics coexist with negative extrapolation differences and displayed sensitivity intervals below zero; these checks do not prove parallel trends.
- Samples require consistently reported pre-rollout traffic. Suppressed post-rollout counts are assigned rather than observed; alternative assignments preserve similar declines.
- The averaged traffic benchmark is not a pooled causal estimate and covers direct search referrals, not total activity. Annual traffic and advertising equivalents assume persistence. Advertising equivalents use other publishers’ benchmarks and strong monetization assumptions, including CPM inputs spanning connected television; they are not general web-display bounds, Wikipedia revenue losses or revenue transferred to Google. Downstream navigation, subscriptions, leads and commerce are excluded.
Evidence behind this briefing
[c1] The study uses article–language–month panels and PPML difference-in-differences, comparing 499,927 English–German and 530,873 English–French matched article pairs. English is a geographic exposure proxy, not a record of individual AIO exposure; identification requires proportional parallel trends absent AIO.
Section 1, panel construction and estimation; Sections 2 and 3.2 · Read the study
[c2] The estimated reduced-form decline in monthly English Wikipedia external-search referrals is 5.45% relative to German and 4.82% relative to French. Table 2 reports PPML coefficients −0.0561 and −0.0494, respectively, with article-and-month clustered standard errors 0.0117 and 0.0136; both are marked p<0.001. These are publisher traffic estimates, not answer accuracy or user preference measurements.
Section 3.2, Table 2 and interpretation · Read the study
[c3] The estimates concern default AIO availability, not the effect conditional on an AIO display. Exposure varies across users, queries and months; the outcome also combines search engines. Attenuation relative to actually exposed Google searches is conditional on sign-consistent effects.
Section 3.2, reduced-form interpretation; Section 2, search referrals · Read the study
[c4] Sample selection requires articles to exist since November 2023 and both editions to have reported counts of at least 10 in every pre-treatment month. Wikimedia suppresses counts below 10. Main post-treatment imputations deliberately work against a negative English effect; reversing them yields similar reported declines, 5.5% versus German and 4.9% versus French.
Section 3.3, censoring sensitivity; Section 2, article selection · Read the study
[c5] Some pre-treatment event-study coefficients are statistically different from zero, despite nonsignificant linear pre-trend tests. Honest DiD sensitivity intervals for the equally weighted May–December effect remain negative through relative-magnitude bounds of 0.40 for German and 0.35 for French; this is conditional robustness, not proof of parallel trends.
Appendix B.2–B.3, event-study qualifications and Honest DiD sensitivity · Read the study
[c6] A back-of-the-envelope calculation averages the two PPML coefficients, giving an approximately 5.14% decline, and estimates 100.27 million fewer direct search-originated visits per month within the comparison-sample calculation. The 1.20 billion annual figure assumes persistence for a year; neither figure measures total Wikipedia traffic losses or downstream internal navigation.
Section 4, Economic magnitude and Scope of the estimates · Read the study
[c7] Post-treatment censoring sensitivity checks preserve the negative English-relative-control estimates. Across four assignments, PPML coefficients range from -0.0561 to -0.0566 against German and -0.0494 to -0.0502 against French, with standard errors of 0.0117 and 0.0136 respectively and p<0.001. The samples contain 499,927 and 530,873 matched article QIDs, respectively; standard errors are two-way clustered by article and month. These checks address imputation sensitivity, not all causal-identification concerns.
Appendix C, Table A3 and concluding paragraph · Read the study
[c8] A timing placebo using only December 2023–April 2024, with March–April designated placebo post-treatment, found no statistically significant English-relative-control shift: -0.0026 against German (SE 0.0094) and 0.0162 against French (SE 0.0179). This is a retained-sample timing check, not proof that parallel trends hold.
Appendix D, paragraph preceding Table A4; Table A4 and concluding paragraph · Read the study
[c9] The separate English–Japanese PPML comparison estimates a 16.53% relative decline beginning May 2024 (coefficient -0.1806; SE 0.0110; p<0.001), using 315,054 matched article pairs and 5,040,864 observations over December 2023–July 2024. July is the endpoint because Japan entered the public AIO expansion in August. Both editions must have published counts throughout the five pre-treatment months. Diagnostics are mixed in detail: the linear pre-treatment test is insignificant, but some event-study pre-period intervals exclude zero and anchored OLS/WLS extrapolations have significant negative pre-trend slopes.
Appendix E opening; E.2, Table A6; E.3, Tables A7–A8 · Read the study
[c10] The traffic benchmark averages the separately estimated German and French coefficients to -0.0527, approximately a 5.14% decline, and pairs this with average observed May–December English search traffic of about 1.852 billion visits per month. Multiplicative counterfactual reconstruction gives about 100.27 million fewer direct search-originated visits monthly. The 1.20 billion annual figure assumes persistence for a year; it is not an observed twelve-month treatment estimate or a separate pooled causal estimate.
Appendix F opening; F.1, Equations A8–A11 and concluding paragraph · Read the study
[c11] Advertising figures are hypothetical equivalents for an otherwise comparable ad-supported publisher, not Wikipedia revenue losses or revenue transferred to Google. Assumptions include one landing-page view per lost referral, 5.22 ads per page, 75%–90% fill, a 70% publisher share, and $3.28/$9.37 CPM inputs. These produce $8.99–$30.81 page RPM, $0.901–$3.090 million monthly, and $10.82–$37.08 million annually only if sustained. The CPM source spans delivery environments including connected television, so the inputs are not general web-display bounds.
Appendix F.2, Equations A12–A16 and concluding paragraph · Read the study
[c12] Non-search referrals are not used as an untreated comparison because AIO could affect the referring websites and therefore their outbound Wikipedia traffic.
Footnotes, footnote 4 · Read the study
[c13] The linear-extrapolation diagnostic leaves factual post-treatment English-versus-control traffic below the extrapolated counterfactual for both controls and both fitting methods. The differences are −0.0510 and −0.0505 for German under OLS and WLS, and −0.1163 and −0.1100 for French. Thus, this diagnostic supports rather than overturns the negative post-AIO finding under a linear continuation of pre-treatment differences.
Appendix B.4, Linear Extrapolation of Pre-Treatment Event-Study Coefficients, paragraph preceding Figures A5 and A6 · Read the study
[c14] The event-study diagnostic identifies a positive French pre-treatment trend, opposite to the estimated negative treatment effect, whereas German pre-treatment deviations fluctuate around zero rather than showing a negative trend. This control-specific qualification should not be conflated with the separate insignificant parametric linear pre-treatment tests.
Appendix B.2, PPML Event Studies, discussion of Figure A3 · Read the study
[c15] The extrapolation fits four nonreference pre-treatment event-study coefficients, anchored at the omitted April 2024 coefficient, and continues the fitted line over May–December 2024. WLS uses inverse marginal-variance weights; this is an anchored event-study diagnostic, not the separate article-level PPML linear-trend test.
Appendix B.4, Figure A6 caption · Read the study
[c16] For the French control, the anchored pre-treatment slopes are significantly positive under OLS (0.0092, SE 0.0012) and WLS (0.0078, SE 0.0011), both p<0.001. Nevertheless, factual-minus-extrapolated-counterfactual post-treatment differences remain negative: −0.1163 (SE 0.0072) under OLS and −0.1100 (SE 0.0070) under WLS, both p<0.001. This control-specific diagnostic qualifies the parallel-trends interpretation rather than overturning the negative finding.
Appendix B, Table A2, French OLS and WLS rows · Read the study
[c17] For the German control, the anchored pre-treatment slopes are statistically insignificant: −0.0009 (SE 0.0013) under OLS and −0.0010 (SE 0.0013) under WLS. The factual-minus-extrapolated-counterfactual differences remain negative and significant at p<0.001: −0.0510 (SE 0.0079) and −0.0505 (SE 0.0078), respectively.
Appendix B, Table A2, German OLS and WLS rows · Read the study
[c18] Table A2 reports PPML log coefficients, not percentage-point changes. Both extrapolations fit four nonreference pre-treatment coefficients and anchor the fitted line at April 2024; OLS weights them equally, whereas WLS uses inverse marginal-variance weights. The reported difference compares the factual May–December 2024 mean with its extrapolated counterfactual, with standard errors computed using the unaltered event-study covariance matrix.
Appendix B, Table A2, notes · Read the study
[c19] For the English–Japanese comparison, all displayed Honest DiD 95% confidence intervals for the equally weighted May–July 2024 average are below zero under the relative-magnitude restriction in Equation (A4). This is a conditional sensitivity result, not proof of parallel trends; the supplied text does not specify numerical sensitivity bounds.
Appendix E.3, Honest DiD paragraph; Figure A9 · Read the study
[c20] The Japanese-control Honest DiD sensitivity intervals use relative-magnitude restrictions and the original PPML event-study covariance matrix.
Appendix E.3, Figure A9 caption · Read the study
[c21] For the English–Japanese comparison, the factual May–July post-treatment mean remains significantly below the extrapolated counterfactual under both OLS and WLS. The factual-minus-counterfactual differences are −0.1577 and −0.1541 PPML log-coefficient units, respectively. This supports a negative shift conditional on linear continuation of the pre-treatment differential, rather than establishing parallel trends.
Appendix E.3, concluding paragraph following Table A8 · Read the study
[c22] Table A8 defines the difference as the factual May–July 2024 mean minus the extrapolated counterfactual mean, in PPML log coefficients. OLS equally weights the four nonreference pre-treatment coefficients; WLS uses inverse marginal-variance weights. Both fits are anchored at April 2024, and standard errors use the unaltered event-study covariance matrix.
Appendix E.3, Table A8 notes · Read the study