AI Overviews: Wikipedia Loses 5% of Traffic — the E-commerce Signal
Summarize this article with AI
A client calls me on a Tuesday morning
A client calls me on a Tuesday morning. He invested $8,000 in buying guides. Result: 47 informational queries lose 2.4 positions on average. Blog traffic drops 3% per month since May 2024.
No manual penalty. Site speed stayed flat. Transactional keywords hold steady.
Content wasn’t the issue. Architecture was.
Then I come across a University of Washington study. It estimates AI Overviews cut monthly referrals to English Wikipedia by 5.45% compared with German. 4.82% compared with French. 100.27 million clicks lost every month. 1.20 billion per year.
Wikipedia.
The most resilient site on the web. The one capturing informational demand since 2001.
Since May 2024, Google has shown AI-generated answers at the top of U.S. results. The user reads the answer on Google. They don’t click. For informational queries, the click becomes optional. For transactional queries, the effect is slower, but it exists.
Across my e-commerce clients, guide pages often make up 25% to 45% of organic traffic. This traffic assists conversions. It isn’t the last click. But it prepares the decision.
AI Overviews steal a click. They also reshape the journey. The user asks a question, reads the answer, then refines the query. The second query becomes more transactional. The first click no longer exists.
If Wikipedia loses 5% of its referrals, what happens to your buying guide pages?
The University of Washington study, decoded
The study is authored by Mehrzad Khosravi and Hema Yoganarasimhan. Updated on September 2, 2026. It hasn’t been peer-reviewed yet. That detail matters. The signal remains.
The researchers used the monthly click data Wikipedia publishes for each article. They compared external referrals to the English edition with those to the German and French editions.
- 499,927 English-German article pairs.
- 530,873 English-French article pairs.
- Poisson pseudo-maximum likelihood difference-in-differences model.
Result: a 5.45% reduction in monthly referrals to English Wikipedia compared with German. 4.82% compared with French.
In volume, that works out to about 100.27 million lost referrals every month. 1.20 billion over a year. These figures are modeled. They don’t measure all Wikipedia traffic, only referrals from external search engines.
The first version of the study, published in February, reported a 15% drop. The authors revised their methodology. They moved from daily pageviews to monthly referrals. They added German and French control groups. The drop went from 15% to 5%.
Not a retreat. A refinement.
Google disputes the metric. The company says Wikipedia data aggregates all external search engines. You can’t isolate Google. True. But Yoganarasimhan points out that other engines account for a marginal share of traffic. The order of magnitude holds.
An English-Japanese check shows a 16.53% drop. The authors call it directional only. The window is shorter. The control group differs. That strengthens the idea of a real effect, but one that is hard to quantify precisely.
Why these editions? About 40% of English-edition traffic comes from the United States. German and French editions mostly come from countries without default AI Overviews during the analysis period. The method compares trajectories before and after May 2024. English and German articles usually move together. When the gap widens, the effect becomes attributable to AI Overviews.
The study says nothing about the exact breakdown by engine. It says the trend is measurable.
Why Wikipedia is a unique laboratory
Wikipedia has everything needed for a natural experiment.
First, the data is open. Every month, the foundation publishes click files, article by article. You see referral volume from engines. You track change over time.
Second, language editions act as control groups. English undergoes the AI Overviews rollout in the United States in May 2024. German and French don’t. On the same article, the trajectory difference measures the treatment effect.
Finally, volume. Hundreds of thousands of article pairs. Random variation cancels out. Wikipedia isn’t a fringe site. It’s a global reference.
The availability bias, which Guillaume Attias teaches at BMO Academy in the DOSE framework, explains why this case changes everything. A diffuse risk remains abstract. Nobody measures it. Then an emblematic case emerges. Wikipedia loses 5%. Suddenly everyone can picture it.
I learned this while deploying semantic cocoons: my clients respond to concrete examples. One Wikipedia figure is worth ten theoretical warnings.
Wikipedia is the canary in the coal mine. Not an e-commerce site, but a site whose survival depends on informational demand. Like your buying guides.
The difference: Wikipedia has millions of pages, millions of links, off-the-charts domain authority. If erosion reaches 5% on this giant, it can exceed 10% on an average e-commerce site. Why? Your guide pages don’t have the same coverage depth. They answer a specific question. AI answers it too. The click disappears.
Across my clients, the pattern is identical: guide pages lose impressions on long-tail queries. Those queries move into AI Overviews, cited or not. The traffic doesn’t come back.
Careful: SEO isn’t doomed. The field is changing. Pages that feed AI answers gain a new kind of visibility. It doesn’t show up in clicks.
Availability bias makes you believe that as long as your product pages hold, all is well. False. Guide pages prepare the purchase. If they lose 5% reach, the average basket fills more slowly.
Avant de chiffrer la perte, plantons le décor. Voici comment se répartit le trafic organique d’un site e-commerce type de 200 000 sessions mensuelles.
La vraie répartition de votre trafic organique
35% guides, 45% produits, 20% catégories sur 200 000 sessions/mois
5% on Wikipedia, how much on your product pages?
Take an e-commerce site with 200,000 organic sessions per month.
Classic structure: 35% guide pages, 45% product pages, 20% category pages. Guide pages draw 70,000 sessions per month.
Apply 5% erosion to informational traffic. That gives 3,500 lost sessions every month. 42,000 per year.
These sessions are not direct conversions. They assist the decision. Across my clients, guide-page assist rate ranges from 12% to 22% depending on the sector. At 15%, those 42,000 lost sessions mean 6,300 fewer assisted sessions. With an average basket of €85 and a 2% conversion rate, the annual impact exceeds €10,700.
This calculation remains an order of magnitude. It shows the risk. You don’t need an academic study to verify it. Cross-reference your Search Console data by page type. Compare May 2024 to today.
| Indicator | Value |
|---|---|
| Monthly organic sessions | 200,000 |
| Informational share | 35% |
| Estimated erosion | 5% |
| Lost sessions per month | 3,500 |
| Lost sessions per year | 42,000 |
| Lost assisted sessions (15% assist rate) | 6,300 |
| Annual impact (basket €85, conversion 2%) | €10,710 |
Traffic doesn’t collapse. It leaks, slowly. No alert in Google Analytics. Direct conversions hold. Revenue doesn’t move right away. Then baskets fill more slowly. The brand loses its place in the research phase.
The problem isn’t the product page. It’s the informational coverage. If AI answers instead of your guide, it doesn’t cite your product page. It cites more neutral sources. Your authority erodes.
5% on Wikipedia means 5% on the web’s best informational asset. On an e-commerce site, erosion can be stronger. Why? AI Overviews trigger mainly on queries where your content isn’t unique. Buying guides are often lists of criteria. AI summarizes them without effort.
Treat informational traffic as an asset. Not a production cost. A margin lever.
Le chiffre fait mal, mais il faut le regarder en face. Appliquez 5% d’érosion à vos guides : voilà ce qui disparaît chaque mois.
La cascade d’érosion sur vos guides
70 000 sessions guides par mois, -5% sous AI Overviews
Google disputes. Wikimedia nuances. You measure.
Google disputes. Wikimedia nuances. You measure.
Google’s dispute is technical. The study takes total external referrals without isolating Google Search. The Wikimedia Foundation acknowledges that many readers access its content via AI without ever visiting wikipedia.org.
The signal remains. The dispute only makes it sharper. The study doesn’t say Google stole 5% of Wikipedia’s traffic. It observes that default AI Overviews availability goes hand in hand with a 5% drop in external referrals.
The authors moved from 15% to 5%. They took time to control variables. A more spectacular estimate was abandoned in favor of a stronger one.
For your e-commerce, the Google-Wikipedia debate is secondary. What matters is your own traffic curve. Do you have a break point in May 2024? Have your guide pages lost impressions on long-tail queries? Have assisted clicks declined?
Tools don’t yet provide a specific « AI Overviews » filter in Search Console. You have to cross-reference. Look at impressions per page. Compare guide pages with product pages. If guides drop faster than the rest, the hypothesis strengthens.
Treat this signal as a leading indicator. Not a verdict. AI shifts SEO value more than it replaces it. Pages that get cited in an AI Overview gain visibility. Pages neither clicked nor cited disappear.
Google says the study can’t isolate its impact. Wikimedia says AI distributes its content without a visit. Both are right. That’s the problem. Classic organic traffic measurement no longer captures the full value of a page.
Semantic cocoons matter more and more. A content architecture that covers a topic in depth has a better chance of being cited. An isolated page gets summarized without credit.
The Wikipedia signal isn’t a short-term alarm. It’s an invitation to redefine your metrics. Direct traffic is no longer the only result. Citation becomes an asset.
Your P&L doesn’t depend on perfect statistical validity. It depends on your ability to anticipate a drop in informational traffic. The methodological debate is healthy. But while Google and researchers argue, your guide pages may be losing ground.
Build your detection system
Build your detection system.
Step one: segment your pages in Search Console by type. One filter for guides, one for the blog, one for product pages. Track impressions and clicks month by month.
Step two: spot the tipping point. In the United States, AI Overviews became the default in May 2024. In Europe, rollout was gradual. Compare your curves before and after that date. No need for a single timestamp. Look at the slope.
Step three: measure citations. Use a rank-tracking tool that flags SERPs with AI Overviews. Note whether your pages appear there. A citation is sometimes worth more than a click. It establishes authority for the transactional queries that follow.
Step four: restructure informational content. Don’t produce more. Produce better. A semantic cocoon covers an entire search intent. AI cites structured sources, not hollow lists.
DOSE, the framework taught by Guillaume Attias at BMO Academy, starts with a perception diagnosis: which bias keeps you from seeing the signal? Here, availability bias. Wikipedia makes the risk visible. Apply it to your own catalog.
Across my clients, some sites lose 12% of informational traffic in six months without an alert. Others gain 22% visibility in citations. The difference: semantic architecture.
I’m not selling the method. I’m showing you the pages.
Stop chasing volume. Seek a place in the answer. Your guide pages need to become the source Google cites. Not the page waiting for the click.
Train yourself. Read Marie Haynes, Aleyda Solis, Lily Ray. They document the impact of AI Overviews on traffic. Search Engine Land and Search Engine Journal publish regular updates. Search Engine Journal covered this study in September 2026.
And you, how many guide pages are already invisible in AI Overviews?
Live SEO audit: measure your cannibalization
An audit of your guide pages, your Search Console, and your semantic architecture. We identify cannibalization before it reaches your margin.
Book a strategic call — 45 minFrequently Asked Questions
Why is Wikipedia a good indicator for e-commerce?
Wikipedia captures a large volume of searches. Your buying guides do the same. If a giant loses 5% of referrals, your guide pages can absorb more.
Google disputes the study. Should I take it seriously?
Yes, it’s a directional signal. The revision from 15% to 5% gives a more reliable method. Look at your informational traffic before and after May 2024.
How do I know if AI Overviews are cannibalizing my pages?
In Search Console, I segment by page type. Then I compare impressions and clicks for guide pages before and after May 2024. If the drop is faster than on product pages, there’s cannibalization.
What impact on my revenue?
Guide pages rarely convert directly. They prepare the sale. With 200,000 sessions per month and 35% informational traffic, 5% erosion means more than €10,000 per year.
What should I do concretely?
I break my pages into blocks. I count how often I’m cited in AI Overviews. Then I reorganize content into semantic cocoons. A citation is something you keep.
