552,676 AI citations: what Ithy’s search data actually shows

What we measured

This is a first-party case study of Ithy, an AI-assisted research and article service. The observation window was September 25 through October 22, 2026, inclusive. We exported the reports on October 26. All figures below describe that historical window; they are not results attributed to a later product release.

Bing defines a citation as content visibly referenced or shown as a source in a supported AI answer. Its average cited-pages measure is a daily average, rather than the number of distinct pages across the whole period. The report covers supported Microsoft and partner experiences; it does not establish visibility in every AI app. [1]

The chart and detailed export deliberately have different totals. Microsoft describes the detailed data as sampled. We have not multiplied the sample to estimate missing pages or treated 232,739 as a second set of citations to add to 552,676. [2]

Cited pages and clicked pages were different groups

The Google Web page export contained 281 article paths with positive clicks. The complete Bing Web page table contained 505. They shared 33 paths, giving 281 + 505 − 33 = 753 distinct clicked article paths. We compared URL paths, without treating different report rows as different people.

The 1,970 figure is easy to overinterpret. It means no click was recorded for those paths in these particular search-page exports. It does not mean that no visitor ever reached them, that AI answers caused a loss of clicks, or that they received no direct or other referral traffic. The reports do not provide the evidence needed for those conclusions.

We also cannot divide total clicks by total citations to obtain a meaningful AI click-through rate. The counts have different scopes and are not linked user journeys. A citation can help establish that a page is being used as a source; a visit is a separate outcome.

English accounted for the most citations; other languages accounted for more identified pages

We matched 2,371 of the 2,374 cited article paths to stored language metadata. Of those, 92 had an undetermined language. Among the remaining 2,279 language-identified paths, 1,312 were in languages other than English: 57.6%.

English nevertheless had 152,767 of the 224,914 sample citations attached to language-identified articles, or 67.9%. A larger number of pages is not the same as a larger number of citations. Spanish contributed 541 cited article paths; French 189; German 114; and Japanese 92. These observations make native-language content worth investigating, but they do not establish that translating an English article will produce the same demand.

This table covers 2,371 metadata-matched article paths and 230,394 sample citations. The three unmatched article paths contributed 210 citations; the non-article page contributed 2,135. Together they reconcile to the detailed export’s 232,739 citations. Language labels describe pages, not visitors’ countries or the languages of all questions. We did not manually audit every language label.

What this changes for our publishing decisions

A page can be useful to an answer engine even when it rarely appears in a conventional click report. For Ithy, that argues for maintaining articles already being cited, checking whether their explanations remain accurate, and keeping their sources accessible. It does not justify publishing every saved draft or treating a citation as proof that the article is correct.

The language comparison gives us another practical direction: review existing native-language explanations on their own merits, instead of making English traffic the only selection criterion. Different subjects, article ages and question patterns could explain part of the difference. This observation is a reason to investigate, not a controlled translation experiment.

We are keeping search clicks, AI citations and product use as separate measures. Clear explanations, stable public URLs and useful source links can serve readers arriving through either route. Whether those readers become regular users is a further question, and this dataset does not answer it.

Daily aggregate data

The table below publishes the complete 28-row aggregate chart export used for the citation total and daily average. It contains no individual prompts, account records or visitor identifiers. The language and overlap tables above publish the other aggregates used in this article.

Method, limitations and reuse

We summed the 28 exported daily citation values, averaged the daily cited-page values, deduplicated article URL paths in the detailed page export, and compared those paths with positive-click paths in the matched Google Web and Bing Web reports. All positive-click rows fit in the saved Google export; the Bing page table was captured completely. We then joined cited article paths to Ithy’s public catalogue metadata for language counts. We did not publish the underlying query phrases or individual article-query mappings.

This is self-reported observational data from one publisher, with AI-assisted drafting and arithmetic checks against the retained exports. It is not an independent audit, a benchmark for the wider web, a ranking experiment or a claim about ChatGPT coverage. The detailed sample may omit activity, and stored metadata can be imperfect. No causal effect of a content, model or indexing change has been established.

You may reuse the aggregate numbers with attribution to Ithy and a link to this case study. Please retain the observation dates and distinguish citations from clicks. Questions or corrections can be sent to wluk@ithy.com.