Parsing Bing AI Performance Exports Into Database Rows
Bing Webmaster Tools’ AI Performance report summarizes when a site’s pages appear as sources in AI-generated answers. Its downloads contain aggregated report…
Bing Webmaster Tools’ AI Performance report shows when Microsoft AI products cite pages from a website. A citation means an AI-generated answer names a page as a source. It does not mean someone clicked the link or visited the site. The report covers Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations, as described in Microsoft’s report announcement.
The exports help you inspect reported citation activity by page, date, and grounding query. A grounding query is a grouped phrase associated with content used to support an AI answer. It is not necessarily the exact prompt a person typed. Treat the files as summaries based on sampled data. They do not establish search rankings, traffic, revenue, or why a page appeared.
Bing offers CSV and Excel exports for grounding queries, page-level citations, and time-series metrics. The active report filter also affects exported data, so record the selected date range and any query or page filter alongside each file. Microsoft’s AI Performance help describes the report and its export behavior, but does not promise a fixed, versioned CSV schema.
Exports reviewed so far have had three formats. A third-party inspection of the files reports the names and headers below. Use them as examples, not as a guaranteed format. Check your own header row before writing a parser.
| Observed file shape | Example fields | What each row represents |
|---|---|---|
Page statistics, sometimes named AIPageStatsReport |
Page, Citations |
Citation activity for a URL |
Overview or time series, sometimes named AIPerformanceOverviewStats |
Date, Citations, Cited Pages |
Daily citation activity and the number of cited pages |
Search queries, sometimes named AISearchQueriesReport |
Grounding Query, Intent, Topic, Citations, Citation Share |
A grounding phrase and its reported measures or classifications |
Each row type describes something different: a page row represents a URL, a time-series row represents a date, and a query row represents a phrase. Keep the files separate unless the data provides a reliable way to connect their rows. Adding their counts together or treating each row as a citation event mixes different measures.
Before parsing, save the original files unchanged. In a filename or separate note, record the site property, export date, selected period, report view, and active filters. For each file, check its delimiter and encoding. Also check the date format, blank fields, and number formatting. Bing does not guarantee that every export will use the observed field names or formatting.
A grounding-query row names a phrase linked to reported citation activity, but it does not show the person’s exact prompt or the AI’s answer. One query can be linked to multiple pages, and one page can appear under multiple queries. If a query file has no page column, it cannot show which URL goes with each query count.
Use the report’s interactive filters to examine query-page relationships. When you export filtered results, note which query or page you selected. If the file does not include that value in a column, record it separately as analyst-supplied information. Do not assign a query’s count to the top page row or infer a one-to-one match from totals that happen to be the same.
When present, the Intent and Topic fields group or classify query activity using AI-generated labels. Microsoft introduced these features as previews. Treat the labels as useful categories, not verified descriptions of what a person intended. A blank label means the row has no label; it does not show that the query has no intent or topic. Microsoft’s feature announcement also describes Citation Share and period comparisons. Its Topics Portfolio beta announcement describes a broader thematic view of grounding queries, not a complete prompt-to-page export.
A citation count shows reported citation activity during the selected period. It does not count clicks, visitors, conversions, or necessarily distinct answers. Compare counts only within the same report view, date range, and filter settings.
Microsoft says the report uses sampled and summarized data. Some low-volume activity may be missing, and totals can differ between views while Microsoft processes the data. As a result, the page-file total, time-series total, and sum of query rows do not have to match. If they differ, check the date ranges, filters, row types, and export times. Do not duplicate or redistribute counts to make the totals match.
Page-level rows show which URLs have reported citation activity and how their counts compare in the selected view. A higher count does not establish that a page ranked higher, appeared more prominently in an answer, or performed better for the business. Preserve the exported URL as-is. If you later group URL variants by removing tracking parameters or merging slash variants, keep the original URL and document the normalization rule.
When available, Citation Share shows what share of citations for a grounding query went to the site. It is not traffic share or a ranking score. Its percentage is calculated separately for each query, so do not add percentages across queries. The measure also does not identify which sites received the other citations.
First identify what each file covers and what each row represents. Compare counts only when the date ranges and filters match. To investigate a query-page relationship, apply the report’s interactive filter and record the selected value with the export. This adds context to the rows, but separate summary files still do not provide a complete record of individual citation events.
For example, suppose a page export reports 12 citations for /guide/ and a query export reports 7 citations for a phrase related to that guide. Those values do not show that the phrase produced 7 of the page’s 12 citations. Without a supported query-to-page mapping in the data being reviewed, that connection remains unproven.
When comparing periods, use equivalent date ranges and filters, and record content changes or other relevant events separately. A change in citation counts after a page edit is an observation, not proof that the edit caused the change. Microsoft’s comparison feature helps show period differences; it does not explain their cause.
AI Performance does not measure search rankings, page authority, or a URL’s importance. It does not show every prompt or answer, or explain why an AI experience selected a source. A citation also does not prove that anyone saw the answer, clicked through, or converted. Use separate analytics to measure visits and business outcomes, and interpret attribution cautiously.
The report also covers the Microsoft and partner experiences included in its stated scope. It does not represent citations across every AI service. Use the export to identify reported citation patterns and pages worth examining, then keep conclusions within what the data actually records.
Give Vroni a GitHub issue, bug report, spec, or rough idea. It reads the repo, plans the change, writes code, runs checks, and works toward a review-ready pull request.
Take a look at vroni.com