Currently Available: Need a skilled Software Developer for your next project?
Categories
SEO, GEO and AEO

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 data, such as citation counts for grounding queries, pages, or dates. A grounding query is a phrase Bing groups with retrieval and citation activity; it is not necessarily a complete user prompt.

To load these exports reliably, record the report view and filters. Preserve the original file and its headers, then store each kind of report row at its own level of detail. Repeated downloads are snapshots of selected periods, so a later export should not automatically be added to an earlier one. Microsoft documents the report’s measures and export categories, but does not publish a definitive CSV schema with guaranteed headers or row keys. Treat the warehouse schema below as an implementation design, then map the headers in each actual download.

Keep the report’s row meanings separate

Bing reports citation activity across Microsoft Copilot, AI-generated summaries in Bing, and selected partner experiences. Its core views include grounding-query counts, page-level citation counts, and time-series metrics. Each view answers a different question: a query row summarizes activity associated with a phrase, a page row summarizes activity for a URL, and a time-series row records a metric for a date. The AI Performance documentation describes the measures and export categories.

These summaries do not provide a complete record of individual answers or exact user prompts. Bing says the data is sampled, and low-volume activity may not appear. Totals can differ between report views. Treat the numbers as indicators of citation activity, not as clicks, traffic, rank, or a census of every AI citation. For a fuller explanation of what the report does and does not measure, see what Bing AI Performance reports.

Microsoft has also added fields and views over time. The June 2026 preview announcement describes Intent, Topic, and Citation Share fields, along with period comparison. Citation Share is the percentage of citations attributed to a site for a grounding query; it does not represent traffic share, ranking, or quality. The Topics Portfolio beta adds a separate topic-level view, including opportunity queries whose citation volumes appear as ranges. Store that view separately from grounding-query counts, and keep ranges as ranges rather than treating them as exact counts.

Inspect the downloaded file before mapping it

Bing offers CSV and Excel exports, but its public documentation does not guarantee header names or specify data types. It also leaves null rules, delimiters, and any primary key unspecified. Start by recording the file’s actual headers and identifying its report view. Do not write a parser that assumes one permanent header set.

For each file, capture the site, view type, selected start and end dates, applied filters, filename, file hash, ingestion time, and export time when the file or download record provides it. Bing says applied filters affect exports, so a filtered file needs to remain distinguishable from an unfiltered one. Mark unavailable context as unknown rather than reconstructing it from memory.

Preserve the original file and each raw header and cell value before converting anything. This lets you reprocess older exports when your mapping changes or Bing adds a column. Keep unknown columns in staging or an extensible JSON field, and flag them for review rather than dropping them.

Map each view to its own table

The following is a recommended warehouse design, not a list of verified Bing headers. Map the names in your downloaded file to these canonical fields and retain the original names alongside them.

Table One row represents Useful fields
bwt_ai_export_run One downloaded file or report export export_id, site_id, report_type, view_type, selected dates, filters, export and ingestion timestamps, filename, file hash, parser version, raw headers
bwt_ai_query_snapshot One grounding-query row in one export export_id, raw grounding query, citations, raw Intent, raw Topic, raw Citation Share
bwt_ai_page_snapshot One page row in one export export_id, raw page URL, citations
bwt_ai_daily_snapshot One date and metric in one time-series export export_id, metric date, citations, cited pages where present
bwt_ai_query_page_snapshot One query–page pair in one export, only when the file includes the pair export_id, grounding query, page URL, pair-level citations where present

An export_id ties rows to the file and its context. Keep query, page, and daily facts in separate tables: a page-level total does not identify which query produced its citations, and a query summary does not identify the cited page. Create a query–page row only when the exported data explicitly contains that pair. Do not join separate query and page summaries to invent one.

For example, store the source phrase in grounding_query_raw and a parsed count in citations, while retaining the original count cell for auditing. If Bing supplies Intent, Topic, or Citation Share, add those values without assuming they appear in every export. The June feature announcement describes these as report features, but does not guarantee a fixed export header for each one.

Normalize without erasing source values

Keep the exact grounding-query text. To compare wording variants, create a separate comparison key; do not replace the phrase Bing exported. Bing groups phrases associated with activity, so the text should not be treated as an exact prompt.

Keep the exported URL in a raw field. A separate normalized URL can support analysis, but lowercasing paths, removing query parameters, or merging trailing-slash variants can collapse distinct source values. Measure collisions in the normalized field before using it to deduplicate rows.

Parse counts and dates only after checking the file’s actual number and date formats. Store the raw cell when parsing fails. For Citation Share, inspect whether the export expresses a value as a percentage such as 12.5% or as a ratio such as 0.125 before converting it. Keep both the raw value and the normalized numeric value, and document the conversion rule.

Treat Intent and Topic as Bing’s labels, not as a taxonomy your warehouse controls. Microsoft describes them as AI/ML classifications, and the labels can evolve. Preserve the exported label exactly so later changes do not rewrite older snapshots.

Preserve each export as a snapshot

Assign every imported file a unique export_id, even when two files cover the same site and date range. Include the selected period and filters in the export manifest, along with the report view and file hash. Two rows with the same query text are not necessarily the same observation if their date ranges, filters, or report views differ.

Do not add overlapping date-range totals together as though each export were a new increment. Bing refreshes the report daily and says processing can take time; a later export for the same period can change as data is refined. Keep both snapshots, then compare them only if they cover the same site and view. Their period and filters must also match. For daily chart data, a downstream table can select the latest accepted snapshot for each site and date while preserving earlier loads for audit.

Version the parser as well as the data. Store a parser version and a schema fingerprint based on the ordered source headers and relevant file characteristics. When a new file changes the fingerprint, review the added, missing, or renamed headers before changing the mapping. This makes schema changes visible and allows old files to be reprocessed.

Validate rows before loading them

Check that the file is readable and that the detected view contains the fields your mapping requires. Flag duplicate or renamed headers and retain extra columns. Quarantine values that fail numeric or date parsing. Citation counts should parse as nonnegative numbers. If the same dimension row appears more than once within one export, flag it for review rather than summing it automatically.

Also verify that each load has a site, report view, date range, and filter context, or explicitly mark any missing context as unknown. Preserve raw URLs and measure collisions introduced by URL normalization. Compare query, page, and time-series totals as diagnostics only. Sampled reporting means those totals are not guaranteed to match.

Illustrative schema label: article setup: technical explainer: direct mechanism (private trial 2026-10-04). The schema and field mappings here are hypothetical implementation guidance, not verified Bing headers or private-trial results. To compare the exported measures with ordinary search reporting, see how to read Bing AI Performance exports.

What I'm building

Delegate tasks. Get software.

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

Email updates

Usually a new article and a few links I found interesting.

No spam. Unsubscribe with one click.

Leave a Reply

Your email address will not be published. Required fields are marked *