In March, Bing added Grounding Query↔Page Mapping to the AI Performance report in its Webmaster Tools, so you can see which queries trigger citations to which specific URLs in the answers from Copilot, partners and the APIs that serve third-party products. To understand it properly, these are the 2 concepts you need to know:
- Citations: how many times your content is referenced by AI systems.
- Grounding queries: These are the queries the AI used internally for retrieval, not necessarily what the user typed. The AI rewrites, expands and normalizes the query before searching.
The problem is that if you want to get the grounding queries for each page, or the other way around, you have to go one by one, and the exports don't include this mapping. For sites with hundreds of queries and thousands of pages, that's useless.
A while back I built myself a workaround that did the matching by semantic similarity, but it's not the same thing.
I've built a Chrome extension (and a bookmarklet with the same code, for anyone who doesn't want to install anything) that extracts the whole mapping and downloads it as a ZIP with 3 CSVs plus an Excel file with the same 3 tables. It runs in your browser with your session, without sending data to any external server. If the account hits Bing's rate limit, it waits and retries on its own, so don't worry, you just need a bit of patience. You can leave the tab open and keep doing your thing in another one.
Update (October 2026): it's now a Chrome extension too
The bookmarklet is now also a Chrome extension, Bing AI Performance Exporter by Natzir, which updates itself. It's already live on the Chrome Web Store. Both versions share the same code and come with these changes:
- One ZIP and one Excel file instead of three separate downloads. The ZIP holds the 3 CSVs, and the Excel file has the same 3 tables, one per sheet, with a frozen header row and filters.
- The CSVs follow your browser's language. In Spanish, for example, they use semicolons as separators and a decimal comma, which is how Excel in that locale opens them without any tweaking.
- The file names include the dates of the range, for example
bing-aip_yourdomain-com_2026-07-09_2026-10-08_mapping.csv.
- It exports the same range the Bing report uses. All the ranges end yesterday: 7 D and 30 D count days, and 3 M and 6 M count calendar months back from today. With Custom it reads the dates you've set, and with Compare it asks you which dates you want.
- You can stop it and pick up later. Progress is saved for 24 hours per site and range, and queries that fail can be retried separately.
If you already had the bookmarklet: bookmarklets don't update themselves. Delete the old bookmark and create the new one from the button below.
New from Bing (October 13): Topics Portfolio, in beta, groups citations by topic (Topic Size, Sites Cited and Queries, with 1, 2 or 3-month windows) and adds Top Opportunity Queries: queries where Bing cites other sites but not yours, with volume shown in ranges. It's an official prompt gap report. That said, Bing asks you to treat the numbers as directional signals, not exact accounting.
Update (July 2026)
Bing announced on June 16 that it was expanding the AI Performance report: each grounding query now comes labeled with Intent (Informational, Commercial, Research…), Topic and Citation Share (the % of that query's citations your site gets). According to Microsoft, the labels are AI-generated. There's also a new 6-month range and a Compare mode. What hasn't changed: the native export still doesn't include the query ↔ page mapping, so you still need the bookmarklet.
I've updated the bookmarklet to pull all of this in: mapping.csv and queries-summary.csv now include the intent, topic and citation_share columns (in %, same as in the UI), so you can now pivot the mapping by search intent or by topic. It also detects the 6-month range if you have it selected.
Installation
The Chrome extension is called Bing AI Performance Exporter by Natzir.
For the bookmarklet, drag this button to your bookmarks bar:
Bing AI Performance Exporter
Drag the button to your bookmarks bar. Clicking it here does nothing (bookmarklets only work on the target page).
If you can't see your bookmarks bar, turn it on with Cmd+Shift+B (Mac) or Ctrl+Shift+B (Windows).
How to use it
- Go to Bing Webmaster Tools and open the AI Performance report for the property you want to analyze.
- Pick the date range: 7 days, 30 days, 3 months, 6 months or Custom. It exports the same range you see in the report. If you have Compare on, the panel asks you which dates you want.
- Click the extension icon or the bookmark in your bookmarks bar.
- You'll see a panel with a spinner and live progress. When it's done, it gives you a ZIP with the 3 CSVs and an Excel file.
- If Bing rate-limits you midway, it waits 5 min and retries on its own. You don't have to do anything.
Expected times:
| Site size |
Approximate time |
| 5-20 queries / pages |
~5 seconds |
| ~100 queries / ~250 pages |
~2 minutes |
| ~500 queries / 1,000 pages |
~8-10 minutes |
| ~1,000 queries / 2,000 pages |
~15-20 minutes |
| 2,500+ queries (enterprise) |
30-60 min, with automatic waits for rate limits |
What it does if Bing rate-limits you
It runs at a pace Bing can handle, with a 45-second pause every 150 queries. If Bing still detects too many requests:
- It shows "Bing is pausing requests. Resuming in 5:00. You can leave this tab in the background."
- It waits 5 minutes, or however long Bing asks for if it specifies a time.
- It sends a light probe to the API.
- If the API responds OK, it carries on from where it left off.
- If it's still throttled, it waits another 5 minutes and tries again. It repeats this for up to 12 cycles (1 hour of waiting max) and then stops with a Resume button so you can continue whenever you want.
Everything you've mapped so far stays saved in the browser for 24 hours, per site and range. If for whatever reason you close the tab or reload, open the same report, click again and it resumes automatically from where you left off. And if you want the data before it finishes, Stop and download what I have gives you whatever it has so far. You never lose the data 😀
The only thing you have to do is not reload the tab while it's waiting. If you reload while Bing is being strict with the rate limit, you may see a page with a JSON "too many requests" message: that's normal... reload in 10 min, the site comes back and you can carry on with the process, but it's better not to reload in the first place.
What you get
A ZIP with three CSVs and an Excel file with the same three tables, one per sheet. The files use the prefix bing-aip_<domain>_<start>_<end>_, with the dates of the range. The CSVs follow your browser's language: in Spanish they're semicolon-separated with a decimal comma, which is what Excel expects in that locale.
mapping.csv
The full mapping in long format. One row per query+URL combination with its citations. It works for both a "queries per page" and a "pages per query" pivot in Excel or Sheets, with no joins needed.
queries-summary.csv
One row per query: total citations and how many pages are mapped to it.
pages-summary.csv
One row per page with total, mapped and unmapped citations, plus the number of associated queries. This is the most interesting table because it exposes Bing's sample bias in the unmapped_citations column: pages with total citations but no queries identified in the mapping (the API doesn't expose them).
Bing's sampling
Bing's own UI warns that this isn't all the data, it's a sample of it. In other words, the query↔page mapping Bing exposes is partial. For each URL you can see the total citations associated with it (how many times AI systems referenced it in total), but Bing can only attribute part of those citations to a specific grounding query through the sample. The rest is left as "citations with no identified query". I call the portion that does have an associated query "mapped citations".
I tested it on 6 sites with very different profiles:
| Site |
Type |
Pages |
Total citations |
Mapped citations |
Coverage |
| A |
Professional site, single-locale, micro |
13 |
142 |
35 |
24.6% |
| B |
Vertical online store, single-locale, small |
263 |
15,224 |
10,974 |
72.1% |
| C |
Transactional platform, multi-locale, medium |
2,380 |
77,164 |
31,185 |
40.4% |
| D |
B2C services platform, single-locale, medium |
1,219 |
213,841 |
83,492 |
39.0% |
| E |
General online store, single-locale, large |
4,912 |
1,820,471 |
845,247 |
46.4% |
| F |
Transactional platform, multi-locale, large |
4,751 |
9,628,043 |
9,297,849 |
96.6% |
Takeaways from the coverage comparison:
- Coverage ranges from 24% to 97% depending on the site. There's no obvious "more/less volume = better coverage" pattern: site E (1.8M citations) has 46% while site F (9.6M) has 97%. The behavior looks erratic, but I share some hypotheses later on in the findings.
- The % of pages with at least one mapped query is even lower: typically between 10% and 25% on medium/large sites. Most cited URLs don't have a single associated query in the sample.
- For small sites (fewer than 200 total citations) the data is practically useless. Only the top queries get mapped.
The pages-summary.csv file gives you that information explicitly in the unmapped_citations column. If a URL has 28 citations but 0 mapped queries, that page is getting "ghost" generative traffic, and Bing doesn't show you where it comes from.
Do grounding queries match real search demand?
One question I've always wanted to answer: do those queries exist as real searches, or is AI grounding generating queries that no human actually searches for?
To answer it, I cross-referenced the grounding data with two sources:
- Bing Webmaster Tools API: Bing search clicks and impressions per query, same date range. The API caps out at 6,700 records per aggregate call, so for sites with thousands of queries you have to look them up one by one with
GetQueryPageStats(query), one call per grounding query.
- Google Search Console: impressions per query for the same period, exported to CSV.
What % of grounding queries can I find in my Search Console or Webmaster Tools? And how many are unique to grounding, invisible in my usual reports?
For each query that shows up in the grounding report, I check whether that same query also gets impressions in Google Search Console (regular Google search) and in Bing Webmaster Tools (regular Bing search):
| Site |
Type |
Total grounding Q |
∩ all 3 sources |
Bing only |
Google only |
Grounding only |
| A |
Professional site, single-locale, micro |
3 |
0 (0.0%) |
0 (0.0%) |
2 (66.7%) |
1 (33.3%) |
| B |
Vertical online store, single-locale, small |
68 |
43 (63.2%) |
7 (10.3%) |
13 (19.1%) |
5 (7.4%) |
| C |
Transactional platform, multi-locale, medium |
569 |
79 (13.9%) |
25 (4.4%) |
277 (48.7%) |
188 (33.0%) |
| D |
B2C services platform, single-locale, medium |
1,098 |
325 (29.6%) |
38 (3.5%) |
462 (42.1%) |
273 (24.9%) |
| E |
General online store, single-locale, large |
2,546 |
322 (12.6%) |
28 (1.1%) |
1,588 (62.4%) |
608 (23.9%) |
| F |
Transactional platform, multi-locale, large |
2,473 |
352 (14.2%) |
44 (1.8%) |
1,631 (66.0%) |
446 (18.0%) |
% of recoverable queries (present in at least one search engine): A 67%, B 93%, C 67%, D 75%, E 76%, F 82%. Median ~76%. Between 7% and 33% of the queries that trigger grounding don't show up in either Bing or Google.
Weighted by citations, the picture changes
The previous table counts queries (each one is worth 1). If we weight by citations (how much each query contributes to the total references to your URLs in AI answers), the "grounding only" queries shrink, because they tend to be long tail with few citations each:
| Site |
Total citations |
∩ all 3 sources |
Bing only |
Google only |
Grounding only |
| A |
35 |
0 (0.0%) |
0 (0.0%) |
28 (80.0%) |
7 (20.0%) |
| B |
11,052 |
10,380 (93.9%) |
322 (2.9%) |
285 (2.6%) |
65 (0.6%) |
| C |
31,185 |
16,740 (53.7%) |
3,676 (11.8%) |
6,552 (21.0%) |
4,217 (13.5%) |
| D |
83,804 |
59,180 (70.6%) |
2,724 (3.3%) |
15,443 (18.4%) |
6,457 (7.7%) |
| E |
851,659 |
644,043 (75.6%) |
86,383 (10.1%) |
95,624 (11.2%) |
25,609 (3.0%) |
| F |
9,304,615 |
8,521,096 (91.6%) |
274,386 (2.9%) |
474,232 (5.1%) |
34,901 (0.4%) |
Note: here the total adds up the citations for each query, which for B, D, E and F is slightly higher than the mapped query↔page citations in the coverage table.
Findings
- The grounding long tail is invisible to regular analytics. By query count, between 7% and 33% of what triggers AI citations doesn't show up in Bing search or GSC. If your plan was to "audit AI queries by looking only at regular Bing Webmaster Tools and Search Console", you're missing anywhere from 7% to a third of the picture.
- But those invisible queries carry little weight in citations. That same long tail is made up of lots of queries with few citations each. That's why, weighted by citations, "grounding only" drops to 0.4%-20%. Depending on what you want to measure (discovering new queries vs understanding what drives your AI volume), one metric or the other is the one that matters.
- There are two cases where "Bing only" weighs heavily in citations: sites E (10.1%) and C (11.8%). I won't risk explaining why. Possible factors: audiences with a stronger Microsoft affinity (Edge users), international markets the user's GSC doesn't capture well, or queries the AI rewrites internally that end up closer to Bing's index than to Google's. I do know those sites have products that are popular outside Spain and get searched in Spanish in Latin America. I'm leaving it as an observation, not a conclusion.
The "grounding only" bucket has several kinds of weird queries
Note: the example queries are made up so I don't expose client data, but the patterns are faithful to what shows up in the real logs.
When you filter the queries that ONLY show up in grounding (not in Bing search or Google), these are the kinds I see:
- Truncated or fragmented. Bing exposes grounding queries that are cut off as if they were complete:
zapatillas running mujer talla pequeñ, zapatillas running mujer tal, zapatillas running muje (Spanish for "women's running shoes small size", cut off at different points). I don't know if these come from users who hit enter before finishing or from some kind of autocomplete.
- Brand-augmented by the model's prior. The user probably asked "mejor robot aspirador para pelo de mascotas" ("best robot vacuum for pet hair"), but the query Bing reports as the source is
roomba pelo mascotas opiniones, xiaomi robot aspirador mascotas, cecotec conga pelo perro (roughly "Roomba pet hair reviews", "Xiaomi robot vacuum pets", "Cecotec Conga dog hair"; Cecotec is a Spanish appliance brand). The model adds brands it assumes are well known in the vertical, and that gets logged as if it were the original query. Your site gets citations even though the user never typed the competitor's brand.
- Multiple languages when the site serves multiple markets. On global sites you see queries in languages the local market's GSC doesn't pick up:
how to descale a coffee machine in English, como limpar máquina de café in Portuguese, come decalcificare macchina caffè in Italian. We already know the AI can run grounding in another language to expand the corpus.
- Very specific regional variations. For example, in professional services:
abogado herencias internacionales la rioja / melilla / ceuta / teruel ("international inheritance lawyer" plus small Spanish regions and cities). Informational long tail, but with an expansion that doesn't make much sense.
That would explain why these queries don't exist in Google: they're things no user would search for, unique to how fan-out works.
What about at the URL level? Are there URLs cited by AI that Google or Bing organic doesn't serve?
Same idea, but weighted by cited URL. For the comparison to be accurate, you have to crawl every grounding URL, follow 301 redirects and canonicals, get the final status, and compare the final destination against GSC. The exact URL won't do, because the AI memorizes old slugs and cites variants with canonical, noindex, 404...
| Site |
Grounding URLs (after redirects and canonicals) |
Total citations |
AI-only URLs (% of grounding) |
AI-only citations (% of total) |
| A |
13 |
142 |
0 (0.0%) |
0 (0.0%) |
| B |
262 |
15,224 |
0 (0.0%) |
0 (0.0%) |
| C |
2,378 |
77,164 |
167 (7.0%) |
709 (0.9%) |
| D |
1,219 |
213,841 |
15 (1.2%) |
172 (0.08%) |
| E |
4,836 |
1,820,471 |
12 (0.2%) |
169 (0.009%) |
| F |
4,744 |
9,628,043 |
0 (0.0%) |
0 (0.0%) |
As you can see, a few AI-only URLs remain, but in my case the reason is purely technical. I don't have access to the full BigQuery data and/or I don't have multiple Search Console properties to get around sampling. I also don't have access to more than 3 months of logs.
Even so, after filtering out technical artifacts, sampling, cross-language equivalents and presence in Bing, true AI-only (URLs invisible in Google and Bing at the same time) comes down to 0.009%-0.9% of citations. The gap between AI and search engines is residual, and most of what looked like "discovery" turned out to be legacy URLs, deleted pages, canonicalized duplicates, data suppressed by sampling, and language variants of pages that Google does index in another locale.
Privacy
- The script runs in your browser, on the BWT page, with your session cookies.
- It doesn't send anything to any external server.
- It doesn't use analytics.
- Partial data is saved in your browser (in the extension, or in Bing's own localStorage with the bookmarklet), only so it can resume after a rate limit. It expires after 24h.
- The source code is on GitHub. You can audit it before installing it, and the extension has its own privacy policy.
How it works under the hood
Internally, Bing uses two REST endpoints that are already authenticated by your session:
POST /webmasters/api/aiperformance/searchqueries/stats returns queries
POST /webmasters/api/aiperformance/pages/stats returns pages
The extension (and the bookmarklet, which uses the same code):
- Requests the CSRF token from
/webmasters/auth/token.
- Lists all queries and all pages (paginated).
- For each query, calls
pages/stats with Query: X and saves the cited pages with their number of citations.
- Applies a preemptive cooldown every 150 queries (45 seconds) so it doesn't burn through the rate-limit quota.
- If it gets a 429 or a body with "too many requests", it pauses for 5 minutes (or however long Bing asks for), probes the API and retries. It loops until it's resolved or it hits the 1-hour limit, and then stops with a button to continue.
- The
pages-summary is derived from the mapping above: unmapped_citations = total - mapped.
- Builds a ZIP with the 3 CSVs and an Excel file in memory, and offers them as downloads.
When Bing returns 404 NoDataFound, it means "that page/query has no mapping in the sample". It's not an error.
Technical note: pulling the data the other way around (page -> query) is redundant. It's the same data, and doing it this way takes less time.
Troubleshooting
"Open the AI Performance report to export its mapping". You opened the panel from another page. The panel's Open AI Performance button takes you to the report for the site you had open.
"Your Bing session has expired". Your Bing session timed out. Reload the page and click again: it picks up where it left off.
"Which dates should it export?". The panel couldn't read the report's date range, usually because you have Compare on. Pick one of the ranges or enter the dates manually.
I see "Bing is pausing requests. Resuming in 5:00". Don't worry, that's normal. Don't reload. Leave the tab open and come back in 5-10 min. It picks up where it left off on its own.
Next steps
If you use it and it helps, let me know. If something blows up, let me know too.