AI Demand Index · query shape · exploratory, then re-tested
How Microsoft Copilot phrases its searches, and which of our first findings held up
When Copilot cites one of our pages, Bing tells the site owner the search it ran. We studied 11,528 of those queries from September, wrote down six hypotheses with their opposite outcomes, then tested them on a fresh export of 8,022 queries from the 30 days to 24 October. Three held, three did not, and one early finding turned out to be a console artifact.
Copilot searches in five to nine words
The median query was 6 words in September and 5 in October, and 68.9% and 65.4% of queries fell between five and nine words. Share of queries by length, September (top bar) and October (lower bar), percent:
Shorter queries earn more citations each
Average citations per query fall as queries get longer, in both exports, and in all 12 properties with data in October. The strict per-property version we registered failed, because the test was too tight: it fails on the September data that produced it. The levels differ between charts because September covers three months and October covers 30 days, so compare direction, not height.
September, 3 months
October, 30 days
One percent of queries carry a fifth to a quarter of citations
- Top 1% of queries hold
- 24.9% / 21.6%
- of citations, September / October
- Gini across queries
- 0.713 / 0.669
- 0 is even, 1 is one query
- Volume vs our share
- 0.051 / 0.132
- Spearman; near zero means a query being popular says little about whether we win it, and it is drifting up
The unlabelled 38% was Bing catching up, not a property of the queries
In September 38.3% of queries had no intent or topic label, and the share of labelled queries rose with volume, so we read that as a pattern. In October only 1.5% are unlabelled. Of 1,745 queries that were unlabelled in September and also appear in October, 1,744 now carry labels. None of the 2,315 that were labelled lost theirs, and the intent is unchanged for 2,310 of 2,315. That finding is withdrawn, and any intent cut on a recent export should carry a note that labels may still be backfilling.
The six registered hypotheses
Registered on 26 October 2026, before the follow-up export was pulled, each with its opposite outcome stated. Thresholds were not changed afterwards.
| # | Registered claim | Verdict | What we found |
|---|---|---|---|
| R1 | Median query length is 5 to 7 words and 60% or more fall in 5 to 9 words. | supported | Median 5; 65.4% in 5 to 9 words. |
| R2 | Citations per query fall strictly with length in at least 10 of 13 properties. | not supported as registered | The test was mis-specified: it also fails on the baseline data (3 of 13). The plain direction, shorter queries earn more citations per query, holds in 12 of 12 properties. |
| R3 | Year-stamping differs 5x or more between properties, highest in Commercial intent. | not supported | The spread fell to 3.5x; Commercial intent is still highest. |
| R4 | The share of labelled queries rises with citation volume. | not supported | Bing backfilled labels; 96 to 100% of queries are labelled in every volume bin. |
| R5 | The top 1% of queries hold 20 to 30% of citations; Gini 0.65 to 0.78. | supported | Top 1% hold 21.6%; Gini 0.669. |
| R6 | Query volume and our citation share are nearly independent (absolute Spearman under 0.15). | supported | Spearman 0.132, up from 0.051, close to the threshold. |
Corrections
- The 38% of queries without an intent label in the September export was a console processing lag. 99.9% of those queries are labelled in the October export, so that finding is withdrawn.
- A criteria / checklist / framework pattern with very high citations per query was one query and is not a finding.
- The published AI Demand Index counted 11 phantom rows for omnirx.org (two multi-line queries split by a line-based parser); corrected in the builder on 2026-10-26.
How to read this
- A row is a query Copilot used when it named one of our pages as a source. It is not a search count and not a click.
- Bing reports only queries for which our own pages were cited, so these are queries that reached us, not the whole market. Bing labels the data a sample.
- The properties belong to one publisher and share templates, so they are not independent observations. Copilot only.
- The two exports are different windows (3 months to 21 September, 30 days, 2026-09-26 to 2026-10-24) and half of the October queries also appear in September, so this is a re-test, not an independent replication.
- One property, lucreya.com, returned no queries in the October window and is excluded from it.
- Only aggregates are published. Cells with fewer than 20 rows are suppressed.
Related: the AI Demand Index · open datasets · instrument integrity
Vincent Wesley Couey, DeepSynthesis. Aggregates released under CC BY 4.0.