What Actually Gets Cited

2. Method

What we measured. Two first-party instruments.

1. Bing Webmaster Tools, AI Performance tab, which reports "Citation sources: Microsoft Copilots and Partners" with a total citation count, a daily trend, a Pages tab and a Grounding Queries tab giving citation share per query. Read per site on 7-day, 30-day and 3-month windows.¹
2. GA4, across five properties in one account, with affiliate_click as the key event; no single property covers the network.² Referrals from chatgpt.com arrive in the "AI Assistant" channel.³

We also pulled Bing's GetRankAndTrafficStats and GetCrawlStats through the JSON API, and fetched live pages with an OAI-SearchBot user agent to count what a crawler receives in the raw server response.⁴⁵

The estate. About 60 live sites, mostly Astro on Cloudflare Pages, carrying affiliate and reference content in consumer and history niches. On 2026-08-08 the split was 11 custom .com sites and 58 .pages.dev platform subdomains, with 69 properties verified in Bing.⁴ Measurements run 2026-07-21 through 2026-09-05.

Honest limits. These are our sites, not a sample of the web: affiliate and reference content in narrow niches, run by one organisation, so unseen site-level factors are shared across the estate. We have no Perplexity data and no Google AI Overview data. Bing's citation numbers are readable only in the web interface, because the Bing API exposes no citation method.⁶ GA4 undercounts outbound clicks: in one 28-day window it logged 142 affiliate_click events while Amazon reported 571, so treat GA4 clicks as a floor.⁷ Direct traffic runs 84 to 86% of sessions at a 4-second average and a 0.45 to 0.72% key-event rate, and we treat it as bot noise, not audience.³⁸ Nothing here is a prediction.

---

3. Findings

(a) The Bing gate: a platform subdomain earned nothing

On 2026-08-08 we pulled GetRankAndTrafficStats across all 69 verified Bing properties. Only 5 had any impressions, and the split was clean: 5 of our 11 custom .com sites (askhistoryai 146, ebikeriderlab 47, panelpicks 43, watchshowdown 15, openscriptorium 1), and 0 of 58 .pages.dev sites, with no exceptions.⁴

Copilot citations sat strictly downstream of that. Read the same day across 31 sites: ebikeriderlab with 45 impressions at positions 2 to 10 earned 24 citations, while brewstationlab and nasdecoded, both at 0 impressions, earned 0.⁹ (That site's impressions read 47 and 45 in two records both dated 2026-08-08, and 66 on 2026-08-15.⁴⁹¹⁰)

The conversion side agreed: on 2026-08-20 all of the roughly 40 .pages.dev hosts on our fleet GA4 property logged 0.00 key events, and for 2026-07-22 to 08-18 about 60 of them drew 23 to 51 sessions each at 0 to 1 second engagement and zero key events.⁸¹¹

A custom domain was necessary and not sufficient: on 2026-08-08 six of our custom-domain sites were also at zero citations.⁴ Our clearest control is a site implementing our own cited-page template closely, 2,902 words with the full JSON-LD stack and an answer capsule naming real models, that earned zero.⁹

Crawl, not sitemap submission, was the mechanism. GetCrawlStats on 2026-08-08 showed watchshowdown 2,220 pages crawled and askhistoryai 1,805, against cleansweeppicks 0 and deskriselab 0, both domained for weeks and both green on every signal our side, GetFeeds included. Bing read a sitemap listing 33 URLs and crawled none of them.⁴

(b) Page count is not the lever

Sitemap <loc> counts against Bing impressions and Copilot citations, read 2026-08-15:¹⁰

The relationship was inverse on our sites. ebikeriderlab earned about 0.92 citations per published page; watchshowdown earned about 0.00 across 2,953, with 421 URLs indexed out of 2,953 submitted, about 14%, 8 impressions and no backlinks.¹⁰

The index-bloat measurement made it concrete. Over 2026-07-30 to 2026-08-20, Bing crawled 3,500 pages of watchshowdown while its InIndex counter stayed pinned at 65, its not-indexed counter climbed from 34 to 1,146, and its crawl rate fell to 0 to 2 pages per day. That is 65 of 3,592 URLs indexed, 1.8%, with 3,404 of those being auto-generated comparisons.¹²

The extreme case is a single page. On 2026-08-15, Bing's AI Performance Pages table for ebikeriderlab contained exactly one row: the homepage, with 56 citations. 60 of the site's 61 pages had earned zero, including about 22 "best X" roundups and about 16 explainers.¹³ That is not a property of page type, because on askhistoryai citations spread across many pages in the same window.¹³ The difference was whether the page ranked in Bing for a real query.

Volume was a lottery: 3 cited slugs out of 2,212 pages, about 0.14%,⁹ and by 2026-08-27, 10 cited out of about 3,678 comparisons, about 0.3%.¹⁴

(c) llms.txt showed no measurable correlation

On 2026-08-08 we checked the three watchshowdown slugs Bing reported as cited against that site's own llms.txt and llms-full.txt. All three appear 0 times in both files.⁹¹⁵ Across the estate we have no measurement linking an llms.txt file to citations, so we keep the files and stopped treating hand-maintaining them as work that earns anything.¹⁶

A separate check did matter: GPTBot is the training crawler, while OAI-SearchBot and ChatGPT-User are tied to citation and live fetch. All 11 of our custom domains list them; one was missing both until 2026-08-12.³¹⁶

(d) Which page archetypes earned citations

Question-explainer guides. On askhistoryai, 3-month window read 2026-08-28: 382 citations across 34 cited pages. Guides accounted for roughly 159; the hubs /key-facts (53) and /reading-list (42) added 95; event pages about 70.¹⁷ The strongest single page on our estate was one guide, /guides/how-accurate-is-the-300-movie-thermopylae, at 60 citations and 55.88% citation share of its grounding query; the query "ancient history facts" ran 24 citations at 40.00% share.¹⁷

Head-to-head comparison pages. On watchshowdown, 3-month window read 2026-08-27: 63 citations across 10 cited pages, of which 8 of 10 were /compare/ pages carrying 46 of the 63. Exactly one product page was cited at all.¹⁴ On a 30-day window read 2026-08-29: 69 citations across 11 pages, 9 of 11 comparisons.¹⁴ Every top Bing query on that site was a comparison ranking at positions 1 to 2 with 0 clicks, several with typos and prompt-shaped phrasing, which reads as assistant retrieval rather than human search.¹⁴

Thin pages earned nothing. On askhistoryai, 129 of 340 event pages run roughly 3,800 to 4,400 words with 16 to 17 sections and metrics tables; the other 211 are roughly 500-word stubs. Zero era-id pages were cited.¹⁷

Length was not the signal. In the cited-page teardown of 2026-08-08, a 492-word comparison page earned 10 citations while a 3,024-word explainer held Bing position 1.0.¹⁵ Two other measured shapes: the title carried the literal part number or acronym (MDV106, GA-B2100, GtG, MPRT), the winning query was a near string match on it, and that token appeared 34, 37, 40 and 32 times respectively across title, meta, OG tags, H1, H2, table and JSON-LD; and the JSON-LD intersection across every cited page was exactly Article plus FAQPage.¹⁵ Against over-fitting: Bing reported panelpicks ranking first for a query whose key terms T-CON and impulse occur 0 times on the page.¹⁵

(e) Server-rendering and prerender

Raw HTML is what most of our traffic sees. On 2026-08-20, Cloudflare recorded 391.43k requests, 98.59k visits and 131.54k page views in 7 days across our zones, against roughly 750 GA4 sessions the same week: about 99% of edge traffic never executes JavaScript.⁸

We fixed a prerender defect on askhistoryai on 2026-08-27: the edge worker injected guide articles off-screen before the app root, so a reader without JavaScript got a blank page across about 50 guides.¹⁸ We could not attribute any Copilot movement to it. Measured 2026-08-29, over the 30 days that include the broken period, /key-facts earned 53 citations and /reading-list 42, the site's second and third most-cited pages, while both served a blank page to non-JS readers. The surge, 363 of that site's 382 all-time citations falling in the same window, began around 2026-08-23 or 08-24, five days before the fix.¹⁸ Bing executes JavaScript, so the cost fell only on crawlers that do not. Also measured: /event/<id>/, the form we advertise, prerendered 0 words on all 341 advertised event URLs, while /?event=<id>, advertised nowhere, prerendered 2,851 words on a sample page.¹⁸

The 2026-09-04 audit fetched 45 live pages with an OAI-SearchBot user agent and counted visible words. All returned 200 except two paths that do not exist, nothing was under 300 words, and every comparison table that exists was in the server HTML. The askhistoryai homepage is a 703-word app shell; the thinnest targets were three comparison pages at 778, 842 and 1,061 words with no links out to buying pages.⁵ One before-and-after, with no causal claim: openscriptorium's /books/ hubs were blank without JavaScript, an About block shipped into every hub on 2026-09-03,¹⁹ and its 3-month Copilot count read 154 on 2026-09-05 against zero on 2026-08-28.¹

(f) ChatGPT versus Copilot: which one paid

Copilot cited us at volume and paid nothing measurable. As of 2026-09-02, askhistoryai was earning roughly 380 Copilot citations per day, producing 3 visits per month and $0.²⁰ On 2026-08-12, bing / organic delivered 41 sessions and 0 key events and google / organic 19 and 0, while chatgpt.com delivered 228 sessions and 32 key events at 8.33%, or 39% of all key events on the fleet property.³

The AI Assistant channel held that position across four readings:^32181122

For 2026-08-04 to 09-01, chatgpt.com referrals were about 10% of sessions, 47% of affiliate clicks and 39% of conversions.²⁰

Within that channel, only buying pages converted. For 2026-08-06 to 09-02, AI Assistant traffic spread across 35 landing pages: /best-moped-style-ebike 70 sessions to 12 key events, /best-standing-desk-on-amazon 6 to 6, /best-electric-bike-on-amazon 2 to 2, against /gaggia-classic-pro-vs-bambino-plus 10 to 0 and /gaming-monitors/qd-oled-vs-woled 4 to 0.²² The same memo records an 8 to 11% click rate on "best X" buying pages and about 0 on explainers.²⁰ Openscriptorium's AI Assistant channel delivered 126 sessions across 54 pages at 28 seconds average and 0 key events.²² Amazon Associates for the trailing 30 days: $24.30 in commissions, 487 clicks, $0.00 in bounties.²²

Copilot citations meanwhile grew steeply. On the 3-month window read 2026-09-05: askhistoryai about 1,500 citations, running 100 to 245 per day since 2026-08-24; panelpicks 615; openscriptorium 154; watchshowdown 131; ebikeriderlab 98; nasdecoded 65; launchmonitorpicks 37; simgearlab 15; brewstationlab 7. Bing search over roughly 40 days, same read: askhistoryai 1,031 impressions and 23 clicks, panelpicks 442 and 5, ebikeriderlab 125 and 4.¹ On our sites, that volume of citation has not turned into revenue.

---

4. What we changed, and the numbers around it

Sitemap discipline. Watchshowdown's sitemap went from 3,592 URLs to 575 on 2026-08-21, with nothing deleted: all 3,404 comparison pages stayed prerendered, linked and indexable, and only the crawl invitation shrank.¹² A second pass on 2026-09-05 took it 684 to 560, comparisons 481 to 332, all 3,828 routes still prerendered.¹⁴

The prune had been cutting our best pages. The heuristic gating inclusion, same category and within 3x on price, excluded 5 of the 10 cited pages, carrying 33 of the 63 citations, including the site's most-cited page at 22. All five were live, linked and prerendered throughout; only the sitemap omitted them. Fixed with an allowlist, sitemap 595 to 600.¹⁴ We claim no citation gain: all five were already cited while missing, which is how we found them.

Curating pages that were already winning. On 2026-08-27 all 8 cited comparison pages were still serving an auto-generated template; curated entries went 10 to 18, rewritten to 1,158 to 1,858 words from facts already in the site's spec files. Two days later that site read 69 citations across 11 pages against 63 across 10, an interval far too short to call a result.¹⁴

Refreshing the one page that earns. ebikeriderlab's homepage carried 100% of its site's citations and the stalest date on it, 2026-06-19, while pages earning zero carried 2026-08-15. It moved to 2026-08-15 alongside a real change: two picks replaced because the products are no longer sold.¹³

Discovery plumbing. On 2026-08-08 Bing knew 12 of ebikeriderlab's 38 pages and 192 of askhistoryai's 396; one feed submission per site recovered 269 URLs in under a minute, at no cost. A crawl of built HTML found 11 live, sitemap-listed pages with zero inbound internal links; the estate now reads 0 unreachable across 63 sites and 1,091 pages.⁴ On 2026-09-04 one homepage linked only 4 of its 11 buying pages and one buying page had zero inbound links from anywhere.⁵¹⁹

Freshness. The 2026-09-04 audit compared each live page's schema dateModified against its source file's last commit. Live pages trailed source across eight sites: one homepage live at 2026-05-28 against source at 2026-08-12; four live at 2026-07-22 against source at 2026-08-28; one site's category pages hardcoded to "Updated May 2026"; every fetched page on askhistoryai hardcoded to "dateModified": "2026-05-28" while that site's own edge worker was stamping 2026-09-05.⁵ All fixed 2026-09-04 and 09-05, along with every coverage gap buildable from claims already in our repositories.¹⁹

---

5. How to check your own page

Each check is tied to something we measured on our own sites. Run them at the tool, or by hand.

1. Canonical host. A platform subdomain? On our sites, 58 of 58 earned zero Bing impressions and zero Copilot citations.
2. Bing impressions. Zero impressions meant zero Copilot citations on our sites, no exceptions. First gate.
3. Bing crawl count, not sitemap acceptance. On our zero-crawl sites every submission signal was green while the crawl count sat at 0.
4. Raw server response. Fetch with an OAI-SearchBot user agent, count visible words, compare against the post-JavaScript page. On our sites those are different documents, and about 99% of edge traffic never ran the JavaScript.
5. Which crawler is allowed. OAI-SearchBot and ChatGPT-User gate search citation and live fetch; GPTBot is training.
6. Title. Does it carry the literal model number or acronym a question would contain? On our cited pages the winning query was a near string match on it.
7. Answer block. Declarative, immediately after the H1, both entities named in full, one hard number with a unit in the first two sentences, no link inside, above the first affiliate link.
8. JSON-LD. Article plus FAQPage was the exact intersection across our cited pages.
9. Dates. A visible updated date matching the schema dateModified. We found live pages months behind source.
10. Reachability. Linked internally at any depth, and in the sitemap. We found live buying pages with zero inbound links, and cited pages missing from the sitemap.
11. Demand. Does it answer a question someone actually asks? On our sites that beat page count and prose quality.
12. llms.txt: reported, not scored. No measurable relationship to citations on our sites.

---

6. Footnotes

Conflicts we are not resolving. Bing impressions for ebikeriderlab are recorded as 47 and as 45 on 2026-08-08⁴⁹, as 66 on 2026-08-15¹⁰ and as 83 on 2026-08-20⁸. Copilot citation totals for panelpicks read 22 on a 30-day window on 2026-08-08¹, about 0 on 2026-08-15¹⁰ and 122 on a 3-month window on 2026-08-20⁸; the windows differ and we did not reconcile them. Citation counts throughout are window-dependent, and 30-day and 3-month readings of the same site are not comparable.

---