0
Copilot citations
Jan–Jun 2026
Jan–Jun 2026
A first-party GEO case study: one Canadian SaaS, 87 experiments, zero paid ads, zero link-building — and an honest look at what the data does and doesn't actually prove.
New to GEO? Start with the field guide to AI search optimization — then come back for the proof.
GrantCompass (grantcompass.ca) earned 135,700 Microsoft Copilot citations between January 1 and June 2026, as measured in Bing Webmaster Tools' AI Performance report. At peak, ~58 pages were being cited on any given day, with daily citations ramping from ~500 to ~2,000. The site reached 25,000+ users in 4 months via organic and AI search — zero paid ads, zero PR, zero link-building. The one-sentence thesis: on-site and structural GEO is the main citation lever; off-site amplification is real but secondary.
Honest caveat (read this): 76 of 87 experiments were confounded — pages hosted multiple simultaneous tests, and the Copilot citation ecosystem expanded platform-wide during this window. An untouched control page was among the top citation gainers. The data shows what rode the wave, not a clean "my tactic caused +X%."
GrantCompass is a Canadian grant-discovery SaaS that helps small businesses and nonprofits find government and private funding. I built and launched it from scratch as a solo founder, starting from zero: no domain authority, no backlinks, no existing audience, no PR relationships to leverage.
The starting point was a clean slate by design. No legacy SEO. No inherited traffic. No partner network to push the brand. This was useful: it meant any citation growth couldn't be explained by existing authority or relationship-driven mentions. If the pages got cited, it would be because of what was on them.
The product addresses a genuinely under-served information gap — Canadian grant programs are fragmented across federal, provincial, and municipal levels, and most grant seekers don't know where to start. That gave the site a shot at being the most useful answer to a real question: "what grants is my business eligible for?" That's the kind of question AI engines want a clear, structured, trustworthy answer to.
The measurement window is January 1 to June 2026. All citation data comes from Bing Webmaster Tools, specifically the "AI Performance / Copilot and partners" report — the same Bing index that underlies ChatGPT's web-browsing mode. No third-party tools or estimated data: this is the raw platform count.
Every lever we pulled was on-site and structural. No PR pitching, no guest-posting, no link-building. The playbook was: build pages that are the best possible answer to a specific grant question, structure them so an AI can extract and cite them easily, and make sure Bing knows they exist.
The full tactical breakdown — with the step-by-step of each lever — lives in the operator's playbook for AI search. This case study focuses on the outcome and the honesty about what it proves.
This is the section most case studies skip. I'm writing it first because intellectual honesty about what a dataset shows is the difference between a finding and a sales pitch. The confounders here are real, and they matter.
1. Platform-wide expansion. Microsoft Copilot was actively expanding its citation footprint during the Jan–Jun 2026 window. Citations roughly doubled platform-wide — not just for GrantCompass. A site that did nothing during this period would have seen citation growth. The question is whether GEO work caused above-baseline capture — which we believe it did, but cannot cleanly isolate.
2. 76 of 87 experiments were confounded. Most pages hosted multiple simultaneous experiments — we'd test answer-first structure and persona-addressed sections on the same page at the same time. You can't attribute outcome cleanly to either. What the data can show is which tactic families correlate with better outcomes across the set, not which single change caused a specific lift.
3. The untouched control page. One page we intentionally left unmodified throughout the experiment window was among our top citation gainers. It picked up substantial Copilot citations without any GEO intervention. This is the clearest evidence that the platform wave was real — and that context matters as much as any specific tactic.
| What we changed | What moved | Honest read |
|---|---|---|
| Answer-first restructure (H2 sections opening with direct answers) | Citations increased on treated pages; best win rate in the set | Correlated with lift, but simultaneous platform expansion makes causation ambiguous. Most confident finding in the dataset — works mechanically with how AI extracts passages. |
| Original grant data pages (eligibility + amounts) | Largest single citation spike in the dataset (one page: 294 → 1,931 citations) | Highest-confidence finding. AI engines cite primary sources because there are no better alternatives. This is replicable regardless of platform waves. |
| Persona-addressed sections ("If you're a first-time applicant in Ontario…") | Best win rate; zero losses in the tested set | Strong directional signal. Pages were treated sequentially rather than simultaneously with other tactics, giving cleaner reads on this one. |
| Bing indexation work (URL submission, sitemap, crawl error fixes) | Pages began appearing in Copilot citations within 1–2 weeks of confirmed indexation | This is a prerequisite, not an experiment — a page in Bing's index can be cited; a page absent from it cannot. Relationship is definitional, not probabilistic. |
| Untouched control page | Citations increased substantially with zero GEO intervention | The clearest evidence of a platform-wide wave. This is the most important data point in the whole set — it keeps the rest honest. |
| Schema markup additions (beyond basic Article/BreadcrumbList) | No measurable citation lift | Consistent with the broader literature (Ahrefs, 1,885-page causal study, May 2026: no positive effect). Schema is hygiene, not a lever. |
Reading this table honestly: the platform wave explains part of the lift; the on-site work explains why GrantCompass captured more of it than a comparable site doing nothing would have. We cannot give you exact percentages for each share.
Strip out the platform tailwind and you're left with a smaller but cleaner set of findings. These are the lessons that hold because they operate on mechanistic grounds — they describe how AI engines extract and cite content, not how big a particular platform is growing.
135,700 citations with zero PR, zero link-building, zero paid promotion. The "off-site is ~85% of citations" orthodoxy doesn't match what happened here. On-site structure and original data were the engine; brand mentions across the web amplify on-site work, but they're not a substitute for it. Brand mentions do correlate with AI visibility ~3× more than backlinks (0.66 vs 0.22, Ahrefs, 75k brands) — but that's a comparison within the off-site category, not proof that off-site beats on-site.
The highest-confidence, highest-impact finding across 87 experiments: when you publish original data that doesn't exist in usable form anywhere else, AI engines have no better option than citing you. One grant-data page went from 294 to 1,931 daily citations after we published structured eligibility data for a provincial program with no other clear web source. This isn't a GEO trick — it's basic information economics.
GrantCompass started with no domain authority. That turned out to be an advantage: GEO tactics lift lower-authority pages more than established ones (Aggarwal et al., ACM KDD 2024: +115% for a rank-5 page, −30% for a rank-1 page using the same techniques). Small sites can move faster in AI search than in Google search. But because citation half-lives are ~4.5 weeks, this window requires continuous freshness investment to stay open.
The ~4.5-week citation half-life (Scrunch × Stacker, 3.5M citation events, Mar 2026) means any citation you earn today is half gone in a month without ongoing work. GrantCompass maintained its citation volume through continuous page refreshes, new grant data, and ongoing answer-first restructures — not through a one-time optimization sprint. Plan accordingly.
The single most common failure mode we found in auditing other sites: pages optimized for GEO that aren't in Bing's index. ChatGPT reads Bing. Copilot reads Bing. A page that's in Google but not Bing is invisible to both. Submit every new page via Bing Webmaster URL submission, validate your sitemap against Bing's index, and check crawl errors — before you spend time on content optimization.
AI answers are non-deterministic — roughly 1 in 9 runs contradicts the others (LumenGEO first-party research). Asking "does ChatGPT cite me?" once is a snapshot of one run. The only honest measurement is share of answers over repeated runs, plus platform-level data like Bing Webmaster AI Performance. If you're making GEO decisions from single-check screenshots, you're optimizing against noise.
The 135,700 figure is not an estimate. It's a platform-counted metric from Bing's own tooling, measured on a verified domain. Here's exactly how the three-layer measurement stack works, so you can replicate it.
Layer 1 — Bing Webmaster AI Performance. This is the canonical data source. Bing Webmaster Tools has an "AI Performance" section that tracks how often Copilot and its partners cite pages from your verified domain. It's the same index that underlies ChatGPT's web-search mode — so it's the closest thing we have to a platform-level citation count across both Copilot and ChatGPT. The 135,700 figure is the aggregate across January 1 to June 2026. The ~58 pages cited on average is the daily mean from the same report.
Layer 2 — Share of answers. Platform data tells you volume; share of answers tells you query-level reliability. The method: pick a set of target queries, run each query multiple times in the same AI engine, record whether your page appears in the citation set each time. The resulting ratio is your share of answers for that query. We tracked this across a sample of 12 high-priority grant-related queries throughout the experiment window.
Layer 3 — The 1-in-9 instability audit. Before making any optimization decision based on a citation check, we ran the same query nine times. Roughly 1 in 9 runs produces a meaningfully different citation set — different sources, different framing, sometimes contradictory answers (LumenGEO first-party research). This is why a single "did AI cite me?" check is unreliable. We used LumenGEO to automate the repeated-run measurement and share-of-answers calculation at scale.
The full pipeline — engines, signals, retrieval gates, and the terminology glossary that ties GEO, AEO, LLMO, and AI search together.
Start here →The manifesto this case study supplements — what the full experiment set found, including the four tactics that actively hurt citation performance.
Read the findings →The step-by-step operator's playbook — each of the six on-site levers documented from the GrantCompass experiment, applied to your pages.
Get the playbook →Score any URL on the 14 structural signals that predict AI citation — with ranked fixes. Built on the same signals this case study tested.
Scan your page →If you want growth leadership or fractional CMO work at an early-stage startup, I'm talking to teams now. If you want to measure your own AI citation share before we talk, LumenGEO is the tool — and the free scanner below takes 30 seconds.
Published June 14, 2026 · First-party data from Bing Webmaster Tools on grantcompass.ca · Built to practice what it preaches: answer-first, entity-dense, passage-chunked, and structured to be cited.