We installed EntityMap on waikay.io in week 12 of our tracking window. No new content. No backlinks. No other changes. This is 19 weeks of Bing Webmaster Tools data, analysed honestly.

Before install (weeks 5 to 11), waikay.io averaged 167 Copilot citations per week. After install (weeks 13 to 19), the average was 614. That is a 3.7× lift measured over the same number of weeks, with a single intervening change.

The more meaningful number is what happened to the ratio of citations per page. Total citations grew 370%. Unique pages cited grew 54%. Those two figures moving at such different rates is not what you would expect from a generic crawl improvement or an external trend. If Bing had simply discovered more of the site, both metrics would scale together. They did not.

One caution on the aggregate numbers: the 3.7× headline is real, but most of the volume lift is carried by multilingual pages (French and Spanish URLs that surged dramatically post-install). The English-only picture is flatter in volume but more interesting in composition. We analyse both separately below.

Stripping out the multilingual pages, total English citations went from approximately 627 pre-install to 577 post-install, a modest 8% decline in raw volume. A sceptic stops here and says the install did nothing for English content.

That reading misses what actually happened. The composition of those citations changed almost entirely.

This is the distinction between raw visibility and commercial visibility. A citation of your homepage or a generic guide is brand awareness. A citation of your product feature page or audit tool, in response to a query from someone evaluating options, is a sales conversation. EntityMap shifted the mix toward the latter, without a single new word published.

Page-by-page breakdown: funnel stage, lift, and hidden value

Each page categorised by funnel stage and its commercial significance.

The 3.7× headline is real, but 75% of the volume lift is carried by multilingual pages (French and Spanish URLs that surged dramatically after install, despite the EntityMap being written entirely in English. The most likely explanation is a crawl depth effect: submitting entitymap.html to Bing Webmaster Tools triggered a deeper domain re-crawl that surfaced language variant pages Copilot had previously underweighted). The mechanism is different from what the spec is designed to produce on individual English pages, and it is worth understanding separately.

The attribution question is legitimate. We cannot run a controlled experiment on a live domain. But there are three aspects of the data that are hard to explain without EntityMap as the cause.

1. The ratio divergence

If an external factor drove the lift (a Bing algorithm update, Copilot’s growing user base, seasonal variation), you would expect citations and pages cited to scale together. They did not. The ratio of citations per page went from 3.65 to 8.55 on average, and peaked at 13.3 in week 19.

To understand why this matters, it helps to separate what each metric actually measures.

When citations grow 7× faster than pages cited, the retrieval system is not just finding more of the site. It is reaching deeper into pages it already knows about, extracting more referencing value from each one. The same URL is being cited across a broader range of queries, not just the one or two it previously matched.

EntityMap is the mechanism that explains this. It does not add new pages. It adds entity-level descriptions, typed relations, and source-tagged chunks, which make individual pages precisely matchable to a wider range of queries. A page about Brand Visibility Tracking, once it is described as an entity with relations to AI Share of Voice, Prompt Tracking, and Competitive Benchmarking, becomes citable in response to queries about any of those related concepts, not just its primary keyword. That is what “used more deeply as a source” means in practice.

2. New pages entered the cited set without new content

Six English pages that received zero citations before install entered the cited set after install. These include the features page, the about page, and a deep guide subpage. No new content was published on any of them. The only mechanism that explains a retrieval system discovering and citing pages it previously ignored (without those pages changing) is a change in how the site’s structure was represented to the retrieval layer.

3. The citation shift follows EntityMap’s entity graph

The pages that surged: /brand-visibility-tracker/, /ai-audit/, /factual-accuracy/ are precisely the pages that EntityMap represents as named entities with source-tagged chunks. The pages that dropped: /aio-guide/, /share-of-voice/, the guide hub index, are broad pages that EntityMap does not specifically surface. The shift was not random. It followed the entity graph.

We are publishing this because the signal is strong enough to document. We are not claiming it is settled.

EntityMap is an open standard. The spec and generation tooling are at entitymap.org. Waikay generates conforming files as a product deliverable. Here is what the data suggests about how to run your own test well.

The replication recipe

If you want to run this experiment cleanly, here is exactly what to do:

The aggregate 3.7× lift is almost entirely driven by non-English pages. Three URLs account for the bulk of post-install volume: /fr/guide-de-visibilite-de-marque-dans-les-ia/chapitre-2-la-part-de-voix-ia/ (1,053 citations), /es/ (655), and /es/metricas-de-visibilidad-ia/ (223). Combined, multilingual pages represent roughly 75% of total post-install citations.

The EntityMap is written entirely in English. It does not reference or describe any French or Spanish URLs. So the simplest explanation (that structured multilingual entity data unlocked multilingual pages) does not hold.

What likely happened

We manually submitted entitymap.html to Bing Webmaster Tools after install. We did not submit the rest of the site. That single submission appears to have been enough to trigger a more thorough re-crawl of the entire domain, surfacing language variant pages that had been underweighted or underindexed in Copilot’s retrieval layer.

The entitymap did not tell Bing “these entities exist in French.” What it did was give Bing’s crawler a structured, authoritative signal about the domain’s content, enough to prompt a deeper pass. The multilingual pages were already there. The entitymap submission unlocked them by making Bing confident enough in the domain to go looking.

This is an important distinction. The multilingual surge is a crawl depth effect triggered by the Bing submission, not a structured data effect. It is still directly attributable to the EntityMap install (the submission is part of the deployment) but the mechanism is different from what the spec is designed to produce on individual pages.

The data in this case study is new. The structural move it represents is not.

In 2012 to 2015, Google made a fundamental transition: from keyword-first retrieval toward entity-first retrieval. Before that shift, Google matched pages to queries by finding documents that contained the right strings of text. “Apple” in your content could mean the fruit, the company, or the Beatles’ record label. Rankings depended on proximity and frequency of the string, not on what the string referred to.

The schema.org sameAs attribute was the publisher-side signal that made the transition possible at scale. By anchoring content to a specific Wikidata or Wikipedia entity, publishers gave Google’s Knowledge Graph the disambiguation layer it needed: this page is not about the word “apple,” it is about this named thing. Relevance scoring became more precise, and rankings followed because the retrieval model had become structurally better at matching intent to content.

The model is not the bottleneck. The retrieval layer is.

LLMs process meaning rather than keywords. Their representations capture semantic relationships, not string patterns. For retrieval-augmented systems like Copilot, that means the model is rarely the weak link. It can reason well over entity-structured input. The question is whether the retrieval layer gives it any.

That retrieval layer is the component that decides which pages or chunks to pull in before the model generates a response. It is where the practical gap sits, and it is where EntityMap acts.

To understand why this matters, it helps to step back and ask what the web was actually built for.

Each layer made the web legible to a new class of reader.

EntityMap does not improve the LLM. It improves what the retrieval layer hands to the LLM. The model was already entity-first. The pipeline feeding it was not. That is the gap EntityMap closes, and it is the same gap sameAs closed for Google’s ranking pipeline a decade ago.

What the timing means

The sameAs analogy is structurally sound but the timing comparison requires more care. The “18 to 24 month window” framing is appealing but it assumes EntityMap will follow the same adoption curve as schema.org, which had Google’s active endorsement and an existing structured data ecosystem to slot into. EntityMap does not have that yet.

What we can say with more confidence is this: the consumption behaviour already exists. Bing is already responding to the entitymap.html signal in ways that are measurable in citation data, before the standard is formally recognised by any major platform. That is the meaningful observation. Not “adopt now before the window closes” but “the infrastructure is responding to this signal today, and the sites that have installed it are already seeing different citation patterns from those that have not.”

Whether that advantage compounds over 18 months or 5 years depends on how quickly AI retrieval systems formalise entity-structured inputs as a ranking signal. We do not know that timeline. What the data suggests is that waiting for certainty means waiting until the signal is already priced in.