AI Performance and how to use it beyond online PR

AI Performance in Bing Webmaster Tools gives us a new perspective. According to the product managers behind this release, it is “the first step towards generative engine optimisation tools in Bing Webmaster Tools”.

The dashboard tracks citation activity in Microsoft Copilot, AI-generated summaries in Bing and selected partner integrations. It lets publishers and website owners see how often their content is mentioned in generative answers, which URLs are cited and how citation activity changes over time. For example, this is my dashboard for one of my content websites, which otherwise generated around 3,000 visits from Google over a similar period.

On the other hand, some client websites that generate millions in monthly revenue have relatively fewer citations.

Microsoft presents this as an extension of traditional tools for webmasters and SEO consultants (indexing, crawl status, recommendations and search performance) into the new world of AI-generated answers.

Visibility today is simply no longer just about links and visits. It is also about whether AI systems reference and cite you when generating answers. It is about default, training or retrieved data.

Here we can also see one difference between SEO and GEO. Reporting will no longer be just about visits from GSC; there will be far more to it.

Google includes AI Overviews data in the standard Search Console performance reports, but offers no dedicated AI dashboard, no URL tracking, no grounding queries and no separate visualisation of AI trends. It only deals with clicks. For now. ChatGPT, Claude and other AI platforms offer SEO consultants and website owners no visibility measurement tools at all. As the platforms look to monetise, they will certainly come up with something.

So with this tool, Bing has once again gained attention after a while. I am glad that I already have it set up and have accumulated data across dozens of projects.

What can we learn in Bing?

To use the data effectively, it is important to understand what each metric actually measures (and what it does not).

Total citations: The total number of times your website appeared as a source in AI-generated answers in Microsoft Copilot or Bing during the selected time period. What it does not measure: position, prominence within specific answers or the quality of your citation’s contribution. Put simply, the first and the last reference are both recorded in this dashboard as the same number, 1.

Average cited pages: The daily average number of unique URLs from your website that appear as sources in AI-generated answers over the selected time period. It shows the diversity of citations across your content. A high number means AI systems draw on many different pages of your website. A low number suggests authority concentrated in a few URLs.

Example: An average of 117 cited pages per day means that AI systems draw on 117 unique URLs each day when generating answers.

This metric shows whether your AI visibility is concentrated or distributed. A website with 5,000 citations but only 5 average cited pages has a dependency problem: if you lose those 5 pages, you lose all your AI visibility.

Grounding Queries: Key phrases that AI systems use when retrieving the content referenced in AI-generated answers.

These are not user search queries as in GSC. Grounding queries are internal search phrases that AI systems use to find your content, not the questions users actually asked.

A user asks Copilot: “How can I improve my website’s and brand’s SEO?” The AI system may then use internal search queries such as “best SEO tips”, “SEO audit” or “recommendations from SEO consultants” to find relevant content. These internal queries are what you see in grounding queries.
But be careful. Microsoft explicitly states that this is a sample of total citation activity. However, it has not disclosed what percentage of grounding queries is shown or how the sample is selected. The product team says it “will continue to refine this metric as more data is processed”.

Grounding queries reveal what AI systems think your content is about. They show you the AI’s view of your content’s positioning, which may differ significantly from your intended target audience.

Pages: the individual pages AI cites most often. This is actionable data: you can see exactly which URLs appear in AI citations and which do not. It does not show a page’s absolute importance, competitive ranking or position within answers.

Identify your best-performing content for AI systems and then reverse-engineer what makes those pages successful. Or adjust and influence how they come across.

How to work with it?

1. Identify the content patterns that earn citations. Bing shows citation counts at page level, so you can see which URLs Microsoft’s AI systems reference most often. It is definitely worth reviewing the most successful pages favoured by Copilot and Bing Chat. Based on that, you can create similar or complementary content and strengthen your position in that area. On the other hand, you will find out which content Microsoft does not see you as an authority on and does not draw from. On at least two websites, I noticed very narrow targeting of the successful pages.
2. Research grounding queries for topic expansion opportunities. Grounding queries are exclusive to Bing. No other AI platform publishes its internal search phrases in such an aggregated form. This is how the AI itself searches for information. You can do a lot with the downloaded queries, from optimisation to changing your focus. You can also map adjacent topics (related topics that AI systems associate with your content) and find semantic gaps (related queries where you do not appear). There are many more scenarios.
3. A hypothesis that was not confirmed. In its announcement, Microsoft highlighted the IndexNow protocol. This protocol notifies search engines about content changes, which helps AI systems reference current information instead of outdated versions. Even for evergreen content, regular updates signal that your content is maintained and trustworthy. This matters for citation decisions. So I simply assumed that the most cited content would be newer rather than older pages. The opposite was true.

Do other models cite and use the same pages we see in the Bing report?

The answer here is most likely no. It would be great if we could multiply the share of Copilot and Bing by total AI traffic and get complete data. But that is not possible. Each model will prefer different content.

What matters, however, is that for OpenAI to sell ads effectively, it will have to build tools for advertisers. These tools will need to show query patterns, volume data and performance metrics.

When Google launched AdWords, it had to create Keyword Planner so that advertisers could see search volume and plan campaigns. SEO specialists who had never bought an ad used this tool very often, because it revealed user search behaviour at scale. So we will see more tools.

“AI citations” ≠ “AI referral traffic”. Similarweb, for example, measures visits sent from AI platforms to websites and shows that in June 2025 there were ~1.13 billion referral visits to the top 1,000 websites, while Google search sent ~191 billion. Either way, this is not exact first-party data.

An extra use. A tool that gives some content websites a new dimension

As direct traffic has fallen, many content websites have started to lose relevance. But what these websites have primarily lost is traffic. People are in fact still reading the information from them, and AI models cite it. So by influencing training or cited data, we also indirectly influence the results.

On two websites, I found that in the Bing ecosystem alone, these websites have tens of thousands of citations, most of them on commercial and competitive topics.

So creating content for this purpose, and above all adjusting these results, is an obvious opportunity. Optimising cited pages. If you have a page describing the best coffee machines, by influencing that page you can most likely also change the answers in the models.

I have not read much about this technique anywhere and it certainly has many pitfalls. One of them is evaluation. But that is also a problem with traditional link building.

Logically, though, there is room for this kind of optimisation. Part of the problem is that you, “as a client”, cannot find out where you appear in AI Performance. On the other hand, for the owner of a content website or a larger magazine, AI Performance practically provides a list of “advertising spaces” that can be monetised. We will see whether this catches on in practice.