GEO in Practice: 3,500 Copilot Citations and 25 Bing Clicks

From August 6 to November 5, 2026, Microsoft Copilot and related Bing AI services cited futurecraft.pro pages 3,536 times. Regular Bing search sent 25 clicks to the site over the same period. That’s a gap of roughly 140x, and it shows how GEO (generative engine optimization) differs from SEO. GEO is the work of getting an AI assistant to use your page as a source for its answer, and you measure it in citations, not clicks.

All numbers below come from Bing Webmaster Tools: a snapshot taken on November 6, 2026, with a 3-month window. Bing still labels the AI Performance report, where the citation data lives, as BETA and calls its data a sample of overall activity. This is one site and one snapshot. The numbers don’t generalize to the market.

The case covers what Copilot actually cites on a technical blog, for which queries, how to separate agent “traffic” from people, what the site did, and what this data doesn’t prove. There’s a checklist at the end for repeating the measurement on your own site.

Where the data comes from: the AI Performance report in Bing Webmaster Tools

Microsoft launched AI Performance in Bing Webmaster Tools in February 2026 as a public preview (announcement on the Bing Webmaster blog). The report shows how often a site’s pages become sources in Microsoft Copilot answers, in Bing’s AI-generated summaries, and in some partner integrations. In the snapshot used for this case, all three sources are combined into one set of numbers.

What the report shows	What the report doesn’t show
Total citations for the period and the daily series	Clicks and visits from the answer to the site
Average number of cited pages per day	Where the link sits inside the answer
Citations per URL	Other sites cited in the same answer
Grounding queries: the phrases the AI used to search for material	Complete data: grounding queries are a sample
Intent label and citation share per query	A breakdown by Copilot, Bing AI summaries, and partners

In Bing’s terms, a citation is a case where a page appears as a source in an AI answer. The count doesn’t tell you whether a person saw the link, or whether they clicked it. So you can’t compare citations with clicks directly: they’re different events. The numbers sit side by side here to show how far apart the two channels are. Their ratio isn’t a conversion rate.

The numbers: citations grow, clicks stay flat

Over 90 days the report shows 3,536 citations (the interface rounds this to 3.5K), 81 pages with citations, an average of 9 cited pages per day, and 41 grounding queries in the sample.

Month	Citations	Days in window	Average per day
August (from the 6th)	356	26	≈14
September	972	30	≈32
October	2,028	31	≈65
November (1-5)	180	5	36

Citations grew 2.1x from September to October. The peak came on October 4 with 265 citations, but that was a single day: in early October the usual level was 60 to 110 a day. The first five days of November average 36 a day, almost half the October average. Five days of sampled data can’t tell you whether that’s a decline, a processing delay, or noise.

Regular Bing web search looks different over the same window:

Channel (window Aug 6 - Nov 5, 2026)	Value	Previous snapshot (Sep 26, 2026, 3 months)
Citations in AI Performance	3,536	1.4K
Pages with citations	81	67
Clicks from Bing search	25	21
Impressions in Bing search	2.9K	5.5K
CTR in Bing search	0.86%	no data

The two snapshot windows overlap by about seven weeks (August to September), so this isn’t a clean period-over-period comparison. The direction is still clear. Citations and the number of cited pages went up, web search impressions fell by almost half, and clicks stayed at around twenty.

One possible explanation: users get some of their answers in Copilot and never reach the classic results page. Bing Webmaster Tools data can’t test this. The report has no clicks from Copilot, and the drop in impressions could have other causes, from seasonality to ranking changes.

Which pages Copilot cites

Citations are unevenly spread. The top page gets 26% of the total, the top four get 49%.

Page	Citations over 3 months
/blog/supabase-startup-guide/	910
/blog/rest-api-design-ai/	329
/blog/ai-agents-complete-guide/	299
/blog/context-engineering-vs-rag/	193
/ru/blog/how-to-create-ai-agent/	113
/ru/blog/mcp-servers-explained/	87
/blog/impact-frequency-matrix/	66
/blog/ai-slop-code-review-methodology/	65
/blog/prompt-library-template/	49
/blog/safe-vs-priced-round/	45

What do the leaders have in common? They’re long guides: the English versions of the top four run from 2.5 to 5.7 thousand words. Each opens with a definition block and a TL;DR.

The Supabase guide starts with a comparison to Firebase. Context Engineering vs RAG is built around comparing the two approaches and includes a decision framework. REST API Design with AI is a set of principles, prompts, and a checklist. There are plenty of tables: the English versions of the Supabase guide and the context engineering article contain dozens of table rows.

Russian-language pages make the top too: the Russian versions of the AI agent tutorial (113) and the MCP servers explainer (87). Further down the list are the Russian versions of the AI agents guide and the RAG guide.

These traits don’t fully explain what Copilot picks. The definition block is on 130 of the blog’s 134 English articles, and the first and third pages in the top have no FAQ section. Structure probably helps a page get into an answer, but topic decides: the Supabase guide gets cited almost three times as often as the next page.

The Markdown version as a separate source

The address /blog/prompt-library-template.md made the top 50 by citations. It’s the Markdown version of the article, and it gets cited separately from the HTML page. Every article on the site has such a version for AI agents.

Responses for these addresses carry X-Robots-Tag: noindex, follow and Link: rel="canonical" pointing to the HTML page (canonical was added in October). The report doesn’t show how Copilot reconciles two versions of the same article, or whether it respects canonical for citations. The practical takeaway is more modest: if a page has alternate versions, you have to add up their citations by hand.

Grounding queries: how Copilot rewrites the question

Grounding queries are the phrases the AI used to search Bing’s index for material before citing a page. The report lists each one with a citation count and a share (citation share). Reviews of the report describe this metric as the site’s percentage of all citations for that query.

Grounding query	Intent (Bing label)	Citations	Site’s share
Supabase evaluation database for startups	Informational	276	40.59%
supavase	Others	202	49.88%
api.dev AI-generated OpenAPI specifications efficiency	Research	104	19.51%
Context engine vs RAG	Comparison	90	19.03%
mcp server	Informational	71	0.06%
enterprise AI prompt library essential components	Informational	58	18.01%
как создать ии агента (Russian for “how to create an AI agent”)	Learn and Solve	30	16.85%
best practices integrating multiple language model apis reliability failover	Research	26	22.41%
tracking user feedback signals with llm traces in production	Research	13	33.33%

Most of these don’t look like anything a person types into a search box. “Best practices integrating multiple language model apis reliability failover” reads like a question the assistant rephrased from a conversation. Queries like this are long and list several aspects at once. For queries labeled Research, the site’s share in the sample is 19% to 33%.

Not every query gets rewritten. “supavase” is a typo that went into the index as is, and the site holds almost half the citations for it. Bing’s classifier assigns the intent labels (“Research”, “Learn and Solve”), so treat them as a hint.

The opposite case is “mcp server”: 71 citations at a 0.06% share. If the share is calculated against all citations for the query, the sample holds on the order of a hundred thousand citations of other sources for it. That’s a rough estimate, but the order of magnitude is clear. On a broad query the site gets about one citation in 1,700. On a narrow research query, it gets anywhere from one in two to one in five.

The takeaway for content strategy: a precise answer to a long, specific question earns more citations than a fight for a broad term. That matches how long-tail works in classic SEO (more on finding such topics in the article on keyword research with AI), with one difference. In GEO, the assistant writes the long tail.

Agent prompts among Bing “keywords”

The Search Performance report (regular web search) lists phrases that no person types into a search box. A few examples, paraphrased and translated (the originals are in Russian):

• a greeting and a request to introduce yourself in one sentence, to check a live connection via MCP: 94 impressions, average position 7.9;
• “too many questions” plus a request to explain how to use LightRAG as a context provider;
• a vibe-coding beginner asking whether it’s legal to build a site on Supabase in Russia.

This looks like prompt text that AI agents and chat assistants send to Bing when they search the web. One such “query” produced 94 of 2.9K impressions for the quarter, about 3% of the total. Most likely, no live person saw that results page.

Signs of an agent query in the report:

• direct address and imperatives (“introduce yourself”, “tell me”); the Russian originals use the informal “you”;
• greetings, service phrases, lines from the middle of a conversation;
• the length of a full sentence or paragraph, with context about the user (“I’m a beginner”);
• a stable position with zero clicks.

Exclude these queries when you calculate CTR and look for topics. Otherwise impressions are inflated and CTR is understated. You can also collect them separately, as a signal of what agents ask about your niche.

What the site did (and what that doesn’t prove)

The table covers only what can be checked in the blog’s repository. Dates come from commit history, and the production rollout may have come later. “March 2026 or earlier” means the measure is already in the repository’s first commit.

Measure	Since	What it does
definition, tldr, faq fields in frontmatter	April 2026	Definition block and TL;DR at the top of the article, FAQ at the end, FAQPage markup in JSON-LD
llms.txt	March 2026 or earlier; generated from frontmatter since July	An article index for AI systems; more in the llms.txt guide
Markdown versions of pages (/blog/<slug>.md)	July 2026	Clean article text for agents; noindex and a canonical to the HTML page in response headers
IndexNow on deploy	March 2026 or earlier	Notifies Bing and Yandex about updated URLs
hreflang en/ru/x-default	March 2026 or earlier	Links the EN and RU versions of the same article

All of these measures were in place before the observation window started. Inside the window, the only additions were headers for the Markdown versions: noindex (August) and canonical (October). New content doesn’t explain the growth either. By frontmatter date, 7 articles came out in August, one in September, and none in October. October still had the most citations.

So the case data doesn’t support saying that any of these measures caused the growth in citations. The rise from 14 to 65 citations a day could come from changes on Copilot’s side, from growth in its usage, or from Bing re-evaluating already indexed pages over time. The site’s own data can’t tell these causes apart.

The accurate way to put it: on a site with these measures, Copilot cites long, structured pages, and the citation count grew over the quarter. Testing the effect of a specific measure takes an experiment. Turn it on for some pages, keep a control group, and track per-URL citations for several months.

How to repeat the measurement on your site

1. Verify the site in Bing Webmaster Tools. If it’s already verified in Google Search Console, you can import it from there.
2. Open AI Performance, pick the 3-month window, and export three slices: the daily citation series, citations per page, and grounding queries.
3. Save a dated snapshot. The window is rolling, and a month later you won’t be able to recover the same numbers. Compare non-overlapping periods.
4. Pull Search Performance for the same window: clicks, impressions, CTR, top pages. Check which pages are strong in one channel and weak in the other.
5. Add up citations across a page’s alternate versions: HTML, .md, AMP, language versions.
6. Tag grounding queries by breadth: narrow research queries versus broad ones. Look at the site’s share, not just the citation count.
7. Filter prompt-like queries out of Search Performance using the signs above, then recalculate CTR.
8. Keep a dated log of site changes (commit history works) so you can line it up with the daily series, without drawing causal conclusions from it.

Limitations of this case

• One site, one niche (a technical blog about AI and development), and one snapshot as of November 6, 2026.
• AI Performance is a preview, and grounding queries are a sample. In the announcement, Bing says it will keep refining the metrics as more data is processed, so the numbers for the same period may change.
• Bing’s classifier assigns the intent labels. Their accuracy is unknown.
• The report doesn’t show clicks from Copilot, the link’s position in the answer, or competing sources in the same answer.
• The comparison with the previous snapshot uses overlapping windows.
• This data doesn’t explain the October 4 peak or the dip in early November.
• Citation share is interpreted from descriptions of the report, and the “mcp server” estimate is arithmetic on sampled data.

Where to start

Step one: connect AI Performance and save a dated snapshot. After that, three things pay off most. Look at pages with a high share on narrow queries and write more material in the same format. Exclude agent queries from your search reports. And test hypotheses about llms.txt or markup on a control group of pages, not on the overall chart.