How Copilot Picks Sources: Lessons From 4,500 AI Citations

AI citations are the moments an assistant like Copilot, ChatGPT, or Perplexity grounds its answer in your page and names you as the source. They are the new referral currency: the reader never searches, never scans results, and often never clicks, but the assistant's recommendation carries your name. This post is what we learned from logging ours.

The dataset is first-party and small enough to be honest about: between May 6 and August 4, 2026, Bing Webmaster Tools logged 4,325 citations of designpixil.com by Microsoft Copilot and its partners, about 4,500 since logging began in April. Those citations came from 68 distinct grounding queries and touched 45 different pages on a site with a domain rating of 6 and roughly a dozen real referring domains. We are not a media company. We are a two-hands design studio whose blog gets cited more often than it gets read, and that gap is exactly what makes the data useful.

One promised caveat before the findings: Bing labels this report a sample, it covers Copilot and partners rather than every assistant, and daily figures get refined after the fact. Treat every number here as directionally solid and decimally humble.

Finding 1: Citations concentrate far harder than traffic does

Search traffic spreads across a portfolio. Citations pile onto a handful of pages. One post, our guide to Figma design handoff for developers, earned 2,900 of the 4,325 citations in the window. That is 67 percent from a single URL. The next tier drops an order of magnitude: our Figma component library guide earned 240, a hiring-questions checklist earned 170, and everything else earned double or single digits.

The share within individual queries is even more concentrated. On "Figma component library best practices," our page held a 42.7 percent citation share, meaning Copilot chose us nearly half the times it answered that question for anyone, anywhere. You do not need a big site to dominate a narrow question. You need the single most extractable page on that question.

The practical rule we took from it: pick questions you can own outright, not topics you can participate in. Ten pages with 30 percent citation share beat a hundred pages with rounding-error share.

Finding 2: Assistants cite what helps, not what sells

Of our 45 cited pages, the top of the list is entirely how-to and checklist content. Our service pages, the ones that describe what we sell, barely appear. The commercial queries that do get grounded ("monthly cost of an embedded senior product designer," "are unlimited design subscriptions worth it for venture-backed startups") pull their answers from posts that answer the question honestly, including the parts that say when the answer is no.

This matches the published research. The Princeton GEO study (Aggarwal et al., KDD 2024) measured that adding citations to sources boosts a page's generative-engine visibility by around 40 percent, with statistics close behind; keyword stuffing measurably hurts. Assistants are selecting for the page that looks like evidence, not the page that looks like an ad.

The uncomfortable implication for founders: your money pages will probably never be your cited pages. The job of cited pages is to put your name inside the answer; the job of money pages is to catch the person who then types your name. Both layers have to exist, and they are not the same pages.

Finding 3: The day our citations went to zero (and why that was good news)

Through late July we were reading 50 to 97 citations a day and feeling great about the trend line. On July 27 the count fell to 4. By August 1 it was 0, and it stayed near zero for a week.

Nothing was broken. The page was still indexed in Bing, Google was actually flipping more of our pages into its index that same week, and Clarity's AI Visibility panel showed real users still receiving answers grounded in our pages. What had ended was a machine. The grounding-query log made it obvious: the top queries in our dataset were phrasings like "Figma Enterprise Lovable design to dev handoff evaluation," repeated at industrial volume and even showing up word-for-word in Google Search Console at the same time. Somebody had been running an automated evaluation workload through Copilot for weeks, their runs kept grounding on our handoff guide, and then their job finished.

Three lessons came out of that cliff:

1. Never project revenue from citations per day. Batch workloads you cannot see can double your numbers and then vanish. Our "organic" baseline was roughly 2 real citations a day, not 60.
2. Measure share, not volume. In the 7 days after the crash, Copilot cited us 15 times on genuinely human queries, and those 15 citations were 22.06 percent of all citations it handed out in our topics. Share of authority survives workload noise; raw counts do not.
3. Machine queries are still readers. Those evaluation runs were somebody's AI agents doing research. Being the source that automated researchers land on is not worthless; it is a preview of how procurement will work.

Finding 4: Assistants ask questions in full sentences, and full-sentence content wins them

The grounding queries in our log do not look like keywords. They look like requests a person types to an assistant: "tips for organizing component library in Figma," "questions to ask a creative agency about collaboration process and deliverable timelines," and in one case a user's literal dashboard sidebar pasted into Copilot with a request to fix the naming and order. Copilot answered that one by citing our SaaS navigation patterns guide.

The same shape shows up on Google's side. In the last week, our pages held position 1 for queries that are unmistakably assistant-fired or assistant-shaped: "who should a startup hire to design an AI product interface," "design subscription vs hiring a full time designer for a saas startup," and "what does unlimited design services actually mean, are there limits or one active request at a time." We rank for those because our FAQ entries phrase the question the way the asker phrases it and answer it in one self-contained paragraph.

The craft rule: write headings and FAQ questions as complete questions, answer them in 40 to 60 self-sufficient words before elaborating, and keep one idea per paragraph. Extraction is a mechanical act; make the mechanics easy.

Finding 5: Freshness and honesty are retrieval features

Two of our best pages spent weeks in Google's "crawled, currently not indexed" purgatory because the crawler had last seen versions from May. We refreshed both with genuinely new sections, requested a recrawl the same day, and both flipped to indexed within days. New posts now enter Copilot's cited pool within days of publishing: a post about vibe-coded products earned its first 7 citations in its first weeks live.

The counterweight matters as much: every date on our site changes only when content genuinely changes. Earlier this year we found template bugs stamping a fresh "last updated" date on every page load and killed them, because a date that is always today is an inconsistency signal to a system that compares its crawl history against your claims. Same with statistics: we deleted every stat we could not trace to a real source and replaced them with verifiable ones. Assistants are, functionally, fact-checkers with a citation budget. Be the page that survives checking.

What we changed because of this data

• Retired citations-per-day as a KPI. Our AI-visibility scoreboard is now share of authority (22 percent last week), distinct pages cited (45 and counting), and sessions referred by assistants (13 last week, from claude.ai, ChatGPT, Perplexity, and NotebookLM).
• Doubled down on ownable questions. The Figma-to-AI-tools territory produced our citation fortress, so it gets the deep guides. Broad topics we cannot own get merged or skipped.
• Put winning phrasings on money pages. The exact questions we hold position 1 for are now FAQ entries on the relevant service pages, so the answer and the offer live one click apart.
• Kept every claim checkable. It is the cheapest ranking factor in this entire discipline.

The commercial punchline, for the skeptics: an assistant's answer is how our largest client found us. They asked an AI who designs AI products well, our name came back, and the first call came from that recommendation. Citations are not a vanity metric. They are the top of a funnel that search consoles cannot fully see.

How to run this audit on your own site (free, about an hour)

1. Bing Webmaster Tools, AI Performance. Free, and the only first-party log of Copilot citations: totals, cited pages, and the grounding queries themselves. Read the queries before the counts; the phrasings tell you what to write next.
2. Microsoft Clarity, AI Visibility. Shows citations on real user queries plus your share of authority against every other cited domain in your topics. This is the number to track monthly.
3. Ahrefs AI responses panel. Free tier includes counts of your pages cited by ChatGPT, Perplexity, Gemini, AI Mode, and Copilot; it is how we caught Perplexity citing us before a single referral arrived.
4. Your analytics referrers. Sessions from chatgpt.com, claude.ai, perplexity.ai, and notebook.google.com are assistant referrals. Small numbers, high intent.
5. Search Console, read queries as sentences. Filter for question-shaped and persona-shaped queries at positions 1 to 10. Those are assistant-adjacent surfaces you already own; put their exact phrasings into your FAQs.

If your counts are zero across all five, start with structure: a self-contained definition at the top of each key page, real statistics with named sources, question-phrased FAQs, honest dates, and open access for AI crawlers in robots.txt. That is the whole entry fee.

Frequently Asked Questions