Data
What Kind of Pages AI Assistants Actually Cite
We run 21 sites that Microsoft Copilot cites about 14,394 times a month. We classified the 25 most-cited pages by shape to see what they have in common — and found something uncomfortable about what all those citations were worth.
Four page shapes account for 80% of our most-cited pages: head-to-head comparisons, definitions, reference tables, and trend explainers — each exactly 20% of the top 25. What they share is that each answers a question with a specific, checkable, quotable answer. Assistants cite what they can lift a sentence out of. And the uncomfortable part: across the same period those 14,394 monthly citations produced zero new referring domains, for a reason that turns out to be structural rather than fixable by volume.
Most writing about AI citation is inference from the outside. This is from the inside, on a portfolio large enough to have a distribution — with the method published separately so you can reproduce it on your own domain.
The Classification
We took the 25 most-cited pages from Bing's AI Performance report and classified each by shape. Counts, and total citations attributable to each shape:
| Shape | Pages | Share | Citations |
|---|---|---|---|
| Trend / event explainer | 5 | 20% | 2,796 |
| Comparison (X vs Y) | 5 | 20% | 1,952 |
| Definition / what-is | 5 | 20% | 1,853 |
| Reference data / table | 5 | 20% | 1,695 |
| Direct question page | 2 | 8% | 753 |
| Listicle / best-of | 1 | 4% | 666 |
| Other | 2 | 8% | 674 |
The even split across the top four is more striking than any single winner. It suggests the shape that matters is not a specific format but a property those formats happen to share.
The Shared Property
Every one of those shapes answers a question that has a specific, checkable answer that fits in a sentence.
- A comparison resolves to this one, because of that.
- A definition resolves to it means this.
- A reference table resolves to for your case, the number is X.
- A trend explainer resolves to here is what it is and why now.
An assistant assembling an answer needs a passage it can lift and attribute. Pages written to be comprehensive rather than answerable do worse, because there is no single passage that resolves the question — the answer is distributed across 3,000 words and the model has nothing crisp to quote.
This matches what we changed and saw work: put the direct answer in the first line in bold, keep a primary source next to each factual claim, and include one table. It is close to the opposite of writing for a word count.
The Part That Should Give You Pause
We accumulated 14,394 citations per 28 days. Over the same period we earned zero new referring domains. Not few. Zero.
The grounding-query table explains it. Our top 25 queries by citation volume, every one of them:
tuscan mom aesthetic · salary overtime law · is myspace still active · federal overtime laws · best AI tools for developer productivity · office siren · employee tax withholding update timing · when to update W4 · w2 example · fastest VPN speeds · cold plunge vs sauna · protein requirements toning vs bulking · fica vs federal tax
These are consumer questions, all of them. Not one is a practitioner question — nobody reaching us asked how to measure AI citations, how to run content operations, or which tool to use.
Which produces the mechanism, and it is not a distribution problem:
We are cited to people who ask questions. Links come from people who publish answers. Consumers do not have websites.
Someone asking whether MySpace is still active gets a good answer partly built from our page, and is satisfied. They have no blog, no newsletter, no editorial calendar. There is no path from that citation to a link however many times it happens — and scaling it tenfold yields tenfold of exactly nothing.
What We Take From It
- Citations and links are different products and need different content. Consumer-question content earns citations and traffic. Only practitioner-question content can earn links, because only practitioners publish.
- Do not judge AI visibility on analytics sessions. 14,394 citations produced 126 assistant-referred sessions — about a hundred citations per visit, because a good citation answers the question inside the chat.
- Do not judge commercial value on citations either. Same numbers, opposite error. They measure attention, not intent and not authority.
- Citations appear to behave like a stock. Our most-cited property has published nothing for a month and still leads the portfolio. Rankings decay when you stop; citations so far have not. We are re-measuring that at 60 and 90 days before believing it.
Frequently Asked Questions
What kind of content do AI assistants cite most?
In our data, four shapes account for 80% of the most-cited pages: head-to-head comparisons (X vs Y), definitions of a term someone just encountered, reference tables of values by category, and explainers of a specific trend or event. Each accounts for 20% of the top 25. What they share is that they answer a question with a specific, checkable, quotable answer — a number, a definition, a direct comparison. Assistants cite what they can lift a sentence out of. Pages built to be comprehensive rather than answerable perform worse, because there is no single passage to quote.
Does publishing more content increase AI citations?
Not reliably, in our experience. We tested volume directly and rejected it — more pages of the same shape did not produce proportionally more citations. What moved the number was writing pages that answer a specific question with a liftable answer, and citing a primary source next to each claim. We also found citations behave more like a stock than a flow: our most-cited property has published nothing for a month and still holds the largest block of citations in the portfolio. Rankings decay when you stop publishing; citations so far have not.
Do AI citations bring traffic?
Some, but far less than the citation count suggests, because a good citation often answers the question inside the chat and produces no click at all. Over an identical 28-day window our portfolio recorded 14,394 citations and 126 assistant-referred sessions — roughly a hundred citations per visit. Both numbers are real and they measure different things. Judging AI visibility by analytics sessions alone understates it by about two orders of magnitude; judging commercial value by citations alone overstates it by about the same.
Do AI citations produce backlinks?
In our data, no — and this was the most uncomfortable thing we found. Over the period we accumulated 14,394 citations per 28 days, new referring domains stayed at zero. The grounding-query table explains why: every one of our top 25 queries is a consumer question, and consumers do not have websites. Links come from people who publish. If your citations are earned on questions asked by readers rather than by writers, scaling that volume produces more of the same nothing.
How do you find out which queries you are cited on?
Bing Webmaster Tools has an AI Performance report showing citations by Copilot and partner models, broken out by grounding query and by page. It is the only first-party citation data any major platform publishes, and there is no API — we checked every plausible endpoint and got 404s, so it has to be read from the interface. Verify your root domain rather than each subdomain and one read covers a whole portfolio. It says nothing about ChatGPT, Claude, Gemini or Perplexity, so treat it as a sample of one ecosystem rather than a total.
Method and Caveats
Data is from Bing Webmaster Tools' AI Performance report (Microsoft Copilot and partner models) for a 21-subdomain portfolio, 3-month window, read 2026-07-26, plus GA4 referrer segmentation and Google Search Console over the 28 days ending 2026-07-24. The full method, including what each measurement misses, is in how to measure whether AI assistants cite your site.
Honest limits. This is one portfolio in one ecosystem — Bing reports only Microsoft-family citations and says nothing about ChatGPT, Claude, Gemini or Perplexity, and Bing itself describes its AI Performance data as a sample. The shape classification is our own judgement applied to 25 pages; a different taxonomy would produce different buckets, and 25 is a small n. Our portfolio skews to calculators, consumer finance, fitness and trend explainers, which plausibly biases which shapes appear at all. Treat the direction as informative and the percentages as indicative. Our own numbers, including the experiments that failed, are in the Thicket Report.