It started with an anomaly. One of our sites dominated search on a query, among the highest impressions in its niche, and yet the generative assistants never seemed to cite it on that question. First in the SERPs, absent from ChatGPT. We first assumed an isolated case, a model quirk. Then we looked for the same signal elsewhere, and found it on two other unrelated verticals.
For years, SEOs assumed that a visible page was an influential page. Ranking first on Google meant being the reference on a topic. With ChatGPT, Perplexity or Copilot, that equation cracks.
We analyzed three sites belonging to three different verticals: health, trades and construction, employment law. The same signals appear in all three cases. The data are anonymized, but they come from real measurements across three distinct sites. They help identify the mechanisms that seem to favor citation by generative engines. Here are the data and the conclusions we drew.
The page that dominates AI citations is not necessarily the one that dominates the SERPs.
This is an investigation into a pattern observed across three independent verticals, not a recipe.
What three independent verticals show
The pages that win AI citations are not necessarily the most visible. They are often the ones that answer a conversational question most directly.
The thesis in three lines
- SEO maximizes the probability of being retrieved.
- AI maximizes the probability of finding the best answer.
- These two objectives sometimes produce different winners.
Retrieval ≠ Citation: being retrieved as a candidate page by a search engine does not guarantee being cited in an AI-generated answer. These are two different selections, with different criteria.
Key Takeaways
- Across three independent verticals, the most visible page in search is not the most cited by AI.
- Hubs often win retrieval. Answer pages often win the citation.
- Intent Match, the direct match to a question, probably explains more variance than the hub versus spoke distinction.
- We propose an exploratory metric, Citation Efficiency (AI citations ÷ SERP impressions), useful mainly to compare similar pages. It is not an industry standard.
- Citation and recommendation remain two different steps: being cited is not being recommended on the right criteria.
Methodology note
Both impressions and AI citations come from Bing Webmaster Tools, the latter via its native AI performance report, per page and over the same observation window. These are exploratory readings specific to these sites, not a representative sample of all engines.
Vertical A: Health (YMYL)
The phenomenon does not appear in a lab. It appears on a real site, in a YMYL topic where answer quality matters particularly.
SERP visibility
| Page type | Impressions |
|---|---|
| Main hub | 204 |
| Secondary guide | 140 |
| Sub-topic A | 77 |
| Answer page A | 43 |
| Sub-topic B | 25 |
| Sub-topic C | 21 |
| Answer page B | 19 |
AI citations
| Page type | Citations |
|---|---|
| Answer page A | 62 |
| Secondary guide | 50 |
| Sub-topic A | 43 |
| Answer page B | 29 |
| Main hub | 18 |
Citation Efficiency (citations ÷ impressions)
| Page type | Citations | Impressions | Efficiency |
|---|---|---|---|
| Answer page B | 29 | 19 | 1.53 |
| Answer page A | 62 | 43 | 1.44 |
| Sub-topic A | 43 | 77 | 0.56 |
| Secondary guide | 50 | 140 | 0.36 |
| Main hub | 18 | 204 | 0.09 |
Plotting impressions on the x-axis and citations on the y-axis, the paradox jumps out: the least visible pages rise, the most visible stays at the bottom right.
AI citations
70 ┤
60 ┤ ● Answer page A
50 ┤ ● Secondary guide
40 ┤ ● Sub-topic A
30 ┤ ● Answer page B
20 ┤ ● Main hub
10 ┤
0 ┼────┬────┬────┬────┬────┬────┬────┬────
0 30 60 90 120 150 180 210
SERP impressionsThe pages that answer a precise situation or question capture a disproportionate share of citations.
Why we looked for a second, then a third site
A result on a single site proves nothing. It could be an artifact: a particular topic, a page structure specific to that site, a quirk of timing. To find out, we had to reproduce the observation elsewhere, on unrelated sites. So we repeated exactly the same reading, SERP impressions versus AI citations, on a second vertical, then on a third.
Vertical B: Trades and profitability
The same phenomenon reappears in a completely different industry.
SERP visibility
| Page type | Impressions |
|---|---|
| Main hub | 402 |
| Sub-topic A | 66 |
| Sub-topic B | 42 |
| Trade page A | 35 |
| Tool A | 20 |
| Sub-topic C | 16 |
| Sub-topic D | 16 |
| Trade page B | 10 |
| Answer page A | 9 |
| Trade page C | 6 |
AI citations
| Page type | Citations |
|---|---|
| Main hub | 172 |
| Trade page C | 57 |
| Trade page A | 39 |
| Answer page A | 31 |
| Sub-topic B | 19 |
| Trade page B | 14 |
| Tool B | 12 |
Citation Efficiency (citations ÷ impressions)
| Page type | Efficiency |
|---|---|
| Trade page C | 9.50 |
| Tool B | 4.00 |
| Answer page A | 3.44 |
| Trade page B | 1.40 |
| Trade page A | 1.11 |
| Main hub | 0.43 |
The most striking number
A trade page with only 6 SERP impressions generates 57 AI citations, a Citation Efficiency of 9.5. Just as powerful as the health example, and impossible for a competitor to exploit.
A ratio above 1 can surprise: how can a page be cited more often than it is seen? Because the two figures are not measured on the same surface. Impressions come from classic SERPs, while generative assistants can draw on their own index or on crawl databases, without generating an impression in Bing or Google. A page that is barely visible in search can therefore be widely reused in answers.
An honest caveat: here, the most cited page in raw volume is the hub, thanks to its massive visibility. But the answer pages dominate by far on efficiency. The pattern does not say they always win in volume, it says they capture a disproportionate share of citations relative to their visibility.
Vertical C: Employment law and psychosocial risks
A third vertical, unrelated to the first two, shows the same signal.
SERP visibility
| Page type | Impressions |
|---|---|
| Statistical study | 128 |
| Workplace situation A | 96 |
| Workplace situation B | 75 |
| Weak signal A | 52 |
| Calculator page | 42 |
AI citations
| Page type | Citations |
|---|---|
| Calculator page | 73 |
| Statistical study | 73 |
| Weak signal A | 34 |
| Recourse guide | 30 |
| Evidence page | 28 |
Citation Efficiency
| Page type | Citations | Impressions | Efficiency |
|---|---|---|---|
| Calculator page | 73 | 42 | 1.74 |
| Weak signal A | 34 | 52 | 0.65 |
| Statistical study | 73 | 128 | 0.57 |
The signal
A calculator page gets nearly twice as many citations as impressions, while the statistical study, far more visible, stays below 0.6.
Why the third site changes everything
A finding on a single site is an anecdote. On two, a hypothesis. On three unrelated verticals, a plausible pattern.
| Number of sites | Status of the finding |
|---|---|
| 1 site | Anecdote, possible artifact |
| 2 sites | Hypothesis to confirm |
| 3 independent verticals | Plausible pattern, to test more broadly |
It is precisely the third vertical that changed our reading. As long as the phenomenon appeared only in health then construction, we could suspect a common bias: a writing style, a way of structuring pages, the same tooling. Employment law, with no editorial or technical link to the other two, ruled out that explanation. The signal seemed less tied to the sites themselves than to how the engines select certain answers.
Why Intent Match explains citations better than classic SEO
A methodological caution. What follows is a model hypothesis, not an absolute truth. The internal pipelines of ChatGPT, Google or Bing vary, change often, and are not public. The model is kept because it predicts well what we observe across the three verticals.
Retrieval
↓
Intent Match
↓
CitationIntent Match: a page’s ability to match the exact natural phrasing of a user’s question.
SEO answers: “Can this page be found?”
Intent Match answers: “Is this the answer we were looking for?”
Classic SEO mainly optimizes retrieval: authority, indexing, coverage. It maximizes the probability of entering the candidate pool. But once that pool is built, the generative engine does not reuse the most authoritative page, it reuses the one that most resembles the question asked. Intent Match is what decides, and it is what best explains our gaps between visibility and citation.
Retrieval pulls several candidates. Intent Match often selects the one that already resembles the question.
| Question type | Selected page |
|---|---|
| Why does my symptom persist? | Answer page |
| How much does an independent professional earn? | Answer page |
| What amount can I obtain? | Answer page |
In all three cases, the system does not need to extract the right portion from the middle of a guide. The page already resembles the question, so it is easier to reuse.
The real pattern: answer pages
What the highest Citation Efficiency pages have in common is not their topic, not their sector, not their traffic volume. It is their ability to answer a question immediately.
| Vertical | Winning page |
|---|---|
| Health | Answer to a symptom |
| Trades / construction | Answer to a trade question |
| Employment law | Answer to a recourse question |
The highest Citation Efficiency pages look more like answers than like content.
Answer page: a specialized page that immediately answers one question, situation or decision, without thematic dilution.
Hubs and spokes: a secondary mechanism
The hub versus spoke distinction keeps explanatory value, but it is secondary.
| Role | Step often won | |
|---|---|---|
| Hub | Broad page, wide coverage, high visibility | Retrieval |
| Spoke | Specialized page, direct answer | Citation |
Hub vs Spoke explains part of the phenomenon. Intent Match probably explains more: a specialized page poorly aligned with a question stays barely cited, while a page, even attached to a hub, that exactly matches a question wins the citation. The hub is not the spoke’s enemy, it captures visibility and feeds the answer pages.
The Citation Efficiency concept
Citation Efficiency: the ratio of a page’s AI citations to its SERP impressions. It measures neither SEO quality nor traffic, but a page’s ability to convert visibility into citations.
Citation Efficiency = AI citations ÷ SERP impressions| Vertical | Best Citation Efficiency |
|---|---|
| Health | 1.53 |
| Trades / construction | 9.50 |
| Employment law | 1.74 |
Even when the numbers vary widely, from 1.53 to 9.50, the pages with high Citation Efficiency are systematically pages with high answer value.
Limits of Citation Efficiency
It is a comparison indicator, not an absolute truth. The ratio:
- depends on the sample: on small numbers it becomes unstable and misleading;
- depends on the observation window;
- depends on the engine measured;
- is mainly useful to compare similar pages.
Use it on sufficient volumes, over a stable window, and by segmenting page types.
How to spot your future Citation Efficiency winners
A page has strong citation potential if it:
- targets a profession, a symptom or a situation;
- targets a precise question or problem;
- contains figures and thresholds;
- has a title close to natural language.
Conversely, a very visible but rarely cited page is often a disguised hub, to break down into answer pages.
How to measure your AI visibility
There is no Search Console equivalent for ChatGPT yet. Measurement is built by cross-referencing several sources, each covering only one angle.
| Source | What it reveals | Limit |
|---|---|---|
| Bing Webmaster Tools | Impressions, queries and AI citations via the AI performance report | Coverage centered on the Bing and Copilot ecosystem |
| Google Search Console | Impressions, queries, appearance in AI Overviews | AI visibility still partial |
| GA4 | Referral traffic from AI assistants | Captures clicks only, not citations without a click |
| Server logs | AI crawler hits: GPTBot, OAI-SearchBot, PerplexityBot, ClaudeBot, Google-Extended | Indicates collection, not citation |
| Dedicated tools | Citation tracking: Profound, Peec AI, Scrunch AI | Coverage and engines vary |
Because the ratio depends on the window and the engine, an isolated measurement is worth little. AI visibility moves: a model or index update can shift your citations from one month to the next. The right reflex is not a one-off audit, but monitoring Citation Efficiency over time.
What the three verticals have in common
| Vertical | SERP winner | Citation Efficiency winner | Signal |
|---|---|---|---|
| Health | Main hub | Answer page | Answer to a symptom |
| Trades / construction | Main hub | Trade page | Answer to a trade question |
| Employment law | Statistical study | Calculator page | Decisional answer |
In this table, the Citation Efficiency winner means the page that captures the most citations relative to its visibility. Despite radically different topics, these pages all belong to the same family: answer pages.
What we observe, what we do not claim
What we observe. The most cited pages answer a question, a decision or a situation.
What we do not claim. We do not prove the internal workings of ChatGPT, Bing or Google. We simply show that the same signal appears across three different verticals.
Going further: citation is not recommendation
Citation Efficiency measures one step, not the whole race. Being cited does not mean being recommended. It helps to read these signals as a ladder of three thresholds.
| Threshold | Question | Metric |
|---|---|---|
| Retrieval | Are we pulled into the candidate pool? | Impressions, indexing, authority |
| Citation | Are we reused in the answer? | Citation Efficiency |
| Recommendation | Are we recommended on the right criteria? | Decision Share of Voice |
For a publisher site, the stakes often stop at the citation and the traffic it brings. For a brand that sells an offer, winning the citation without winning the recommendation can be enough to appear without converting. That is the subject of Why am I cited by ChatGPT but not converting?.
Conclusion
What looked like an anomaly on a single site turned out to be a stable signal. The three verticals analyzed show the same phenomenon: generative engines do not seem to favor only the most visible pages. They seem to favor the pages that best match a question phrased by a user.
SEO remains essential to enter the candidate pool. But once retrieved, a page still has to win the Intent Match, then the citation.
SEO makes you findable. Intent Match makes you citable.
In an environment where generative engines become a layer of access to information, the question is no longer only “can I be found?”, but “am I the easiest answer to reuse?”.
Sources and tools cited
- Google Search Central: AI features and your website
- Google Search Central: structured data intro
- GEO paper, KDD 2024: Generative Engine Optimization
FAQ
Is this limited to one sector?
No. We observe it across three independent verticals: health YMYL, trades and construction, employment law. In all three, pages that answer a precise question capture a disproportionate share of citations relative to their SERP visibility.
Does GEO replace SEO?
No. SEO makes you findable, which remains the entry condition into the candidate pool. GEO adds a question: once retrieved, does your page win the Intent Match and then the citation?
Do backlinks still matter?
Yes, for authority and retrieval. But a strong link profile does not guarantee a citation if the page lacks a self-contained block of text aligned with a question.
How do I know if ChatGPT cites me?
By cross-referencing Bing Webmaster Tools, Google Search Console, GA4, AI crawler server logs and citation-tracking tools, then comparing citations and impressions to compute a Citation Efficiency per page.
What best explains the citation?
In our observations, Intent Match, the direct match between a page and the natural phrasing of a question, explains more variance than the simple hub versus specialized-page distinction.
What question does the buyer ask AI?
Which documentation simplification can lower the standard?
Which technical requirement must be clearly formulated?
Which evidence should be requested or published?
Which criterion excludes an insufficient answer?