The address of a page that an AI engine cites tends to repeat the shape of the question. In the Machine Relations Index answer ledger, 197,384 cited-URL events came from 23,183 answer runs on 1,008 buying questions. On head-to-head questions, 37.2 percent of the cited URLs had "vs" in their last path segment (11,239 of 30,246). On "best X" and top-list questions the figure was 1.5 percent. On "best X" questions, 37.7 percent of cited URLs started with "best" or "top" or carried one of them in the slug (14,636 of 38,847); on head-to-head questions it was 9.6 percent.
On problem-first questions, the buyer describes a situation and names no product: 20.5 percent of cited URLs had a question-form slug that began with "what", "how", "why", "is", "are", "can" or "do" (5,382 of 26,258). On best-X and top-list questions that figure was 2.3 and 2.8 percent.
This page counts what cited addresses look like. It does not say that a slug earns a citation: a page titled "X vs Y" is usually a comparison, and the comparison, not the word in its address, may be what the engine used.
What was counted #
- Population. Successful runs in the Index's "mri_v2" panel, from 2026-07-06 to 2026-11-07, on its 1,008 buying questions: 168 in each of six question shapes (best X, top lists, how to choose, is it worth it, problem first, X vs Y). Runs flagged as naming a brand in the prompt are excluded. The ledger export is dated 2026-11-07 08:10 UTC.
- Engines. ChatGPT, Claude, Gemini, Google AI Mode and Perplexity. Google AI Overviews are left out because this analysis was not run on them.
- Unit. One cited URL in one answer run, counted once however many times the run repeated it. The same URL cited in two runs counts twice. The set holds 197,384 such events, 23,183 runs and 29,282 root domains.
- What was read. Only the address. The last segment of the URL path (the slug) was matched against fixed patterns, listed in the method note. No page was fetched or read.
The slug follows the question shape #
Share of cited URLs whose slug matches each pattern, by the shape of the question that produced the citation (events matching, out of events):
| Question shape | Cited URLs | "vs" in slug | "best" or "top" in slug | Question-form slug |
|---|---|---|---|---|
| X vs Y | 30,246 | 37.2% (11,239) | 9.6% (2,903) | 6.0% (1,802) |
| Best X | 38,847 | 1.5% (594) | 37.7% (14,636) | 2.3% (895) |
| Top lists | 39,901 | 1.5% (600) | 30.0% (11,979) | 2.8% (1,127) |
| How to choose | 31,480 | 9.5% (3,004) | 16.0% (5,029) | 12.1% (3,819) |
| Is it worth it | 30,652 | 10.2% (3,118) | 12.7% (3,879) | 12.3% (3,767) |
| Problem first | 26,258 | 0.6% (151) | 3.6% (939) | 20.5% (5,382) |
Each bold cell is the pattern the question shape would predict, and each is the highest in its column. The two middle shapes, "how to choose" and "is it worth it", split between all three patterns. A buyer weighing options gets pages that are partly comparisons, partly rankings and partly explanations.
Because the 168 questions in a shape are the unit that varies, the intervals below resample questions, with each question's citations kept together (2,000 draws, fixed seed):
| Cell | Share | 95% interval |
|---|---|---|
| "vs" in slug, X vs Y questions | 37.2% | 33.4 to 41.0 |
| "vs" in slug, best-X questions | 1.5% | 1.2 to 1.9 |
| "best" or "top" in slug, best-X questions | 37.7% | 34.6 to 40.7 |
| "best" or "top" in slug, X vs Y questions | 9.6% | 7.3 to 12.3 |
| "best" or "top" in slug, top-list questions | 30.0% | 27.3 to 32.9 |
| Question-form slug, problem-first questions | 20.5% | 18.3 to 22.8 |
| Question-form slug, best-X questions | 2.3% | 1.8 to 2.8 |
None of these pairs overlap. The pattern is not carried by one publisher. The "vs" citations on X-vs-Y questions come from 2,021 root domains, and the largest, nerdwallet.com, holds 105 of the 11,239 (0.9 percent). The "best" or "top" citations on best-X and top-list questions come from 3,811 domains, and the largest, forbes.com, holds 679 of 26,615 (2.6 percent).
The result does not depend on counting repeats. Counting each distinct URL once per question shape, 26.8 percent of the URLs cited on X-vs-Y questions had "vs" in the slug, and 27.0 and 22.9 percent of those cited on best-X and top-list questions had "best" or "top".
Engines differ in how closely they follow it #
Share of cited URLs matching the pattern the question shape predicts, by engine (events matching, out of events):
| Engine | "vs" on X-vs-Y questions | "best" or "top" on best-X and top-list questions | Question-form on problem-first questions |
|---|---|---|---|
| ChatGPT | 48.5% (1,770 of 3,648) | 16.8% (2,987 of 17,809) | 23.1% (1,179 of 5,108) |
| Claude | 46.2% (3,068 of 6,640) | 46.1% (6,183 of 13,416) | 22.3% (345 of 1,546) |
| Google AI Mode | 42.2% (459 of 1,087) | 32.6% (647 of 1,982) | 21.4% (180 of 842) |
| Perplexity | 32.6% (5,499 of 16,858) | 41.8% (14,661 of 35,080) | 19.5% (3,476 of 17,824) |
| Gemini | 22.0% (443 of 2,013) | 20.4% (2,137 of 10,461) | 21.5% (202 of 938) |
The question-form slug on problem-first questions is the steadiest cell, between 19.5 and 23.1 percent on every engine. The other two cells vary. ChatGPT, Claude and Google AI Mode cite "vs" addresses on 42.2 to 48.5 percent of head-to-head citations, Perplexity on 32.6 and Gemini on 22.0. On rankings, Gemini's three rates sit within two points of each other, so it shows the least tilt toward the shape of the question. The engines split on rankings. Claude and Perplexity cite "best" or "top" addresses in 46.1 and 41.8 percent of citations on best-X and top-list questions. ChatGPT does so in 16.8 percent and Gemini in 20.4. The engines cite different amounts: Perplexity supplies 105,295 of the 197,384 events and Google AI Mode 6,124, so the pooled table above is weighted toward Perplexity.
ChatGPT cites homepages; the others almost never do #
Homepages, meaning a bare domain with no path, are 2.2 percent of all 197,384 cited URLs. They are not spread evenly across engines:
| Engine | Cited URLs | Homepages | Share | 95% interval |
|---|---|---|---|---|
| ChatGPT | 35,357 | 2,992 | 8.5% | 7.4 to 9.8 |
| Gemini | 17,416 | 243 | 1.4% | 1.0 to 1.9 |
| Claude | 33,192 | 278 | 0.8% | 0.6 to 1.1 |
| Perplexity | 105,295 | 767 | 0.7% | 0.6 to 0.9 |
| Google AI Mode | 6,124 | 34 | 0.6% | 0.3 to 0.9 |
About one ChatGPT citation in twelve points at the front of a site and not at an article. This page does not say why. On the other four engines, no more than 1.4 percent of citations point at a homepage.
What this changes for how pages are named #
- Name the page for the question it answers. The question "X vs Y" is answered by addresses that say "vs", and "best X" is answered by addresses that say "best". A single page cannot carry both slugs. A buyer researching a purchase asks both, and the pages cited for them differ.
- Problem-first questions are answered by explainers. When the buyer describes a situation and names no product, one cited URL in five is a "what is", "how to" or "why" page. A page titled around a product name has the wrong shape for that question.
- Do not rename existing pages on this evidence alone. The data shows what engines cited, not what a rename would earn. A slug change also breaks inbound links unless it is redirected. If a page already answers one shape well, add the missing shape as a separate page.
- Homepages are mostly a ChatGPT pattern. The 8.5 percent applies to one engine in this panel, and the counts do not separate a company's own homepage from a third party's.
What this page does not show #
- Cause. A slug may be a label for a page that is a comparison or a ranking, and the engine may have chosen the page for its content. No page was read, so this does not separate the two.
- Buyer demand. The 1,008 questions are the Index's own panel, written for coverage of buying decisions. They are not a sample of what buyers type.
- Citation quality. A cited URL is a source an engine linked or listed. Whether the answer relied on it, or named it in the text, is a separate question: the Index finds engines name only 17.6 percent of the sources they cite (AI engines cite sources they do not name).
- Google AI Overviews. Not included.
- Trend. The page does not test whether these shares change over time. One part of the address, the year, is counted in the year-in-the-URL study. Another, the folder the address sits in, is counted in the site-section study.
- The name for this work. What the engines call the work is counted separately in What AI engines call this work.
Method note #
Source: the Index answer ledger's citation-events table, daily export of 2026-11-07 08:10 UTC, read with DuckDB. Filters: panel "mri_v2", runs not flagged as brand-prompted, a non-empty cited URL, engines ChatGPT, Claude, Gemini, Google AI Mode and Perplexity, run dates 2026-07-06 to 2026-11-07. Unit: distinct (answer run, cited URL) pairs. The slug is the last non-empty segment of the URL path, taken before any query string or fragment, and a URL with an empty path is a homepage. Patterns on the slug: "vs" is a hyphen-delimited token; "best or top" is a slug that starts with "best-" or "top-" or contains "-best-" or "-top-" followed by a letter; question-form is a slug that starts with "what-", "how-", "why-", "is-", "are-", "can-", "do-" or "does-". Intervals resample questions within the group, 2,000 draws, fixed seed, with every cited URL of a question kept together. The distinct-URL check counts each URL once per question shape. The tables report the pooled counts, so each cell is a share of events and not an average of per-question shares.
Frequently asked questions #
Do AI engines cite URLs that match the question? #
Often, in this ledger. On head-to-head questions, 37.2 percent of cited URLs had "vs" in the slug (11,239 of 30,246) against 1.5 percent on best-X questions. On best-X questions, 37.7 percent had "best" or "top" in the slug, against 9.6 percent on head-to-head questions. This describes the cited addresses and does not show that the slug earned the citation.
Which engine cites homepages? #
ChatGPT, mostly. 8.5 percent of its 35,357 cited URLs were a homepage (2,992), against 0.6 to 1.4 percent on Claude, Gemini, Perplexity and Google AI Mode.
Should I rename pages to include vs or best in the URL? #
Not on this evidence alone. The ledger shows what engines cited, not what a rename would earn, and a rename breaks inbound links unless it is redirected. Where a page already answers one question shape, the safer reading is to build a separate page for the shape it is missing.
What kind of page do AI engines cite when the buyer names no product? #
For problem-first questions, one cited URL in five (20.5 percent, 5,382 of 26,258) had a question-form slug such as "what", "how" or "why". That is a count of addresses, not a verdict on the pages behind them.