AI search news ·
Do AI engines cite URLs that name the question? 91.8% of cited deep pages carry a question word in the path — and only 5.1% carry a year
Of 3,030 deep-page URLs that four AI engines cited when answering 12 small-business buyer questions across three separate dated runs, 91.8% carry at least one word from the question in their path, 77.1% carry two or more, and 82.2% carry a distinctive question word (a term specific to that question, not just `ai`, `tool` or `business`). Engines are not merely citing deep pages — they are citing pages whose URL literally names the thing that was asked. ChatGPT is again the outlier: 77.2% of its deep citations name the question, against 95.1% for Perplexity and 94.4% for Claude. Dated slugs are rare and never stale: only 5.1% of cited paths carry a year, 152 of those 154 say `2026`, and not one carries a year older than 2025.
Primary source: AskedAbout — slug-anatomy cut of three dated four-engine runs (2026-07-28 · 2026-08-03 · 2026-08-10)
The instrument
Our weekly self-audit asks 12 pinned buyer questions — "What are the best AI visibility tools for small businesses?", "What is a good Profound alternative for a small business?", "Is there a one-time AI visibility audit instead of a monthly subscription?" and nine more — to ChatGPT, Perplexity, Gemini and Claude through their APIs, three samples per question per engine, and stores every answer with its returned citation list. Yesterday's cut established that 92.3% of cited URLs are deep pages, not homepages. This cut reads the words in those deep paths: does the URL an engine cites name the question it was asked — and does it carry a year, a listicle marker, or an alternatives/comparison marker?
The numbers: 91.8% of cited deep pages name the question in the URL
Each cited URL's path was split into tokens and matched against the content words of the question that produced it (question tokens minus a fixed stopword list, light plural stemming — the exact keyword list per question is in the cut artifact). Gemini's citation channel returns bare hostnames and is excluded as an instrument fact; homepages carry no path and are excluded by definition. That leaves 3,030 deep-page citations from ChatGPT, Perplexity and Claude:
| Engine | Deep-page citations | ≥1 question word in path | ≥2 question words | ≥1 distinctive word (excl. ai/tool/business) | Mean question words per path |
|---|---|---|---|---|---|
| ChatGPT | 544 | 420 (77.2%) | 314 (57.7%) | 373 (68.6%) | 1.76 |
| Perplexity | 1,874 | 1,783 (95.1%) | 1,516 (80.9%) | 1,613 (86.1%) | 2.42 |
| Claude | 612 | 578 (94.4%) | 506 (82.7%) | 506 (82.7%) | 2.52 |
| Gemini | 0 of 1,292 (bare hostnames) | — | — | — | — |
| Pooled | 3,030 | 2,781 (91.8%) | 2,336 (77.1%) | 2,492 (82.2%) | 2.32 |
The reading is stable across the three runs — pooled any-word match reads 91.6% → 91.2% → 92.6% and the engine ordering never changes: ChatGPT 79.5% → 73.9% → 78.3%, Perplexity 95.1% → 94.7% → 95.6%, Claude 92.3% → 95.5% → 95.8%. Even after excluding the three tokens that saturate this vertical (`ai`, `tool`, `business`), 82.2% of cited paths still carry a word specific to the question — `chatgpt`, `alternative`, `audit`, `athenahq`, `profound`, `cheapest`, `dentist`. The most-matched question words across the corpus, in order: `ai` (1,439 paths), `visibility` (892), `tool` (785), `chatgpt` (666), `alternative` (513), `business` (508), `audit` (327), `athenahq` (265), `profound` (262), `search` (240).
Years, listicles and alternatives markers
| Path feature | ChatGPT | Perplexity | Claude | Pooled (n=3,030) |
|---|---|---|---|---|
| Carries a year (20xx token) | 3.5% (19) | 3.8% (72) | 10.3% (63) | 5.1% (154) |
| …of which the year is 2026 | 17 of 19 | 72 of 72 | 63 of 63 | 152 of 154 (2 say 2025; 0 older) |
| Listicle marker (best / top / leading number) | 12.7% | 32.5% | 25.2% | 27.5% |
| Alternatives / vs / compare marker | 21.3% | 27.4% | 22.7% | 25.4% |
| How-to / guide / tutorial marker | 14.0% | 10.7% | 10.1% | 11.2% |
| Any of the three intent markers | 43.0% | 59.1% | 50.5% | 54.5% |
| Mean words in the final slug segment | 4.32 | 4.86 | 5.21 | 4.84 |
| Opaque final segment (numeric / hex id) | 0.9% | 1.7% | 0.2% | 1.2% (37) |
Two of these rows carry the practical news. First, the year row: only one cited path in twenty is dated, and when it is, it is dated this year — 152 of 154 say `2026`, two say `2025`, and across 3,030 citations there is not a single `2024` or older. Either engines are not surfacing stale dated posts, or the sites that date their slugs re-mint them yearly; either way a `-2024-` slug is absent from the citation layer of every engine we measure. Claude leans hardest on dated slugs (10.3%, three times ChatGPT and Perplexity). Second, the intent-marker rows: over half of all cited paths (54.5%) carry a `best/top`, `alternatives/vs` or `how-to/guide` marker — the URL announces the page's shape as well as its topic. The listicle share tracks the question: it is 63.6% on "best AI visibility tools" and 0% on the one-time-audit question; the alternatives marker is 96.6% on the Profound-alternative question and 90.2% on the AthenaHQ one.
Per question: where the slug-literal pattern breaks
Eleven of the twelve questions read above 82% any-word match; the exception is the one-time-audit-versus-subscription question at 69.1% — the engines cannot find many pages whose URL says `one-time` or `subscription`, so they cite adjacent `ai-visibility-audit` pages instead. The distinctive-word view is harsher on one question: "What tools track whether AI recommends my business?" reads 89.6% on any word but only 9.2% on a distinctive word — almost nothing cited for it carries `track` or `recommend` in the path; it is answered from generic `ai-visibility-tools` pages. Our own sole citation in the corpus fits the pattern exactly: on 2026-08-10 Perplexity cited `askedabout.com/compare/best-profound-alternatives` in all three samples of the Profound-alternative question — a path carrying `profound`, `alternative` and `best`, i.e. two question words plus both a listicle and an alternatives marker.
Why this matters for anyone chasing AI visibility
- Name the question in the URL. 91.8% of the deep pages engines cite carry a question word in the path and 77.1% carry two. A page that answers a buyer question under a slug like `/blog/our-thoughts` is fighting the corpus; `/blog/best-<thing>-for-<segment>` is the modal shape.
- Distinctive beats generic. 82.2% of cited paths still match after removing `ai`, `tool` and `business` — the slugs carry the specific term (the competitor name, the verb, the vertical). The one question with a 9.2% distinctive rate is the one whose buyers get generic answers.
- Do not date the slug unless you will re-mint it. Only 5.1% of cited paths carry a year, and 152 of 154 say 2026 — a year in the URL is a freshness claim engines appear to hold you to, and no `2024` slug survives in the citation layer.
- ChatGPT rewards the slug less than the others do. At 77.2% it is 18 points below Perplexity and Claude, consistent with its 28.4% homepage share yesterday: ChatGPT leans more on brand/domain-level authority, the other two on the specific answering page. A page can be well-cited by Perplexity and invisible to ChatGPT for reasons the slug will not fix.
- Check which page and which words get cited, not just whether your domain appears. Run the free check to see what the engines cite when asked about your business.
Methodology
Slug-anatomy cut of the three dated four-engine production runs (2026-07-28, 2026-08-03, 2026-08-10; answer rows pulled from the production `aeoSelfAnswers` table by runId; panel version 2026-06-27, unchanged; 12 questions × 4 engines × 3 samples = 432 answers). Unit: distinct cited URL per answer, deduplicated on exact URL after stripping query and fragment. Only path-carrying, non-homepage URLs are scored (4,573 slots → 1,292 bare Gemini hostnames excluded → 251 homepages excluded → 3,030 deep pages), reconciling exactly with yesterday's depth cut. Keyword match: path split on non-alphanumerics; a URL names the question if ≥1 content token of the asked question (question tokens minus a fixed stopword list, possessives stripped, light plural stemming) appears among its path tokens; the distinctive rate excludes `ai`, `tool`, `business`. Year = any `20xx` token; listicle = `best`/`top` or a leading-number segment; alternatives = alternative(s)/vs/versus/compare/comparison/competitor(s); how-to = how-to/guide/tutorial. Selfchecked (15 assertions: slot/engine/run/question sums reconcile, stemmer and classifier unit tests). Full per-engine, per-run, per-question tables, keyword lists and token frequencies in the cut artifact `slug-anatomy-cut-2026-08-16.json`.
Do AI engines cite pages whose URL contains the query keywords?
Overwhelmingly yes. Across 3,030 deep-page citations from ChatGPT, Perplexity and Claude, 91.8% of cited URL paths carried at least one word from the question asked, 77.1% carried two or more, and 82.2% carried a distinctive question word once generic vertical tokens were excluded. The reading held at 91–93% in each of three separate dated runs.
Which AI engine cares least about the URL slug?
ChatGPT. 77.2% of its deep citations name the question in the path, against 95.1% for Perplexity and 94.4% for Claude — an 18-point gap that held in all three runs. Combined with its 28.4% homepage-citation share, ChatGPT leans more on domain-level authority and less on the specific answering page.
Should I put the year in my URL for AI visibility?
Only 5.1% of AI-cited paths carry a year at all, so it is not required. But when a cited path is dated, it is dated this year: 152 of 154 say 2026, 2 say 2025, and none says 2024 or older across 3,030 citations. A year in the slug is a freshness claim — put one in only if you will re-mint the URL each year.
See your number
See which businesses AI names when your client's buyers ask.
Running this for clients? The $249 agency 5-pack audits five businesses, white-labeled.
Who runs this
- Built and operated by Sensara LLC, Atlanta, Georgia — about us and how the audit works.
- See what the report looks like before you run anything — score per engine, the competitors AI names instead of you, and a fix plan.
- We run the same audit on ourselves every week and publish the result: in the latest run AI named AskedAbout in 1 of 144 answers. We report our own numbers the way we report yours.