AI search news ·

Do AI engines cite homepages or deep pages? 92.3% of cited URLs are deep — and the modal citation is a /blog/answer-slug page

Across 3,281 path-carrying URLs that four AI engines cited when answering 12 small-business buyer questions in three separate dated runs, only 7.7% pointed at a site's homepage — 92.3% cited a deep page, at a mean depth of 2.04 path segments. The engines do not link brands; they link answers: the single most common cited shape is a two-segment path like `domain.com/blog/answer-slug` (61.3% of all cited URLs), and `/blog/…` alone carries 43.6% of deep citations. The per-engine spread is the practical finding: ChatGPT cites homepages 28.4% of the time — 17× Perplexity's 1.7% and roughly 90× Claude's 0.3%. If your site's citable surface is a homepage and a services page, three of the four engines have almost nothing they will link to.

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?", and ten 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. Previous cuts of the same corpus measured how many domains compete per question and how consistent each engine's citations are across identical asks. This cut asks a question none of those did: when an engine cites a source, does it link the site's homepage or a specific deep page — and how deep?

The numbers: homepages are 7.7% of what engines link

Pooled across the three runs, the corpus holds 4,573 cited-URL slots (distinct URL per answer). 1,292 of them — every Gemini citation — arrive as bare hostnames with no path and are excluded from depth ratios rather than guessed at (more below). Of the 3,281 URLs that carry a path:

EnginePath-carrying cited URLsHomepageDeep pageMean path segments
ChatGPT760216 (28.4%)544 (71.6%)1.54
Perplexity1,90733 (1.7%)1,874 (98.3%)2.22
Claude6142 (0.3%)612 (99.7%)2.11
Gemini0 of 1,292 (bare hostnames)
Pooled3,281251 (7.7%)3,030 (92.3%)2.04

The reading is stable: pooled homepage share reads 8.2% → 7.5% → 7.3% across the three runs, and no engine changes character between runs (ChatGPT reads 32.1% → 27.0% → 26.2%; Perplexity never exceeds 2.4%; Claude never exceeds 0.5%). Depth clusters hard at exactly two segments — 2,011 of 3,281 cited URLs (61.3%) are two-segment paths, the `/blog/answer-slug` shape — against 13.6% one-segment and 17.5% three-or-more.

Where the deep citations point

Of the 3,030 deep citations, the leading first path segment is `/blog/` with 1,320 — 43.6% of all deep citations, 46.0% adding `/blogs/`. The rest of the top of the table: `/r/` (Reddit threads, 161), locale prefixes like `/en/` (86), `/watch` (YouTube, 62), `/learn/` (58), `/articles/` (54), `/services/` (50), `/products/` (47). Notably commercial shapes also earn citations: `/compare/…` (40), `/alternatives/…` (40) and `/pricing` (37) all appear — engines will link a well-structured comparison or pricing page, not only editorial. Our own only citation in the corpus is itself the modal shape: on 2026-08-10 Perplexity cited `askedabout.com/compare/best-profound-alternatives` — a two-segment deep page — in all 3 samples of the Profound-alternative question. The homepage it never cited.

The Gemini caveat, stated as an instrument fact

All 1,292 Gemini citation entries in the corpus are bare hostnames — `example.com`, no scheme, no path. That is a property of its citation channel as our instrument receives it, not evidence Gemini prefers homepages: the pages it grounded on may well be deep, but the channel strips the path, so Gemini cannot testify in this cut and is excluded from every depth ratio above. Anyone reporting "what Gemini cites" at URL level is working with data the channel does not return.

Why this matters for anyone chasing AI visibility

Methodology

URL-depth 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). Unit: distinct cited URL per answer — each answer's returned citation array is deduplicated on exact URL after stripping query and fragment (tracking parameters such as `utm_source=openai` are noise, not distinct pages). A URL is path-informative only if it carries a scheme or an explicit path; bare hostnames are counted separately and never assumed to be homepages. Homepage = empty path after trailing-slash normalization; depth = number of path segments; first-segment table computed over deep citations only. Selfchecked (22 assertions: per-engine and per-run sums reconcile, classifier unit tests). Full per-engine, per-run and segment tables in the cut artifact `url-depth-cut-2026-08-15.json`.

Do AI engines link to homepages or specific pages?

Overwhelmingly specific pages. In our corpus of 3,281 path-carrying cited URLs across ChatGPT, Perplexity and Claude, only 7.7% pointed at a homepage; 92.3% cited a deep page at a mean depth of about two path segments. The reading held across three separate dated runs (8.2%, 7.5%, 7.3% homepage share).

What kind of page does ChatGPT link to in its answers?

ChatGPT is the most homepage-friendly engine we measure — 28.4% of its path-carrying citations point at a homepage — but deep pages still dominate at 71.6%, at a mean depth of 1.54 segments. Perplexity (1.7% homepage) and Claude (0.3%) almost exclusively cite deep pages.

What URL structure gets cited most by AI engines?

A two-segment path — domain.com/blog/answer-slug — is the modal cited shape at 61.3% of all path-carrying citations, and /blog/ alone carries 43.6% of deep citations. Comparison and pricing paths (/compare/, /alternatives/, /pricing) also earn citations, so citable pages are not limited to editorial content.

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.

Check a client's AI visibility

Begin your check

Free · 60 sec

No account · No card · 3 buyer questions, 2 engines

By running a check you agree to our Terms and Privacy Policy.

Who runs this